Re: Kernel 2.6.9 Multiple Page Allocation Failures

[Marcelo Tosatti <marcelo.tosatti@cyclades.com>]

On Sat, Nov 13, 2004 at 03:47:43PM +0100, Stefan Schmidt wrote:
> On Fri, Nov 12, 2004 at 11:09:29PM +1100, Nick Piggin wrote:
> > OK. Occasional page allocation failures are not a problem, although
> > if it is an order-0 allocation it may be an idea to increase
> > min_free_kbytes a bit more.
> What about more than occasional but more like constant errors? ;)
> 
> > I think you said earlier that you had min_free_kbytes set to 8192?
> > Well after applying my patch, the memory watermarks get squashed
> > down, so you'd want to set it to at least 16384 afterward. Maybe
> > more.
> I took the default from 2.6.10-rc1-bk19 with your patch and doubled it. No
> luck with the following values subsequently applied:
> #vm.min_free_kbytes=3831
> #vm.min_free_kbytes=7662
> #vm.min_free_kbytes=15324
> #vm.min_free_kbytes=61296
> vm.min_free_kbytes=65535
> Did not help against the page allocation errors or boosting up the machines
> performance.

Nick, such high reservations should have protected the system from OOM.

> I got XFS filesystem corruption in the end which (just) perhaps was triggered by
> page allocation errors for it says:
> XFS internal error XFS_WANT_CORRUPTED_RETURN after several page allocation
> errors for the application having XFS in its stack trace leading to:
> kernel: xfs_force_shutdown(sdd5,0x8) called from line 1091 of file
>         fs/xfs/xfs_trans.c.  Return address = 0xc0212e5c
> kernel: Filesystem "sdd5": Corruption of in-memory data detected.
>         Shutting down filesystem: sdd5
> kernel: Please umount the filesystem, and rectify the problem(s)
> 
> Corruption of in-memory data sounds like something we would not want to
> happen, right?

Definately. I suspect XFS is unable to handle OOM graciously, or some other
problem.

> I attached full traces to http://oss.sgi.com/bugzilla/show_bug.cgi?id=345 which
> is an rather old ticket dealing with page allocation errors and xfs but
> perhaps it does fit.
> 
> 	Stefan