Re: deadlock with latest xfs

[Dave Chinner <david@fromorbit.com>]

On Mon, Oct 27, 2008 at 12:42:09PM +1100, Lachlan McIlroy wrote:
> Dave Chinner wrote:
>> On Sun, Oct 26, 2008 at 11:53:51AM +1100, Dave Chinner wrote:
>>> On Fri, Oct 24, 2008 at 05:48:04PM +1100, Dave Chinner wrote:
>>>> OK, I just hung a single-threaded rm -rf after this completed:
>>>>
>>>> # fsstress -p 1024 -n 100 -d /mnt/xfs2/fsstress
>>>>
>>>> It has hung with this trace:
....
>> Got it now. I can reproduce this in a couple of minutes now that both
>> the test fs and the fs hosting the UML fs images are using lazy-count=1
>> (and the frequent 10s long host system freezes have gone away, too).
>>
>> Looks like *another* new memory allocation problem [1]:
.....
>> We've entered memory reclaim inside the xfsdatad while trying to do
>> unwritten extent completion during I/O completion, and that memory
>> reclaim is now blocked waiting for I/o completion that cannot make
>> progress.
>>
>> Nasty.
>>
>> My initial though is to make _xfs_trans_alloc() able to take a KM_NOFS argument
>> so we don't re-enter the FS here. If we get an ENOMEM in this case, we should
>> then re-queue the I/O completion at the back of the workqueue and let other
>> I/o completions progress before retrying this one. That way the I/O that
>> is simply cleaning memory will make progress, hence allowing memory
>> allocation to occur successfully when we retry this I/O completion...
> It could work - unless it's a synchronous I/O in which case the I/O is not
> complete until the extent conversion takes place.

Right. Pushing unwritten extent conversion onto a different
workqueue is probably the only way to handle this easily.
That's the same solution Irix has been using for a long time
(the xfsc thread)....

> Could we allocate the memory up front before the I/O is issued?

Possibly, but that will create more memory pressure than
allocation in I/O completion because now we could need to hold
thousands of allocations across an I/O - think of the case where
we are running low on memory and have a disk subsystem capable of
a few hundred thousand I/Os per second. the allocation failing would
prevent the I/os from being issued, and if this is buffered writes
into unwritten extents we'd be preventing dirty pages from being
cleaned....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com