Re: [PATCH 2/2] cfq-iosched: get rid of the need for __GFP_FAIL in cfq_find_alloc_queue()

[Vivek Goyal <vgoyal@redhat.com>]

On Thu, Jul 09, 2009 at 07:38:23PM +0200, Jens Axboe wrote:
> On Thu, Jul 09 2009, Vivek Goyal wrote:
> > On Sat, Jun 27, 2009 at 08:26:17PM +0200, Jens Axboe wrote:
> > > On Fri, Jun 26 2009, Jeff Moyer wrote:
> > > > Jens Axboe <jens.axboe@oracle.com> writes:
> > > > 
> > > > > Setup an emergency fallback cfqq that we allocate at IO scheduler init
> > > > > time. If the slab allocation fails in cfq_find_alloc_queue(), we'll just
> > > > > punt IO to that cfqq instead. This ensures that cfq_find_alloc_queue()
> > > > > never fails without having to ensure free memory.
> > > > >
> > > > > Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
> > > > > ---
> > > > >  block/cfq-iosched.c |  124 +++++++++++++++++++++++++++-----------------------
> > > > >  1 files changed, 67 insertions(+), 57 deletions(-)
> > > > >
> > > > > diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
> > > > > index c760ae7..91e7e0b 100644
> > > > > --- a/block/cfq-iosched.c
> > > > > +++ b/block/cfq-iosched.c
> > > > > +	/*
> > > > > +	 * Fallback dummy cfqq for extreme OOM conditions
> > > > > +	 */
> > > > > +	struct cfq_queue oom_cfqq;
> > > > 
> > > > OK, so you're embedding a cfqq into the cfqd.  That's 136 bytes, so I
> > > > guess that's not too bad.
> > > > 
> > > > > +	/*
> > > > > +	 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
> > > > > +	 * Grab a permanent reference to it, so that the normal code flow
> > > > > +	 * will not attempt to free it.
> > > > > +	 */
> > > > > +	cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
> > > > > +	atomic_inc(&cfqd->oom_cfqq.ref);
> > > > > +
> > > > 
> > > > I guess this is so we never try to free it, good.  ;)
> > > > 
> > > > One issue I have with this patch is that, if a task happens to run into
> > > > this condition, there is no way out.  It will always have the oom_cfqq
> > > > as it's cfqq.  Can't we fix that if we recover from the OOM condition?
> > > 
> > > Yeah, I fixed that about an hour after posting the patches. See:
> > > 
> > > http://git.kernel.dk/?p=linux-2.6-block.git;a=commit;h=0370bc158cb1d5faa4b8a38c0de3211f0fd5bd64
> > > 
> > 
> > Hi Jens,
> > 
> > I think above patch might not fix the issue of an oom_cfqq getting stuck
> > with an io context. The reason being that once we allocate the cfqq, it
> > will be cached in cic and once next request comes, we will retrieve it
> > from cic and never call cfq_get_queue()/cfq_find_alloc_queue().
> > 
> > I think we probably need to do cfqq == oom_cfqq check in cfq_set_request()
> > also.
> 
> Yes good catch, this is needed too!  Can you please send as a "real"
> patch with signed-off-by added? Thanks!

Sure. Here you go.

In case memory is scarce, we now default to oom_cfqq. Once memory is
available again, we should allocate a new cfqq and stop using oom_cfqq for
a particular io context.

Once a new request comes in, check if we are using oom_cfqq, and if yes,
try to allocate a new cfqq.

Tested the patch by forcing the use of oom_cfqq and upon next request thread
realized that it was using oom_cfqq and it allocated a new cfqq.

Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/cfq-iosched.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Index: linux5/block/cfq-iosched.c
===================================================================
--- linux5.orig/block/cfq-iosched.c	2009-07-04 13:58:48.000000000 -0400
+++ linux5/block/cfq-iosched.c	2009-07-09 15:56:59.000000000 -0400
@@ -2311,7 +2311,7 @@ cfq_set_request(struct request_queue *q,
 		goto queue_fail;
 
 	cfqq = cic_to_cfqq(cic, is_sync);
-	if (!cfqq) {
+	if (!cfqq || cfqq == &cfqd->oom_cfqq) {
 		cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
 		cic_set_cfqq(cic, cfqq, is_sync);
 	}