Merge pull request #12564 from sreecha/exp-cq-fix

Fix TSAN failure in completion queue

src/core/lib/surface/completion_queue.c

@@ -565,13 +565,13 @@ static void cq_check_tag(grpc_completion_queue *cq, void *tag, bool lock_cq) {}
  * true if the increment was successful; false if the counter is zero */
 static bool atm_inc_if_nonzero(gpr_atm *counter) {
   while (true) {
-    gpr_atm count = gpr_atm_no_barrier_load(counter);
+    gpr_atm count = gpr_atm_acq_load(counter);
     /* If zero, we are done. If not, we must to a CAS (instead of an atomic
      * increment) to maintain the contract: do not increment the counter if it
      * is zero. */
     if (count == 0) {
       return false;
-    } else if (gpr_atm_no_barrier_cas(counter, count, count + 1)) {
+    } else if (gpr_atm_full_cas(counter, count, count + 1)) {
       break;
     }
   }
@@ -643,8 +643,12 @@ static void cq_end_op_for_next(grpc_exec_ctx *exec_ctx,
   /* Add the completion to the queue */
   bool is_first = cq_event_queue_push(&cqd->queue, storage);
   gpr_atm_no_barrier_fetch_add(&cqd->things_queued_ever, 1);
-  bool will_definitely_shutdown =
-      gpr_atm_no_barrier_load(&cqd->pending_events) == 1;
+
+  /* Since we do not hold the cq lock here, it is important to do an 'acquire'
+     load here (instead of a 'no_barrier' load) to match with the release store
+     (done via gpr_atm_full_fetch_add(pending_events, -1)) in cq_shutdown_next
+     */
+  bool will_definitely_shutdown = gpr_atm_acq_load(&cqd->pending_events) == 1;
 
   if (!will_definitely_shutdown) {
     /* Only kick if this is the first item queued */
@@ -888,7 +892,7 @@ static grpc_event cq_next(grpc_completion_queue *cq, gpr_timespec deadline,
       }
     }
 
-    if (gpr_atm_no_barrier_load(&cqd->pending_events) == 0) {
+    if (gpr_atm_acq_load(&cqd->pending_events) == 0) {
       /* Before returning, check if the queue has any items left over (since
          gpr_mpscq_pop() can sometimes return NULL even if the queue is not
          empty. If so, keep retrying but do not return GRPC_QUEUE_SHUTDOWN */
@@ -934,7 +938,7 @@ static grpc_event cq_next(grpc_completion_queue *cq, gpr_timespec deadline,
   }
 
   if (cq_event_queue_num_items(&cqd->queue) > 0 &&
-      gpr_atm_no_barrier_load(&cqd->pending_events) > 0) {
+      gpr_atm_acq_load(&cqd->pending_events) > 0) {
     gpr_mu_lock(cq->mu);
     cq->poller_vtable->kick(POLLSET_FROM_CQ(cq), NULL);
     gpr_mu_unlock(cq->mu);
@@ -985,6 +989,9 @@ static void cq_shutdown_next(grpc_exec_ctx *exec_ctx,
     return;
   }
   cqd->shutdown_called = true;
+  /* Doing a full_fetch_add (i.e acq/release) here to match with
+   * cq_begin_op_for_next and and cq_end_op_for_next functions which read/write
+   * on this counter without necessarily holding a lock on cq */
   if (gpr_atm_full_fetch_add(&cqd->pending_events, -1) == 1) {
     cq_finish_shutdown_next(exec_ctx, cq);
   }