public inbox for [email protected]
help / color / mirror / Atom feedFrom: Nathan Bossart <[email protected]>
To: Andres Freund <[email protected]>
Cc: David Geier <[email protected]>
Cc: PostgreSQL-development <[email protected]>
Subject: Re: Fix assertion in autovacuum worker
Date: Tue, 28 Nov 2023 20:42:47 -0600
Message-ID: <20231129024247.GA479372@nathanxps13> (raw)
In-Reply-To: <[email protected]>
References: <[email protected]>
<20231128220516.GA475168@nathanxps13>
<[email protected]>
On Tue, Nov 28, 2023 at 04:03:49PM -0800, Andres Freund wrote:
> On 2023-11-28 16:05:16 -0600, Nathan Bossart wrote:
>> From a glance, it looks to me like the problem is that pgstat_shutdown_hook
>> is registered as a before_shmem_exit callback, while ProcKill is registered
>> as an on_shmem_exit callback.
>
> That's required, as pgstat_shutdown_hook() needs to acquire lwlocks, which you
> can't after ProcKill(). It's also not unique to pgstat, several other
> before_shmem_exit() callbacks acquire lwlocks (e.g. AtProcExit_Twophase(),
> ParallelWorkerShutdown(), do_pg_abort_backup(), the first three uses of
> before_shmem_exit when git grepping) - which makes sense, they are presumably
> before_shmem_exit() because they need to manage shared state, which often
> needs locks.
>
> In normal backends this is fine-ish, because ShutdownPostgres() is registered
> very late (and thus is called early in the shutdown sequence), and the
> AbortOutOfAnyTransaction() contained therein indirectly calls
> LWLockReleaseAll() and very little happens outside of the transaction context.
Right. Perhaps we could add a LWLockReleaseAll() to
pgstat_shutdown_hook() instead of the autovacuum code, but I'm afraid that
is still just a hack.
>> I would expect your patch to fix this particular issue, but I'm wondering
>> whether there's a bigger problem here.
>
> Yes, there is - our subsystem initialization, shutdown, error recovery
> infrastructure is a mess. We've interwoven transaction handling far too
> tightly with error handling, the order of subystem initialization is basically
> random and differs between operating systems (due to EXEC_BACKEND) and "mode"
> of execution (the order differs when using single user mode) and we've
> distributed error recovery into ~10 places (all the sigsetjmp()s in backend
> code, xact.c and and a few other places like WalSndErrorCleanup()).
:(
I do remember looking into uniting all the various sigsetjmp() calls
before. That could be worth another try. The rest will probably require
additional thought...
--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com
view thread (7+ messages) latest in thread
reply
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: [email protected]
Cc: [email protected], [email protected], [email protected]
Subject: Re: Fix assertion in autovacuum worker
In-Reply-To: <20231129024247.GA479372@nathanxps13>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox