Re: Anti-critical-section assertion failure in mcxt.c reached by walsender - Mailing list pgsql-hackers

From Noah Misch
Subject Re: Anti-critical-section assertion failure in mcxt.c reached by walsender
Date
Msg-id 20210508165507.GB3082635@rfd.leadboat.com
Whole thread Raw
In response to Re: Anti-critical-section assertion failure in mcxt.c reached by walsender  (Thomas Munro <thomas.munro@gmail.com>)
List pgsql-hackers
On Sat, May 08, 2021 at 04:57:54PM +1200, Thomas Munro wrote:
> On Sat, May 8, 2021 at 2:30 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> > May 07 03:31:39 gcc202 kernel: sunvdc: vdc_tx_trigger() failure, err=-11
> 
> That's -EAGAIN (assuming errnos match x86) and I guess it indicates
> that VDC_MAX_RETRIES is exceeded here:
> 
> https://github.com/torvalds/linux/blob/master/drivers/block/sunvdc.c#L451
> https://github.com/torvalds/linux/blob/master/drivers/block/sunvdc.c#L526
> 
> One theory is that the hypervisor/host is occasionally too swamped to
> service the request queue fast enough over a ~10ms period, given that
> vio_ldc_send() itself retries 1000 times with a 1us sleep, the outer
> loop tries ten times, and ldc.c's write_nonraw() reports -EAGAIN when
> there is no space for the message.  (Alternatively, it's trying to
> send a message that's too big for the channel, the channel is
> corrupted by bugs, or my fly-by of this code I'd never heard of before
> now is just way off...)

Nice discovery.  From
https://github.com/torvalds/linux/commit/a11f6ca9aef989b56cd31ff4ee2af4fb31a172ec
I see those details are 2.5 years old, somewhat young relative to the driver
as a whole.  I don't know which part should change, though.



pgsql-hackers by date:

Previous
From: Tom Lane
Date:
Subject: Re: [PATCH] force_parallel_mode and GUC categories
Next
From: Tom Lane
Date:
Subject: Re: [PATCH] Identify LWLocks in tracepoints