Fix possible crash for hedged requests

Previously, it was possible for hedged requests to continue choosing
replica even after the query had been cancelled
(RemoteQueryExecutor::tryCancel()), and not only this does not make
sense, but could also lead to a crash, due to use-after-free of
current_thread (ThreadStatus), since fiber had been created on a
different thread (thread for query pipeline), but will be destroyed from
another thread (that calls QueryPipeline dtor), and the query pipeline's
thread could be already destroyed by that time (especially under threads
pressure).

v0: IConnection::cancelAsync()
v2: remove it, since the query is sent in a deferred manner for hedged
requests, so that said that modifying HedgedConnections::sendCancel()
should be enough

Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
This commit is contained in:
Azat Khuzhin 2024-06-13 14:23:45 +02:00
parent 923caa52ba
commit ff6d1d09d6

View File

@ -255,6 +255,17 @@ void HedgedConnections::sendCancel()
if (!sent_query || cancelled) if (!sent_query || cancelled)
throw Exception(ErrorCodes::LOGICAL_ERROR, "Cannot cancel. Either no query sent or already cancelled."); throw Exception(ErrorCodes::LOGICAL_ERROR, "Cannot cancel. Either no query sent or already cancelled.");
/// All hedged connections should be stopped, since otherwise before the
/// HedgedConnectionsFactory will be destroyed (that will happen from
/// QueryPipeline dtor) they could still do some work.
/// And not only this does not make sense, but it also could lead to
/// use-after-free of the current_thread, since the thread from which they
/// had been created differs from the thread where the dtor of
/// QueryPipeline will be called and the initial thread could be already
/// destroyed (especially when the system is under pressure).
if (hedged_connections_factory.hasEventsInProcess())
hedged_connections_factory.stopChoosingReplicas();
cancelled = true; cancelled = true;
for (auto & offset_status : offset_states) for (auto & offset_status : offset_states)