Re: [PATCH][5/?] count writeback pages in nr_scanned

[Nick Piggin <nickpiggin@yahoo.com.au>]

Rik van Riel wrote:
> On Wed, 5 Jan 2005, Andrew Morton wrote:
> 
>> Rik van Riel <riel@redhat.com> wrote:
> 
> 
>>> The recent OOM kill problem has been happening:
>>> 1) with cache pressure on lowmem only, due to a block device write
>>> 2) with no block congestion at all
>>> 3) with pretty much all pageable lowmme pages in writeback state
>>
>>
>> You must have a wild number of requests configured in the queue.  Is 
>> this CFQ?
> 
> 
> Yes, it is with CFQ.  Around 650MB of lowmem is in writeback
> stage, which is over 99% of the active and inactive lowmem
> pages...
> 
>> I've done testing with "all of memory under writeback" before and it 
>> went OK.  It's certainly a design objective to handle this well.  But 
>> that testing was before we broke it.
> 
> 
> I suspect something might still be broken.  It may take a few
> days of continuous testing to trigger the bug, though ...
> 

It is possible to be those blk_congestion_wait paths, because
the queue simply won't be congested. So doing io_schedule_timeout
might help.

I wonder if reducing the size of the write queue in CFQ would help
too? IIRC, it only really wants a huge read queue.