[asan/win] Use SRW locks to fix a race in BlockingMutex
authorReid Kleckner <rnk@google.com>
Mon, 30 Jul 2018 23:32:33 +0000 (23:32 +0000)
committerReid Kleckner <rnk@google.com>
Mon, 30 Jul 2018 23:32:33 +0000 (23:32 +0000)
Summary:
Before my change, BlockingMutex used Windows critial sections. Critical
sections can only be initialized by calling InitializeCriticalSection,
dynamically.

The primary sanitizer allocator expects to be able to reinterpret zero
initialized memory as a BlockingMutex and immediately lock it.
RegionInfo contains a mutex, and it placement new is never called for
it. These objects are accessed via:
  RegionInfo *GetRegionInfo(uptr class_id) const {
    DCHECK_LT(class_id, kNumClasses);
    RegionInfo *regions = reinterpret_cast<RegionInfo *>(SpaceEnd());
    return &regions[class_id];
  }
The memory comes from the OS without any other initialization.

For various reasons described in the comments, BlockingMutex::Lock would
check if the object appeared to be zero-initialized, and it would lazily
call the LinkerInitialized constructor to initialize the critical
section. This pattern is obviously racy, and the code had a bunch of
FIXMEs about it.

The best fix here is to use slim reader writer locks, which can start
out zero-initialized. They are available starting in Windows Vista. I
think it's safe to go ahead and use them today.

Reviewers: kcc, vitalybuka

Subscribers: kubamracek, llvm-commits

Differential Revision: https://reviews.llvm.org/D49893

llvm-svn: 338331

compiler-rt/lib/sanitizer_common/sanitizer_mutex.h
compiler-rt/lib/sanitizer_common/sanitizer_win.cc

index 4bb878e..9041f3a 100644 (file)
@@ -73,13 +73,8 @@ class SpinMutex : public StaticSpinMutex {
 
 class BlockingMutex {
  public:
-#if SANITIZER_WINDOWS
-  // Windows does not currently support LinkerInitialized
-  explicit BlockingMutex(LinkerInitialized);
-#else
   explicit constexpr BlockingMutex(LinkerInitialized)
-      : opaque_storage_ {0, }, owner_(0) {}
-#endif
+      : opaque_storage_ {0, }, owner_{0} {}
   BlockingMutex();
   void Lock();
   void Unlock();
index b8060b2..38e567d 100644 (file)
@@ -767,43 +767,22 @@ void *internal_start_thread(void (*func)(void *arg), void *arg) { return 0; }
 void internal_join_thread(void *th) { }
 
 // ---------------------- BlockingMutex ---------------- {{{1
-const uptr LOCK_UNINITIALIZED = 0;
-const uptr LOCK_READY = (uptr)-1;
-
-BlockingMutex::BlockingMutex(LinkerInitialized li) {
-  // FIXME: see comments in BlockingMutex::Lock() for the details.
-  CHECK(li == LINKER_INITIALIZED || owner_ == LOCK_UNINITIALIZED);
-
-  CHECK(sizeof(CRITICAL_SECTION) <= sizeof(opaque_storage_));
-  InitializeCriticalSection((LPCRITICAL_SECTION)opaque_storage_);
-  owner_ = LOCK_READY;
-}
 
 BlockingMutex::BlockingMutex() {
-  CHECK(sizeof(CRITICAL_SECTION) <= sizeof(opaque_storage_));
-  InitializeCriticalSection((LPCRITICAL_SECTION)opaque_storage_);
-  owner_ = LOCK_READY;
+  CHECK(sizeof(SRWLOCK) <= sizeof(opaque_storage_));
+  internal_memset(this, 0, sizeof(*this));
 }
 
 void BlockingMutex::Lock() {
-  if (owner_ == LOCK_UNINITIALIZED) {
-    // FIXME: hm, global BlockingMutex objects are not initialized?!?
-    // This might be a side effect of the clang+cl+link Frankenbuild...
-    new(this) BlockingMutex((LinkerInitialized)(LINKER_INITIALIZED + 1));
-
-    // FIXME: If it turns out the linker doesn't invoke our
-    // constructors, we should probably manually Lock/Unlock all the global
-    // locks while we're starting in one thread to avoid double-init races.
-  }
-  EnterCriticalSection((LPCRITICAL_SECTION)opaque_storage_);
-  CHECK_EQ(owner_, LOCK_READY);
+  AcquireSRWLockExclusive((PSRWLOCK)opaque_storage_);
+  CHECK_EQ(owner_, 0);
   owner_ = GetThreadSelf();
 }
 
 void BlockingMutex::Unlock() {
-  CHECK_EQ(owner_, GetThreadSelf());
-  owner_ = LOCK_READY;
-  LeaveCriticalSection((LPCRITICAL_SECTION)opaque_storage_);
+  CheckLocked();
+  owner_ = 0;
+  ReleaseSRWLockExclusive((PSRWLOCK)opaque_storage_);
 }
 
 void BlockingMutex::CheckLocked() {