[vm, mac] Disable TSAN during profiler sample collection.

The suspended thread might hold a TSAN-internal lock to a location that will be read by the sampling thread, which would cause a dead lock.

TEST=tsan
Bug: https://github.com/dart-lang/sdk/issues/62332
Change-Id: I0ac6a05c25067749b52ed83b14f80ab1850013a3
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/471321
Commit-Queue: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Alexander Aprelev <aam@google.com>

runtime/platform/thread_sanitizer.h

@@ -5,6 +5,7 @@
 #ifndef RUNTIME_PLATFORM_THREAD_SANITIZER_H_
 #define RUNTIME_PLATFORM_THREAD_SANITIZER_H_
 
+#include "platform/allocation.h"
 #include "platform/globals.h"
 
 #if __SANITIZE_THREAD__
@@ -54,6 +55,8 @@ extern "C" void __tsan_write8_pc(void* addr, void* pc);
 extern "C" void __tsan_write16_pc(void* addr, void* pc);
 extern "C" void __tsan_func_entry(void* pc);
 extern "C" void __tsan_func_exit();
+extern "C" void __tsan_ignore_thread_begin();
+extern "C" void __tsan_ignore_thread_end();
 constexpr uintptr_t kExternalPCBit = 1ULL << 60;
 #else
 #define NO_SANITIZE_THREAD
@@ -72,4 +75,22 @@ constexpr uintptr_t kExternalPCBit = 1ULL << 60;
 #define DO_IF_NOT_TSAN(CODE) CODE
 #endif
 
+namespace dart {
+
+class TsanIgnoreScope : public ValueObject {
+ public:
+  TsanIgnoreScope() {
+#if defined(USING_THREAD_SANITIZER)
+    __tsan_ignore_thread_begin();
+#endif
+  }
+  ~TsanIgnoreScope() {
+#if defined(USING_THREAD_SANITIZER)
+    __tsan_ignore_thread_end();
+#endif
+  }
+};
+
+}  // namespace dart
+
 #endif  // RUNTIME_PLATFORM_THREAD_SANITIZER_H_

runtime/vm/profiler.cc

@@ -199,25 +199,10 @@ class ProfilerStackWalker : public ValueObject {
 // MSAN/ASAN are unaware of frames initialized by generated code.
 NO_SANITIZE_ADDRESS
 NO_SANITIZE_MEMORY
-#if defined(DART_HOST_OS_MACOS)
-// Mac profiling is cross-thread and TSAN doesn't know that thread_suspend
-// establishes synchronization.
-NO_SANITIZE_THREAD
-#endif
 static uword* LoadStackSlot(uword* ptr) {
   return reinterpret_cast<uword*>(*ptr);
 }
 
-#if defined(DART_HOST_OS_MACOS)
-// Mac profiling is cross-thread and TSAN doesn't know that thread_suspend
-// establishes synchronization.
-#define IGNORE_RACE(x) x##_ignore_race
-#define IGNORE_RACE2(x) x##IgnoreRace
-#else
-#define IGNORE_RACE(x) x
-#define IGNORE_RACE2(x) x
-#endif
-
 // The layout of C stack frames.
 #if defined(HOST_ARCH_IA32) || defined(HOST_ARCH_X64) ||                       \
     defined(HOST_ARCH_ARM) || defined(HOST_ARCH_ARM64)
@@ -388,8 +373,7 @@ static bool GetAndValidateThreadStackBounds(OSThread* os_thread,
 
 #if defined(DART_INCLUDE_SIMULATOR)
   const bool use_simulator_stack_bounds =
-      FLAG_use_simulator && thread != nullptr &&
-      thread->IGNORE_RACE2(IsExecutingDartCode)();
+      FLAG_use_simulator && thread != nullptr && thread->IsExecutingDartCode();
   if (use_simulator_stack_bounds) {
     Isolate* isolate = thread->isolate();
     ASSERT(isolate != nullptr);
@@ -1073,13 +1057,12 @@ class ProfilerDartStackWalker : public ProfilerStackWalker {
 
   void walk() {
     RELEASE_ASSERT(StubCode::HasBeenInitialized());
-    if (thread_->IGNORE_RACE2(IsDeoptimizing)()) {
+    if (thread_->IsDeoptimizing()) {
       sample_->set_ignore_sample(true);
       return;
     }
 
-    uword* exit_fp =
-        reinterpret_cast<uword*>(thread_->IGNORE_RACE(top_exit_frame_info)());
+    uword* exit_fp = reinterpret_cast<uword*>(thread_->top_exit_frame_info());
     bool has_exit_frame = exit_fp != nullptr;
     if (has_exit_frame) {
       // Exited from compiled code or interpreter.
@@ -1090,14 +1073,13 @@ class ProfilerDartStackWalker : public ProfilerStackWalker {
       pc_ = CallerPC();
       fp_ = CallerFP();
     } else {
-      if (thread_->IGNORE_RACE(vm_tag)() == VMTag::kDartTagId) {
+      if (thread_->vm_tag() == VMTag::kDartTagId) {
         // Running compiled code.
         // Use the FP and PC from the thread interrupt or simulator; already set
         // in the constructor.
 
 #if defined(DART_DYNAMIC_MODULES)
-      } else if (thread_->IGNORE_RACE(vm_tag)() ==
-                 VMTag::kDartInterpretedTagId) {
+      } else if (thread_->vm_tag() == VMTag::kDartInterpretedTagId) {
         // Running interpreter.
         pc_ = reinterpret_cast<uword*>(thread_->interpreter()->get_pc());
         fp_ = reinterpret_cast<uword*>(thread_->interpreter()->get_fp());
@@ -1300,7 +1282,7 @@ static Sample* SetupSample(Thread* thread,
                            bool allocation_sample,
                            ThreadId tid) {
   ASSERT(thread != nullptr);
-  Isolate* isolate = thread->IGNORE_RACE(isolate)();
+  Isolate* isolate = thread->isolate();
   SampleBlockBuffer* buffer = Profiler::sample_block_buffer();
   Sample* sample = allocation_sample ? buffer->ReserveAllocationSample(isolate)
                                      : buffer->ReserveCPUSample(isolate);
@@ -1308,7 +1290,7 @@ static Sample* SetupSample(Thread* thread,
     return nullptr;
   }
   sample->Init(isolate->main_port(), OS::GetCurrentMonotonicMicros(), tid);
-  uword vm_tag = thread->IGNORE_RACE(vm_tag)();
+  uword vm_tag = thread->vm_tag();
 #if defined(DART_INCLUDE_SIMULATOR)
   // When running in the simulator, the runtime entry function address
   // (stored as the vm tag) is the address of a redirect function.
@@ -1321,7 +1303,7 @@ static Sample* SetupSample(Thread* thread,
   }
 #endif
   sample->set_vm_tag(vm_tag);
-  sample->set_user_tag(thread->IGNORE_RACE(user_tag)());
+  sample->set_user_tag(thread->user_tag());
   sample->set_thread_task(thread->task_kind());
   return sample;
 }
@@ -1405,7 +1387,7 @@ void Profiler::SampleThreadSingleFrame(Thread* thread,
   ASSERT(Profiler::sample_block_buffer() != nullptr);
 
 #if !defined(PRODUCT)
-  Isolate* isolate = thread->IGNORE_RACE(isolate)();
+  Isolate* isolate = thread->isolate();
 
   // Increment counter for vm tag.
   VMTagCounters* counters = isolate->vm_tag_counters();
@@ -1441,9 +1423,9 @@ void ReleaseToCurrentBlock(Isolate* isolate) {
 void Profiler::SampleThread(Thread* thread,
                             const InterruptedThreadState& state) {
   ASSERT(thread != nullptr);
-  OSThread* os_thread = thread->IGNORE_RACE(os_thread)();
+  OSThread* os_thread = thread->os_thread();
   ASSERT(os_thread != nullptr);
-  Isolate* isolate = thread->IGNORE_RACE(isolate)();
+  Isolate* isolate = thread->isolate();
 
   // Double check if interrupts are disabled
   // after the thread interrupter decided to send a signal.
@@ -1465,7 +1447,7 @@ void Profiler::SampleThread(Thread* thread,
     return;
   }
 
-  const bool in_dart_code = thread->IGNORE_RACE2(IsExecutingDartCode)();
+  const bool in_dart_code = thread->IsExecutingDartCode();
 
   uintptr_t sp = 0;
   uintptr_t fp = state.fp;
@@ -1513,7 +1495,7 @@ void Profiler::SampleThread(Thread* thread,
   }
 
   if (thread->IsDartMutatorThread()) {
-    if (thread->IGNORE_RACE2(IsDeoptimizing)()) {
+    if (thread->IsDeoptimizing()) {
       counters_.single_frame_sample_deoptimizing.fetch_add(1);
       SampleThreadSingleFrame(thread, sample, pc);
       ReleaseToCurrentBlock(isolate);
@@ -1547,7 +1529,7 @@ void Profiler::SampleThread(Thread* thread,
   Dart_Port port = (isolate != nullptr) ? isolate->main_port() : ILLEGAL_PORT;
   ProfilerNativeStackWalker native_stack_walker(
       &counters_, port, sample, isolate, stack_lower, stack_upper, pc, fp, sp);
-  const bool exited_dart_code = thread->IGNORE_RACE2(HasExitedDartCode)();
+  const bool exited_dart_code = thread->HasExitedDartCode();
   ProfilerDartStackWalker dart_stack_walker(thread, port, sample, isolate, pc,
                                             fp, sp, lr,
                                             /*allocation_sample=*/false);

runtime/vm/profiler_test.cc

@@ -2461,6 +2461,20 @@ ISOLATE_UNIT_TEST_CASE(Profiler_EnterExitIsolate) {
   }
 }
 
+// Poke a lot at OSThread::thread_interrupt_disabled_, which will be read by the
+// sampling thread.
+// https://github.com/dart-lang/sdk/issues/62332
+ISOLATE_UNIT_TEST_CASE(Profiler_ThreadEnableDisableProfiler) {
+  EnableProfiler();
+  Profiler::UpdateFlagProfilePeriod(50);  // Microseconds.
+  Profiler::UpdateSamplePeriod();
+
+  for (intptr_t i = 0; i < 100000; i++) {
+    Dart_ThreadDisableProfiling();
+    Dart_ThreadEnableProfiling();
+  }
+}
+
 ISOLATE_UNIT_TEST_CASE(Profiler_UpdateRunningState) {
   Isolate* isolate = Isolate::Current();
   SampleFilter filter(isolate->main_port(), Thread::kMutatorTask, -1, -1);

runtime/vm/thread.cc

@@ -1056,20 +1056,10 @@ bool Thread::IsExecutingDartCode() const {
   return (top_exit_frame_info() == 0) && VMTag::IsDartTag(vm_tag());
 }
 
-bool Thread::IsExecutingDartCodeIgnoreRace() const {
-  return (top_exit_frame_info_ignore_race() == 0) &&
-         VMTag::IsDartTag(vm_tag_ignore_race());
-}
-
 bool Thread::HasExitedDartCode() const {
   return (top_exit_frame_info() != 0) && !VMTag::IsDartTag(vm_tag());
 }
 
-bool Thread::HasExitedDartCodeIgnoreRace() const {
-  return (top_exit_frame_info_ignore_race() != 0) &&
-         !VMTag::IsDartTag(vm_tag_ignore_race());
-}
-
 template <class C>
 C* Thread::AllocateReusableHandle() {
   C* handle = reinterpret_cast<C*>(reusable_handles_.AllocateScopedHandle());

runtime/vm/thread.h

@@ -580,8 +580,6 @@ class Thread : public ThreadState, public IntrusiveDListEntry<Thread> {
 
   // The isolate that this thread is operating on, or nullptr if none.
   Isolate* isolate() const { return isolate_; }
-  NO_SANITIZE_THREAD
-  Isolate* isolate_ignore_race() const { return isolate_; }
   static intptr_t isolate_offset() { return OFFSET_OF(Thread, isolate_); }
   static intptr_t isolate_group_offset() {
     return OFFSET_OF(Thread, isolate_group_);
@@ -637,11 +635,9 @@ class Thread : public ThreadState, public IntrusiveDListEntry<Thread> {
 
   // Is |this| executing Dart code?
   bool IsExecutingDartCode() const;
-  bool IsExecutingDartCodeIgnoreRace() const;
 
   // Has |this| exited Dart code?
   bool HasExitedDartCode() const;
-  bool HasExitedDartCodeIgnoreRace() const;
 
   bool HasCompilerState() const { return compiler_state_ != nullptr; }
 
@@ -749,8 +745,6 @@ class Thread : public ThreadState, public IntrusiveDListEntry<Thread> {
   }
 
   uword top_exit_frame_info() const { return top_exit_frame_info_; }
-  NO_SANITIZE_THREAD
-  uword top_exit_frame_info_ignore_race() const { return top_exit_frame_info_; }
   void set_top_exit_frame_info(uword top_exit_frame_info) {
     top_exit_frame_info_ = top_exit_frame_info;
   }
@@ -881,8 +875,6 @@ class Thread : public ThreadState, public IntrusiveDListEntry<Thread> {
 #endif
 
   uword vm_tag() const { return vm_tag_; }
-  NO_SANITIZE_THREAD
-  uword vm_tag_ignore_race() const { return vm_tag_; }
   void set_vm_tag(uword tag) { vm_tag_ = tag; }
   static intptr_t vm_tag_offset() { return OFFSET_OF(Thread, vm_tag_); }
 
@@ -1352,8 +1344,6 @@ class Thread : public ThreadState, public IntrusiveDListEntry<Thread> {
   }
 
   bool IsDeoptimizing() const { return deopt_context_ != nullptr; }
-  NO_SANITIZE_THREAD
-  bool IsDeoptimizingIgnoreRace() const { return deopt_context_ != nullptr; }
   DeoptContext* deopt_context() const { return deopt_context_; }
   void set_deopt_context(DeoptContext* value) {
     ASSERT(value == nullptr || deopt_context_ == nullptr);
@@ -1372,8 +1362,6 @@ class Thread : public ThreadState, public IntrusiveDListEntry<Thread> {
   }
 
   uword user_tag() const { return user_tag_; }
-  NO_SANITIZE_THREAD
-  uword user_tag_ignore_race() const { return user_tag_; }
   static intptr_t user_tag_offset() { return OFFSET_OF(Thread, user_tag_); }
   static intptr_t current_tag_offset() {
     return OFFSET_OF(Thread, current_tag_);

runtime/vm/thread_interrupter_macos.cc

@@ -76,6 +76,11 @@ class ThreadInterrupterMacOS {
       return;
     }
     ThreadInterruptScope signal_handler_scope;
+    // Mac profiling is cross-thread and TSAN doesn't know that thread_suspend
+    // establishes synchronization. Also, the suspended thread might hold a
+    // TSAN-internal lock for a location the sampling thread might read, which
+    // would cause a deadlock.
+    TsanIgnoreScope tsan_ignore_scope;
     Profiler::SampleThread(thread, ProcessState(state));
   }
 

runtime/vm/thread_state.h

@@ -32,8 +32,6 @@ class ThreadState : public BaseThread {
 
   // OSThread corresponding to this thread.
   OSThread* os_thread() const { return os_thread_; }
-  NO_SANITIZE_THREAD
-  OSThread* os_thread_ignore_race() const { return os_thread_; }
   void set_os_thread(OSThread* os_thread) { os_thread_ = os_thread; }
 
   // The topmost zone used for allocation in this thread.