Thread: Is this a bug in auto vacuum worker?

Is this a bug in auto vacuum worker?

From
CNG L
Date:
In AutoVacWorkerMain(), the code calls on_shmem_exit(FreeWorkInfo,0) and then kill() to send SIGUSR2 to autovacuum launcher process. In FreeWorkInfo() it sets AutoVacuumShmem->av_signal[AutoVacRebalance] = true. But, is it possible that when autovacuum launcher’s SIGUSR2 handler checks av_signal[AutoVacRebalance], FreeWorkInfo() haven’t set this flag yet and thus autovacuum launcher will miss the AutoVacRebalance event?

Re: Is this a bug in auto vacuum worker?

From
Michael Paquier
Date:
On Tue, Jan 08, 2019 at 10:47:01AM -0800, CNG L wrote:
> In AutoVacWorkerMain(), the code calls on_shmem_exit(FreeWorkInfo,0) and
> then kill() to send SIGUSR2 to autovacuum launcher process. In
> FreeWorkInfo() it sets AutoVacuumShmem->av_signal[AutoVacRebalance] = true.
> But, is it possible that when autovacuum launcher’s SIGUSR2 handler checks
> av_signal[AutoVacRebalance], FreeWorkInfo() haven’t set this flag yet and
> thus autovacuum launcher will miss the AutoVacRebalance event?

No, I don't think so.  There are a couple of instructions between the
moment SIGUSR2 is sent from the worker to the launcher for rebalancing
and the moment where on_shmem_exit() callbacks are called, still the
launcher gets aware of the death of the worker in ProcKill, so the
order of the events is right (double-checked manually as well):
1) Send SIGUSR2 from the worker to the launcher, asking for
rebalancing.
2) shmem callback is called.
3) launcher noticies the death of the worker, sees the signal, and
triggers the rebalancing.
--
Michael

Attachment