On 2014-04-16 14:11:22 +0300, Heikki Linnakangas wrote:
> On 04/16/2014 01:59 PM, Andres Freund wrote:
> >On 2014-04-16 13:50:35 +0300, Heikki Linnakangas wrote:
> >>On 04/16/2014 11:20 AM, digoal@126.com wrote:
> >>>HI,
> >>> I can reproduce this.
> >>
> >>Ok. Can you try with the attached patch applied, please?
> >
> >Perhaps this should also support partial writes, akin to
> >79ce29c734c6a652b2f7193bda537cff0c8eb8c1? I don't find it all that
> >unreasonble for the OS to refuse 16MB - or even bigger - writes in some
> >scenarios.
>
> The write in XLogFileInit is always 8kB (XLOG_BLCKSZ).
Oh, then disregard that. I just remembered you fixing it for normal wal
writes...
A strace (or the platform's equivalent) might be interesting. Digoal?
> But yeah, it might
> make sense to retry. Although I wonder why XLogFileInit would be
> particularly vulnerable to short writes. I'm afraid if we fix that, next
> we'll start seeing failures from other writes, to data files.
I personally think we indeed should fix that generally, but that's not a
small change. We'd probably need some wrappers to make that happen
sensibly.
Greetings,
Andres Freund
--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
