From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pf0-f179.google.com (mail-pf0-f179.google.com [209.85.192.179]) by kanga.kvack.org (Postfix) with ESMTP id 5A11E6B0256 for ; Sun, 6 Mar 2016 23:39:56 -0500 (EST) Received: by mail-pf0-f179.google.com with SMTP id 129so28230671pfw.1 for ; Sun, 06 Mar 2016 20:39:56 -0800 (PST) Received: from lgeamrelo12.lge.com (LGEAMRELO12.lge.com. [156.147.23.52]) by mx.google.com with ESMTP id q75si26012099pfq.207.2016.03.06.20.39.54 for ; Sun, 06 Mar 2016 20:39:55 -0800 (PST) Date: Mon, 7 Mar 2016 13:40:31 +0900 From: Joonsoo Kim Subject: Re: Suspicious error for CMA stress test Message-ID: <20160307044031.GC24602@js1304-P5Q-DELUXE> References: <56D6F008.1050600@huawei.com> <56D79284.3030009@redhat.com> <56D832BD.5080305@huawei.com> <20160304020232.GA12036@js1304-P5Q-DELUXE> <56D9325B.9060709@huawei.com> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline Content-Transfer-Encoding: 8bit In-Reply-To: <56D9325B.9060709@huawei.com> Sender: owner-linux-mm@kvack.org List-ID: To: Hanjun Guo Cc: Laura Abbott , "linux-arm-kernel@lists.infradead.org" , "linux-kernel@vger.kernel.org" , Andrew Morton , Sasha Levin , Laura Abbott , qiuxishi , Catalin Marinas , Will Deacon , Arnd Bergmann , "thunder.leizhen@huawei.com" , dingtinahong , chenjie6@huawei.com, "linux-mm@kvack.org" On Fri, Mar 04, 2016 at 02:59:39PM +0800, Hanjun Guo wrote: > On 2016/3/4 10:02, Joonsoo Kim wrote: > > On Thu, Mar 03, 2016 at 08:49:01PM +0800, Hanjun Guo wrote: > >> On 2016/3/3 15:42, Joonsoo Kim wrote: > >>> 2016-03-03 10:25 GMT+09:00 Laura Abbott : > >>>> (cc -mm and Joonsoo Kim) > >>>> > >>>> > >>>> On 03/02/2016 05:52 AM, Hanjun Guo wrote: > >>>>> Hi, > >>>>> > >>>>> I came across a suspicious error for CMA stress test: > >>>>> > >>>>> Before the test, I got: > >>>>> -bash-4.3# cat /proc/meminfo | grep Cma > >>>>> CmaTotal: 204800 kB > >>>>> CmaFree: 195044 kB > >>>>> > >>>>> > >>>>> After running the test: > >>>>> -bash-4.3# cat /proc/meminfo | grep Cma > >>>>> CmaTotal: 204800 kB > >>>>> CmaFree: 6602584 kB > >>>>> > >>>>> So the freed CMA memory is more than total.. > >>>>> > >>>>> Also the the MemFree is more than mem total: > >>>>> > >>>>> -bash-4.3# cat /proc/meminfo > >>>>> MemTotal: 16342016 kB > >>>>> MemFree: 22367268 kB > >>>>> MemAvailable: 22370528 kB > >> [...] > >>>> I played with this a bit and can see the same problem. The sanity > >>>> check of CmaFree < CmaTotal generally triggers in > >>>> __move_zone_freepage_state in unset_migratetype_isolate. > >>>> This also seems to be present as far back as v4.0 which was the > >>>> first version to have the updated accounting from Joonsoo. > >>>> Were there known limitations with the new freepage accounting, > >>>> Joonsoo? > >>> I don't know. I also played with this and looks like there is > >>> accounting problem, however, for my case, number of free page is slightly less > >>> than total. I will take a look. > >>> > >>> Hanjun, could you tell me your malloc_size? I tested with 1 and it doesn't > >>> look like your case. > >> I tested with malloc_size with 2M, and it grows much bigger than 1M, also I > >> did some other test: > > Thanks! Now, I can re-generate erronous situation you mentioned. > > > >> - run with single thread with 100000 times, everything is fine. > >> > >> - I hack the cam_alloc() and free as below [1] to see if it's lock issue, with > >> the same test with 100 multi-thread, then I got: > > [1] would not be sufficient to close this race. > > > > Try following things [A]. And, for more accurate test, I changed code a bit more > > to prevent kernel page allocation from cma area [B]. This will prevent kernel > > page allocation from cma area completely so we can focus cma_alloc/release race. > > > > Although, this is not correct fix, it could help that we can guess > > where the problem is. > > > > Thanks. > > > > [A] > > diff --git a/mm/cma.c b/mm/cma.c > > index c003274..43ed02d 100644 > > --- a/mm/cma.c > > +++ b/mm/cma.c > > @@ -496,7 +496,9 @@ bool cma_release(struct cma *cma, const struct page *pages, unsigned int count) > > > > VM_BUG_ON(pfn + count > cma->base_pfn + cma->count); > > > > + mutex_lock(&cma_mutex); > > free_contig_range(pfn, count); > > + mutex_unlock(&cma_mutex); > > cma_clear_bitmap(cma, pfn, count); > > trace_cma_release(pfn, pages, count); > > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > > index c6c38ed..1ce8a59 100644 > > --- a/mm/page_alloc.c > > +++ b/mm/page_alloc.c > > @@ -2192,7 +2192,8 @@ void free_hot_cold_page(struct page *page, bool cold) > > * excessively into the page allocator > > */ > > if (migratetype >= MIGRATE_PCPTYPES) { > > - if (unlikely(is_migrate_isolate(migratetype))) { > > + if (is_migrate_cma(migratetype) || > > + unlikely(is_migrate_isolate(migratetype))) { > > free_one_page(zone, page, pfn, 0, migratetype); > > goto out; > > } > > As I replied in previous email, the solution will fix the problem, the Cma freed memory and > system freed memory is in sane state after apply above patch. > > I also tested this situation which only apply the code below: > > if (migratetype >= MIGRATE_PCPTYPES) { > - if (unlikely(is_migrate_isolate(migratetype))) { > + if (is_migrate_cma(migratetype) || > + unlikely(is_migrate_isolate(migratetype))) { > free_one_page(zone, page, pfn, 0, migratetype); > goto out; > } > > > This will not fix the problem, but will reduce the errorous freed number of memory, > hope this helps. > > > > > > > [B] > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > > index f2dccf9..c6c38ed 100644 > > --- a/mm/page_alloc.c > > +++ b/mm/page_alloc.c > > @@ -1493,6 +1493,7 @@ static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, > > int alloc_flags) > > { > > int i; > > + bool cma = false; > > > > for (i = 0; i < (1 << order); i++) { > > struct page *p = page + i; > > @@ -1500,6 +1501,9 @@ static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, > > return 1; > > } > > > > + if (is_migrate_cma(get_pcppage_migratetype(page))) > > + cma = true; > > + > > set_page_private(page, 0); > > set_page_refcounted(page); > > > > @@ -1528,6 +1532,12 @@ static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, > > else > > clear_page_pfmemalloc(page); > > > > + if (cma) { > > + page_ref_dec(page); > > mm/page_alloc.c: In function a??prep_new_pagea??: > mm/page_alloc.c:1407:3: error: implicit declaration of function a??page_ref_deca?? [-Werror=implicit-function-declaration] > page_ref_dec(page); > ^ I tested with linux-next and there is new mechanism to manipulate page reference count and this is that. You can have same effect with atomic_dec(&page->_count) in mainline kernel. Thanks. -- 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: email@kvack.org