Re: shmem_getpage_gfp VM_BUG_ON triggered. [3.7rc2]

[Hugh Dickins]

On Thu, 25 Oct 2012, Johannes Weiner wrote:
> On Wed, Oct 24, 2012 at 09:36:27PM -0700, Hugh Dickins wrote:
> > On Wed, 24 Oct 2012, Dave Jones wrote:
> > 
> > > Machine under significant load (4gb memory used, swap usage fluctuating)
> > > triggered this...
> > > 
> > > WARNING: at mm/shmem.c:1151 shmem_getpage_gfp+0xa5c/0xa70()
> > > Pid: 29795, comm: trinity-child4 Not tainted 3.7.0-rc2+ #49
> > > 
> > > 1148                         error = shmem_add_to_page_cache(page, mapping, index,
> > > 1149                                                 gfp, swp_to_radix_entry(swap));
> > > 1150                         /* We already confirmed swap, and make no allocation */
> > > 1151                         VM_BUG_ON(error);
> > > 1152                 }
> > 
> > That's very surprising.  Easy enough to handle an error there, but
> > of course I made it a VM_BUG_ON because it violates my assumptions:
> > I rather need to understand how this can be, and I've no idea.
> 
> Could it be concurrent truncation clearing out the entry between
> shmem_confirm_swap() and shmem_add_to_page_cache()?  I don't see
> anything preventing that.
> 
> The empty slot would not match the expected swap entry this call
> passes in and the returned error would be -ENOENT.

Excellent notion, many thanks Hannes, I believe you've got it.

I've hit that truncation problem in swapoff (and commented on it
in shmem_unuse_inode), but never hit it or considered it here.
I think of the page lock as holding it stable, but truncation's
free_swap_and_cache only does a trylock on the swapcache page,
so we're not secured against that possibility.

So I'd like to change it to VM_BUG_ON(error && error != -ENOENT),
but there's a little tidying up to do in the -ENOENT case, which
needs more thought.  A delete_from_swap_cache(page) - though we
can be lazy and leave that to reclaim for such a rare occurrence -
and probably a mem_cgroup uncharge; but the memcg hooks are always
the hardest to get right, I'll have think about that one carefully.

Hugh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/