Re: [PATCH] cma: cached pageblock type fixup

[Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>]

On Tuesday 19 June 2012 14:00:45 Mel Gorman wrote:
> On Tue, Jun 19, 2012 at 01:28:50PM +0200, Bartlomiej Zolnierkiewicz wrote:
> > On Tuesday 19 June 2012 11:00:15 Mel Gorman wrote:
> > > On Wed, May 23, 2012 at 09:22:00AM +0200, Bartlomiej Zolnierkiewicz wrote:
> > > > From: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
> > > > Subject: [PATCH] cma: cached pageblock type fixup
> > > > 
> > > > CMA pages added to per-cpu pages lists in free_hot_cold_page()
> > > > have private field set to MIGRATE_CMA pageblock type .  If this
> > > > happes just before start_isolate_page_range() in alloc_contig_range()
> > > > changes pageblock type of the page to MIGRATE_ISOLATE it may result
> > > > in the cached pageblock type being stale in free_pcppages_bulk()
> > > > (which may be triggered by drain_all_pages() in alloc_contig_range()),
> > > 
> > > So what?
> > 
> > "page being added to MIGRATE_CMA free list instead of MIGRATE_ISOLATE
> > one in __free_one_page() and (if the page is reused just before
> > test_pages_isolated() check) causing alloc_contig_range() failure."
> > 
> 
> And.... so what?
> 
> If it's on the wrong free list then it can still be isolated from that
> list if it has not been allocated. If it has been allocated then it must
> be a MIGRATE_CMA-compatible allocation so the page can be migrated and
> the operation retried.
> 
> > > The pages get freed to the MIGRATE_CMA region. At worst they will be
> > > used for an allocation request that is compatible with being migrated by
> > > CMA. This will delay the allocation time of alloc_contig_range() but is
> > > hardly critical.
> > 
> > There is no waiting on MIGRATE_CMA pages in alloc_contig_range() to
> > become free again.  If the conversion to MIGRATE_ISOLATE fails then
> > the whole alloc_contig_range() allocation fails on test_pages_isolated()
> > check.
> > 
> 
> Then retry the operation if alloc_contig_range() finds it raced.

Right, my other patch is trying to go in this direction:
http://marc.info/?l=linux-mm&m=133775796822644&w=2

> > > Your fix on the other hand adds another call to get_pageblock_type() to
> > > free_pcppages_bulk which is expensive. The change made to
> > > buffered_rmqueue() is horrific. It takes the per-cpu page allocation
> > > path and adds a spin lock to it which completely defeats the purpose of
> > > having the per-cpu allocation avoid taking locks. This will have a very
> > > heavy impact on performance, particularly on parallel workloads.
> > > 
> > > As the impact of the race should be marginal and the cost of the fix is
> > > so unbelivably high I'm nacking this patch. If this race is a problem then
> > > it should be handled in alloc_contig_range() not in the allocator fast paths.
> > 
> > Do you have any idea how this race can be handled in alloc_contig_range()?
> > 
> 
> If the page is on the wrong free list, just isolate it or move it to the
> MIGRATE_ISOLATE free list at that point. If it has been allocated then
> migrate it and move the resulting free page to the MIGRATE_ISOLATE list.

Thanks, this makes sense but still leaves us with some page allocation vs
alloc_contig_range() races (i.e. pages "in-flight" state being added/removed
to/from pcp lists so not being on the freelists and not being allocated).

Best regards,
--
Bartlomiej Zolnierkiewicz
Samsung Poland R&D Center

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>