[tsan] Lock/Unlock allocator and stacks on fork

[vitalybuka]

We do that for other Sanitizers, and we
should do the same for TSAN.
There are know deadlocks reports here.

[llvmbot]

@llvm/pr-subscribers-compiler-rt-sanitizer

Author: Vitaly Buka (vitalybuka)

Changes

We do that for other Sanitizers, and we
should do the same for TSAN.
There are know deadlocks reports here.

---

Full diff: https://github.com/llvm/llvm-project/pull/96600.diff

2 Files Affected:

• (modified) compiler-rt/lib/tsan/rtl/tsan_mman.cpp (+10)
• (modified) compiler-rt/test/sanitizer_common/TestCases/Posix/fork_threaded.c (-3)

diff --git a/compiler-rt/lib/tsan/rtl/tsan_mman.cpp b/compiler-rt/lib/tsan/rtl/tsan_mman.cpp
index 1e579c6a020e9..e129e9af272f5 100644
--- a/compiler-rt/lib/tsan/rtl/tsan_mman.cpp
+++ b/compiler-rt/lib/tsan/rtl/tsan_mman.cpp
@@ -17,6 +17,7 @@
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_errno.h"
 #include "sanitizer_common/sanitizer_placement_new.h"
+#include "sanitizer_common/sanitizer_stackdepot.h"
 #include "tsan_flags.h"
 #include "tsan_interface.h"
 #include "tsan_report.h"
@@ -119,9 +120,18 @@ ScopedGlobalProcessor::~ScopedGlobalProcessor() {
 void AllocatorLockBeforeFork() SANITIZER_NO_THREAD_SAFETY_ANALYSIS {
   global_proc()->internal_alloc_mtx.Lock();
   InternalAllocatorLock();
+#if !SANITIZER_APPLE
+  // OS X allocates from hooks, see 6a3958247a.
+  allocator()->ForceLock();
+  StackDepotLockBeforeFork();
+#endif
 }
 
 void AllocatorUnlockAfterFork(bool child) SANITIZER_NO_THREAD_SAFETY_ANALYSIS {
+#if !SANITIZER_APPLE
+  StackDepotUnlockAfterFork(child);
+  allocator()->ForceUnlock();
+#endif
   InternalAllocatorUnlock();
   global_proc()->internal_alloc_mtx.Unlock();
 }
diff --git a/compiler-rt/test/sanitizer_common/TestCases/Posix/fork_threaded.c b/compiler-rt/test/sanitizer_common/TestCases/Posix/fork_threaded.c
index 13240234a1c79..27b67db0c0a38 100644
--- a/compiler-rt/test/sanitizer_common/TestCases/Posix/fork_threaded.c
+++ b/compiler-rt/test/sanitizer_common/TestCases/Posix/fork_threaded.c
@@ -6,9 +6,6 @@
 // FIXME: It probably hangs on this platform.
 // UNSUPPORTED: ppc
 
-// FIXME: TSAN does not lock allocator.
-// UNSUPPORTED: tsan
-
 // FIXME: False stack overflow report
 // UNSUPPORTED: android && asan
 

[dvyukov]

Note that tsan simply disables itself after multithreaded fork, so it produces false sense to operation with die_after_fork=0. And it still can deadlock on mutex/atomic operations and maybe more.

[fmayer]

nit: and stack depot? "and stacks" is confusing a bit.