Re: lot of VM problem with 2.4.23

[Andrea Arcangeli <andrea@suse.de>]

On Mon, Dec 22, 2003 at 09:03:44AM +0200, Ville Herva wrote:
> On Sun, Dec 21, 2003 at 11:23:38PM +0100, you [Willy Tarreau] wrote:
> > >
> > > Dec 21 22:08:44 stock kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0)
> > > Dec 21 22:08:44 stock kernel: OOM: nr_swap_pages=0cd865e6c c012e1e8 c0262e3c 00000000 000001d2 00000000 00000001 cd863c00 
> > 
> > OK, so there's no available swap anymore (nr_swap_pages=0, Marcelo forgot the
> > '\n' in the patch). I simply think that with other kernels, you're very short
> > of memory, but it runs, while with this one, all memory gets consumed, and
> > since there's no smart oom killer, one process has to get killed.
> 
> BTW, I have a box with 128MB ram and 512MB swap running 2.4.21-jam1 (it has
> the -aa vm). I can't shut it down cleanly, because trying it goes into
> endless loop trying to free memory when turning off swap. Nothing but
> alt-sysrq-b seems to work.
> 
> I don't know if there is a kernel memory leak, since all user level
> processes should be killed at that point, right? Unfortunately I didn't have
> time to dig deeper, as the box is in (sort of) production.

if this is a leak, I doubt it has been introduced recently, the only
swap accounting related change was in the shm layer, and it was supposed
to be a race fix. You may want to check if you've some shm allocated
in /dev/shm or ipcs, while the machine reboots.