* [RFC PATCH 0/2] avoid external fragmentation related to migration fallback @ 2016-01-29 19:23 ChengYi He 2016-01-29 19:24 ` [RFC PATCH 1/2] mm/page_alloc: let migration fallback support pages of requested order ChengYi He ` (2 more replies) 0 siblings, 3 replies; 8+ messages in thread From: ChengYi He @ 2016-01-29 19:23 UTC (permalink / raw) To: Andrew Morton Cc: Mel Gorman, Michal Hocko, Vlastimil Babka, David Rientjes, Joonsoo Kim, Yaowei Bai, Xishi Qiu, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel, chengyihetaipei The patchset tries to avoid external fragmentation related to migration fallback. The symptom of this external fragmentation is that there are lots of allocation failures of order-2 and order-3 pages and many threads enter allocation slowpath to compact and direct reclaim pages. This will degrade system performance and make it less responsive to users. This symptom could be observed in devices which have only one memory zone, such as some arm64 android devices. Below are two examples in which processes fail to allocate order-2 and order-3 pages and enter direct reclaim. 1) In arm64, THREAD_SIZE_ORDER is 2. The thread enters direct reclaim while allocating an order-2 page as kernel stack. [<ffffffc0002077dc>] dump_backtrace+0x0/0x134 [<ffffffc000207920>] show_stack+0x10/0x1c [<ffffffc000cedc64>] dump_stack+0x1c/0x28 [<ffffffc0002cb6bc>] try_to_free_pages+0x5d8/0x704 [<ffffffc0002c219c>] __alloc_pages_nodemask+0x544/0x834 [<ffffffc00021a1e4>] copy_process.part.58+0xf4/0xdfc [<ffffffc00021b000>] do_fork+0xe0/0x358 [<ffffffc00021b310>] SyS_clone+0x10/0x1c 2) The thread enters direct reclaim while allocating an order-3 page. [<ffffffc0002077dc>] dump_backtrace+0x0/0x134 [<ffffffc000207920>] show_stack+0x10/0x1c [<ffffffc000cedc64>] dump_stack+0x1c/0x28 [<ffffffc0002cb6bc>] try_to_free_pages+0x5d8/0x704 [<ffffffc0002c219c>] __alloc_pages_nodemask+0x544/0x834 [<ffffffc0002f0c68>] new_slab+0x80/0x23c [<ffffffc0002f2780>] __slab_alloc.isra.50.constprop.55+0x26c/0x300 [<ffffffc0002f2abc>] kmem_cache_alloc+0x94/0x1d4 [<ffffffc0003e3a1c>] fuse_alloc_inode+0x20/0xb8 [<ffffffc00030f468>] alloc_inode+0x1c/0x90 [<ffffffc000310134>] iget5_locked+0xa0/0x1c0 [<ffffffc0003e3d00>] fuse_iget+0x60/0x1bc [<ffffffc0003de13c>] fuse_lookup_name+0x140/0x194 [<ffffffc0003de1c4>] fuse_lookup+0x34/0x110 [<ffffffc00030236c>] lookup_real+0x30/0x54 [<ffffffc000302efc>] __lookup_hash+0x30/0x48 [<ffffffc00030377c>] lookup_slow+0x44/0xbc [<ffffffc000304dd8>] path_lookupat+0x104/0x710 [<ffffffc00030540c>] filename_lookup.isra.32+0x28/0x74 [<ffffffc0003072cc>] user_path_at_empty+0x58/0x88 [<ffffffc000307308>] user_path_at+0xc/0x18 [<ffffffc0002f8764>] SyS_faccessat+0xc0/0x1bc Below is the root cause of this external fragmentation which could be observed in devices which have only one memory zone, such as some arm64 android devices: 1) In arm64, the first 4GB physical address space is of DMA zone. If the size of physical memory is less than 4GB and the whole memory is in the first 4GB address space, then the system will have only one DMA zone. 2) By default, all pageblocks are Movable. 3) Allocators such as slab, ion, graphics preferably allocate pages of Unmvoable migration type. It might fallback to allocate Movable pages and changes Movable pageblocks into Unmovable ones. 4) Movable pagesblocks will become less and less due to above reason. However, in android system, AnonPages request is always high. The Movable pages will be easily exhausted. 5) While Movable pages are exhausted, the Movable allocations will frequently fallback to allocate the largest feasiable pages of the other migration types. The order-2 and order-3 Unmovable pages will be split into smaller ones easily. This symptom doesn't appear in arm32 android which usually has two memory zones including Highmem and Normal. The slab, ion, and graphics allocators allocate pages with flag GFP_KERNEL. Only Movable pageblocks in Normal zone become less, and the Movable pages in Highmem zone are still a lot. Thus, the Movable pages will not be easily exhausted, and there will not be frequent fallbacks. Since the root cause is that fallbacks might frequently split order-2 and order-3 pages of the other migration types. This patch tweaks fallback mechanism to avoid splitting order-2 and order-3 pages. while fallbacks happen, if the largest feasible pages are less than or queal to COSTLY_ORDER, i.e. 3, then try to select the smallest feasible pages. The reason why fallbacks prefer the largest feasiable pages is to increase fallback efficiency since fallbacks are likely to happen again. By stealing the largest feasible pages, it could reduce the oppourtunities of antoher fallback. Besides, it could make consecutive allocations more approximate to each other and make system less fragment. However, if the largest feasible pages are less than or equal to order-3, fallbacks might split it and make the upcoming order-3 page allocations fail. My test is against arm64 android devices with kernel 3.10.49. I set the same account and install the same applications in both deivces and use them synchronously. Test result: 1) Test without this patch: Most free pages are order-0 Unmovable ones. allocstall and compact_stall in /proc/vmstat are relatively high. And most occurances of allocstall are due to order-2 and order-3 allocations. 2) Test with this patch: There are more order-2 and order-3 free pages. allocstall and compact_stall in /proc/vmstat are relatively low. And most occurances of allocstall are due to order-0 allocations. Log: 1) Test without this patch: ------ TIME (date) ------ Fri Jul 3 16:52:55 CST 2015 ------ UPTIME (uptime) ------ up time: 2 days, 12:06:52, idle time: 8 days, 14:48:55, sleep time: 16:43:56 ------ MEMORY INFO (/proc/meminfo) ------ MemTotal: 2792568 kB MemFree: 194524 kB Buffers: 3788 kB Cached: 380872 kB ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ Free pages count per migrate type at order 0 1 2 3 4 Node 0, zone DMA, type Unmovable 43852 701 0 0 0 Node 0, zone DMA, type Reclaimable 3357 0 0 0 0 Node 0, zone DMA, type Movable 0 5 0 0 0 Node 0, zone DMA, type Reserve 0 1 5 0 0 Node 0, zone DMA, type CMA 2 0 0 0 0 Node 0, zone DMA, type Isolate 0 0 0 0 0 Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate Node 0, zone DMA 362 80 170 2 113 0 ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ pgsteal_kswapd_dma 31755040 pgsteal_direct_dma 34597394 pgscan_kswapd_dma 36427664 pgscan_direct_dma 39490711 kswapd_low_wmark_hit_quickly 201929 kswapd_high_wmark_hit_quickly 4858 allocstall 664269 allocstall_order_0 9738 allocstall_order_1 1787 allocstall_order_2 637608 allocstall_order_3 15136 pgmigrate_success 2941956 pgmigrate_fail 1033 compact_migrate_scanned 142985157 compact_free_scanned 4734040109 compact_isolated 7720362 compact_stall 65978 compact_fail 46084 compact_success 11717 2) Test with this patch: ------ TIME (date) ------ Fri Jul 3 16:52:31 CST 2015 ------ UPTIME (uptime) ------ up time: 2 days, 12:06:30 ------ MEMORY INFO (/proc/meminfo) ------ MemTotal: 2792568 kB MemFree: 47612 kB Buffers: 3732 kB Cached: 387048 kB ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ Free pages count per migrate type at order 0 1 2 3 4 Node 0, zone DMA, type Unmovable 272 243 126 1 0 Node 0, zone DMA, type Reclaimable 0 361 168 46 0 Node 0, zone DMA, type Movable 4103 1782 130 3 0 Node 0, zone DMA, type Reserve 0 0 0 0 0 Node 0, zone DMA, type CMA 563 2 0 0 0 Node 0, zone DMA, type Isolate 0 0 0 0 0 Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate Node 0, zone DMA 183 12 417 2 113 0 ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ pgsteal_kswapd_dma 50710868 pgsteal_direct_dma 1756780 pgscan_kswapd_dma 58281837 pgscan_direct_dma 2022049 kswapd_low_wmark_hit_quickly 37599 kswapd_high_wmark_hit_quickly 13564 allocstall 27510 allocstall_order_0 26101 allocstall_order_1 23 allocstall_order_2 1224 allocstall_order_3 162 pgmigrate_success 63751 pgmigrate_fail 7 compact_migrate_scanned 278170 compact_free_scanned 6155410 compact_isolated 140762 compact_stall 749 compact_fail 54 compact_success 22 unevictable_pgs_culled 794 Below is the status of another device with this patch. /proc/pagetypeinfo shows that even if there are no Movable pages, there are lots of order-2 and order-3 Unmovable pages. For this case, if the patch is not applied, then order-2 and order-3 Unmovable pages will be split easily. It's likely that system perforamnce will become low due to severe external fragmentation. ------ UPTIME (uptime) ------ up time: 33 days, 08:10:58 ------ MEMORY INFO (/proc/meminfo) ------ MemTotal: 2792568 kB MemFree: 37340 kB Buffers: 13412 kB Cached: 655456 kB ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ Free pages count per migrate type at order 0 1 2 3 4 Node 0, zone DMA, type Unmovable 718 628 1116 301 0 Node 0, zone DMA, type Reclaimable 198 93 0 0 0 Node 0, zone DMA, type Movable 0 0 0 0 0 Node 0, zone DMA, type Reserve 0 0 0 0 0 Node 0, zone DMA, type CMA 89 11 3 0 0 Node 0, zone DMA, type Isolate 0 0 0 0 0 Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate Node 0, zone DMA 377 115 120 2 113 0 ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ pgsteal_direct_dma 28575192 pgsteal_kswapd_dma 378357910 pgscan_kswapd_dma 422765699 pgscan_direct_dma 31860747 kswapd_low_wmark_hit_quickly 947979 kswapd_high_wmark_hit_quickly 139901 allocstall 592989 compact_migrate_scanned 149884903 compact_free_scanned 6629299888 compact_isolated 7699012 compact_stall 52550 compact_fail 45155 compact_success 6057 ChengYi He (2): mm/page_alloc: let migration fallback support pages of requested order mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback mm/page_alloc.c | 92 ++++++++++++++++++++++++++++++++++++--------------------- 1 file changed, 59 insertions(+), 33 deletions(-) -- 1.9.1 ^ permalink raw reply [flat|nested] 8+ messages in thread
* [RFC PATCH 1/2] mm/page_alloc: let migration fallback support pages of requested order 2016-01-29 19:23 [RFC PATCH 0/2] avoid external fragmentation related to migration fallback ChengYi He @ 2016-01-29 19:24 ` ChengYi He 2016-01-29 19:25 ` [RFC PATCH 2/2] mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback ChengYi He 2016-01-29 21:03 ` [RFC PATCH 0/2] avoid external fragmentation related to " Vlastimil Babka 2 siblings, 0 replies; 8+ messages in thread From: ChengYi He @ 2016-01-29 19:24 UTC (permalink / raw) To: Andrew Morton Cc: Mel Gorman, Michal Hocko, Vlastimil Babka, David Rientjes, Joonsoo Kim, Yaowei Bai, Xishi Qiu, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel, chengyihetaipei This helper function only factors out the code flow within each order during fallback. There is no function change. Signed-off-by: ChengYi He <chengyihetaipei@gmail.com> --- mm/page_alloc.c | 79 +++++++++++++++++++++++++++++++++------------------------ 1 file changed, 46 insertions(+), 33 deletions(-) diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 63358d9..50c325a 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -1752,51 +1752,64 @@ static void unreserve_highatomic_pageblock(const struct alloc_context *ac) } } -/* Remove an element from the buddy allocator from the fallback list */ static inline struct page * -__rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) +__rmqueue_fallback_order(struct zone *zone, unsigned int order, + int start_migratetype, int current_order) { struct free_area *area; - unsigned int current_order; struct page *page; int fallback_mt; bool can_steal; - /* Find the largest possible block of pages in the other list */ - for (current_order = MAX_ORDER-1; - current_order >= order && current_order <= MAX_ORDER-1; - --current_order) { - area = &(zone->free_area[current_order]); - fallback_mt = find_suitable_fallback(area, current_order, - start_migratetype, false, &can_steal); - if (fallback_mt == -1) - continue; + area = &(zone->free_area[current_order]); + fallback_mt = find_suitable_fallback(area, current_order, + start_migratetype, false, &can_steal); + if (fallback_mt == -1) + return NULL; - page = list_first_entry(&area->free_list[fallback_mt], - struct page, lru); - if (can_steal) - steal_suitable_fallback(zone, page, start_migratetype); + page = list_first_entry(&area->free_list[fallback_mt], + struct page, lru); + if (can_steal) + steal_suitable_fallback(zone, page, start_migratetype); - /* Remove the page from the freelists */ - area->nr_free--; - list_del(&page->lru); - rmv_page_order(page); + /* Remove the page from the freelists */ + area->nr_free--; + list_del(&page->lru); + rmv_page_order(page); - expand(zone, page, order, current_order, area, - start_migratetype); - /* - * The pcppage_migratetype may differ from pageblock's - * migratetype depending on the decisions in - * find_suitable_fallback(). This is OK as long as it does not - * differ for MIGRATE_CMA pageblocks. Those can be used as - * fallback only via special __rmqueue_cma_fallback() function - */ - set_pcppage_migratetype(page, start_migratetype); + expand(zone, page, order, current_order, area, + start_migratetype); + /* + * The pcppage_migratetype may differ from pageblock's + * migratetype depending on the decisions in + * find_suitable_fallback(). This is OK as long as it does not + * differ for MIGRATE_CMA pageblocks. Those can be used as + * fallback only via special __rmqueue_cma_fallback() function + */ + set_pcppage_migratetype(page, start_migratetype); - trace_mm_page_alloc_extfrag(page, order, current_order, - start_migratetype, fallback_mt); + trace_mm_page_alloc_extfrag(page, order, current_order, + start_migratetype, fallback_mt); - return page; + return page; +} + +/* Remove an element from the buddy allocator from the fallback list */ +static inline struct page * +__rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) +{ + unsigned int current_order; + struct page *page; + + /* Find the largest possible block of pages in the other list */ + for (current_order = MAX_ORDER-1; + current_order >= order && current_order <= MAX_ORDER-1; + --current_order) { + page = __rmqueue_fallback_order(zone, order, start_migratetype, + current_order); + + if (page) + return page; } return NULL; -- 1.9.1 ^ permalink raw reply related [flat|nested] 8+ messages in thread
* [RFC PATCH 2/2] mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback 2016-01-29 19:23 [RFC PATCH 0/2] avoid external fragmentation related to migration fallback ChengYi He 2016-01-29 19:24 ` [RFC PATCH 1/2] mm/page_alloc: let migration fallback support pages of requested order ChengYi He @ 2016-01-29 19:25 ` ChengYi He 2016-01-30 10:41 ` Xishi Qiu 2016-01-29 21:03 ` [RFC PATCH 0/2] avoid external fragmentation related to " Vlastimil Babka 2 siblings, 1 reply; 8+ messages in thread From: ChengYi He @ 2016-01-29 19:25 UTC (permalink / raw) To: Andrew Morton Cc: Mel Gorman, Michal Hocko, Vlastimil Babka, David Rientjes, Joonsoo Kim, Yaowei Bai, Xishi Qiu, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel, chengyihetaipei While buddy system fallbacks to allocate different migration type pages, it prefers the largest feasible pages and might split the chosen page into smalller ones. If the largest feasible pages are less than or equal to orde-3 and migration fallback happens frequently, then order-2 and order-3 pages can be exhausted easily. This patch aims to allocate the smallest feasible pages for the fallback mechanism under this condition. Signed-off-by: ChengYi He <chengyihetaipei@gmail.com> --- mm/page_alloc.c | 19 ++++++++++++++++--- 1 file changed, 16 insertions(+), 3 deletions(-) diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 50c325a..3fcb653 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -1802,9 +1802,22 @@ __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) struct page *page; /* Find the largest possible block of pages in the other list */ - for (current_order = MAX_ORDER-1; - current_order >= order && current_order <= MAX_ORDER-1; - --current_order) { + for (current_order = MAX_ORDER - 1; + current_order >= max_t(unsigned int, PAGE_ALLOC_COSTLY_ORDER + 1, order); + --current_order) { + page = __rmqueue_fallback_order(zone, order, start_migratetype, + current_order); + + if (page) + return page; + } + + /* + * While current_order <= PAGE_ALLOC_COSTLY_ORDER, find the smallest + * feasible pages in the other list to avoid splitting high order pages + */ + for (current_order = order; current_order <= PAGE_ALLOC_COSTLY_ORDER; + ++current_order) { page = __rmqueue_fallback_order(zone, order, start_migratetype, current_order); -- 1.9.1 ^ permalink raw reply related [flat|nested] 8+ messages in thread
* Re: [RFC PATCH 2/2] mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback 2016-01-29 19:25 ` [RFC PATCH 2/2] mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback ChengYi He @ 2016-01-30 10:41 ` Xishi Qiu 0 siblings, 0 replies; 8+ messages in thread From: Xishi Qiu @ 2016-01-30 10:41 UTC (permalink / raw) To: ChengYi He Cc: Andrew Morton, Mel Gorman, Michal Hocko, Vlastimil Babka, David Rientjes, Joonsoo Kim, Yaowei Bai, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel On 2016/1/30 3:25, ChengYi He wrote: > While buddy system fallbacks to allocate different migration type pages, > it prefers the largest feasible pages and might split the chosen page > into smalller ones. If the largest feasible pages are less than or equal > to orde-3 and migration fallback happens frequently, then order-2 and > order-3 pages can be exhausted easily. This patch aims to allocate the > smallest feasible pages for the fallback mechanism under this condition. > > Signed-off-by: ChengYi He <chengyihetaipei@gmail.com> > --- > mm/page_alloc.c | 19 ++++++++++++++++--- > 1 file changed, 16 insertions(+), 3 deletions(-) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 50c325a..3fcb653 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -1802,9 +1802,22 @@ __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) > struct page *page; > > /* Find the largest possible block of pages in the other list */ > - for (current_order = MAX_ORDER-1; > - current_order >= order && current_order <= MAX_ORDER-1; > - --current_order) { > + for (current_order = MAX_ORDER - 1; > + current_order >= max_t(unsigned int, PAGE_ALLOC_COSTLY_ORDER + 1, order); > + --current_order) { > + page = __rmqueue_fallback_order(zone, order, start_migratetype, > + current_order); > + > + if (page) > + return page; > + } > + > + /* > + * While current_order <= PAGE_ALLOC_COSTLY_ORDER, find the smallest > + * feasible pages in the other list to avoid splitting high order pages > + */ > + for (current_order = order; current_order <= PAGE_ALLOC_COSTLY_ORDER; > + ++current_order) { > page = __rmqueue_fallback_order(zone, order, start_migratetype, > current_order); > Hi Chengyi, So you mean use the largest block first, if no large block left, the use the smallest block, right? I have an idea, how about set two migrate types(movable and unmovable) when doing init work? The function is memmap_init_zone(). I don't know how to set the ratio, maybe unmovable takes 1/10 memory, and left 9/10 memory to movable? I think this effect is a little like the two zones (normal and movable). Another two ideas https://lkml.org/lkml/2015/8/14/67 7d348b9ea64db0a315d777ce7d4b06697f946503, maybe this patch is not applied on your 3.10 Thanks, Xishi Qiu ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH 0/2] avoid external fragmentation related to migration fallback 2016-01-29 19:23 [RFC PATCH 0/2] avoid external fragmentation related to migration fallback ChengYi He 2016-01-29 19:24 ` [RFC PATCH 1/2] mm/page_alloc: let migration fallback support pages of requested order ChengYi He 2016-01-29 19:25 ` [RFC PATCH 2/2] mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback ChengYi He @ 2016-01-29 21:03 ` Vlastimil Babka 2016-01-30 10:21 ` Xishi Qiu 2016-02-01 13:53 ` Mel Gorman 2 siblings, 2 replies; 8+ messages in thread From: Vlastimil Babka @ 2016-01-29 21:03 UTC (permalink / raw) To: ChengYi He, Andrew Morton Cc: Mel Gorman, Michal Hocko, David Rientjes, Joonsoo Kim, Yaowei Bai, Xishi Qiu, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel On 01/29/2016 08:23 PM, ChengYi He wrote: [...] > Below is the root cause of this external fragmentation which could be > observed in devices which have only one memory zone, such as some arm64 > android devices: > > 1) In arm64, the first 4GB physical address space is of DMA zone. If the > size of physical memory is less than 4GB and the whole memory is in the > first 4GB address space, then the system will have only one DMA zone. > 2) By default, all pageblocks are Movable. > 3) Allocators such as slab, ion, graphics preferably allocate pages of > Unmvoable migration type. It might fallback to allocate Movable pages > and changes Movable pageblocks into Unmovable ones. > 4) Movable pagesblocks will become less and less due to above reason. > However, in android system, AnonPages request is always high. The > Movable pages will be easily exhausted. > 5) While Movable pages are exhausted, the Movable allocations will > frequently fallback to allocate the largest feasiable pages of the other > migration types. The order-2 and order-3 Unmovable pages will be split > into smaller ones easily. > > This symptom doesn't appear in arm32 android which usually has two > memory zones including Highmem and Normal. The slab, ion, and graphics > allocators allocate pages with flag GFP_KERNEL. Only Movable pageblocks > in Normal zone become less, and the Movable pages in Highmem zone are > still a lot. Thus, the Movable pages will not be easily exhausted, and > there will not be frequent fallbacks. Hm, this 1 zone vs 2 zones shouldn't make that much difference, unless a) you use zone reclaim mode, or b) you have an old kernel without fair zone allocation policy? > Since the root cause is that fallbacks might frequently split order-2 > and order-3 pages of the other migration types. This patch tweaks > fallback mechanism to avoid splitting order-2 and order-3 pages. while > fallbacks happen, if the largest feasible pages are less than or queal to > COSTLY_ORDER, i.e. 3, then try to select the smallest feasible pages. The > reason why fallbacks prefer the largest feasiable pages is to increase > fallback efficiency since fallbacks are likely to happen again. By > stealing the largest feasible pages, it could reduce the oppourtunities > of antoher fallback. Besides, it could make consecutive allocations more > approximate to each other and make system less fragment. However, if the > largest feasible pages are less than or equal to order-3, fallbacks might > split it and make the upcoming order-3 page allocations fail. In theory I don't see immediately why preferring smaller pages for fallback should be a clear win. If it's Unmovable allocations stealing from Movable pageblocks, the allocations will spread over larger areas instead of being grouped together. Maybe, for Movable allocations stealing from Unmovable allocations, preferring smallest might make sense and be safe, as any extra fragmentation is fixable bycompaction. Maybe it was already tried (by Joonsoo?) at some point, I'm not sure right now. > My test is against arm64 android devices with kernel 3.10.49. I set the > same account and install the same applications in both deivces and use > them synchronously. 3.10 is wayyyyyy old. There were numerous patches to compaction and anti-fragmentation since then. IIRC the fallback decisions were quite suboptimal at that point. I'm not even sure how you could apply your patches to both recent kernel for posting them, and 3.10 for testing? Is it possible to test on 4.4? > > Test result: > 1) Test without this patch: > Most free pages are order-0 Unmovable ones. allocstall and compact_stall > in /proc/vmstat are relatively high. And most occurances of allocstall > are due to order-2 and order-3 allocations. > 2) Test with this patch: > There are more order-2 and order-3 free pages. allocstall and > compact_stall in /proc/vmstat are relatively low. And most occurances of > allocstall are due to order-0 allocations. > > Log: > 1) Test without this patch: > ------ TIME (date) ------ > Fri Jul 3 16:52:55 CST 2015 > ------ UPTIME (uptime) ------ > up time: 2 days, 12:06:52, idle time: 8 days, 14:48:55, sleep time: 16:43:56 > ------ MEMORY INFO (/proc/meminfo) ------ > MemTotal: 2792568 kB > MemFree: 194524 kB > Buffers: 3788 kB > Cached: 380872 kB > ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ > Free pages count per migrate type at order 0 1 2 3 4 > Node 0, zone DMA, type Unmovable 43852 701 0 0 0 > Node 0, zone DMA, type Reclaimable 3357 0 0 0 0 > Node 0, zone DMA, type Movable 0 5 0 0 0 > Node 0, zone DMA, type Reserve 0 1 5 0 0 > Node 0, zone DMA, type CMA 2 0 0 0 0 > Node 0, zone DMA, type Isolate 0 0 0 0 0 > Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate > Node 0, zone DMA 362 80 170 2 113 0 > ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ > pgsteal_kswapd_dma 31755040 > pgsteal_direct_dma 34597394 > pgscan_kswapd_dma 36427664 > pgscan_direct_dma 39490711 > kswapd_low_wmark_hit_quickly 201929 > kswapd_high_wmark_hit_quickly 4858 > allocstall 664269 > allocstall_order_0 9738 > allocstall_order_1 1787 > allocstall_order_2 637608 > allocstall_order_3 15136 > pgmigrate_success 2941956 > pgmigrate_fail 1033 > compact_migrate_scanned 142985157 > compact_free_scanned 4734040109 > compact_isolated 7720362 > compact_stall 65978 > compact_fail 46084 > compact_success 11717 > > 2) Test with this patch: > ------ TIME (date) ------ > Fri Jul 3 16:52:31 CST 2015 > ------ UPTIME (uptime) ------ > up time: 2 days, 12:06:30 > ------ MEMORY INFO (/proc/meminfo) ------ > MemTotal: 2792568 kB > MemFree: 47612 kB > Buffers: 3732 kB > Cached: 387048 kB > ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ > Free pages count per migrate type at order 0 1 2 3 4 > Node 0, zone DMA, type Unmovable 272 243 126 1 0 > Node 0, zone DMA, type Reclaimable 0 361 168 46 0 > Node 0, zone DMA, type Movable 4103 1782 130 3 0 > Node 0, zone DMA, type Reserve 0 0 0 0 0 > Node 0, zone DMA, type CMA 563 2 0 0 0 > Node 0, zone DMA, type Isolate 0 0 0 0 0 > Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate > Node 0, zone DMA 183 12 417 2 113 0 > ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ > pgsteal_kswapd_dma 50710868 > pgsteal_direct_dma 1756780 > pgscan_kswapd_dma 58281837 > pgscan_direct_dma 2022049 > kswapd_low_wmark_hit_quickly 37599 > kswapd_high_wmark_hit_quickly 13564 > allocstall 27510 > allocstall_order_0 26101 > allocstall_order_1 23 > allocstall_order_2 1224 > allocstall_order_3 162 > pgmigrate_success 63751 > pgmigrate_fail 7 > compact_migrate_scanned 278170 > compact_free_scanned 6155410 > compact_isolated 140762 > compact_stall 749 > compact_fail 54 > compact_success 22 > unevictable_pgs_culled 794 > > Below is the status of another device with this patch. > /proc/pagetypeinfo shows that even if there are no Movable pages, there > are lots of order-2 and order-3 Unmovable pages. For this case, if the > patch is not applied, then order-2 and order-3 Unmovable pages will be > split easily. It's likely that system perforamnce will become low due to > severe external fragmentation. > > ------ UPTIME (uptime) ------ > up time: 33 days, 08:10:58 > ------ MEMORY INFO (/proc/meminfo) ------ > MemTotal: 2792568 kB > MemFree: 37340 kB > Buffers: 13412 kB > Cached: 655456 kB > ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ > Free pages count per migrate type at order 0 1 2 3 4 > Node 0, zone DMA, type Unmovable 718 628 1116 301 0 > Node 0, zone DMA, type Reclaimable 198 93 0 0 0 > Node 0, zone DMA, type Movable 0 0 0 0 0 > Node 0, zone DMA, type Reserve 0 0 0 0 0 > Node 0, zone DMA, type CMA 89 11 3 0 0 > Node 0, zone DMA, type Isolate 0 0 0 0 0 > Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate > Node 0, zone DMA 377 115 120 2 113 0 > ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ > pgsteal_direct_dma 28575192 > pgsteal_kswapd_dma 378357910 > pgscan_kswapd_dma 422765699 > pgscan_direct_dma 31860747 > kswapd_low_wmark_hit_quickly 947979 > kswapd_high_wmark_hit_quickly 139901 > allocstall 592989 > compact_migrate_scanned 149884903 > compact_free_scanned 6629299888 > compact_isolated 7699012 > compact_stall 52550 > compact_fail 45155 > compact_success 6057 > > ChengYi He (2): > mm/page_alloc: let migration fallback support pages of requested order > mm/page_alloc: avoid splitting pages of order 2 and 3 in migration > fallback > > mm/page_alloc.c | 92 ++++++++++++++++++++++++++++++++++++--------------------- > 1 file changed, 59 insertions(+), 33 deletions(-) > ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH 0/2] avoid external fragmentation related to migration fallback 2016-01-29 21:03 ` [RFC PATCH 0/2] avoid external fragmentation related to " Vlastimil Babka @ 2016-01-30 10:21 ` Xishi Qiu 2016-02-01 13:53 ` Mel Gorman 1 sibling, 0 replies; 8+ messages in thread From: Xishi Qiu @ 2016-01-30 10:21 UTC (permalink / raw) To: Vlastimil Babka Cc: ChengYi He, Andrew Morton, Mel Gorman, Michal Hocko, David Rientjes, Joonsoo Kim, Yaowei Bai, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel On 2016/1/30 5:03, Vlastimil Babka wrote: > On 01/29/2016 08:23 PM, ChengYi He wrote: > > [...] > >> Below is the root cause of this external fragmentation which could be >> observed in devices which have only one memory zone, such as some arm64 >> android devices: >> >> 1) In arm64, the first 4GB physical address space is of DMA zone. If the >> size of physical memory is less than 4GB and the whole memory is in the >> first 4GB address space, then the system will have only one DMA zone. >> 2) By default, all pageblocks are Movable. >> 3) Allocators such as slab, ion, graphics preferably allocate pages of >> Unmvoable migration type. It might fallback to allocate Movable pages >> and changes Movable pageblocks into Unmovable ones. >> 4) Movable pagesblocks will become less and less due to above reason. >> However, in android system, AnonPages request is always high. The >> Movable pages will be easily exhausted. >> 5) While Movable pages are exhausted, the Movable allocations will >> frequently fallback to allocate the largest feasiable pages of the other >> migration types. The order-2 and order-3 Unmovable pages will be split >> into smaller ones easily. >> >> This symptom doesn't appear in arm32 android which usually has two >> memory zones including Highmem and Normal. The slab, ion, and graphics >> allocators allocate pages with flag GFP_KERNEL. Only Movable pageblocks >> in Normal zone become less, and the Movable pages in Highmem zone are >> still a lot. Thus, the Movable pages will not be easily exhausted, and >> there will not be frequent fallbacks. > > Hm, this 1 zone vs 2 zones shouldn't make that much difference, unless > a) you use zone reclaim mode, or b) you have an old kernel without fair > zone allocation policy? > Hi Vlastimil, I agree with you. I think if we have a normal zone and a movable zone, then the effect of compaction will be better. e.g. U: unmovable page M: movable page F: free page one zone(DMA): paddr: 0 max ZONE_DMA: U M F U M F ... U M F U M F after compact: U F F U F F ... U M M U M M two zone(DMA and MOVABLE) paddr: 0 max ZONE_DMA: the same as above ZONE_MOVABLE: M F M F M F ... M F M F M F after compact: F F F F F F ... M M M M M M // we get large block than above >> Since the root cause is that fallbacks might frequently split order-2 >> and order-3 pages of the other migration types. This patch tweaks >> fallback mechanism to avoid splitting order-2 and order-3 pages. while >> fallbacks happen, if the largest feasible pages are less than or queal to >> COSTLY_ORDER, i.e. 3, then try to select the smallest feasible pages. The >> reason why fallbacks prefer the largest feasiable pages is to increase >> fallback efficiency since fallbacks are likely to happen again. By >> stealing the largest feasible pages, it could reduce the oppourtunities >> of antoher fallback. Besides, it could make consecutive allocations more >> approximate to each other and make system less fragment. However, if the >> largest feasible pages are less than or equal to order-3, fallbacks might >> split it and make the upcoming order-3 page allocations fail. > > In theory I don't see immediately why preferring smaller pages for > fallback should be a clear win. If it's Unmovable allocations stealing > from Movable pageblocks, the allocations will spread over larger areas > instead of being grouped together. Maybe, for Movable allocations > stealing from Unmovable allocations, preferring smallest might make > sense and be safe, as any extra fragmentation is fixable bycompaction. > Maybe it was already tried (by Joonsoo?) at some point, I'm not sure > right now. > >> My test is against arm64 android devices with kernel 3.10.49. I set the >> same account and install the same applications in both deivces and use >> them synchronously. > > 3.10 is wayyyyyy old. There were numerous patches to compaction and > anti-fragmentation since then. IIRC the fallback decisions were quite > suboptimal at that point. I'm not even sure how you could apply your > patches to both recent kernel for posting them, and 3.10 for testing? > Is it possible to test on 4.4? > I think it's hard to update the drivers on android smart phone. Thanks, Xishi Qiu >> >> Test result: >> 1) Test without this patch: >> Most free pages are order-0 Unmovable ones. allocstall and compact_stall >> in /proc/vmstat are relatively high. And most occurances of allocstall >> are due to order-2 and order-3 allocations. >> 2) Test with this patch: >> There are more order-2 and order-3 free pages. allocstall and >> compact_stall in /proc/vmstat are relatively low. And most occurances of >> allocstall are due to order-0 allocations. >> >> Log: >> 1) Test without this patch: >> ------ TIME (date) ------ >> Fri Jul 3 16:52:55 CST 2015 >> ------ UPTIME (uptime) ------ >> up time: 2 days, 12:06:52, idle time: 8 days, 14:48:55, sleep time: 16:43:56 >> ------ MEMORY INFO (/proc/meminfo) ------ >> MemTotal: 2792568 kB >> MemFree: 194524 kB >> Buffers: 3788 kB >> Cached: 380872 kB >> ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ >> Free pages count per migrate type at order 0 1 2 3 4 >> Node 0, zone DMA, type Unmovable 43852 701 0 0 0 >> Node 0, zone DMA, type Reclaimable 3357 0 0 0 0 >> Node 0, zone DMA, type Movable 0 5 0 0 0 >> Node 0, zone DMA, type Reserve 0 1 5 0 0 >> Node 0, zone DMA, type CMA 2 0 0 0 0 >> Node 0, zone DMA, type Isolate 0 0 0 0 0 >> Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate >> Node 0, zone DMA 362 80 170 2 113 0 >> ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ >> pgsteal_kswapd_dma 31755040 >> pgsteal_direct_dma 34597394 >> pgscan_kswapd_dma 36427664 >> pgscan_direct_dma 39490711 >> kswapd_low_wmark_hit_quickly 201929 >> kswapd_high_wmark_hit_quickly 4858 >> allocstall 664269 >> allocstall_order_0 9738 >> allocstall_order_1 1787 >> allocstall_order_2 637608 >> allocstall_order_3 15136 >> pgmigrate_success 2941956 >> pgmigrate_fail 1033 >> compact_migrate_scanned 142985157 >> compact_free_scanned 4734040109 >> compact_isolated 7720362 >> compact_stall 65978 >> compact_fail 46084 >> compact_success 11717 >> >> 2) Test with this patch: >> ------ TIME (date) ------ >> Fri Jul 3 16:52:31 CST 2015 >> ------ UPTIME (uptime) ------ >> up time: 2 days, 12:06:30 >> ------ MEMORY INFO (/proc/meminfo) ------ >> MemTotal: 2792568 kB >> MemFree: 47612 kB >> Buffers: 3732 kB >> Cached: 387048 kB >> ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ >> Free pages count per migrate type at order 0 1 2 3 4 >> Node 0, zone DMA, type Unmovable 272 243 126 1 0 >> Node 0, zone DMA, type Reclaimable 0 361 168 46 0 >> Node 0, zone DMA, type Movable 4103 1782 130 3 0 >> Node 0, zone DMA, type Reserve 0 0 0 0 0 >> Node 0, zone DMA, type CMA 563 2 0 0 0 >> Node 0, zone DMA, type Isolate 0 0 0 0 0 >> Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate >> Node 0, zone DMA 183 12 417 2 113 0 >> ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ >> pgsteal_kswapd_dma 50710868 >> pgsteal_direct_dma 1756780 >> pgscan_kswapd_dma 58281837 >> pgscan_direct_dma 2022049 >> kswapd_low_wmark_hit_quickly 37599 >> kswapd_high_wmark_hit_quickly 13564 >> allocstall 27510 >> allocstall_order_0 26101 >> allocstall_order_1 23 >> allocstall_order_2 1224 >> allocstall_order_3 162 >> pgmigrate_success 63751 >> pgmigrate_fail 7 >> compact_migrate_scanned 278170 >> compact_free_scanned 6155410 >> compact_isolated 140762 >> compact_stall 749 >> compact_fail 54 >> compact_success 22 >> unevictable_pgs_culled 794 >> >> Below is the status of another device with this patch. >> /proc/pagetypeinfo shows that even if there are no Movable pages, there >> are lots of order-2 and order-3 Unmovable pages. For this case, if the >> patch is not applied, then order-2 and order-3 Unmovable pages will be >> split easily. It's likely that system perforamnce will become low due to >> severe external fragmentation. >> >> ------ UPTIME (uptime) ------ >> up time: 33 days, 08:10:58 >> ------ MEMORY INFO (/proc/meminfo) ------ >> MemTotal: 2792568 kB >> MemFree: 37340 kB >> Buffers: 13412 kB >> Cached: 655456 kB >> ------ PAGETYPEINFO (/proc/pagetypeinfo) ------ >> Free pages count per migrate type at order 0 1 2 3 4 >> Node 0, zone DMA, type Unmovable 718 628 1116 301 0 >> Node 0, zone DMA, type Reclaimable 198 93 0 0 0 >> Node 0, zone DMA, type Movable 0 0 0 0 0 >> Node 0, zone DMA, type Reserve 0 0 0 0 0 >> Node 0, zone DMA, type CMA 89 11 3 0 0 >> Node 0, zone DMA, type Isolate 0 0 0 0 0 >> Number of blocks type Unmovable Reclaimable Movable Reserve CMA Isolate >> Node 0, zone DMA 377 115 120 2 113 0 >> ------ VIRTUAL MEMORY STATS (/proc/vmstat) ------ >> pgsteal_direct_dma 28575192 >> pgsteal_kswapd_dma 378357910 >> pgscan_kswapd_dma 422765699 >> pgscan_direct_dma 31860747 >> kswapd_low_wmark_hit_quickly 947979 >> kswapd_high_wmark_hit_quickly 139901 >> allocstall 592989 >> compact_migrate_scanned 149884903 >> compact_free_scanned 6629299888 >> compact_isolated 7699012 >> compact_stall 52550 >> compact_fail 45155 >> compact_success 6057 >> >> ChengYi He (2): >> mm/page_alloc: let migration fallback support pages of requested order >> mm/page_alloc: avoid splitting pages of order 2 and 3 in migration >> fallback >> >> mm/page_alloc.c | 92 ++++++++++++++++++++++++++++++++++++--------------------- >> 1 file changed, 59 insertions(+), 33 deletions(-) >> > > > . > ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH 0/2] avoid external fragmentation related to migration fallback 2016-01-29 21:03 ` [RFC PATCH 0/2] avoid external fragmentation related to " Vlastimil Babka 2016-01-30 10:21 ` Xishi Qiu @ 2016-02-01 13:53 ` Mel Gorman 2016-02-02 1:01 ` Xishi Qiu 1 sibling, 1 reply; 8+ messages in thread From: Mel Gorman @ 2016-02-01 13:53 UTC (permalink / raw) To: Vlastimil Babka Cc: ChengYi He, Andrew Morton, Michal Hocko, David Rientjes, Joonsoo Kim, Yaowei Bai, Xishi Qiu, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel On Fri, Jan 29, 2016 at 10:03:52PM +0100, Vlastimil Babka wrote: > > Since the root cause is that fallbacks might frequently split order-2 > > and order-3 pages of the other migration types. This patch tweaks > > fallback mechanism to avoid splitting order-2 and order-3 pages. while > > fallbacks happen, if the largest feasible pages are less than or queal to > > COSTLY_ORDER, i.e. 3, then try to select the smallest feasible pages. The > > reason why fallbacks prefer the largest feasiable pages is to increase > > fallback efficiency since fallbacks are likely to happen again. By > > stealing the largest feasible pages, it could reduce the oppourtunities > > of antoher fallback. Besides, it could make consecutive allocations more > > approximate to each other and make system less fragment. However, if the > > largest feasible pages are less than or equal to order-3, fallbacks might > > split it and make the upcoming order-3 page allocations fail. > > In theory I don't see immediately why preferring smaller pages for > fallback should be a clear win. If it's Unmovable allocations stealing > from Movable pageblocks, the allocations will spread over larger areas > instead of being grouped together. Maybe, for Movable allocations > stealing from Unmovable allocations, preferring smallest might make > sense and be safe, as any extra fragmentation is fixable bycompaction. I strongly agree that spreading the fallback allocations over a larger area is likely to have a negative impact. Given the age of the kernel being tested, it would make sense to either rebase or at the very last backport the patches that affect watermark calculations and the treatment of high-order pages. -- Mel Gorman SUSE Labs ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH 0/2] avoid external fragmentation related to migration fallback 2016-02-01 13:53 ` Mel Gorman @ 2016-02-02 1:01 ` Xishi Qiu 0 siblings, 0 replies; 8+ messages in thread From: Xishi Qiu @ 2016-02-02 1:01 UTC (permalink / raw) To: Mel Gorman Cc: Vlastimil Babka, ChengYi He, Andrew Morton, Michal Hocko, David Rientjes, Joonsoo Kim, Yaowei Bai, Alexander Duyck, 'Kirill A . Shutemov', Johannes Weiner, linux-mm, linux-kernel On 2016/2/1 21:53, Mel Gorman wrote: > On Fri, Jan 29, 2016 at 10:03:52PM +0100, Vlastimil Babka wrote: >>> Since the root cause is that fallbacks might frequently split order-2 >>> and order-3 pages of the other migration types. This patch tweaks >>> fallback mechanism to avoid splitting order-2 and order-3 pages. while >>> fallbacks happen, if the largest feasible pages are less than or queal to >>> COSTLY_ORDER, i.e. 3, then try to select the smallest feasible pages. The >>> reason why fallbacks prefer the largest feasiable pages is to increase >>> fallback efficiency since fallbacks are likely to happen again. By >>> stealing the largest feasible pages, it could reduce the oppourtunities >>> of antoher fallback. Besides, it could make consecutive allocations more >>> approximate to each other and make system less fragment. However, if the >>> largest feasible pages are less than or equal to order-3, fallbacks might >>> split it and make the upcoming order-3 page allocations fail. >> >> In theory I don't see immediately why preferring smaller pages for >> fallback should be a clear win. If it's Unmovable allocations stealing >> from Movable pageblocks, the allocations will spread over larger areas >> instead of being grouped together. Maybe, for Movable allocations >> stealing from Unmovable allocations, preferring smallest might make >> sense and be safe, as any extra fragmentation is fixable bycompaction. > > I strongly agree that spreading the fallback allocations over a larger > area is likely to have a negative impact. Given the age of the kernel > being tested, it would make sense to either rebase or at the very last > backport the patches that affect watermark calculations and the > treatment of high-order pages. > Is it the feature of MIGRATE_HIGHATOMIC? Thanks, Xishi Qiu ^ permalink raw reply [flat|nested] 8+ messages in thread
end of thread, other threads:[~2016-02-02 1:04 UTC | newest] Thread overview: 8+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2016-01-29 19:23 [RFC PATCH 0/2] avoid external fragmentation related to migration fallback ChengYi He 2016-01-29 19:24 ` [RFC PATCH 1/2] mm/page_alloc: let migration fallback support pages of requested order ChengYi He 2016-01-29 19:25 ` [RFC PATCH 2/2] mm/page_alloc: avoid splitting pages of order 2 and 3 in migration fallback ChengYi He 2016-01-30 10:41 ` Xishi Qiu 2016-01-29 21:03 ` [RFC PATCH 0/2] avoid external fragmentation related to " Vlastimil Babka 2016-01-30 10:21 ` Xishi Qiu 2016-02-01 13:53 ` Mel Gorman 2016-02-02 1:01 ` Xishi Qiu
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).