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

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

Marcelo Tosatti wrote:
>>The caller would need to wait on all the zones which can satisfy the
>>caller's allocation request.  A bit messy, although not rocket science. 
>>One would have to be careful to avoid additional CPU consumption due to
>>delivery of multiple wakeups at each I/O completion.
>>
>>We should be able to demonstrate that such a change really fixes some
>>problem though.  Otherwise, why bother?
> 
> 
> Agreed. The current scheme works well enough, we dont have spurious OOM kills
> anymore, which is the only "problem" such change ought to fix. 
> 
> You might have performance increase in some situations I believe (because you 
> have perzone waitqueues), but I agree its does not seem to be worth the
> trouble.

I think what Andrea is worried about is that blk_congestion_wait is
fairly vague, and can be a source of instability in the scanning
implementation.

For example, if you have a heavy IO workload that is saturating your
disks, blk_congestion_wait may do the right thing and sleep until
they become uncongested and writeout can continue.

But at 2:00 am, when your backup job is trickling writes into another
block device, blk_congestion_wait returns much earlier, and before
many pages have been cleaned.

Bad example? Yeah maybe, but I think this is what Andrea is getting
at. Would it be a problem to replace those blk_congestion_waits with
unconditional io_schedule_timeout()s? That would be the dumb-but-more
-deterministic solution.