agora inbox for pgsql-hackers@postgresql.org  

help / color / mirror / Atom feed

[PATCH v1 4/7] Row pattern recognition patch (executor).
213+ messages / 2 participants
[nested] [flat]

* [PATCH v1 4/7] Row pattern recognition patch (executor).
@ 2023-06-25 11:48  Tatsuo Ishii <ishii@postgresql.org>
  0 siblings, 0 replies; 213+ messages in thread

From: Tatsuo Ishii @ 2023-06-25 11:48 UTC (permalink / raw)

---
 src/backend/executor/nodeWindowAgg.c | 223 +++++++++++++++++++++-
 src/backend/utils/adt/windowfuncs.c  | 274 ++++++++++++++++++++++++++-
 src/include/catalog/pg_proc.dat      |   9 +
 src/include/nodes/execnodes.h        |  12 ++
 src/include/windowapi.h              |   9 +
 5 files changed, 517 insertions(+), 10 deletions(-)

diff --git a/src/backend/executor/nodeWindowAgg.c b/src/backend/executor/nodeWindowAgg.c
index 310ac23e3a..ae997dff86 100644
--- a/src/backend/executor/nodeWindowAgg.c
+++ b/src/backend/executor/nodeWindowAgg.c
@@ -48,6 +48,7 @@
 #include "utils/acl.h"
 #include "utils/builtins.h"
 #include "utils/datum.h"
+#include "utils/fmgroids.h"
 #include "utils/expandeddatum.h"
 #include "utils/lsyscache.h"
 #include "utils/memutils.h"
@@ -159,6 +160,14 @@ typedef struct WindowStatePerAggData
 	bool		restart;		/* need to restart this agg in this cycle? */
 } WindowStatePerAggData;
 
+/*
+ * Map between Var attno in a target list and the parsed attno.
+ */
+typedef struct AttnoMap {
+	List		*attno;			/* att number in target list (list of AttNumber) */
+	List		*attnosyn;		/* parsed att number (list of AttNumber) */
+} AttnoMap;
+
 static void initialize_windowaggregate(WindowAggState *winstate,
 									   WindowStatePerFunc perfuncstate,
 									   WindowStatePerAgg peraggstate);
@@ -195,9 +204,9 @@ static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
 
 static bool are_peers(WindowAggState *winstate, TupleTableSlot *slot1,
 					  TupleTableSlot *slot2);
-static bool window_gettupleslot(WindowObject winobj, int64 pos,
-								TupleTableSlot *slot);
 
+static void attno_map(Node *node, AttnoMap *map);
+static bool attno_map_walker(Node *node, void *context);
 
 /*
  * initialize_windowaggregate
@@ -2388,6 +2397,12 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
 	TupleDesc	scanDesc;
 	ListCell   *l;
 
+	TargetEntry	*te;
+	Expr		*expr;
+	Var			*var;
+	int			nargs;
+	AttnoMap	attnomap;
+
 	/* check for unsupported flags */
 	Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
 
@@ -2483,6 +2498,16 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
 	winstate->temp_slot_2 = ExecInitExtraTupleSlot(estate, scanDesc,
 												   &TTSOpsMinimalTuple);
 
+	winstate->prev_slot = ExecInitExtraTupleSlot(estate, scanDesc,
+												 &TTSOpsMinimalTuple);
+
+	winstate->next_slot = ExecInitExtraTupleSlot(estate, scanDesc,
+												 &TTSOpsMinimalTuple);
+
+	winstate->null_slot = ExecInitExtraTupleSlot(estate, scanDesc,
+												 &TTSOpsMinimalTuple);
+	winstate->null_slot = ExecStoreAllNullTuple(winstate->null_slot);
+
 	/*
 	 * create frame head and tail slots only if needed (must create slots in
 	 * exactly the same cases that update_frameheadpos and update_frametailpos
@@ -2667,6 +2692,69 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
 	winstate->inRangeAsc = node->inRangeAsc;
 	winstate->inRangeNullsFirst = node->inRangeNullsFirst;
 
+	/* Set up row pattern recognition PATTERN clause */
+	winstate->patternVariableList = node->patternVariable;
+	winstate->patternRegexpList = node->patternRegexp;
+
+	/* Set up row pattern recognition DEFINE clause */
+
+	/*
+	 * Collect mapping between varattno and varattnosyn in the targetlist.
+	 * XXX: For now we only check RPR's argument. Eventually we have to
+	 * recurse the targetlist to find out all mappings in Var nodes.
+	 */
+	attnomap.attno = NIL;
+	attnomap.attnosyn = NIL;
+
+	foreach (l, node->plan.targetlist)
+	{
+		te = lfirst(l);
+		if (IsA(te->expr, WindowFunc))
+		{
+			WindowFunc	*func = (WindowFunc *)te->expr;
+			if (func->winfnoid != F_RPR)
+				continue;
+
+			/* sanity check */
+			nargs = list_length(func->args);
+			if (list_length(func->args) != 1)
+				elog(ERROR, "RPR must have 1 argument but function %d has %d args", func->winfnoid, nargs);
+
+			expr = (Expr *) lfirst(list_head(func->args));
+			if (!IsA(expr, Var))
+				elog(ERROR, "RPR's arg is not Var");
+
+			var = (Var *)expr;
+			elog(DEBUG1, "resname: %s varattno: %d varattnosyn: %d",
+				 te->resname, var->varattno, var->varattnosyn);
+			attnomap.attno = lappend_int(attnomap.attno, var->varattno);
+			attnomap.attnosyn = lappend_int(attnomap.attnosyn, var->varattnosyn);
+		}
+	}
+
+	winstate->defineVariableList = NIL;
+	winstate->defineClauseList = NIL;
+	if (node->defineClause != NIL)
+	{
+		foreach(l, node->defineClause)
+		{
+			char		*name;
+			ExprState	*exps;
+
+			te = lfirst(l);
+			name = te->resname;
+			expr = te->expr;
+
+			elog(DEBUG1, "defineVariable name: %s", name);
+			winstate->defineVariableList = lappend(winstate->defineVariableList,
+												   makeString(pstrdup(name)));
+			/* tweak expr so that it referes to outer slot */
+			attno_map((Node *)expr, &attnomap);
+			exps = ExecInitExpr(expr, (PlanState *) winstate);
+			winstate->defineClauseList = lappend(winstate->defineClauseList, exps);
+		}
+	}
+
 	winstate->all_first = true;
 	winstate->partition_spooled = false;
 	winstate->more_partitions = false;
@@ -2674,6 +2762,76 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
 	return winstate;
 }
 
+/*
+ * Rewrite Var node's varattno to the varattno which is used in the target
+ * list using AttnoMap.  We also rewrite varno so that it sees outer tuple
+ * (PREV) or inner tuple (NEXT).
+ */
+static void
+attno_map(Node *node, AttnoMap *map)
+{
+	(void) expression_tree_walker(node, attno_map_walker, (void *) map);
+}
+
+static bool
+attno_map_walker(Node *node, void *context)
+{
+	FuncExpr	*func;
+	int			nargs;
+	Expr		*expr;
+	Var			*var;
+	AttnoMap	*attnomap;
+	ListCell	*lc1, *lc2;
+	
+	if (node == NULL)
+		return false;
+
+	attnomap = (AttnoMap *) context;
+
+	if (IsA(node, FuncExpr))
+	{
+		func = (FuncExpr *)node;
+
+		if (func->funcid == F_PREV || func->funcid == F_NEXT)
+		{
+			/* sanity check */
+			nargs = list_length(func->args);
+			if (list_length(func->args) != 1)
+				elog(ERROR, "PREV/NEXT must have 1 argument but function %d has %d args", func->funcid, nargs);
+
+			expr = (Expr *) lfirst(list_head(func->args));
+			if (!IsA(expr, Var))
+				elog(ERROR, "PREV/NEXT's arg is not Var");	/* XXX: is it possible that arg type is Const? */
+			var = (Var *)expr;
+
+			if (func->funcid == F_PREV)
+				var->varno = OUTER_VAR;
+			else
+				var->varno = INNER_VAR;
+		}
+		return expression_tree_walker(node, attno_map_walker, (void *) context);
+	}
+	else if (IsA(node, Var))
+	{
+		var = (Var *)node;	 
+
+		elog(DEBUG1, "original varno: %d varattno: %d", var->varno, var->varattno);
+
+		forboth(lc1, attnomap->attno, lc2, attnomap->attnosyn)
+		{
+			int	attno = lfirst_int(lc1);
+			int	attnosyn = lfirst_int(lc2);
+
+			if (var->varattno == attnosyn)
+			{
+				elog(DEBUG1, "loc: %d rewrite varattno from: %d to %d", var->location, attnosyn, attno);
+				var->varattno = attno;
+			}
+		}
+	}
+	return expression_tree_walker(node, attno_map_walker, (void *) context);
+}
+
 /* -----------------
  * ExecEndWindowAgg
  * -----------------
@@ -2691,6 +2849,8 @@ ExecEndWindowAgg(WindowAggState *node)
 	ExecClearTuple(node->agg_row_slot);
 	ExecClearTuple(node->temp_slot_1);
 	ExecClearTuple(node->temp_slot_2);
+	ExecClearTuple(node->prev_slot);
+	ExecClearTuple(node->next_slot);
 	if (node->framehead_slot)
 		ExecClearTuple(node->framehead_slot);
 	if (node->frametail_slot)
@@ -2740,6 +2900,8 @@ ExecReScanWindowAgg(WindowAggState *node)
 	ExecClearTuple(node->agg_row_slot);
 	ExecClearTuple(node->temp_slot_1);
 	ExecClearTuple(node->temp_slot_2);
+	ExecClearTuple(node->prev_slot);
+	ExecClearTuple(node->next_slot);
 	if (node->framehead_slot)
 		ExecClearTuple(node->framehead_slot);
 	if (node->frametail_slot)
@@ -3080,7 +3242,7 @@ are_peers(WindowAggState *winstate, TupleTableSlot *slot1,
  *
  * Returns true if successful, false if no such row
  */
-static bool
+bool
 window_gettupleslot(WindowObject winobj, int64 pos, TupleTableSlot *slot)
 {
 	WindowAggState *winstate = winobj->winstate;
@@ -3420,14 +3582,53 @@ WinGetFuncArgInFrame(WindowObject winobj, int argno,
 	WindowAggState *winstate;
 	ExprContext *econtext;
 	TupleTableSlot *slot;
-	int64		abs_pos;
-	int64		mark_pos;
 
 	Assert(WindowObjectIsValid(winobj));
 	winstate = winobj->winstate;
 	econtext = winstate->ss.ps.ps_ExprContext;
 	slot = winstate->temp_slot_1;
 
+	if (WinGetSlotInFrame(winobj, slot,
+						  relpos, seektype, set_mark,
+						  isnull, isout) == 0)
+	{
+		econtext->ecxt_outertuple = slot;
+		return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
+							econtext, isnull);
+	}
+
+	if (isout)
+		*isout = true;
+	*isnull = true;
+	return (Datum) 0;
+}
+
+/*
+ * WinGetSlotInFrame
+ * slot: TupleTableSlot to store the result
+ * relpos: signed rowcount offset from the seek position
+ * seektype: WINDOW_SEEK_HEAD or WINDOW_SEEK_TAIL
+ * set_mark: If the row is found/in frame and set_mark is true, the mark is
+ *		moved to the row as a side-effect.
+ * isnull: output argument, receives isnull status of result
+ * isout: output argument, set to indicate whether target row position
+ *		is out of frame (can pass NULL if caller doesn't care about this)
+ *
+ * Returns 0 if we successfullt got the slot. false if out of frame.
+ * (also isout is set)
+ */
+int
+WinGetSlotInFrame(WindowObject winobj, TupleTableSlot *slot,
+					 int relpos, int seektype, bool set_mark,
+					 bool *isnull, bool *isout)
+{
+	WindowAggState *winstate;
+	int64		abs_pos;
+	int64		mark_pos;
+
+	Assert(WindowObjectIsValid(winobj));
+	winstate = winobj->winstate;
+
 	switch (seektype)
 	{
 		case WINDOW_SEEK_CURRENT:
@@ -3583,15 +3784,13 @@ WinGetFuncArgInFrame(WindowObject winobj, int argno,
 		*isout = false;
 	if (set_mark)
 		WinSetMarkPosition(winobj, mark_pos);
-	econtext->ecxt_outertuple = slot;
-	return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
-						econtext, isnull);
+	return 0;
 
 out_of_frame:
 	if (isout)
 		*isout = true;
 	*isnull = true;
-	return (Datum) 0;
+	return -1;
 }
 
 /*
@@ -3622,3 +3821,9 @@ WinGetFuncArgCurrent(WindowObject winobj, int argno, bool *isnull)
 	return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
 						econtext, isnull);
 }
+
+WindowAggState *
+WinGetAggState(WindowObject winobj)
+{
+	return winobj->winstate;
+}
diff --git a/src/backend/utils/adt/windowfuncs.c b/src/backend/utils/adt/windowfuncs.c
index b87a624fb2..63b225271c 100644
--- a/src/backend/utils/adt/windowfuncs.c
+++ b/src/backend/utils/adt/windowfuncs.c
@@ -13,6 +13,9 @@
  */
 #include "postgres.h"
 
+#include "catalog/pg_collation_d.h"
+#include "executor/executor.h"
+#include "nodes/execnodes.h"
 #include "nodes/supportnodes.h"
 #include "utils/builtins.h"
 #include "windowapi.h"
@@ -39,7 +42,9 @@ typedef struct
 static bool rank_up(WindowObject winobj);
 static Datum leadlag_common(FunctionCallInfo fcinfo,
 							bool forward, bool withoffset, bool withdefault);
-
+static bool get_slots(WindowObject winobj, WindowAggState *winstate, int current_pos);
+static int evaluate_pattern(WindowObject winobj, WindowAggState *winstate,
+							int relpos, char *vname, char *quantifier, bool *result);
 
 /*
  * utility routine for *_rank functions.
@@ -713,3 +718,270 @@ window_nth_value(PG_FUNCTION_ARGS)
 
 	PG_RETURN_DATUM(result);
 }
+
+/*
+ * rpr
+ * allow to use "Row pattern recognition: WINDOW clause" (SQL:2016 R020) in
+ * the target list.
+ * Usage: SELECT rpr(colname) OVER (..)
+ * where colname is defined in PATTERN clause.
+ */
+Datum
+window_rpr(PG_FUNCTION_ARGS)
+{
+#define	MAX_PATTERNS	16	/* max variables in PATTERN clause */
+
+	WindowObject winobj = PG_WINDOW_OBJECT();
+	WindowAggState	*winstate = WinGetAggState(winobj);
+	Datum		result;
+	bool		expression_result;
+	bool		isnull;
+	int			relpos;
+	int64		curr_pos, markpos;
+	ListCell	*lc, *lc1;
+
+	curr_pos = WinGetCurrentPosition(winobj);
+	elog(DEBUG1, "rpr is called. row: " INT64_FORMAT, curr_pos);
+
+	/*
+	 * Evaluate PATTERN until one of expressions is not true or out of frame.
+	 */
+	relpos = 0;
+
+	forboth(lc, winstate->patternVariableList, lc1, winstate->patternRegexpList)
+	{
+		char	*vname = strVal(lfirst(lc));
+		char	*quantifier = strVal(lfirst(lc1));
+
+		elog(DEBUG1, "relpos: %d pattern vname: %s quantifier: %s", relpos, vname, quantifier);
+
+		/* evaluate row pattern against current row */
+		relpos = evaluate_pattern(winobj, winstate, relpos, vname, quantifier, &expression_result);
+
+		/*
+		 * If the expression did not match, we are done.
+		 */
+		if (!expression_result)
+			break;
+
+		/* out of frame? */
+		if (relpos < 0)
+			break;
+
+		/* count up relative row position */
+		relpos++;
+	}
+
+	elog(DEBUG1, "relpos: %d", relpos);
+
+	/*
+	 * If current row satified the pattern, return argument expression.
+	 */
+	if (expression_result)
+	{
+		result = WinGetFuncArgInFrame(winobj, 0,
+									  0, WINDOW_SEEK_HEAD, false,
+									  &isnull, NULL);
+	}
+
+	/*
+	 * At this point we can set mark down to current pos -2.
+	 */
+	markpos = curr_pos -2;
+	elog(DEBUG1, "markpos: " INT64_FORMAT, markpos);
+	if (markpos >= 0)
+		WinSetMarkPosition(winobj, markpos);
+
+	if (expression_result)
+		PG_RETURN_DATUM(result);
+
+	PG_RETURN_NULL();
+}
+
+/*
+ * Evaluate expression associated with PATTERN variable vname.
+ * relpos is relative row position in a frame (starting from 0).
+ * "quantifier" is the quatifier part of the PATTERN regular expression.
+ * Currently only '+' is allowed.
+ * result is out paramater representing the expression evaluation result
+ * is true of false.
+ * Return values are:
+ * >=0: the last match relative row position
+ * -1: current row is out of frame
+ */
+static
+int evaluate_pattern(WindowObject winobj, WindowAggState *winstate,
+					 int relpos, char *vname, char *quantifier, bool *result)
+{
+	ExprContext	*econtext = winstate->ss.ps.ps_ExprContext;
+	ListCell	*lc1, *lc2;
+	ExprState	*pat;
+	Datum		eval_result;
+	int			sts;
+	bool		out_of_frame = false;
+	bool		isnull;
+	StringInfo	encoded_str = makeStringInfo();
+	char		pattern_str[128];
+
+	forboth (lc1, winstate->defineVariableList, lc2, winstate->defineClauseList)
+	{
+		char	*name = strVal(lfirst(lc1));
+		bool	second_try_match = false;
+
+		if (strcmp(vname, name))
+			continue;
+
+		/* set expression to evaluate */
+		pat = lfirst(lc2);
+
+		for (;;)
+		{
+			if (!get_slots(winobj, winstate, relpos))
+			{
+				out_of_frame = true;
+				break;	/* current row is out of frame */
+			}
+
+			/* evaluate the expression */
+			eval_result = ExecEvalExpr(pat, econtext, &isnull);
+			if (isnull)
+			{
+				/* expression is NULL */
+				elog(DEBUG1, "expression for %s is NULL at row: %d", vname, relpos);
+				break;
+			}
+			else
+			{
+				if (!DatumGetBool(eval_result))
+				{
+					/* expression is false */
+					elog(DEBUG1, "expression for %s is false at row: %d", vname, relpos);
+					break;
+				}
+				else
+				{
+					/* expression is true */
+					elog(DEBUG1, "expression for %s is true at row: %d", vname, relpos);
+					appendStringInfoChar(encoded_str, vname[0]);
+
+					/* If quantifier is "+", we need to look for more matching row */
+					if (quantifier && !strcmp(quantifier, "+"))
+					{
+						/* remember that we want to try another row */
+						second_try_match = true;
+						relpos++;
+					}
+					else
+						break;
+				}
+			}
+		}
+		if (second_try_match)
+			relpos--;
+
+		if (out_of_frame)
+		{
+			*result = false;
+			return -1;
+		}
+
+		/* build regular expression */
+		snprintf(pattern_str, sizeof(pattern_str), "%c%s", vname[0], quantifier);
+
+		/*
+		 * Do regular expression matching against sequence of rows satisfying
+		 * the expression using regexp_instr().
+		 */
+		sts = DatumGetInt32(DirectFunctionCall2Coll(regexp_instr, DEFAULT_COLLATION_OID,
+													PointerGetDatum(cstring_to_text(encoded_str->data)),
+													PointerGetDatum(cstring_to_text(pattern_str))));
+		elog(DEBUG1, "regexp_instr returned: %d. str: %s regexp: %s",
+			 sts, encoded_str->data, pattern_str);
+		*result = (sts > 0)? true : false;
+	}
+	return relpos;
+}
+
+/*
+ * Get current, previous and next tuple.
+ * Returns true if still within frame.
+ */
+static bool
+get_slots(WindowObject winobj, WindowAggState *winstate, int current_pos)
+{
+	TupleTableSlot *slot;
+	bool	isnull, isout;
+	int		sts;
+	ExprContext *econtext;
+
+	econtext = winstate->ss.ps.ps_ExprContext;
+
+	/* for current row */
+	slot = winstate->temp_slot_1;
+	sts = WinGetSlotInFrame(winobj, slot,
+							current_pos, WINDOW_SEEK_HEAD, false,
+							&isnull, &isout);
+	if (sts < 0)
+	{
+		elog(DEBUG1, "current row is out of frame");
+		econtext->ecxt_scantuple = winstate->null_slot;
+		return false;
+	}
+	else
+		econtext->ecxt_scantuple = slot;
+
+	/* for PREV */
+	if (current_pos > 0)
+	{
+		slot = winstate->prev_slot;
+		sts = WinGetSlotInFrame(winobj, slot,
+								current_pos - 1, WINDOW_SEEK_HEAD, false,
+								&isnull, &isout);
+		if (sts < 0)
+		{
+			elog(DEBUG1, "previous row out of frame at: %d", current_pos);
+			econtext->ecxt_outertuple = winstate->null_slot;
+		}
+		econtext->ecxt_outertuple = slot;
+	}
+	else
+		econtext->ecxt_outertuple = winstate->null_slot;
+
+	/* for NEXT */
+	slot = winstate->next_slot;
+	sts = WinGetSlotInFrame(winobj, slot,
+							current_pos + 1, WINDOW_SEEK_HEAD, false,
+							&isnull, &isout);
+	if (sts < 0)
+	{
+		elog(DEBUG1, "next row out of frame at: %d", current_pos);
+		econtext->ecxt_innertuple = winstate->null_slot;
+	}
+	else
+		econtext->ecxt_innertuple = slot;
+
+	return true;
+}
+
+/*
+ * prev
+ * Dummy function to invoke RPR's navigation operator "PREV".
+ * This is *not* a window function.
+ */
+Datum
+window_prev(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_DATUM(PG_GETARG_DATUM(0));
+}
+
+/*
+ * next
+ * Dummy function to invoke RPR's navigation operation "NEXT".
+ * This is *not* a window function.
+ */
+Datum
+window_next(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_DATUM(PG_GETARG_DATUM(0));
+}
+
diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat
index 6996073989..e3a9e0ffeb 100644
--- a/src/include/catalog/pg_proc.dat
+++ b/src/include/catalog/pg_proc.dat
@@ -10397,6 +10397,15 @@
 { oid => '3114', descr => 'fetch the Nth row value',
   proname => 'nth_value', prokind => 'w', prorettype => 'anyelement',
   proargtypes => 'anyelement int4', prosrc => 'window_nth_value' },
+{ oid => '6122', descr => 'row pattern recognition in window',
+  proname => 'rpr', prokind => 'w', prorettype => 'anyelement',
+  proargtypes => 'anyelement', prosrc => 'window_rpr' },
+{ oid => '6123', descr => 'previous value',
+  proname => 'prev', provolatile => 's', prorettype => 'anyelement',
+  proargtypes => 'anyelement', prosrc => 'window_prev' },
+{ oid => '6124', descr => 'next value',
+  proname => 'next', provolatile => 's', prorettype => 'anyelement',
+  proargtypes => 'anyelement', prosrc => 'window_next' },
 
 # functions for range types
 { oid => '3832', descr => 'I/O',
diff --git a/src/include/nodes/execnodes.h b/src/include/nodes/execnodes.h
index cb714f4a19..bc95c5fe9b 100644
--- a/src/include/nodes/execnodes.h
+++ b/src/include/nodes/execnodes.h
@@ -2519,6 +2519,13 @@ typedef struct WindowAggState
 	int64		groupheadpos;	/* current row's peer group head position */
 	int64		grouptailpos;	/* " " " " tail position (group end+1) */
 
+	/* these fields are used in Row pattern recognition: */
+	List	   *patternVariableList;	/* list of row pattern variables names (list of String) */
+	List	   *patternRegexpList;	/* list of row pattern regular expressions ('*', '+' or '?'. list of String) */
+	List	   *defineVariableList;	/* list of row pattern definition variables (list of String) */
+	List	   *defineClauseList;	/* expression for row pattern definition
+									 * search conditions ExprState list */
+
 	MemoryContext partcontext;	/* context for partition-lifespan data */
 	MemoryContext aggcontext;	/* shared context for aggregate working data */
 	MemoryContext curaggcontext;	/* current aggregate's working data */
@@ -2555,6 +2562,11 @@ typedef struct WindowAggState
 	TupleTableSlot *agg_row_slot;
 	TupleTableSlot *temp_slot_1;
 	TupleTableSlot *temp_slot_2;
+
+	/* temporary slots for RPR */
+	TupleTableSlot *prev_slot;	/* PREV row navigation operator */
+	TupleTableSlot *next_slot;	/* NEXT row navigation operator */
+	TupleTableSlot *null_slot;	/* all NULL slot */
 } WindowAggState;
 
 /* ----------------
diff --git a/src/include/windowapi.h b/src/include/windowapi.h
index b8c2c565d1..a0facf38fe 100644
--- a/src/include/windowapi.h
+++ b/src/include/windowapi.h
@@ -58,7 +58,16 @@ extern Datum WinGetFuncArgInFrame(WindowObject winobj, int argno,
 								  int relpos, int seektype, bool set_mark,
 								  bool *isnull, bool *isout);
 
+extern int WinGetSlotInFrame(WindowObject winobj, TupleTableSlot *slot,
+							 int relpos, int seektype, bool set_mark,
+							 bool *isnull, bool *isout);
+
 extern Datum WinGetFuncArgCurrent(WindowObject winobj, int argno,
 								  bool *isnull);
 
+extern WindowAggState *WinGetAggState(WindowObject winobj);
+
+extern bool window_gettupleslot(WindowObject winobj, int64 pos, TupleTableSlot *slot);
+
+
 #endif							/* WINDOWAPI_H */
-- 
2.25.1


----Next_Part(Sun_Jun_25_21_05_09_2023_126)--
Content-Type: Text/X-Patch; charset=us-ascii
Content-Transfer-Encoding: 7bit
Content-Disposition: inline;
 filename="v1-0005-Row-pattern-recognition-patch-docs.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread

* [PATCH v4 2/8] Address space reservation for shared memory
@ 2024-10-16 18:21  Dmitrii Dolgov <9erthalion6@gmail.com>
  0 siblings, 0 replies; 213+ messages in thread

From: Dmitrii Dolgov @ 2024-10-16 18:21 UTC (permalink / raw)

Currently the kernel is responsible to chose an address, where to place each
shared memory mapping, which is the lowest possible address that do not clash
with any other mappings. This is considered to be the most portable approach,
but one of the downsides is that there is no place to resize allocated mappings
anymore. Here is how it looks like for one mapping in /proc/$PID/maps,
/dev/zero represents the anonymous shared memory we talk about:

    00400000-00490000         /path/bin/postgres
    ...
    012d9000-0133e000         [heap]
    7f443a800000-7f470a800000 /dev/zero (deleted)
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
    ...
    7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
    7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)

By specifying the mapping address directly it's possible to place the
mapping in a way that leaves room for resizing. The idea is:

* To reserve some address space via mmap'ing a large chunk of memory
  with PROT_NONE and MAP_NORESERVE. This way we prepare a playground for
  preparing shared memory layout without risking anything interfering
  with that.

* To slice the reserved space up into sections, one to use for each
  shared segment.

* Allocate shared memory segments out of corresponding slices and
  leaving unclaimed space in between them. This is implemented via
  mmap'ing memory at a specified address from the reserved space with
  MAP_FIXED.

The result looks like this:

    012d9000-0133e000         [heap]
    7f443a800000-7f444196c000 /dev/zero (deleted)
    7f444196c000-7f470a800000                     # reserved space
    7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
    7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2

Things like address space randomization should not be a problem in this
context, since the randomization is applied to the mmap base, which is
one per process.

This approach also do not impact the actual memory usage as reported by
the kernel. Here is the output of /proc/$PID/status for the master
version with shared_buffers = 128 MB:

    // Peak virtual memory size, which is described as total pages
    // mapped in mm_struct. It corresponds to the mapped reserved space
    // and is the only number that grows with it.
    VmPeak:          2043192 kB
    // Size of memory portions. It contains RssAnon + RssFile + RssShmem
    VmRSS:             22908 kB
    // Size of resident anonymous memory
    RssAnon:             768 kB
    // Size of resident file mappings
    RssFile:           10364 kB
    // Size of resident shmem memory (includes SysV shm, mapping of tmpfs and
    // shared anonymous mappings)
    RssShmem:          11776 kB

Here is the same for the patch when reserving 20GB of space:

    VmPeak:         21250648 kB
    VmRSS:             22948 kB
    RssAnon:             768 kB
    RssFile:           10404 kB
    RssShmem:          11776 kB

Cgroup v2 doesn't have any problems with that as well. To verify a new cgroup
was created with the memory limit 256 MB, then PostgreSQL was launched withing
this cgroup with shared_buffers = 128 MB:

    $ cd /sys/fs/cgroup
    $ mkdir postgres
    $ cd postres
    $ echo 268435456 > memory.max

    $ echo $MASTER_PID_SHELL > cgroup.procs
    # postgres from the master branch has being successfully launched
    #  from that shell
    $ cat memory.current
    17465344 (~16.6 MB)
    # stop postgres

    $ echo $PATCH_PID_SHELL > cgroup.procs
    # postgres from the patch has being successfully launched from that shell
    $ cat memory.current
    17637376 (~16.8 MB)

To control the amount of space reserved a new GUC max_available_memory
is introduced. Ideally it should be based on the maximum available
memory, hense the name.
---
 src/backend/port/sysv_shmem.c       | 284 ++++++++++++++++++++++++----
 src/backend/port/win32_shmem.c      |   2 +-
 src/backend/storage/ipc/ipci.c      |   5 +-
 src/backend/utils/init/globals.c    |   1 +
 src/backend/utils/misc/guc_tables.c |  14 ++
 src/include/storage/pg_shmem.h      |   4 +-
 6 files changed, 271 insertions(+), 39 deletions(-)

diff --git a/src/backend/port/sysv_shmem.c b/src/backend/port/sysv_shmem.c
index 56af0231d24..a0f03ff868f 100644
--- a/src/backend/port/sysv_shmem.c
+++ b/src/backend/port/sysv_shmem.c
@@ -108,6 +108,66 @@ static AnonymousMapping Mappings[ANON_MAPPINGS];
 /* Keeps track of used mapping segments */
 static int next_free_segment = 0;
 
+/*
+ * Anonymous mapping placing (/dev/zero (deleted) below) looks like this:
+ *
+ * 00400000-00490000         /path/bin/postgres
+ * ...
+ * 012d9000-0133e000         [heap]
+ * 7f443a800000-7f470a800000 /dev/zero (deleted)
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * ...
+ * 7f471aef2000-7f471aef9000 /dev/shm/PostgreSQL.3859891842
+ * 7f471aef9000-7f471aefa000 /SYSV007dbf7d (deleted)
+ * ...
+ *
+ * We would like to place multiple mappings in such a way, that there will be
+ * enough space between them in the address space to be able to resize up to
+ * certain size, but without counting towards the total memory consumption.
+ *
+ * To achieve that we first reserve some shared memory address space by
+ * mmap'ing a segment of MaxAvailableMemory size with PROT_NONE and
+ * MAP_NORESERVE (these flags allow to make sure this space will not be used by
+ * anything else, yet do not count against memory limits). Having the reserved
+ * space, we allocate out of it actual chunks of shared memory as usual,
+ * updating a pointer to the current available reserved space for the next
+ * allocation with the gap between segments in mind.
+ *
+ * The result would look like this:
+ *
+ * 012d9000-0133e000         [heap]
+ * 7f4426f54000-7f442e010000 /dev/zero (deleted)
+ * 7f442e010000-7f443a800000                     # reserved empty space
+ * 7f443a800000-7f444196c000 /dev/zero (deleted)
+ * 7f444196c000-7f470a800000                     # reserved empty space
+ * 7f470a800000-7f471831d000 /usr/lib/locale/locale-archive
+ * 7f4718400000-7f4718401000 /usr/lib64/libicudata.so.74.2
+ * [...]
+ *
+ * The reserved space pointer is calculated to slice up the total reserved
+ * space into fixed fractions of address space for each segment, as specified
+ * in the SHMEM_RESIZE_RATIO array.
+ */
+static double SHMEM_RESIZE_RATIO[1] = {
+	1.0, 									/* MAIN_SHMEM_SLOT */
+};
+
+/*
+ * Offset from the beginning of the reserved space, which indicates currently
+ * available range. New shared memory segments have to be allocated at this
+ * offset related to the reserved space.
+ */
+static Size reserved_offset = 0;
+
+/*
+ * Flag telling that we have decided to use huge pages.
+ *
+ * XXX: It's possible to use GetConfigOption("huge_pages_status", false, false)
+ * instead, but it feels like an overkill.
+ */
+static bool huge_pages_on = false;
+
 static void *InternalIpcMemoryCreate(IpcMemoryKey memKey, Size size);
 static void IpcMemoryDetach(int status, Datum shmaddr);
 static void IpcMemoryDelete(int status, Datum shmId);
@@ -626,39 +686,198 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
  *
  * This function will modify mapping size to the actual size of the allocation,
  * if it ends up allocating a segment that is larger than requested.
+ *
+ * Note that we do not switch from huge pages to regular pages in this
+ * function, this decision was already made in ReserveAnonymousMemory and we
+ * stick to it.
  */
 static void
-CreateAnonymousSegment(AnonymousMapping *mapping)
+CreateAnonymousSegment(AnonymousMapping *mapping, Pointer base)
 {
 	Size		allocsize = mapping->shmem_size;
 	void	   *ptr = MAP_FAILED;
 	int			mmap_errno = 0;
+	int			mmap_flags = PG_MMAP_FLAGS;
 
 #ifndef MAP_HUGETLB
-	/* PGSharedMemoryCreate should have dealt with this case */
-	Assert(huge_pages != HUGE_PAGES_ON);
+	/* ReserveAnonymousMemory should have dealt with this case */
+	Assert(huge_pages != HUGE_PAGES_ON && !huge_pages_on);
 #else
-	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	if (huge_pages_on)
 	{
-		/*
-		 * Round up the request size to a suitable large value.
-		 */
 		Size		hugepagesize;
-		int			mmap_flags;
 
+		/* Make sure nothing is messed up */
+		Assert(huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY);
+
+		/* Round up the request size to a suitable large value */
 		GetHugePageSize(&hugepagesize, &mmap_flags);
 
 		if (allocsize % hugepagesize != 0)
 			allocsize += hugepagesize - (allocsize % hugepagesize);
 
+		mmap_flags = PG_MMAP_FLAGS | mmap_flags;
+	}
+#endif
+
+	elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p",
+		 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
+	/*
+	 * Try to create mapping at an address out of the reserved range, which
+	 * will allow to extend it later. Use reserved_offset to allocate the
+	 * segment, then update currently available reserved range.
+	 *
+	 * If the last step has failed, fallback to the regular mapping
+	 * creation and signal that shared buffers could not be resized without
+	 * a restart.
+	 */
+	ptr = mmap(base + reserved_offset, allocsize, PROT_READ | PROT_WRITE,
+			   mmap_flags | MAP_FIXED, -1, 0);
+	mmap_errno = errno;
+
+	if (ptr == MAP_FAILED)
+	{
+		DebugMappings();
+		elog(DEBUG1, "segment[%s]: mmap(%zu) at address %p failed: %m, "
+					 "fallback to the non-resizable allocation",
+			 MappingName(mapping->shmem_segment), allocsize, base + reserved_offset);
+
 		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS | mmap_flags, -1, 0);
+						   PG_MMAP_FLAGS, -1, 0);
+		mmap_errno = errno;
+	}
+	else
+	{
+		Size total_reserved = (Size) MaxAvailableMemory * BLCKSZ;
+
+		reserved_offset += total_reserved * SHMEM_RESIZE_RATIO[next_free_segment];
+	}
+
+	if (ptr == MAP_FAILED)
+	{
+		errno = mmap_errno;
+		DebugMappings();
+		ereport(FATAL,
+				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
+						MappingName(mapping->shmem_segment)),
+				 (mmap_errno == ENOMEM) ?
+				 errhint("This error usually means that PostgreSQL's request "
+						 "for a shared memory segment exceeded available memory, "
+						 "swap space, or huge pages. To reduce the request size "
+						 "(currently %zu bytes), reduce PostgreSQL's shared "
+						 "memory usage, perhaps by reducing \"shared_buffers\" or "
+						 "\"max_connections\".",
+						 allocsize) : 0));
+	}
+
+	mapping->shmem = ptr;
+	mapping->shmem_size = allocsize;
+}
+
+/*
+ * ReserveAnonymousMemory
+ *
+ * Reserve shared memory address space, from which shared memory segments are
+ * going to be sliced out. The goal of this exercise is to support segments
+ * resizing, for which we need a reserved space free of potential clashes with
+ * other mmap'd areas that are not under our control. Reservation is done via
+ * mmap, and will not allocate any memory until it will be actually used, and
+ * MAP_NORESERVE allows to make it not counting againt kernel reservation
+ * limits (e.g. in cgroups or for huge pages). Do not get confused because of
+ * MAP_NORESERVE -- we need to reserve some space, but not the actual memory,
+ * and that is that this flag is about.
+ *
+ * Note, that with MAP_NORESERVE a reservation with hugetlb will succeed even
+ * if there is actually not enough huge pages. Hence this function is
+ * responsible for deciding whether to use huge pages or not. To achieve that
+ * we need to probe first and try to allocate needed memory for all segments --
+ * if this succeeds, we unmap the probe segment and use hugetlb; if it fails,
+ * we proceed with the regular memory.
+ */
+void *
+ReserveAnonymousMemory(Size reserve_size)
+{
+	Size		allocsize = reserve_size;
+	void	   *ptr = MAP_FAILED;
+	int			mmap_errno = 0;
+
+	/* Complain if hugepages demanded but we can't possibly support them */
+#if !defined(MAP_HUGETLB)
+	if (huge_pages == HUGE_PAGES_ON)
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("huge pages not supported on this platform")));
+#else
+	if (huge_pages == HUGE_PAGES_ON || huge_pages == HUGE_PAGES_TRY)
+	{
+		Size		hugepagesize, total_size = 0;
+		int			mmap_flags;
+
+		GetHugePageSize(&hugepagesize, &mmap_flags);
+
+		/*
+		 * Figure out how much memory is needed for all segments, keeping in
+		 * mind that for every segment this value will be rounding up by the
+		 * huge page size. The resulting value will be used to probe memory and
+		 * decide whether we will allocate huge pages or not.
+		 *
+		 * We could actually have a mix and match of segments with and without
+		 * huge pages. But in that case we need to have multiple reservation
+		 * spaces to use corresponding memory (hugetlb adress space reserved
+		 * for hugetlb segments, regular memory for others), and it doesn't
+		 * seem to worth the complexity for now.
+		 */
+		for(int segment = 0; segment < ANON_MAPPINGS; segment++)
+		{
+			int	numSemas;
+			Size segment_size = CalculateShmemSize(&numSemas, segment);
+
+			if (segment_size % hugepagesize != 0)
+				segment_size += hugepagesize - (segment_size % hugepagesize);
+
+			total_size += segment_size;
+		}
+
+		/* Map total amount of memory to test its availability. */
+		elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB",
+					 total_size);
+		ptr = mmap(NULL, total_size, PROT_NONE,
+				   PG_MMAP_FLAGS | MAP_ANONYMOUS | mmap_flags, -1, 0);
 		mmap_errno = errno;
 		if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
 		{
-			DebugMappings();
-			elog(DEBUG1, "segment[%s]: mmap(%zu) with MAP_HUGETLB failed, huge pages disabled: %m",
-				 MappingName(mapping->shmem_segment), allocsize);
+			/* No huge pages, we will go with the regular page size */
+			elog(DEBUG1, "reserving space: probe mmap(%zu) with MAP_HUGETLB "
+						 "failed, huge pages disabled: %m", total_size);
+		}
+		else
+		{
+			/*
+			 * All fine, unmap the temporary segment and proceed with reserving
+			 * using huge pages.
+			 */
+			if (munmap(ptr, total_size) < 0)
+				elog(LOG, "reservice space: munmap(%p, %zu) failed: %m",
+					 ptr, total_size);
+
+			/* Round up the requested size to a suitable large value. */
+			if (allocsize % hugepagesize != 0)
+				allocsize += hugepagesize - (allocsize % hugepagesize);
+
+			elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB",
+						 allocsize);
+			ptr = mmap(NULL, allocsize, PROT_NONE,
+					   PG_MMAP_FLAGS | MAP_ANONYMOUS | MAP_NORESERVE | mmap_flags,
+					   -1, 0);
+			mmap_errno = errno;
+
+			/* This should not happen, but handle errors anyway */
+			if (huge_pages == HUGE_PAGES_TRY && ptr == MAP_FAILED)
+			{
+				elog(DEBUG1, "reserving space: mmap(%zu) with MAP_HUGETLB "
+							 "failed, huge pages disabled: %m", allocsize);
+			}
 		}
 	}
 #endif
@@ -666,10 +885,12 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 	/*
 	 * Report whether huge pages are in use.  This needs to be tracked before
 	 * the second mmap() call if attempting to use huge pages failed
-	 * previously.
+	 * previously. At this point ptr is either pointing to the probe segment,
+	 * if we couldn't mmap it, or the reservation space.
 	 */
 	SetConfigOption("huge_pages_status", (ptr == MAP_FAILED) ? "off" : "on",
 					PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
+	huge_pages_on = ptr != MAP_FAILED;
 
 	if (ptr == MAP_FAILED && huge_pages != HUGE_PAGES_ON)
 	{
@@ -677,10 +898,11 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		 * Use the original size, not the rounded-up value, when falling back
 		 * to non-huge pages.
 		 */
-		allocsize = mapping->shmem_size;
-		ptr = mmap(NULL, allocsize, PROT_READ | PROT_WRITE,
-				   PG_MMAP_FLAGS, -1, 0);
-		mmap_errno = errno;
+		allocsize = reserve_size;
+
+		elog(DEBUG1, "reserving space: mmap(%zu)", allocsize);
+		ptr = mmap(NULL, allocsize, PROT_NONE,
+				   MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
 	}
 
 	if (ptr == MAP_FAILED)
@@ -688,20 +910,18 @@ CreateAnonymousSegment(AnonymousMapping *mapping)
 		errno = mmap_errno;
 		DebugMappings();
 		ereport(FATAL,
-				(errmsg("segment[%s]: could not map anonymous shared memory: %m",
-						MappingName(mapping->shmem_segment)),
+				(errmsg("reserving space: could not map anonymous shared "
+						"memory: %m"),
 				 (mmap_errno == ENOMEM) ?
 				 errhint("This error usually means that PostgreSQL's request "
-						 "for a shared memory segment exceeded available memory, "
-						 "swap space, or huge pages. To reduce the request size "
-						 "(currently %zu bytes), reduce PostgreSQL's shared "
-						 "memory usage, perhaps by reducing \"shared_buffers\" or "
-						 "\"max_connections\".",
+						 "for a reserved shared memory address space exceeded "
+						 "available memory, swap space, or huge pages. To "
+						 "reduce the request reservation size (currently %zu "
+						 "bytes), reduce PostgreSQL's \"maximum_shared_buffers\".",
 						 allocsize) : 0));
 	}
 
-	mapping->shmem = ptr;
-	mapping->shmem_size = allocsize;
+	return ptr;
 }
 
 /*
@@ -740,7 +960,7 @@ AnonymousShmemDetach(int status, Datum arg)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	IpcMemoryKey NextShmemSegID;
 	void	   *memAddress;
@@ -760,14 +980,6 @@ PGSharedMemoryCreate(Size size,
 				 errmsg("could not stat data directory \"%s\": %m",
 						DataDir)));
 
-	/* Complain if hugepages demanded but we can't possibly support them */
-#if !defined(MAP_HUGETLB)
-	if (huge_pages == HUGE_PAGES_ON)
-		ereport(ERROR,
-				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-				 errmsg("huge pages not supported on this platform")));
-#endif
-
 	/* For now, we don't support huge pages in SysV memory */
 	if (huge_pages == HUGE_PAGES_ON && shared_memory_type != SHMEM_TYPE_MMAP)
 		ereport(ERROR,
@@ -782,7 +994,7 @@ PGSharedMemoryCreate(Size size,
 	if (shared_memory_type == SHMEM_TYPE_MMAP)
 	{
 		/* On success, mapping data will be modified. */
-		CreateAnonymousSegment(mapping);
+		CreateAnonymousSegment(mapping, base);
 
 		next_free_segment++;
 
diff --git a/src/backend/port/win32_shmem.c b/src/backend/port/win32_shmem.c
index 4dee856d6bd..ce719f1b412 100644
--- a/src/backend/port/win32_shmem.c
+++ b/src/backend/port/win32_shmem.c
@@ -205,7 +205,7 @@ EnableLockPagesPrivilege(int elevel)
  */
 PGShmemHeader *
 PGSharedMemoryCreate(Size size,
-					 PGShmemHeader **shim)
+					 PGShmemHeader **shim, Pointer base)
 {
 	void	   *memAddress;
 	PGShmemHeader *hdr;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 8b38e985327..076888c0172 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -203,9 +203,12 @@ CreateSharedMemoryAndSemaphores(void)
 	PGShmemHeader *seghdr;
 	Size		size;
 	int			numSemas;
+	void 		*base;
 
 	Assert(!IsUnderPostmaster);
 
+	base = ReserveAnonymousMemory((Size) MaxAvailableMemory * BLCKSZ);
+
 	for(int segment = 0; segment < ANON_MAPPINGS; segment++)
 	{
 		/* Compute the size of the shared-memory block */
@@ -217,7 +220,7 @@ CreateSharedMemoryAndSemaphores(void)
 		 *
 		 * XXX: Do multiple shims are needed, one per segment?
 		 */
-		seghdr = PGSharedMemoryCreate(size, &shim);
+		seghdr = PGSharedMemoryCreate(size, &shim, base);
 
 		/*
 		 * Make sure that huge pages are never reported as "unknown" while the
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 2152aad97d9..1d42a5856c0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -140,6 +140,7 @@ int			max_parallel_maintenance_workers = 2;
  * register background workers.
  */
 int			NBuffers = 16384;
+int			MaxAvailableMemory = 131072;
 int			MaxConnections = 100;
 int			max_worker_processes = 8;
 int			max_parallel_workers = 8;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 4eaeca89f2c..dede37f7905 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2364,6 +2364,20 @@ struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"max_available_memory", PGC_SIGHUP, RESOURCES_MEM,
+			gettext_noop("Sets the upper limit for the shared_buffers value."),
+			gettext_noop("Shared memory could be resized at runtime, this "
+						 "parameters sets the upper limit for it, beyond which "
+						 "resizing would not be supported. Normally this value "
+						 "would be the same as the total available memory."),
+			GUC_UNIT_BLOCKS
+		},
+		&MaxAvailableMemory,
+		131072, 16, INT_MAX / 2,
+		NULL, NULL, NULL
+	},
+
 	{
 		{"vacuum_buffer_usage_limit", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the buffer pool size for VACUUM, ANALYZE, and autovacuum."),
diff --git a/src/include/storage/pg_shmem.h b/src/include/storage/pg_shmem.h
index 138078c29c5..4a83e255652 100644
--- a/src/include/storage/pg_shmem.h
+++ b/src/include/storage/pg_shmem.h
@@ -60,6 +60,7 @@ extern PGDLLIMPORT ShmemSegment Segments[ANON_MAPPINGS];
 extern PGDLLIMPORT int shared_memory_type;
 extern PGDLLIMPORT int huge_pages;
 extern PGDLLIMPORT int huge_page_size;
+extern PGDLLIMPORT int MaxAvailableMemory;
 
 /* Possible values for huge_pages and huge_pages_status */
 typedef enum
@@ -100,10 +101,11 @@ extern void PGSharedMemoryNoReAttach(void);
 #endif
 
 extern PGShmemHeader *PGSharedMemoryCreate(Size size,
-										   PGShmemHeader **shim);
+										   PGShmemHeader **shim, Pointer base);
 extern bool PGSharedMemoryIsInUse(unsigned long id1, unsigned long id2);
 extern void PGSharedMemoryDetach(void);
 extern void GetHugePageSize(Size *hugepagesize, int *mmap_flags);
+void *ReserveAnonymousMemory(Size reserve_size);
 
 /* The main segment, contains everything except buffer blocks and related data. */
 #define MAIN_SHMEM_SEGMENT 0
-- 
2.45.1


--vninua6xybvzgrci
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
	filename="v4-0003-Introduce-multiple-shmem-segments-for-shared-buff.patch"



^ permalink  raw  reply  [nested|flat] 213+ messages in thread