Re: [PATCH v2] tmpfs: fix race between umount and writepage

[Hugh Dickins <hughd@google.com>]

On Sat, 7 May 2011, Konstantin Khlebnikov wrote:
> Hugh Dickins wrote:
> 
> > Here's the patch I was testing last night, but I do want to test it
> > some more (I've not even tried your unmounting case yet), and I do want
> > to make some changes to it (some comments, and see if I can move the
> > mem_cgroup_cache_charge outside of the mutex, making it GFP_KERNEL
> > rather than GFP_NOFS - at the time that mem_cgroup charging went in,
> > we did not know here if it was actually a shmem swap page, whereas
> > nowadays we can be sure, since that's noted in the swap_map).
> > 
> > In shmem_unuse_inode I'm widening the shmem_swaplist_mutex to protect
> > against shmem_evict_inode; and in shmem_writepage adding to the list
> > earlier, while holding lock on page still in pagecache to protect it.
> > 
> > But testing last night showed corruption on this laptop (no problem
> > on other machines): I'm guessing it's unrelated, but I can't be sure
> > of that without more extended testing.
> 
> This patch fixed my problem, I didn't catch any crashes on my test-case:
> swapout-unmount.

Thank you, Konstantin, for testing that and reporting back.

I tried using your script on Thursday, but couldn't get the tuning right
for this machine: with numbers too big everything would go OOM, with
numbers too small it wouldn't even go to swap, with numbers on the edge
it would soon settle into a steady state with almost nothing in swap.

Just once, without the patch, I did get to "Self-destruct in 5 seconds",
but that was not reproducible enough for me to test that the patch would
be fixing anything.

I was going to try today on other machines with more cpus and more memory,
though not as much as yours; but now I'll let your report save me the time,
and just add your Tested-by.  Big thank you for that!

Besides adding comments, I have changed the patch around since then, at
the shmem_unuse_inode end: to avoid any memory allocation while holding
the mutex (and then we no longer need to drop and retake info->lock,
so it gets a little simpler).  It would be dishonest of me to claim your
Tested-by for the changed code (and your mount/write/umount loop would
not have been testing swapoff): since it is an independent fix with a
different justification, I'll split that part off into a 2/3.

3/3 being the fix to the "corruption" I noticed while testing, corruption
being on the filesystem I had on /dev/loop0, over a tmpfs file filling its
filesystem: when I wrote, I'd missed the "I/O" errors in /var/log/messages.

It was another case of a long-standing but largely theoretical race,
now made easily reproducible by recent changes (the preallocation in
between find_lock_page and taking info->lock): when the filesystem is
full, you could get ENOSPC from a race in bringing back a previously
allocated page from swap.

I'll write these three up now and send off to Andrew.

Hugh