Date: Fri, 12 Apr 2024 11:01:24 +0900 (JST)
Message-Id: <20240412.110124.977560148390249226.horikyota.ntt@gmail.com>
To: michael@paquier.xyz
Cc: reshkekirill@gmail.com, nathandbossart@gmail.com, antholeu@amazon.com,
 x4mmm@yandex-team.ru, pgsql-hackers@postgresql.org
Subject: Re: Allow non-superuser to cancel superuser tasks.
From: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
In-Reply-To: <Zhh7-1JkKr-BMPCP@paquier.xyz>
References: <ZhcQ8Ki3Ck6VlYqu@paquier.xyz>
	<CALdSSPiOLADNe1ZS-1dnQXWVXgGAC6=3TBKABu9C3_97YGOoMg@mail.gmail.com>
	<Zhh7-1JkKr-BMPCP@paquier.xyz>
User-Agent: Mew version 6.8 on Emacs 27.2
Mime-Version: 1.0
Content-Type: Text/Plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Archived-At: <https://www.postgresql.org/message-id/20240412.110124.977560148390249226.horikyota.ntt%40gmail.com>
Precedence: bulk

At Fri, 12 Apr 2024 09:10:35 +0900, Michael Paquier <michael@paquier.xyz> wrote in 
> On Thu, Apr 11, 2024 at 04:55:59PM +0500, Kirill Reshke wrote:
> > The test doesn't fail because pg_terminate_backend actually meets his
> > point: autovac is killed. But while dying, autovac also receives
> > segfault. Thats because of injections points:
> > 
> > (gdb) bt
> > #0  0x000056361c3379ea in tas (lock=0x7fbcb9632224 <error: Cannot
> > access memory at address 0x7fbcb9632224>) at
> > ../../../../src/include/storage/s_lock.h:228
> > #1  ConditionVariableCancelSleep () at condition_variable.c:238
...
> > #3  0x000056361c330a40 in CleanupProcSignalState (status=<optimized
out>, arg=<optimized out>) at procsignal.c:240
> > #4  0x000056361c328801 in shmem_exit (code=code@entry=1) at ipc.c:276
> > #9  0x000056361c3378d7 in ConditionVariableTimedSleep
> > (cv=0x7fbcb9632224, timeout=timeout@entry=-1,
...
> I can see this stack trace as well.  Capturing a bit more than your
> own stack, this is crashing in the autovacuum worker while waiting on
> a condition variable when processing a ProcessInterrupts().
> 
> That may point to a legit bug with condition variables in this
> context, actually?  From what I can see, issuing a signal on a backend
> process waiting with a condition variable is able to process the
> interrupt correctly.

ProcSignalInit sets up CleanupProcSignalState to be called via
on_shmem_exit. If the CV is allocated in a dsm segment, shmem_exit
should have detached the region for the CV.  CV cleanup code should be
invoked via before_shmem_exit.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center