[core] mutable_object_test flaky tsan fix ray-project#44764

mutable_object test is still flaky. In the recent buildkite runs, it appears that there is a data race on the access to num_readers in SetErrorUnlocked(). This PR fixes that data race.

Signed-off-by: Jack Humphries <[email protected]>

src/ray/object_manager/common.cc

@@ -81,9 +81,11 @@ Status PlasmaObjectHeader::TryToAcquireSemaphore(sem_t *sem) const {
   // section after `SetErrorUnlocked()` has been called. One thread could be in the
   // critical section when that is called, but no additional thread will enter the
   // critical section.
-  RAY_RETURN_NOT_OK(CheckHasError());
-
-  return Status::OK();
+  Status s = CheckHasError();
+  if (!s.ok()) {
+    RAY_CHECK_EQ(sem_post(sem), 0);
+  }
+  return s;
 }
 
 void PlasmaObjectHeader::SetErrorUnlocked(Semaphores &sem) {
@@ -97,11 +99,8 @@ void PlasmaObjectHeader::SetErrorUnlocked(Semaphores &sem) {
   // be more than one writer.
   RAY_CHECK_EQ(sem_post(sem.object_sem), 0);
 
-  // Increment `sem.header_sem` by `num_readers` since there could potentially be that
-  // many readers blocked on `sem_wait()`.
-  for (int64_t i = 0; i < num_readers; i++) {
-    RAY_CHECK_EQ(sem_post(sem.header_sem), 0);
-  }
+  // Increment `header_sem` to unblock any readers and/or the writer.
+  RAY_CHECK_EQ(sem_post(sem.header_sem), 0);
 }
 
 Status PlasmaObjectHeader::WriteAcquire(Semaphores &sem,