[sanitizers] Optimize locking StackDepotBase for fork

compiler-rt/lib/sanitizer_common/sanitizer_flat_map.h

@@ -109,6 +109,10 @@ class TwoLevelMap {
     return *AddressSpaceView::LoadWritable(&map2[idx % kSize2]);
   }
 
+  void Lock() SANITIZER_NO_THREAD_SAFETY_ANALYSIS { mu_.Lock(); }
+
+  void Unlock() SANITIZER_NO_THREAD_SAFETY_ANALYSIS { mu_.Unlock(); }
+
  private:
   constexpr uptr MmapSize() const {
     return RoundUpTo(kSize2 * sizeof(T), GetPageSizeCached());

compiler-rt/lib/sanitizer_common/sanitizer_stackdepotbase.h

@@ -161,18 +161,32 @@ StackDepotBase<Node, kReservedBits, kTabSizeLog>::Get(u32 id) {
 
 template <class Node, int kReservedBits, int kTabSizeLog>
 void StackDepotBase<Node, kReservedBits, kTabSizeLog>::LockBeforeFork() {
-  for (int i = 0; i < kTabSize; ++i) {
-    lock(&tab[i]);
-  }
+  // Do not lock hash table. It's very expensive, but it's not rely needed. The
+  // parent process will neither lock nor unlock. Child process risks to be
+  // deadlocked on already locked buckets. To avoid deadlock we will unlock
+  // every locked buckets in `UnlockAfterFork`. This may affect consistency of
+  // the hash table, but the only issue is a few items inserted by parent
+  // process will be not found by child, and the child may insert them again,
+  // wasting some space in `stackStore`.
+
+  // We still need to lock nodes.
+  nodes.Lock();
 }
 
 template <class Node, int kReservedBits, int kTabSizeLog>
 void StackDepotBase<Node, kReservedBits, kTabSizeLog>::UnlockAfterFork(
     bool fork_child) {
+  nodes.Unlock();
+
+  // Only unlock in child process to avoid deadlock. See `LockBeforeFork`.
+  if (!fork_child)
+    return;
+
   for (int i = 0; i < kTabSize; ++i) {
     atomic_uint32_t *p = &tab[i];
     uptr s = atomic_load(p, memory_order_relaxed);
-    unlock(p, s & kUnlockMask);
+    if (s & kLockMask)
+      unlock(p, s & kUnlockMask);
   }
 }