Re: [PATCH 12/12] mm: Throttle direct reclaimers if PF_MEMALLOC reserves are low and swap is backed by network storage

[Mel Gorman <mgorman@suse.de>]

On Mon, Apr 18, 2011 at 10:32:51PM +1000, NeilBrown wrote:
> On Thu, 14 Apr 2011 11:41:38 +0100 Mel Gorman <mgorman@suse.de> wrote:
> 
> > If swap is backed by network storage such as NBD, there is a risk that a
> > large number of reclaimers can hang the system by consuming all
> > PF_MEMALLOC reserves. To avoid these hangs, the administrator must tune
> > min_free_kbytes in advance. This patch will throttle direct reclaimers
> > if half the PF_MEMALLOC reserves are in use as the system is at risk of
> > hanging. A message will be displayed so the administrator knows that
> > min_free_kbytes should be tuned to a higher value to avoid the
> > throttling in the future.
> > 
> 
> (I knew there was something else).
> 
> I understand that there are suggestions that direct reclaim should always be
> serialised as this reduces lock contention and improve data patterns (or
> something like that).
> 

AFAIK, this suggestion never got much beyond the "hand-waving" stage
of development. It tended to trip up on the fact that such a feature
could also throttle processes on machines with plenty of free clean
unmapped pagecache which would be undesirable.

> Would that make this patch redundant? 

Depends on how it was being serialised but ....

> Or does this provide some extra
> guarantee that the other proposal would not?
> 

This patch could be extended to serialise direct reclaims in situations
other than PFMEMALLOC is low if someone demonstrated the benefit.

-- 
Mel Gorman
SUSE Labs