Re: [HACKERS] backend freezeing on win32 fixed (I hope ;-) ) - Mailing list pgsql-hackers

From Tom Lane
Subject Re: [HACKERS] backend freezeing on win32 fixed (I hope ;-) )
Date
Msg-id 601.934812420@sss.pgh.pa.us
Whole thread Raw
In response to backend freezeing on win32 fixed (I hope ;-) )  (Horak Daniel <horak@mmp.plzen-city.cz>)
Responses Re: [HACKERS] backend freezeing on win32 fixed (I hope ;-) )
List pgsql-hackers
Horak Daniel <horak@mmp.plzen-city.cz> writes:
> I think I have fixed the freezing of the postgres backend on Windows NT. Now
> it survives 5 regression test in a cycle with some concurrent connections
> during running the tests.
> It have looked like a problem with initializing the same semaphore for
> second time (they are "preinitialized" for performance reasons in
> InitProcGlobal() in storage/lmgr/proc.c)

They should never be "initialized a second time".  And the preallocation
is *not* for performance reasons, it is to make sure we can actually get
enough semaphores (rather than dying under load when we fail to get the
N+1'st semaphore when starting the N+1'st backend).

> The fix (made for v6.5.1) is here:
> [ Fix consists of diking out preallocation of semaphores by postmaster ]

I do not like this patch one bit --- I think it is voodoo that doesn't
really have anything to do with the true problem.  I don't know what
the true problem is, mind you, but I don't think this is the way to
fix it.

Is it possible that the CygWin environment doesn't have a correct
emulation of IPC semaphores, such that a sema allocated by one process
(the postmaster) is not available to other procs (the backends)?
That would explain preallocation not working --- but if that's it then
we have major problems in other places, since the code assumes that a
sema once allocated will remain available to later backends.
        regards, tom lane


pgsql-hackers by date:

Previous
From: The Hermit Hacker
Date:
Subject: Re: [CORE] Re: tomorrow
Next
From: Horak Daniel
Date:
Subject: RE: [HACKERS] backend freezeing on win32 fixed (I hope ;-) )