Hang in semop() in 'waiting' backends

Поиск
Список
Период
Сортировка
Искать

Hang in semop() in 'waiting' backends

От:
Robert Fielding <rob@dsvr.net>
Дата:
Hi,

I'm afraid information is a little sketchy with this one, but I'll 
present to you quickly what's going on as best as I can. Your 
assistance is appreciated:

platform: Linux 2.4.20 #19 SMP RedHat 9 base, PG 7.4.5 (also 7.4.2)
problem: postgres backends waiting forever

strace shows:

semop(19529737, 0xbfffe020, 1

SELECT domain_add ('robf', 'testdomain123456.uk', '212.69.194.46' )  ;\


domain_add is a very complex function which calls lots of the SELECT 
functions and performs string functions on the input.

I've tried reindexing both domains_pkey and domains_vs_idx indexes 
attached to the domains table with no effect. I've also restarted the 
service several times. This only results as a temporary fix in clearing 
the waiting processes, but not adding domains.

Instead I have suspended service, dumped and recreated the DNS 
database. I have also upgraded to PG 7.4.5 (I know not the latest, but 
this machine doesn't have external access). The dump and restore 
resolved the issue temporarily.

ipcs shows all shared memory use clears when the service is stopped, 
and resumes when restarted.

Individual calls to domain_add() work; multiple calls within a 
transaction block on the first call. The individual calls from a psql 
shell also fail after this. The only fix is to kill the backends which 
are waiting.

This is a production system which has suddenly gone wrong. Last schema 
update was posted to CVS 2005/03/21 which included UPDATE and DELETE 
constraints and foreign key updates; It seems strange updates from 23rd 
of last month would start causing problems now. I will need to talk to 
the developer responsible if any updates haven't been posted to CVS - 
however this is not our practice. Given the hang, could it still be 
described as a bug anyway?

Best regards,

Rob Fielding
Designer Servers
Business Serve plc

Re: Hang in semop() in 'waiting' backends

От:
Tom Lane <tgl@sss.pgh.pa.us>
Дата:
Robert Fielding  writes:
> strace shows:
> semop(19529737, 0xbfffe020, 1

Can you attach to a few of the stuck processes with gdb and get stack
traces?  Are they all doing exactly the same query when they hang?

			regards, tom lane
FAQ