agora inbox for pgsql-hackers@postgresql.org
help / color / mirror / Atom feedFrom: Alvaro Herrera <alvherre@alvh.no-ip.org>
Subject: [PATCH v10] Make XLogCtl->LogwrtResult accessible with atomics
Date: Tue, 2 Feb 2021 14:03:43 -0300
Currently, access to LogwrtResult is protected by a spinlock. This
becomes severely contended in some scenarios, such as with a largish
replication flock: walsenders all calling GetFlushRecPtr repeatedly
cause the processor heat up to the point where eggs can be fried on top.
This can be reduced to a non-problem by replacing XLogCtl->LogwrtResult
with a struct containing a pair of atomically accessed variables. Do so.
In a few places, we can adjust the exact location where the locals are
updated to account for the fact that we no longer need the spinlock.
Author: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/20200831182156.GA3983@alvherre.pgsql
---
src/backend/access/transam/xlog.c | 106 ++++++++++++++----------------
src/include/port/atomics.h | 29 ++++++++
src/tools/pgindent/typedefs.list | 1 +
3 files changed, 78 insertions(+), 58 deletions(-)
diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c
index 15ab8d90d4..29ebd38103 100644
--- a/src/backend/access/transam/xlog.c
+++ b/src/backend/access/transam/xlog.c
@@ -287,16 +287,13 @@ static bool doPageWrites;
*
* LogwrtRqst indicates a byte position that we need to write and/or fsync
* the log up to (all records before that point must be written or fsynced).
- * LogwrtResult indicates the byte positions we have already written/fsynced.
- * These structs are identical but are declared separately to indicate their
- * slightly different functions.
+ * LogWrtResult indicates the byte positions we have already written/fsynced.
+ * These structs are similar but are declared separately to indicate their
+ * slightly different functions; in addition, the latter is read and written
+ * using atomic operations.
*
- * To read XLogCtl->LogwrtResult, you must hold either info_lck or
- * WALWriteLock. To update it, you need to hold both locks. The point of
- * this arrangement is that the value can be examined by code that already
- * holds WALWriteLock without needing to grab info_lck as well. In addition
- * to the shared variable, each backend has a private copy of LogwrtResult,
- * which is updated when convenient.
+ * In addition to the shared variable, each backend has a private copy of
+ * LogwrtResult, each member of which is separately updated when convenient.
*
* The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
* (protected by info_lck), but we don't need to cache any copies of it.
@@ -319,6 +316,12 @@ static bool doPageWrites;
*----------
*/
+typedef struct XLogwrtAtomic
+{
+ pg_atomic_uint64 Write; /* last byte + 1 written out */
+ pg_atomic_uint64 Flush; /* last byte + 1 flushed */
+} XLogwrtAtomic;
+
typedef struct XLogwrtRqst
{
XLogRecPtr Write; /* last byte + 1 to write out */
@@ -456,6 +459,7 @@ typedef struct XLogCtlData
{
XLogCtlInsert Insert;
+ XLogwrtAtomic LogwrtResult; /* uses atomics */
/* Protected by info_lck: */
XLogwrtRqst LogwrtRqst;
XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
@@ -473,12 +477,6 @@ typedef struct XLogCtlData
pg_time_t lastSegSwitchTime;
XLogRecPtr lastSegSwitchLSN;
- /*
- * Protected by info_lck and WALWriteLock (you must hold either lock to
- * read it, but both to update)
- */
- XLogwrtResult LogwrtResult;
-
/*
* Latest initialized page in the cache (last byte position + 1).
*
@@ -598,7 +596,7 @@ static ControlFileData *ControlFile = NULL;
static int UsableBytesInSegment;
/*
- * Private, possibly out-of-date copy of shared LogwrtResult.
+ * Private, possibly out-of-date copy of shared XLogCtl->LogwrtResult.
* See discussion above.
*/
static XLogwrtResult LogwrtResult = {0, 0};
@@ -907,8 +905,6 @@ XLogInsertRecord(XLogRecData *rdata,
/* advance global request to include new block(s) */
if (XLogCtl->LogwrtRqst.Write < EndPos)
XLogCtl->LogwrtRqst.Write = EndPos;
- /* update local result copy while I have the chance */
- LogwrtResult = XLogCtl->LogwrtResult;
SpinLockRelease(&XLogCtl->info_lck);
}
@@ -1786,6 +1782,7 @@ AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
* Now that we have the lock, check if someone initialized the page
* already.
*/
+ LogwrtResult.Write = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
while (upto >= XLogCtl->InitializedUpTo || opportunistic)
{
nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
@@ -1805,17 +1802,18 @@ AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
if (opportunistic)
break;
- /* Before waiting, get info_lck and update LogwrtResult */
+ /* Advance shared memory write request position */
SpinLockAcquire(&XLogCtl->info_lck);
if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
- LogwrtResult = XLogCtl->LogwrtResult;
SpinLockRelease(&XLogCtl->info_lck);
/*
- * Now that we have an up-to-date LogwrtResult value, see if we
- * still need to write it or if someone else already did.
+ * Before waiting, update LogwrtResult.Write and see if we still need
+ * to write it or if someone else already did.
*/
+ LogwrtResult.Write =
+ pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
if (LogwrtResult.Write < OldPageRqstPtr)
{
/*
@@ -1830,7 +1828,8 @@ AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
- LogwrtResult = XLogCtl->LogwrtResult;
+ LogwrtResult.Write =
+ pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
if (LogwrtResult.Write >= OldPageRqstPtr)
{
/* OK, someone wrote it already */
@@ -2076,7 +2075,7 @@ XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
/*
* Update local LogwrtResult (caller probably did this already, but...)
*/
- LogwrtResult = XLogCtl->LogwrtResult;
+ LogwrtResult.Write = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
/*
* Since successive pages in the xlog cache are consecutively allocated,
@@ -2291,6 +2290,10 @@ XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
Assert(npages == 0);
+ /* Publish current write result position */
+ pg_atomic_monotonic_advance_u64(&XLogCtl->LogwrtResult.Write,
+ LogwrtResult.Write);
+
/*
* If asked to flush, do so
*/
@@ -2327,22 +2330,9 @@ XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
LogwrtResult.Flush = LogwrtResult.Write;
}
- /*
- * Update shared-memory status
- *
- * We make sure that the shared 'request' values do not fall behind the
- * 'result' values. This is not absolutely essential, but it saves some
- * code in a couple of places.
- */
- {
- SpinLockAcquire(&XLogCtl->info_lck);
- XLogCtl->LogwrtResult = LogwrtResult;
- if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
- XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
- if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
- XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
- SpinLockRelease(&XLogCtl->info_lck);
- }
+ /* Publish current flush result position */
+ pg_atomic_monotonic_advance_u64(&XLogCtl->LogwrtResult.Flush,
+ LogwrtResult.Flush);
}
/*
@@ -2356,8 +2346,8 @@ XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
XLogRecPtr WriteRqstPtr = asyncXactLSN;
bool sleeping;
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
SpinLockAcquire(&XLogCtl->info_lck);
- LogwrtResult = XLogCtl->LogwrtResult;
sleeping = XLogCtl->WalWriterSleeping;
if (XLogCtl->asyncXactLSN < asyncXactLSN)
XLogCtl->asyncXactLSN = asyncXactLSN;
@@ -2558,14 +2548,14 @@ XLogFlush(XLogRecPtr record)
{
XLogRecPtr insertpos;
- /* read LogwrtResult and update local state */
+ /* Skip ahead write request pointer to latest written globally */
SpinLockAcquire(&XLogCtl->info_lck);
if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
- LogwrtResult = XLogCtl->LogwrtResult;
SpinLockRelease(&XLogCtl->info_lck);
/* done already? */
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
if (record <= LogwrtResult.Flush)
break;
@@ -2593,7 +2583,7 @@ XLogFlush(XLogRecPtr record)
}
/* Got the lock; recheck whether request is satisfied */
- LogwrtResult = XLogCtl->LogwrtResult;
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
if (record <= LogwrtResult.Flush)
{
LWLockRelease(WALWriteLock);
@@ -2717,7 +2707,6 @@ XLogBackgroundFlush(void)
/* read LogwrtResult and update local state */
SpinLockAcquire(&XLogCtl->info_lck);
- LogwrtResult = XLogCtl->LogwrtResult;
WriteRqst = XLogCtl->LogwrtRqst;
SpinLockRelease(&XLogCtl->info_lck);
@@ -2725,8 +2714,10 @@ XLogBackgroundFlush(void)
WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
/* if we have already flushed that far, consider async commit records */
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
if (WriteRqst.Write <= LogwrtResult.Flush)
{
+ pg_memory_barrier();
SpinLockAcquire(&XLogCtl->info_lck);
WriteRqst.Write = XLogCtl->asyncXactLSN;
SpinLockRelease(&XLogCtl->info_lck);
@@ -2742,6 +2733,7 @@ XLogBackgroundFlush(void)
{
if (openLogFile >= 0)
{
+ LogwrtResult.Write = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
wal_segment_size))
{
@@ -2801,7 +2793,8 @@ XLogBackgroundFlush(void)
/* now wait for any in-progress insertions to finish and get write lock */
WaitXLogInsertionsToFinish(WriteRqst.Write);
LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
- LogwrtResult = XLogCtl->LogwrtResult;
+ LogwrtResult.Write = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
if (WriteRqst.Write > LogwrtResult.Write ||
WriteRqst.Flush > LogwrtResult.Flush)
{
@@ -2889,9 +2882,7 @@ XLogNeedsFlush(XLogRecPtr record)
return false;
/* read LogwrtResult and update local state */
- SpinLockAcquire(&XLogCtl->info_lck);
- LogwrtResult = XLogCtl->LogwrtResult;
- SpinLockRelease(&XLogCtl->info_lck);
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
/* check again */
if (record <= LogwrtResult.Flush)
@@ -5487,10 +5478,13 @@ StartupXLOG(void)
XLogCtl->InitializedUpTo = EndOfLog;
}
+ /*
+ * Update local and shared status. This is OK to do without any locks
+ * because no other process can be reading or writing WAL yet.
+ */
LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
-
- XLogCtl->LogwrtResult = LogwrtResult;
-
+ pg_atomic_write_u64(&XLogCtl->LogwrtResult.Write, EndOfLog);
+ pg_atomic_write_u64(&XLogCtl->LogwrtResult.Flush, EndOfLog);
XLogCtl->LogwrtRqst.Write = EndOfLog;
XLogCtl->LogwrtRqst.Flush = EndOfLog;
@@ -5925,9 +5919,7 @@ GetFlushRecPtr(TimeLineID *insertTLI)
{
Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
- SpinLockAcquire(&XLogCtl->info_lck);
- LogwrtResult = XLogCtl->LogwrtResult;
- SpinLockRelease(&XLogCtl->info_lck);
+ LogwrtResult.Flush = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Flush);
/*
* If we're writing and flushing WAL, the time line can't be changing, so
@@ -8832,9 +8824,7 @@ GetXLogInsertRecPtr(void)
XLogRecPtr
GetXLogWriteRecPtr(void)
{
- SpinLockAcquire(&XLogCtl->info_lck);
- LogwrtResult = XLogCtl->LogwrtResult;
- SpinLockRelease(&XLogCtl->info_lck);
+ LogwrtResult.Write = pg_atomic_read_u64(&XLogCtl->LogwrtResult.Write);
return LogwrtResult.Write;
}
diff --git a/src/include/port/atomics.h b/src/include/port/atomics.h
index f7cd0f6f20..9601e94220 100644
--- a/src/include/port/atomics.h
+++ b/src/include/port/atomics.h
@@ -515,6 +515,35 @@ pg_atomic_sub_fetch_u64(volatile pg_atomic_uint64 *ptr, int64 sub_)
return pg_atomic_sub_fetch_u64_impl(ptr, sub_);
}
+/*
+ * Monotonically advance the given variable using only atomic operations until
+ * it's at least the target value.
+ *
+ * Full barrier semantics (even when value is unchanged).
+ */
+static inline void
+pg_atomic_monotonic_advance_u64(volatile pg_atomic_uint64 *ptr, uint64 target_)
+{
+ uint64 currval;
+
+#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
+ AssertPointerAlignment(ptr, 8);
+#endif
+
+ currval = pg_atomic_read_u64(ptr);
+ if (currval >= target_)
+ {
+ pg_memory_barrier();
+ return;
+ }
+
+ while (currval < target_)
+ {
+ if (pg_atomic_compare_exchange_u64(ptr, &currval, target_))
+ break;
+ }
+}
+
#undef INSIDE_ATOMICS_H
#endif /* ATOMICS_H */
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 35c9f1efce..8207ac6c4b 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -3053,6 +3053,7 @@ XLogRedoAction
XLogSegNo
XLogSource
XLogStats
+XLogwrtAtomic
XLogwrtResult
XLogwrtRqst
XPV
--
2.30.2
--6zo5pkxyl4hg4egn--
view thread (3+ messages) latest in thread
reply
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: pgsql-hackers@postgresql.org
Cc: alvherre@alvh.no-ip.org
Subject: Re: [PATCH v10] Make XLogCtl->LogwrtResult accessible with atomics
In-Reply-To: <no-message-id-645654@localhost>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox