Enable jemalloc pool test with BA_GLOBAL_PROVIDER

[vinser52]

Description

Enable Jemalloc pool test with BA_GLOBAL_PROVIDER

Ref: #903

Checklist

• Code compiles without errors locally
• All tests pass locally
• CI workflows execute properly

[vinser52]

@bratpiorka @lplewa I created this PR as a reproducer for the issue with jemalloc (See #903 ). It works only from a certain version of jemalloc, but looks like our CI needs to be fixed to use the "proper" version of jemalloc.

[ldorau]

@vinser52 Please test this PR on top of #941

[ldorau]

@vinser52 Please test this PR on top of #941

@vinser52 I've checked - this PR works on top of #941

[vinser52]

@vinser52 Please test this PR on top of #941

@vinser52 I've checked - this PR works on top of #941

thank you, @ldorau, Sorry I was busy yesterday and wanted to check it today.

[ldorau]

@vinser52 #941 has been merged, so you can rebase this PR now

[vinser52]

@vinser52 #941 has been merged, so you can rebase this PR now

Done. Works locally on my machine. Waiting for the CI.

[bratpiorka]

I guess this would require to add the wrapper for hwloc_topology_init() with attribute((disable_sanitizer_instrumentation)) or similar

[ldorau]

@vinser52 I've tried to fix that, but compiler attributes disabling AddressSanitizer do not work (build still fails):

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 70ac087..8835c40 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -452,6 +452,8 @@ endif()
 # Sanitizer flags
 if(UMF_USE_ASAN)
     add_sanitizer_flag(address)
+    set(UMF_COMMON_COMPILE_DEFINITIONS ${UMF_COMMON_COMPILE_DEFINITIONS}
+                                       UMF_USE_ASAN=1)
 endif()
 if(UMF_USE_UBSAN)
     add_sanitizer_flag(undefined)
diff --git a/src/provider/provider_os_memory.c b/src/provider/provider_os_memory.c
index 1fe4679..c26d3ed 100644
--- a/src/provider/provider_os_memory.c
+++ b/src/provider/provider_os_memory.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2022-2024 Intel Corporation
+ * Copyright (C) 2022-2025 Intel Corporation
  *
  * Under the Apache License v2.0 with LLVM Exceptions. See LICENSE.TXT.
  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
@@ -17,6 +17,20 @@
 #include <umf/memory_provider_ops.h>
 #include <umf/providers/provider_os_memory.h>

+#ifdef UMF_USE_ASAN
+#ifdef _WIN32
+#define ATTR_NO_ASAN __declspec(noinline) __declspec(no_sanitize_address)
+#else /* !_WIN32 */
+#ifdef __clang__
+#define ATTR_NO_ASAN __attribute__((no_sanitize("address"), noinline))
+#else /* !__clang__ */
+#define ATTR_NO_ASAN __attribute__((no_sanitize_address, noinline))
+#endif /* !__clang__ */
+#endif /* !_WIN32 */
+#else  /* !UMF_USE_ASAN */
+#define ATTR_NO_ASAN
+#endif /* !UMF_USE_ASAN */
+
 // OS Memory Provider requires HWLOC
 #if defined(UMF_NO_HWLOC)

@@ -568,6 +582,10 @@ static umf_result_t translate_params(umf_os_memory_provider_params_t *in_params,
     return UMF_RESULT_SUCCESS;
 }

+static int ATTR_NO_ASAN hwloc_topology_init_no_asan(hwloc_topology_t *topo) {
+    return hwloc_topology_init(topo);
+}
+
 static umf_result_t os_initialize(void *params, void **provider) {
     umf_result_t ret;

@@ -594,7 +612,7 @@ static umf_result_t os_initialize(void *params, void **provider) {

     memset(os_provider, 0, sizeof(*os_provider));

-    int r = hwloc_topology_init(&os_provider->topo);
+    int r = hwloc_topology_init_no_asan(&os_provider->topo);
     if (r) {
         LOG_ERR("HWLOC topology init failed");
         ret = UMF_RESULT_ERROR_OUT_OF_HOST_MEMORY;

[ldorau]

BTW: on Ubuntu 24.04 it works well

[ldorau]

@vinser52 @bratpiorka Should this PR be still opened?

[vinser52]

@vinser52 @bratpiorka Should this PR be still opened?

So I just have no enough time to resolve the CI issues. if someone can take over I would be happy. But in general I think it makes sense to have such tests in CI

[ldorau]

@vinser52 @bratpiorka This test will not work because of an issue in jemalloc - jemalloc does not free the allocated memory extents on which the arena_extent_split() failed:

Jemalloc splits allocated memory extents - it calls arena_extent_split() that calls umfMemoryProviderAllocationSplit() but BA_GLOBAL_PROVIDER does not support split() op and arena_extent_split() always fails. Then jemalloc allocates a new smaller memory extent, but it never frees the first extent on which the split() failed and - as a result - the assert in the tracker is always triggered, because the tracker is not empty:

$ ./test/test_jemalloc_pool --gtest_filter="*.allocFree/1*"
[PID:4827   TID:4827   INFO  UMF] utils_log_init: Logger enabled (UMF version: 0.11.0-dev4.git1.g8c111e8e, level: DEBUG, flush: DEBUG, pid: yes, timestamp: no)
[PID:4827   TID:4827   DEBUG UMF] umf_ba_create_global: UMF base allocator created
[PID:4827   TID:4827   DEBUG UMF] umfMemoryTrackerCreate: tracker created, handle=0x7f1ae2ab0068, alloc_segments_map=0x7f1ae2ab0070
[PID:4827   TID:4827   DEBUG UMF] umfInit: UMF tracker created
[PID:4827   TID:4827   DEBUG UMF] umfInit: UMF IPC cache initialized
[PID:4827   TID:4827   DEBUG UMF] umfInit: UMF library initialized
[PID:4827   TID:4827   INFO  UMF] utils_log_init: Logger enabled (UMF version: 0.11.0-dev4.git1.g8c111e8e, level: DEBUG, flush: DEBUG, pid: yes, timestamp: no)
Running main() from /home/ldorau/work/unified-memory-framework/build/_deps/googletest-src/googletest/src/gtest_main.cc
Note: Google Test filter = *allocFree/1*
[==========] Running 6 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 6 tests from jemallocPoolTest/umfPoolTest
[ RUN      ] jemallocPoolTest/umfPoolTest.allocFree/1
[PID:4827   TID:4827   DEBUG UMF] umfTrackingMemoryProviderCreate: upstream=0x7f1ae2ab0168, tracker=0x7f1ae2ab0068, pool=0x7f1ae2ab0268, ipcCache=0x7f1ae2a9e008, hIpcMappedCache=0x7f1ae2ac1068
[PID:4827   TID:4827   INFO  UMF] umfPoolCreateInternal: Memory pool created: 0x7f1ae2ab0268
[PID:4827   TID:4827   DEBUG UMF] umfMemoryTrackerAddAtLevel: memory region is added, tracker=0x7f1ae2ab0068, level=0, pool=0x7f1ae2ab0268, ptr=0x7f1ae1bff000, size=2097152
[PID:4827   TID:4827   ERROR UMF] trackingAllocationSplit: upstream provider failed to split the region
[PID:4827   TID:4827   ERROR UMF] trackingAllocationSplit: failed to split memory region: ptr=0x7f1ae1bff000, totalSize=2097152, firstSize=4096
[PID:4827   TID:4827   DEBUG UMF] umfMemoryTrackerAddAtLevel: memory region is added, tracker=0x7f1ae2ab0068, level=0, pool=0x7f1ae2ab0268, ptr=0x7f1ae2a88000, size=4096
[PID:4827   TID:4827   DEBUG UMF] umfMemoryTrackerRemove: memory region removed: tracker=0x7f1ae2ab0068, level=0, pool=0x7f1ae2ab0268, ptr=0x7f1ae2a88000, size=4096
[PID:4827   TID:4827   ERROR UMF] check_if_tracker_is_empty: [0] ptr = 0x7f1ae1bff000, pool = 0x7f1ae2ab0268, size = 2097152
[PID:4827   TID:4827   ERROR UMF] check_if_tracker_is_empty: tracking provider of pool 0x7f1ae2ab0268 is not empty! (1 items left)
test_jemalloc_pool: /home/ldorau/work/unified-memory-framework/src/provider/provider_tracking.c:827: check_if_tracker_is_empty: Assertion `n_items == 0 && "tracking provider is not empty!"' failed.
Aborted

[bratpiorka]

@vinser52 @bratpiorka This test will not work because of an issue in jemalloc - jemalloc does not free the allocated memory extents on which the arena_extent_split() failed:

Jemalloc splits allocated memory extents - it calls arena_extent_split() that calls umfMemoryProviderAllocationSplit() but BA_GLOBAL_PROVIDER does not support split() op and arena_extent_split() always fails. Then jemalloc allocates a new smaller memory extent, but it never frees the first extent on which the split() failed and - as a result - the assert in the tracker is always triggered, because the tracker is not empty

Thanks for looking at this. I think that there would be no problem with adding support for split and merge for BA but the question is if we want to do this because of these tests

[ldorau]

I think that there would be no problem with adding support for split and merge for BA but the question is if we want to do this because of these tests

@vinser52 @bratpiorka
In my opinion adding support for split() and merge() to BA_GLOBAL_PROVIDER would require too much effort now and it is not worth doing that just for enabling only one test.

In my opinion this PR can be closed for now.

[ldorau]

This test will not work because of an issue in jemalloc - see: #902 (comment)

[vinser52]

I think that there would be no problem with adding support for split and merge for BA but the question is if we want to do this because of these tests

@vinser52 @bratpiorka In my opinion adding support for split() and merge() to BA_GLOBAL_PROVIDER would require too much effort now and it is not worth doing that just for enabling only one test.

In my opinion this PR can be closed for now.

I think we should not waste time on adding split/merge support to BA allocator.

did I get correctly that there is a bug in jemalloc? It should support scenario when split/merge fails but in fact, there is a memory leak, right? Should we create an issue in their GitHub?

[ldorau]

I think that there would be no problem with adding support for split and merge for BA but the question is if we want to do this because of these tests

@vinser52 @bratpiorka In my opinion adding support for split() and merge() to BA_GLOBAL_PROVIDER would require too much effort now and it is not worth doing that just for enabling only one test.
In my opinion this PR can be closed for now.

I think we should not waste time on adding split/merge support to BA allocator.

did I get correctly that there is a bug in jemalloc? It should support scenario when split/merge fails but in fact, there is a memory leak, right? Should we create an issue in their GitHub?

Right. I have submitted the issue in jemalloc: jemalloc/jemalloc#2815

@vinser52 Can we close this PR?

[ldorau]

@vinser52 I have submitted #1193 instead of this PR