pg.ddx.io pgsql-bugs@postgresql.org mailing list archive
help / color / mirror / Atom feed[PATCH v3 4/4] wip: llvmjit: Plug memory leak when performing inlining.
1+ messages / 1 participants
[nested] [flat]
* [PATCH v3 4/4] wip: llvmjit: Plug memory leak when performing inlining.
@ 2022-11-01 05:35 Andres Freund <andres@anarazel.de>
0 siblings, 0 replies; 1+ messages in thread
From: Andres Freund @ 2022-11-01 05:35 UTC (permalink / raw)
When performing inlining LLVM unfortunately "leaks" types (the types
survive and are usable, but a new round of inlining will recreate new
types that are structurally equivalent). I could not come up with a
backpatchable fix for that, so address the problem by dropping &
recreating the LLVMContextRef that contains all IR related data -
which releases all the types.
That incurs some overhead, so only do so after 10000 JITed queries.
Needs a good bit more work:
- confusing naming between different types of contexts (PG side
LLVMJitContext, and LLVM side LLVMContextRef)
- add detailed explanation of leak issue
- consider only incrementing the counter triggering the recreation of
the memory context during inlining.
- several FIXMEs in code
Reported-By: Justin Pryzby <pryzby@telsasoft.com>
Reported-By: Kurt Roeckx <kurt@roeckx.be>
Reported-By: Jaime Casanova <jcasanov@systemguards.com.ec>
Author: Andres Freund
Discussion: https://postgr.es/m/7acc8678-df5f-4923-9cf6-e843131ae89d@www.fastmail.com
Discussion: https://postgr.es/m/20201218235607.GC30237@telsasoft.com
---
src/include/jit/llvmjit.h | 8 +++
src/backend/jit/llvm/llvmjit.c | 76 ++++++++++++++++++++++++-
src/backend/jit/llvm/llvmjit_inline.cpp | 12 ++++
3 files changed, 94 insertions(+), 2 deletions(-)
inline patch 1: src/include/jit/llvmjit.h, src/backend/jit/llvm/llvmjit.c, src/backend/jit/llvm/llvmjit_inline.cpp (+94/-2 lines) download inline-1.patch
diff --git a/src/include/jit/llvmjit.h b/src/include/jit/llvmjit.h
index f4c01c28672..d90d7b2e347 100644
--- a/src/include/jit/llvmjit.h
+++ b/src/include/jit/llvmjit.h
@@ -42,6 +42,13 @@ typedef struct LLVMJitContext
/* number of modules created */
size_t module_generation;
+ /*
+ * The LLVM Context used by this JIT context. An LLVM context is reused
+ * across many compilations, but occasionally reset to prevent it using
+ * too much memory due to more and more types accumulating.
+ */
+ LLVMContextRef llvm_context;
+
/* current, "open for write", module */
LLVMModuleRef module;
@@ -99,6 +106,7 @@ extern LLVMValueRef llvm_function_reference(LLVMJitContext *context,
LLVMModuleRef mod,
FunctionCallInfo fcinfo);
+extern void llvm_inline_reset_caches(void);
extern void llvm_inline(LLVMModuleRef mod);
/*
diff --git a/src/backend/jit/llvm/llvmjit.c b/src/backend/jit/llvm/llvmjit.c
index 3625edceb3e..a7e19a69a8b 100644
--- a/src/backend/jit/llvm/llvmjit.c
+++ b/src/backend/jit/llvm/llvmjit.c
@@ -42,6 +42,10 @@
#include "utils/memutils.h"
#include "utils/resowner_private.h"
+
+#define LLVMJIT_LLVM_CONTEXT_REUSE_MAX 100
+
+
/* Handle of a module emitted via ORC JIT */
typedef struct LLVMJitHandle
{
@@ -81,6 +85,10 @@ static LLVMModuleRef llvm_types_module = NULL;
static bool llvm_session_initialized = false;
static size_t llvm_generation = 0;
+/* number of LLVMJitContexts that currently are in use */
+static size_t llvm_jit_context_in_use_count = 0;
+/* how many times has the current LLVMContextRef been used */
+static size_t llvm_llvm_context_reuse_count = 0;
static const char *llvm_triple = NULL;
static const char *llvm_layout = NULL;
static LLVMContextRef llvm_context;
@@ -141,6 +149,55 @@ llvm_create_context(int jitFlags)
llvm_session_initialize();
+ /*
+ * Every now and then create a new LLVMContextRef. Unfortunately, during
+ * every round of inlining, types may "leak" (they can still be found/used
+ * via the context, but new types will be created the next time in
+ * inlining is performed). To prevent that from slowly accumulating
+ * problematic amounts of memory, recreate the LLVMContextRef we use. We
+ * don't want to do so too often, as that implies some overhead
+ * (particularly re-loading the module summaries / modules is fairly
+ * expensive).
+ *
+ * We can only safely recreate the LLVM context if no other code is being
+ * JITed, otherwise we'd release the types in use for that.
+ *
+ * FIXME: Extract into helper function.
+ */
+ if (llvm_jit_context_in_use_count == 0 &&
+ llvm_llvm_context_reuse_count > LLVMJIT_LLVM_CONTEXT_REUSE_MAX)
+ {
+ llvm_llvm_context_reuse_count = 0;
+
+ Assert(llvm_context != NULL);
+
+ /*
+ * Need to reset the modules that the inlining code caches before
+ * disposing of the context. LLVM modules exist within a specific LLVM
+ * context, therefore disposing of the context before resetting the
+ * cache would lead to dangling pointers to modules.
+ */
+ llvm_inline_reset_caches();
+
+ if (llvm_context != NULL)
+ LLVMContextDispose(llvm_context);
+ llvm_context = LLVMContextCreate();
+
+ /*
+ * Re-build cached type information, so code generation code can rely
+ * on that information to be present (also prevents the variables to
+ * be dangling references).
+ *
+ * FIXME: should split the handling of llvm_triple / llvm_layout out
+ * of llvm_create_types() - that doesn't need to be redone.
+ */
+ llvm_create_types();
+ }
+ else
+ {
+ llvm_llvm_context_reuse_count++;
+ }
+
ResourceOwnerEnlargeJIT(CurrentResourceOwner);
context = MemoryContextAllocZero(TopMemoryContext,
@@ -151,6 +208,8 @@ llvm_create_context(int jitFlags)
context->base.resowner = CurrentResourceOwner;
ResourceOwnerRememberJIT(CurrentResourceOwner, PointerGetDatum(context));
+ llvm_jit_context_in_use_count++;
+
return context;
}
@@ -163,6 +222,12 @@ llvm_release_context(JitContext *context)
LLVMJitContext *llvm_context = (LLVMJitContext *) context;
ListCell *lc;
+ /*
+ * Consider as cleaned up even if we skip doing so below, that way we can
+ * verify the tracking is correct (see llvm_shutdown()).
+ */
+ llvm_jit_context_in_use_count--;
+
/*
* When this backend is exiting, don't clean up LLVM. As an error might
* have occurred from within LLVM, we do not want to risk reentering. All
@@ -901,6 +966,10 @@ llvm_shutdown(int code, Datum arg)
return;
}
+ if (llvm_jit_context_in_use_count != 0)
+ elog(PANIC, "LLVMJitContext in use count not 0 at exit (is %zu)",
+ llvm_jit_context_in_use_count);
+
#if LLVM_VERSION_MAJOR > 11
{
if (llvm_opt3_orc)
@@ -1004,8 +1073,11 @@ llvm_create_types(void)
* Load triple & layout from clang emitted file so we're guaranteed to be
* compatible.
*/
- llvm_triple = pstrdup(LLVMGetTarget(llvm_types_module));
- llvm_layout = pstrdup(LLVMGetDataLayoutStr(llvm_types_module));
+ if (llvm_triple == NULL)
+ {
+ llvm_triple = pstrdup(LLVMGetTarget(llvm_types_module));
+ llvm_layout = pstrdup(LLVMGetDataLayoutStr(llvm_types_module));
+ }
TypeSizeT = llvm_pg_var_type("TypeSizeT");
TypeParamBool = load_return_type(llvm_types_module, "FunctionReturningBool");
diff --git a/src/backend/jit/llvm/llvmjit_inline.cpp b/src/backend/jit/llvm/llvmjit_inline.cpp
index 73fe253f2ec..6b46e141fcb 100644
--- a/src/backend/jit/llvm/llvmjit_inline.cpp
+++ b/src/backend/jit/llvm/llvmjit_inline.cpp
@@ -152,6 +152,18 @@ summaries_for_guid(const InlineSearchPath& path, llvm::GlobalValue::GUID guid);
#define ilog(...) (void) 0
#endif
+/*
+ * Reset inlining related state. This needs to be called before the currently
+ * used LLVMContextRef is disposed (and a new one create), otherwise we would
+ * have dangling references to deleted modules.
+ */
+void
+llvm_inline_reset_caches(void)
+{
+ module_cache->clear();
+ summary_cache->clear();
+}
+
/*
* Perform inlining of external function references in M based on a simple
* cost based analysis.
--
2.38.0
only message in thread
Thread overview: 1+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2022-11-01 05:35 [PATCH v3 4/4] wip: llvmjit: Plug memory leak when performing inlining. Andres Freund <andres@anarazel.de>
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox