Thread (35 messages) 35 messages, 5 authors, 2012-11-18

Re: shmem_getpage_gfp VM_BUG_ON triggered. [3.7rc2]

From: Hugh Dickins <hughd@google.com>
Date: 2012-10-25 21:48:40
Also in: lkml

On Thu, 25 Oct 2012, Johannes Weiner wrote:
On Wed, Oct 24, 2012 at 09:36:27PM -0700, Hugh Dickins wrote:
quoted
On Wed, 24 Oct 2012, Dave Jones wrote:
quoted
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, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help