Date: Fri, 8 Nov 2024 15:36:49 -0800
From: Noah Misch <noah@leadboat.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Francesco Degrassi <francesco.degrassi@optionfactory.net>,
	pgsql-bugs@lists.postgresql.org
Subject: Re: Leader backend hang on IPC/ParallelFinish when LWLock held at
 parallel query start
Message-ID: <20241108233649.01.nmisch@google.com>
References: 
 <CAC-SaSzHUKT=vZJ8MPxYdC_URPfax+yoA1hKTcF4ROz_Q6z0_Q@mail.gmail.com>
 <20240918030159.2a.nmisch@google.com>
 <1737570.1726633422@sss.pgh.pa.us>
 <20240918050434.13.nmisch@google.com>
 <3441950.1730999841@sss.pgh.pa.us>
 <20241107192233.94.nmisch@google.com>
 <3459938.1731007759@sss.pgh.pa.us>
 <20241107193823.dc.nmisch@google.com>
 <3694355.1731084396@sss.pgh.pa.us>
 <3793541.1731088615@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <3793541.1731088615@sss.pgh.pa.us>
User-Agent: Mutt/2.2.12 (2023-09-09)
Archived-At: 
 <https://www.postgresql.org/message-id/20241108233649.01.nmisch%40google.com>
Precedence: bulk

On Fri, Nov 08, 2024 at 12:56:55PM -0500, Tom Lane wrote:
> I wrote:
> > Here's a proposed patch along that line.  I left the test case from
> > ac04aa84a alone, since it works perfectly well to test this way too.
> 
> I'd modeled that on the existing recovery code for DSM segment creation
> failure, just below.  But a look at the code coverage report shows
> (unsurprisingly) that that path is never exercised in our regression
> tests, so I wondered if it actually works ... and it doesn't work
> very well.  To test, I lobotomized InitializeParallelDSM to always
> force pcxt->nworkers = 0.  That results in a bunch of unsurprising
> regression test diffs, plus a couple of
> 
> +ERROR:  could not find key 4 in shm TOC at 0x229f138
> 
> which turns out to be the fault of ExecHashJoinReInitializeDSM:
> it's not accounting for the possibility that we didn't really
> start a parallel hash join.
> 
> I'm also not happy about ReinitializeParallelWorkers'
> 
> 	Assert(pcxt->nworkers >= nworkers_to_launch);
> 
> The one existing caller manages not to trigger that because it's
> careful to reduce its request based on pcxt->nworkers, but it
> doesn't seem to me that callers should be expected to have to.
> 
> So I end with the attached.  There might still be some more issues
> that the regression tests don't reach, but I think this is the
> best we can do for today.

Looks good.