]> git.ipfire.org Git - thirdparty/postgresql.git/commit
Fix possibility of self-deadlock in ResolveRecoveryConflictWithBufferPin().
authorAndres Freund <andres@anarazel.de>
Tue, 3 May 2022 01:25:00 +0000 (18:25 -0700)
committerAndres Freund <andres@anarazel.de>
Tue, 3 May 2022 01:30:15 +0000 (18:30 -0700)
commit9cda785b4ea35f1bc6ae61945b31dc1e09b5aec5
tree311664bfc9e5cf4e7007a2ab02d681b3e28f4183
parent2adb8debe5aa30caee9810a987f264c4f1515395
Fix possibility of self-deadlock in ResolveRecoveryConflictWithBufferPin().

The tests added in 9f8a050f68d failed nearly reliably on FreeBSD in CI, and
occasionally on the buildfarm. That turns out to be caused not by a bug in the
test, but by a longstanding bug in recovery conflict handling.

The standby timeout handler, used by ResolveRecoveryConflictWithBufferPin(),
executed SendRecoveryConflictWithBufferPin() inside a signal handler. A bad
idea, because the deadlock timeout handler (or a spurious latch set) could
have interrupted ProcWaitForSignal(). If unlucky that could cause a
self-deadlock on ProcArrayLock, if the deadlock check is in
SendRecoveryConflictWithBufferPin()->CancelDBBackends().

To fix, set a flag in StandbyTimeoutHandler(), and check the flag in
ResolveRecoveryConflictWithBufferPin().

Subsequently the recovery conflict tests will be backpatched.

Discussion: https://postgr.es/m/20220413002626.udl7lll7f3o7nre7@alap3.anarazel.de
Backpatch: 10-
src/backend/storage/ipc/standby.c