Hi Gioh, On Fri, Jul 18, 2014 at 03:45:36PM +0900, Gioh Kim wrote: > > Hi, > > For page migration of CMA, buffer-heads of lru should be dropped. > Please refer to https://lkml.org/lkml/2014/7/4/101 for the history. Just nit: Please write *problem* in description instead of URL link. > > I have two solution to drop bhs. > One is invalidating entire lru. You mean? All of percpu bh_lrus so if the system has N cpu, it invalidates N * 8? > Another is searching the lru and dropping only one bh that Laura proposed > at https://lkml.org/lkml/2012/8/31/313. > > I'm not sure which has better performance. For whom? system or requestor of CMA? > So I did performance test on my cortex-a7 platform with Lmbench > that has "File & VM system latencies" test. > I am attaching the results. > The first line is of invalidating entire lru and the second is dropping selected bh. You mean you did Lmbench with background CMA allocation? Could you describe in detail? > > File & VM system latencies in microseconds - smaller is better > ------------------------------------------------------------------------------- > Host OS 0K File 10K File Mmap Prot Page 100fd > Create Delete Create Delete Latency Fault Fault selct > --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- > 10.178.33 Linux 3.10.19 25.1 19.6 32.6 19.7 5098.0 0.666 3.45880 6.506 > 10.178.33 Linux 3.10.19 24.9 19.5 32.3 19.4 5059.0 0.563 3.46380 6.521 > > > I tried several times but the result tells that they are the same under 1% gap > except Protection Fault. > But the latency of Protection Fault is very small and I think it has little effect. > > Therefore we can choose anything but I choose invalidating entire lru. Not sure we can conclude like that. A few weeks ago, I saw a patch which increases bh_lrus's size. https://lkml.org/lkml/2014/7/4/107 IOW, some of workloads really affects by percpu bh_lrus so it would be better to be careful to drain, I think. You want to argue CMA allocation is rare so the cost is marginable. It might but some of usecase might call it frequently with small request (ie, 8K, 16K). Anyway, why cannot CMA have the cost without affecting other subsystem? I mean it's okay for CMA to consume more time to shoot out the bh instead of simple all bh_lru invalidation because big order allocation is kinds of slow thing in the VM and everybody already know that and even sometime get failed so it's okay to add more code that extremly slow path. > The try_to_free_buffers() which is calling drop_buffers() is called by many filesystem code. > So I think inserting codes in drop_buffers() can affect the system. > And also we cannot distinguish migration type in drop_buffers(). > > In alloc_contig_range() we can distinguish migration type and invalidate lru if it needs. > I think alloc_contig_range() is proper to deal with bh like following patch. > > Laura, can I have you name on Acked-by line? > Please let me represent my thanks. > > Thanks for any feedback. > > ------------------------------- 8< ---------------------------------- > > >From 33c894b1bab9bc26486716f0c62c452d3a04d35d Mon Sep 17 00:00:00 2001 > From: Gioh Kim <gioh.kim@xxxxxxx> > Date: Fri, 18 Jul 2014 13:40:01 +0900 > Subject: [PATCH] CMA/HOTPLUG: clear buffer-head lru before page migration > > The bh must be free to migrate a page at which bh is mapped. > The reference count of bh is increased when it is installed > into lru so that the bh of lru must be freed before migrating the page. > > This frees every bh of lru. We could free only bh of migrating page. > But searching lru costs more than invalidating entire lru. > > Signed-off-by: Gioh Kim <gioh.kim@xxxxxxx> > Acked-by: Laura Abbott <lauraa@xxxxxxxxxxxxxx> > --- > mm/page_alloc.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index b99643d4..3b474e0 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -6369,6 +6369,9 @@ int alloc_contig_range(unsigned long start, unsigned long end, > if (ret) > return ret; > > + if (migratetype == MIGRATE_CMA || migratetype == MIGRATE_MOVABLE) > + invalidate_bh_lrus(); > + Q1. It's a only CMA problem? Memory-Hotplug is not a problem? Or other places? I mean it would be better to handle in generic way. Q2. Why do you call it right before calling __alloc_contig_migrate_range? Some of pages will go bh_lrus by __alloc_contig_migrate_ranges. In that case, it is useless without caller's retry logic. Even you do it from caller's retrial logic, it's not a good idea because you makes new binding alloc_contig_range and uppder layer. So, IMHO, it would be better to handle it in migrate_pages. Maybe we could define new API try_to_drop_buffers which calls try_to_free_buffers and then only if the function fails due to percpu lru count, we could drain only the bh in percpu lru list instead of all bh draining. And places in migration path should use it rather than try_to_relese_page. But the problem from this approach invents new API which should be maintained so not sure Andrew think it's worth. Maybe we should see the code and diffstat. Overenginnering? > ret = __alloc_contig_migrate_range(&cc, start, end); > if (ret) > goto done; > -- > 1.7.9.5 > > -- > To unsubscribe, send a message with 'unsubscribe linux-mm' in > the body to majordomo@xxxxxxxxx. For more info on Linux MM, > see: http://www.linux-mm.org/ . > Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a> -- Kind regards, Minchan Kim -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
