Roman Puchkovskiy created IGNITE-19142:
------------------------------------------
Summary: IncomingSnapshotCopier.cancel() blocks forever if called
from multiple threads
Key: IGNITE-19142
URL: https://issues.apache.org/jira/browse/IGNITE-19142
Project: Ignite
Issue Type: Bug
Reporter: Roman Puchkovskiy
Assignee: Roman Puchkovskiy
One of the test runs hang forever. Thread dump and process state analysis shown
the following:
# Some thread A invoked IncomingSnapshotCopier.cancel() (twice) on the same
copier, which made its busyLock block any operations
# Another thread B, trying to process an InstallSnapshotRequest, found that
the leader has changed (probably, due to the cluster being shut down) and
triggered 'interrupt download snapshots'
# As a result, this thread B called IncomingSnapshotCopier.cancel() on the
same copier, but its busyLock.block() (which internally just takes a write
lock) blocks this thread forever (as the lock is taken by thread A on step 1)
# Before trying to cancel the snapshot downloading, thread B took
Node.writeLock. As the thread is now blocked forever, it cannot release it, so
every operation on the JRaft Node is blocked, including shutdown
# So the cluster hangs forever on stop, making the tests hang forever as well
We should make IncomingSnapshotCopier.cancel() idempotent even when called from
different threads.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
