Re: Race conditions with WAL sender PID lookups

From: Michael Paquier <michael(dot)paquier(at)gmail(dot)com>
To: Thomas Munro <thomas(dot)munro(at)enterprisedb(dot)com>
Cc: PostgreSQL mailing lists <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: Race conditions with WAL sender PID lookups
Date: 2017-05-19 02:05:45
Message-ID: CAB7nPqRa3EpFedbW8RuHpmSn8u_fdzV3oCRhRndFqWfELSvHVg@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

On Thu, May 18, 2017 at 1:43 PM, Thomas Munro
<thomas(dot)munro(at)enterprisedb(dot)com> wrote:
> On Thu, May 11, 2017 at 1:48 PM, Michael Paquier
> <michael(dot)paquier(at)gmail(dot)com> wrote:
>> I had my eyes on the WAL sender code this morning, and I have noticed
>> that walsender.c is not completely consistent with the PID lookups it
>> does in walsender.c. In two code paths, the PID value is checked
>> without holding the WAL sender spin lock (WalSndRqstFileReload and
>> pg_stat_get_wal_senders), which looks like a very bad idea contrary to
>> what the new WalSndWaitStopping() does and what InitWalSenderSlot() is
>> doing for ages.
>
> There is also code that accesses shared walsender state without
> spinlocks over in syncrep.c. I think that file could use a few words
> of explanation for why it's OK to access pid, state and flush without
> synchronisation.

Yes, that is read during the quorum and priority sync evaluation.
Except sync_standby_priority, all the other variables should be
protected using the spin lock of the WAL sender. walsender_private.h
is clear regarding that. So the current coding is inconsistent even
there. Attached is an updated patch.
--
Michael

Attachment Content-Type Size
walsnd-pid-races-v2.patch application/octet-stream 4.0 KB

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Michael Paquier 2017-05-19 02:08:41 Re: [bug fix] PG10: libpq doesn't connect to alternative hosts when some errors occur
Previous Message Peter Eisentraut 2017-05-19 02:02:45 Re: different column orders in regression test database