[HBASE-27093] AsyncNonMetaRegionLocator：put Complete CompletableFuture outside lock block

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.0.0-alpha-2, 2.4.12
• Fix Version/s: 2.5.0, 3.0.0-alpha-3, 2.4.13
• Component/s: asyncclient, Client
• Labels: deadlock

Environment

2.4.12

• Hadoop Flags: Reviewed
• Assignee: zhiwang
• Reporter: zhiwang
• Created: 07/Jun/22 03:02
• Updated: 08/Jun/22 06:08
• Resolved: 07/Jun/22 04:26

Description

 

synchronized (tableCache) {
  tableCache.pendingRequests.remove(req);
  // fail the request itself, no matter whether it is a DoNotRetryIOException, as we have
  // already retried several times
  CompletableFuture<?> future = tableCache.allRequests.remove(req);
  if (future != null) {
    future.completeExceptionally(error);
  }
  tableCache.clearCompletedRequests(null);
  // Remove a complete locate request in a synchronized block, so the table cache must have
  // quota to send a candidate request.
  toSend = tableCache.getCandidate();
  toSend.ifPresent(r -> tableCache.send(r));
}

Complete or completeExceptionally in a lock block may cause deadlock. We can put this action outside the lock block to avoid deadlock.

 

Issue Links

• links to: GitHub Pull Request #4496