From: | Robert Haas <robertmhaas(at)gmail(dot)com> |
---|---|
To: | Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> |
Cc: | pgsql-hackers <pgsql-hackers(at)postgresql(dot)org> |
Subject: | Re: Group clear xid can leak semaphore count |
Date: | 2017-01-05 19:43:40 |
Message-ID: | CA+TgmoZY4f=XKtaVPd_XJFYAGWu7Pz5wtL_mpbi-n_VBCbqiPw@mail.gmail.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
On Sat, Dec 31, 2016 at 12:44 AM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> During the review of Group update Clog patch [1], Dilip noticed an
> issue with the patch where it can leak the semaphore count in one of
> the corner case. I have checked and found that similar issue exists
> for Group clear xid (ProcArrayGroupClearXid) as well. I think the
> reason why this problem is not noticed by anyone till now is that it
> can happen only in a rare scenario when the backend waiting for xid
> clear is woken up due to some unrelated signal. This problem didn't
> exist in the original commit
> (0e141c0fbb211bdd23783afa731e3eef95c9ad7a) of the patch, but later
> while fixing some issues in the committed patch, it got introduced in
> commit 4aec49899e5782247e134f94ce1c6ee926f88e1c. Patch to fix the
> issue is attached.
Yeah, that looks like a bug. Thanks for the detailed analysis;
committed and back-patched to 9.6.
I suppose in the worst case it's possible that we'd leak a semaphore
count and then every future time we enter a PGSemaphoreLock using that
PGPROC we have to eat up the leaked count (or counts) and then put it
(or them) back after we really wait. That would suck. But I wasn't
able to observe any leaks in a high-concurrency pgbench test on hydra,
so it's either very unlikely or requires some additional condition to
trigger the problem.
I think we have run into this kind of issue before. I wonder if
there's any way to insert some kind of a guard - e.g. detect at
backend startup time that the semaphore has a non-zero value and fix
it, issuing a warning along the way... maybe something like:
while (sem_trywait(sem) == 0)
++bogus;
if (bogus > 0)
elog(WARNING, "uh oh");
I'm not sure if that would be prone to false positives though.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
From | Date | Subject | |
---|---|---|---|
Next Message | Vitaly Burovoy | 2017-01-05 19:46:28 | Re: Re: [BUGS][PATCH] BUG #14486: Inserting and selecting interval have different constraints |
Previous Message | Andres Freund | 2017-01-05 19:10:12 | Re: [PATCH] Rename pg_switch_xlog to pg_switch_wal |