- 26 Sep, 2014 40 commits
-
-
Mel Gorman authored
commit f7b5d647 upstream. The purpose of numa_zonelist_order=zone is to preserve lower zones for use with 32-bit devices. If locality is preferred then the numa_zonelist_order=node policy should be used. Unfortunately, the fair zone allocation policy overrides this by skipping zones on remote nodes until the lower one is found. While this makes sense from a page aging and performance perspective, it breaks the expected zonelist policy. This patch restores the expected behaviour for zone-list ordering. Signed-off-by:
Mel Gorman <mgorman@suse.de> Acked-by:
Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by:
Jiri Slaby <jslaby@suse.cz>
-
Mel Gorman authored
commit bb0b6dff upstream. When kswapd is awake reclaiming, the per-cpu stat thresholds are lowered to get more accurate counts to avoid breaching watermarks. This threshold update iterates over all possible CPUs which is unnecessary. Only online CPUs need to be updated. If a new CPU is onlined, refresh_zone_stat_thresholds() will set the thresholds correctly. Signed-off-by:
-
Mel Gorman authored
commit 0d5d823a upstream. zone->pages_scanned is a write-intensive cache line during page reclaim and it's also updated during page free. Move the counter into vmstat to take advantage of the per-cpu updates and do not update it in the free paths unless necessary. On a small UMA machine running tiobench the difference is marginal. On a 4-node machine the overhead is more noticable. Note that automatic NUMA balancing was disabled for this test as otherwise the system CPU overhead is unpredictable. 3.16.0-rc3 3.16.0-rc3 3.16.0-rc3 vanillarearrange-v5 vmstat-v5 User 746.94 759.78 774.56 System 65336.22 58350.98 32847.27 Elapsed 27553.52 27282.02 27415.04 Note that the overhead reduction will vary depending on where exactly pages are allocated and freed. Signed-off-by:
-
Mel Gorman authored
commit 3484b2de upstream. The arrangement of struct zone has changed over time and now it has reached the point where there is some inappropriate sharing going on. On x86-64 for example o The zone->node field is shared with the zone lock and zone->node is accessed frequently from the page allocator due to the fair zone allocation policy. o span_seqlock is almost never used by shares a line with free_area o Some zone statistics share a cache line with the LRU lock so reclaim-intensive and allocator-intensive workloads can bounce the cache line on a stat update This patch rearranges struct zone to put read-only and read-mostly fields together and then splits the page allocator intensive fields, the zone statistics and the page reclaim intensive fields into their own cache lines. Note that the type of lowmem_reserve changes due to the watermark calculations being signed and avoiding a signed/unsigned conversion there. On the test configuration I used the overall size of struct zone shrunk by one cache line. On smaller machines, this is not likely to be noticable. However, on a 4-node NUMA machine running tiobench the system CPU overhead is reduced by this patch. 3.16.0-rc3 3.16.0-rc3 vanillarearrange-v5r9 User 746.94 759.78 System 65336.22 58350.98 Elapsed 27553.52 27282.02 Signed-off-by:
-
Mel Gorman authored
commit 24b7e581 upstream. This was formerly the series "Improve sequential read throughput" which noted some major differences in performance of tiobench since 3.0. While there are a number of factors, two that dominated were the introduction of the fair zone allocation policy and changes to CFQ. The behaviour of fair zone allocation policy makes more sense than tiobench as a benchmark and CFQ defaults were not changed due to insufficient benchmarking. This series is what's left. It's one functional fix to the fair zone allocation policy when used on NUMA machines and a reduction of overhead in general. tiobench was used for the comparison despite its flaws as an IO benchmark as in this case we are primarily interested in the overhead of page allocator and page reclaim activity. On UMA, it makes little difference to overhead 3.16.0-rc3 3.16.0-rc3 vanilla lowercost-v5 User 383.61 386.77 System 403.83 401.74 Elapsed 5411.50 5413.11 On a 4-socket NUMA machine it's a bit more noticable 3.16.0-rc3 3.16.0-rc3 vanilla lowercost-v5 User 746.94 802.00 System 65336.22 40852.33 Elapsed 27553.52 27368.46 This patch (of 6): The LRU insertion and activate tracepoints take PFN as a parameter forcing the overhead to the caller. Move the overhead to the tracepoint fast-assign method to ensure the cost is only incurred when the tracepoint is active. Signed-off-by:
-
Jerome Marchand authored
commit 2ab051e1 upstream. When memory cgoups are enabled, the code that decides to force to scan anonymous pages in get_scan_count() compares global values (free, high_watermark) to a value that is restricted to a memory cgroup (file). It make the code over-eager to force anon scan. For instance, it will force anon scan when scanning a memcg that is mainly populated by anonymous page, even when there is plenty of file pages to get rid of in others memcgs, even when swappiness == 0. It breaks user's expectation about swappiness and hurts performance. This patch makes sure that forced anon scan only happens when there not enough file pages for the all zone, not just in one random memcg. [hannes@cmpxchg.org: cleanups] Signed-off-by:
Jerome Marchand <jmarchan@redhat.com> Acked-by:
Michal Hocko <mhocko@suse.cz> Acked-by:
Rik van Riel <riel@redhat.com> Cc: Mel Gorman <mgorman@suse.de> Signed-off-by:
-
Joonsoo Kim authored
commit 474750ab upstream. Richard Yao reported a month ago that his system have a trouble with vmap_area_lock contention during performance analysis by /proc/meminfo. Andrew asked why his analysis checks /proc/meminfo stressfully, but he didn't answer it. https://lkml.org/lkml/2014/4/10/416 Although I'm not sure that this is right usage or not, there is a solution reducing vmap_area_lock contention with no side-effect. That is just to use rcu list iterator in get_vmalloc_info(). rcu can be used in this function because all RCU protocol is already respected by writers, since Nick Piggin commit db64fe02 ("mm: rewrite vmap layer") back in linux-2.6.28 Specifically : insertions use list_add_rcu(), deletions use list_del_rcu() and kfree_rcu(). Note the rb tree is not used from rcu reader (it would not be safe), only the vmap_area_list has full RCU protection. Note that __purge_vmap_area_lazy() already uses this rcu protection. rcu_read_lock(); list_for_each_entry_rcu(va, &vmap_area_list, list) { if (va->flags & VM_LAZY_FREE) { if (va->va_start < *start) *start = va->va_start; if (va->va_end > *end) *end = va->va_end; nr += (va->va_end - va->va_start) >> PAGE_SHIFT; list_add_tail(&va->purge_list, &valist); va->flags |= VM_LAZY_FREEING; va->flags &= ~VM_LAZY_FREE; } } rcu_read_unlock(); Peter: : While rcu list traversal over the vmap_area_list is safe, this may : arrive at different results than the spinlocked version. The rcu list : traversal version will not be a 'snapshot' of a single, valid instant : of the entire vmap_area_list, but rather a potential amalgam of : different list states. Joonsoo: : Yes, you are right, but I don't think that we should be strict here. : Meminfo is already not a 'snapshot' at specific time. While we try to get : certain stats, the other stats can change. And, although we may arrive at : different results than the spinlocked version, the difference would not be : large and would not make serious side-effect. [edumazet@google.com: add more commit description] Signed-off-by:
Joonsoo Kim <iamjoonsoo.kim@lge.com> Reported-by:
Richard Yao <ryao@gentoo.org> Acked-by:
Eric Dumazet <edumazet@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Zhang Yanfei <zhangyanfei.yes@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Andi Kleen <andi@firstfloor.org> Signed-off-by:
-
Jerome Marchand authored
commit 21bda264 upstream. Commit 71e3aac0 ("thp: transparent hugepage core") adds copy_pte_range prototype to huge_mm.h. I'm not sure why (or if) this function have been used outside of memory.c, but it currently isn't. This patch makes copy_pte_range() static again. Signed-off-by:
David Rientjes <rientjes@google.com> Signed-off-by:
-
David Rientjes authored
commit 14a4e214 upstream. Commit 9f1b868a ("mm: thp: khugepaged: add policy for finding target node") improved the previous khugepaged logic which allocated a transparent hugepages from the node of the first page being collapsed. However, it is still possible to collapse pages to remote memory which may suffer from additional access latency. With the current policy, it is possible that 255 pages (with PAGE_SHIFT == 12) will be collapsed remotely if the majority are allocated from that node. When zone_reclaim_mode is enabled, it means the VM should make every attempt to allocate locally to prevent NUMA performance degradation. In this case, we do not want to collapse hugepages to remote nodes that would suffer from increased access latency. Thus, when zone_reclaim_mode is enabled, only allow collapsing to nodes with RECLAIM_DISTANCE or less. There is no functional change for systems that disable zone_reclaim_mode. Signed-off-by:
Vlastimil Babka <vbabka@suse.cz> Acked-by:
-
Hugh Dickins authored
commit c0d73261 upstream. Use ACCESS_ONCE() in handle_pte_fault() when getting the entry or orig_pte upon which all subsequent decisions and pte_same() tests will be made. I have no evidence that its lack is responsible for the mm/filemap.c:202 BUG_ON(page_mapped(page)) in __delete_from_page_cache() found by trinity, and I am not optimistic that it will fix it. But I have found no other explanation, and ACCESS_ONCE() here will surely not hurt. If gcc does re-access the pte before passing it down, then that would be disastrous for correct page fault handling, and certainly could explain the page_mapped() BUGs seen (concurrent fault causing page to be mapped in a second time on top of itself: mapcount 2 for a single pte). Signed-off-by:
Hugh Dickins <hughd@google.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Signed-off-by:
-
Hugh Dickins authored
commit 66d2f4d2 upstream. Under shmem swapping load, I sometimes hit the VM_BUG_ON_PAGE(!PageLRU) in isolate_lru_pages() at mm/vmscan.c:1281! Commit 2457aec6 ("mm: non-atomically mark page accessed during page cache allocation where possible") looks like interrupted work-in-progress. mm/filemap.c's call to init_page_accessed() is fine, but not mm/shmem.c's - shmem_write_begin() is clearly wrong to use it after shmem_getpage(), when the page is always visible in radix_tree, and often already on LRU. Revert change to shmem_write_begin(), and use init_page_accessed() or mark_page_accessed() appropriately for SGP_WRITE in shmem_getpage_gfp(). SGP_WRITE also covers shmem_symlink(), which did not mark_page_accessed() before; but since many other filesystems use [__]page_symlink(), which did and does mark the page accessed, consider this as rectifying an oversight. Signed-off-by:
-
Mel Gorman authored
commit 888cf2db upstream. If a page is marked for immediate reclaim then it is moved to the tail of the LRU list. This occurs when the system is under enough memory pressure for pages under writeback to reach the end of the LRU but we test for this using atomic operations on every writeback. This patch uses an optimistic non-atomic test first. It'll miss some pages in rare cases but the consequences are not severe enough to warrant such a penalty. While the function does not dominate profiles during a simple dd test the cost of it is reduced. 73048 0.7428 vmlinux-3.15.0-rc5-mmotm-20140513 end_page_writeback 23740 0.2409 vmlinux-3.15.0-rc5-lessatomic end_page_writeback Signed-off-by:
-
Mel Gorman authored
commit 2457aec6 upstream. aops->write_begin may allocate a new page and make it visible only to have mark_page_accessed called almost immediately after. Once the page is visible the atomic operations are necessary which is noticable overhead when writing to an in-memory filesystem like tmpfs but should also be noticable with fast storage. The objective of the patch is to initialse the accessed information with non-atomic operations before the page is visible. The bulk of filesystems directly or indirectly use grab_cache_page_write_begin or find_or_create_page for the initial allocation of a page cache page. This patch adds an init_page_accessed() helper which behaves like the first call to mark_page_accessed() but may called before the page is visible and can be done non-atomically. The primary APIs of concern in this care are the following and are used by most filesystems. find_get_page find_lock_page find_or_create_page grab_cache_page_nowait grab_cache_page_write_begin All of them are very similar in detail to the patch creates a core helper pagecache_get_page() which takes a flags parameter that affects its behavior such as whether the page should be marked accessed or not. Then old API is preserved but is basically a thin wrapper around this core function. Each of the filesystems are then updated to avoid calling mark_page_accessed when it is known that the VM interfaces have already done the job. There is a slight snag in that the timing of the mark_page_accessed() has now changed so in rare cases it's possible a page gets to the end of the LRU as PageReferenced where as previously it might have been repromoted. This is expected to be rare but it's worth the filesystem people thinking about it in case they see a problem with the timing change. It is also the case that some filesystems may be marking pages accessed that previously did not but it makes sense that filesystems have consistent behaviour in this regard. The test case used to evaulate this is a simple dd of a large file done multiple times with the file deleted on each iterations. The size of the file is 1/10th physical memory to avoid dirty page balancing. In the async case it will be possible that the workload completes without even hitting the disk and will have variable results but highlight the impact of mark_page_accessed for async IO. The sync results are expected to be more stable. The exception is tmpfs where the normal case is for the "IO" to not hit the disk. The test machine was single socket and UMA to avoid any scheduling or NUMA artifacts. Throughput and wall times are presented for sync IO, only wall times are shown for async as the granularity reported by dd and the variability is unsuitable for comparison. As async results were variable do to writback timings, I'm only reporting the maximum figures. The sync results were stable enough to make the mean and stddev uninteresting. The performance results are reported based on a run with no profiling. Profile data is based on a separate run with oprofile running. async dd 3.15.0-rc3 3.15.0-rc3 vanilla accessed-v2 ext3 Max elapsed 13.9900 ( 0.00%) 11.5900 ( 17.16%) tmpfs Max elapsed 0.5100 ( 0.00%) 0.4900 ( 3.92%) btrfs Max elapsed 12.8100 ( 0.00%) 12.7800 ( 0.23%) ext4 Max elapsed 18.6000 ( 0.00%) 13.3400 ( 28.28%) xfs Max elapsed 12.5600 ( 0.00%) 2.0900 ( 83.36%) The XFS figure is a bit strange as it managed to avoid a worst case by sheer luck but the average figures looked reasonable. samples percentage ext3 86107 0.9783 vmlinux-3.15.0-rc4-vanilla mark_page_accessed ext3 23833 0.2710 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed ext3 5036 0.0573 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed ext4 64566 0.8961 vmlinux-3.15.0-rc4-vanilla mark_page_accessed ext4 5322 0.0713 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed ext4 2869 0.0384 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed xfs 62126 1.7675 vmlinux-3.15.0-rc4-vanilla mark_page_accessed xfs 1904 0.0554 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed xfs 103 0.0030 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed btrfs 10655 0.1338 vmlinux-3.15.0-rc4-vanilla mark_page_accessed btrfs 2020 0.0273 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed btrfs 587 0.0079 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed tmpfs 59562 3.2628 vmlinux-3.15.0-rc4-vanilla mark_page_accessed tmpfs 1210 0.0696 vmlinux-3.15.0-rc4-accessed-v3r25 init_page_accessed tmpfs 94 0.0054 vmlinux-3.15.0-rc4-accessed-v3r25 mark_page_accessed [akpm@linux-foundation.org: don't run init_page_accessed() against an uninitialised pointer] Signed-off-by:
Prabhakar Lad <prabhakar.csengg@gmail.com> Signed-off-by:
-
Mel Gorman authored
commit e7470ee8 upstream. Discarding buffers uses a bunch of atomic operations when discarding buffers because ...... I can't think of a reason. Use a cmpxchg loop to clear all the necessary flags. In most (all?) cases this will be a single atomic operations. [akpm@linux-foundation.org: move BUFFER_FLAGS_DISCARD into the .c file] Signed-off-by:
-
Mel Gorman authored
commit 6fb81a17 upstream. When adding pages to the LRU we clear the active bit unconditionally. As the page could be reachable from other paths we cannot use unlocked operations without risk of corruption such as a parallel mark_page_accessed. This patch tests if is necessary to clear the active flag before using an atomic operation. This potentially opens a tiny race when PageActive is checked as mark_page_accessed could be called after PageActive was checked. The race already exists but this patch changes it slightly. The consequence is that that the page may be promoted to the active list that might have been left on the inactive list before the patch. It's too tiny a race and too marginal a consequence to always use atomic operations for. Signed-off-by:
-
Mel Gorman authored
commit e3741b50 upstream. There should be no references to it any more and a parallel mark should not be reordered against us. Use non-locked varient to clear page active. Signed-off-by:
-
Mel Gorman authored
commit 07a42788 upstream. shmem_getpage_gfp uses an atomic operation to set the SwapBacked field before it's even added to the LRU or visible. This is unnecessary as what could it possible race against? Use an unlocked variant. Signed-off-by:
-
Mel Gorman authored
commit cfc47a28 upstream. get_pageblock_migratetype() is called during free with IRQs disabled. This is unnecessary and disables IRQs for longer than necessary. Signed-off-by:
-
Mel Gorman authored
commit b745bc85 upstream. cold is a bool, make it one. Make the likely case the "if" part of the block instead of the else as according to the optimisation manual this is preferred. Signed-off-by:
-
Mel Gorman authored
commit dc4b0caf upstream. In the free path we calculate page_to_pfn multiple times. Reduce that. Signed-off-by:
-
Mel Gorman authored
commit 7aeb09f9 upstream. X86 prefers the use of unsigned types for iterators and there is a tendency to mix whether a signed or unsigned type if used for page order. This converts a number of sites in mm/page_alloc.c to use unsigned int for order where possible. Signed-off-by:
-
Mel Gorman authored
commit 5dab2911 upstream. ALLOC_NO_WATERMARK is set in a few cases. Always by kswapd, always for __GFP_MEMALLOC, sometimes for swap-over-nfs, tasks etc. Each of these cases are relatively rare events but the ALLOC_NO_WATERMARK check is an unlikely branch in the fast path. This patch moves the check out of the fast path and after it has been determined that the watermarks have not been met. This helps the common fast path at the cost of making the slow path slower and hitting kswapd with a performance cost. It's a reasonable tradeoff. Signed-off-by:
-
Mel Gorman authored
commit a6e21b14 upstream. Currently it's calculated once per zone in the zonelist. Signed-off-by:
-
Mel Gorman authored
commit d34c5fa0 upstream. A node/zone index is used to check if pages are compatible for merging but this happens unconditionally even if the buddy page is not free. Defer the calculation as long as possible. Ideally we would check the zone boundary but nodes can overlap. Signed-off-by:
-
Mel Gorman authored
commit d8846374 upstream. There is no need to calculate zone_idx(preferred_zone) multiple times or use the pgdat to figure it out. Signed-off-by:
-
Mel Gorman authored
commit 664eedde upstream. If cpusets are not in use then we still check a global variable on every page allocation. Use jump labels to avoid the overhead. Signed-off-by:
-
Mel Gorman authored
commit ea5e9539 upstream. This patch exposes the jump_label reference count in preparation for the next patch. cpusets cares about both the jump_label being enabled and how many users of the cpusets there currently are. Signed-off-by:
Peter Zijlstra <peterz@infradead.org> Signed-off-by:
-
Mel Gorman authored
commit 800a1e75 upstream. If a zone cannot be used for a dirty page then it gets marked "full" which is cached in the zlc and later potentially skipped by allocation requests that have nothing to do with dirty zones. Signed-off-by:
-
Mel Gorman authored
commit 65bb3719 upstream. The zlc is used on NUMA machines to quickly skip over zones that are full. However it is always updated, even for the first zone scanned when the zlc might not even be active. As it's a write to a bitmap that potentially bounces cache line it's deceptively expensive and most machines will not care. Only update the zlc if it was active. Signed-off-by:
-
Jianyu Zhan authored
commit 2329d375 upstream. In mm/swap.c, __lru_cache_add() is exported, but actually there are no users outside this file. This patch unexports __lru_cache_add(), and makes it static. It also exports lru_cache_add_file(), as it is use by cifs and fuse, which can loaded as modules. Signed-off-by:
Jianyu Zhan <nasa4836@gmail.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Rafael Aquini <aquini@redhat.com> Cc: Mel Gorman <mgorman@suse.de> Acked-by:
Zhang Yanfei <zhangyanfei@cn.fujitsu.com> Signed-off-by:
-
Vlastimil Babka authored
commit 5bcc9f86 upstream. For the MIGRATE_RESERVE pages, it is useful when they do not get misplaced on free_list of other migratetype, otherwise they might get allocated prematurely and e.g. fragment the MIGRATE_RESEVE pageblocks. While this cannot be avoided completely when allocating new MIGRATE_RESERVE pageblocks in min_free_kbytes sysctl handler, we should prevent the misplacement where possible. Currently, it is possible for the misplacement to happen when a MIGRATE_RESERVE page is allocated on pcplist through rmqueue_bulk() as a fallback for other desired migratetype, and then later freed back through free_pcppages_bulk() without being actually used. This happens because free_pcppages_bulk() uses get_freepage_migratetype() to choose the free_list, and rmqueue_bulk() calls set_freepage_migratetype() with the *desired* migratetype and not the page's original MIGRATE_RESERVE migratetype. This patch fixes the problem by moving the call to set_freepage_migratetype() from rmqueue_bulk() down to __rmqueue_smallest() and __rmqueue_fallback() where the actual page's migratetype (e.g. from which free_list the page is taken from) is used. Note that this migratetype might be different from the pageblock's migratetype due to freepage stealing decisions. This is OK, as page stealing never uses MIGRATE_RESERVE as a fallback, and also takes care to leave all MIGRATE_CMA pages on the correct freelist. Therefore, as an additional benefit, the call to get_pageblock_migratetype() from rmqueue_bulk() when CMA is enabled, can be removed completely. This relies on the fact that MIGRATE_CMA pageblocks are created only during system init, and the above. The related is_migrate_isolate() check is also unnecessary, as memory isolation has other ways to move pages between freelists, and drain pcp lists containing pages that should be isolated. The buffered_rmqueue() can also benefit from calling get_freepage_migratetype() instead of get_pageblock_migratetype(). Signed-off-by:
Yong-Taek Lee <ytk.lee@samsung.com> Reported-by:
Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Suggested-by:
Minchan Kim <minchan@kernel.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Michal Nazarewicz <mina86@mina86.com> Cc: "Wang, Yalin" <Yalin.Wang@sonymobile.com> Signed-off-by:
-
Mel Gorman authored
commit 1a501907 upstream. Commit "mm: vmscan: obey proportional scanning requirements for kswapd" ensured that file/anon lists were scanned proportionally for reclaim from kswapd but ignored it for direct reclaim. The intent was to minimse direct reclaim latency but Yuanhan Liu pointer out that it substitutes one long stall for many small stalls and distorts aging for normal workloads like streaming readers/writers. Hugh Dickins pointed out that a side-effect of the same commit was that when one LRU list dropped to zero that the entirety of the other list was shrunk leading to excessive reclaim in memcgs. This patch scans the file/anon lists proportionally for direct reclaim to similarly age page whether reclaimed by kswapd or direct reclaim but takes care to abort reclaim if one LRU drops to zero after reclaiming the requested number of pages. Based on ext4 and using the Intel VM scalability test 3.15.0-rc5 3.15.0-rc5 shrinker proportion Unit lru-file-readonce elapsed 5.3500 ( 0.00%) 5.4200 ( -1.31%) Unit lru-file-readonce time_range 0.2700 ( 0.00%) 0.1400 ( 48.15%) Unit lru-file-readonce time_stddv 0.1148 ( 0.00%) 0.0536 ( 53.33%) Unit lru-file-readtwice elapsed 8.1700 ( 0.00%) 8.1700 ( 0.00%) Unit lru-file-readtwice time_range 0.4300 ( 0.00%) 0.2300 ( 46.51%) Unit lru-file-readtwice time_stddv 0.1650 ( 0.00%) 0.0971 ( 41.16%) The test cases are running multiple dd instances reading sparse files. The results are within the noise for the small test machine. The impact of the patch is more noticable from the vmstats 3.15.0-rc5 3.15.0-rc5 shrinker proportion Minor Faults 35154 36784 Major Faults 611 1305 Swap Ins 394 1651 Swap Outs 4394 5891 Allocation stalls 118616 44781 Direct pages scanned 4935171 4602313 Kswapd pages scanned 15921292 16258483 Kswapd pages reclaimed 15913301 16248305 Direct pages reclaimed 4933368 4601133 Kswapd efficiency 99% 99% Kswapd velocity 670088.047 682555.961 Direct efficiency 99% 99% Direct velocity 207709.217 193212.133 Percentage direct scans 23% 22% Page writes by reclaim 4858.000 6232.000 Page writes file 464 341 Page writes anon 4394 5891 Note that there are fewer allocation stalls even though the amount of direct reclaim scanning is very approximately the same. Signed-off-by:
Yuanhan Liu <yuanhan.liu@linux.intel.com> Cc: Bob Liu <bob.liu@oracle.com> Cc: Jan Kara <jack@suse.cz> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by:
-
Tim Chen authored
commit d23da150 upstream. We remove the call to grab_super_passive in call to super_cache_count. This becomes a scalability bottleneck as multiple threads are trying to do memory reclamation, e.g. when we are doing large amount of file read and page cache is under pressure. The cached objects quickly got reclaimed down to 0 and we are aborting the cache_scan() reclaim. But counting creates a log jam acquiring the sb_lock. We are holding the shrinker_rwsem which ensures the safety of call to list_lru_count_node() and s_op->nr_cached_objects. The shrinker is unregistered now before ->kill_sb() so the operation is safe when we are doing unmount. The impact will depend heavily on the machine and the workload but for a small machine using postmark tuned to use 4xRAM size the results were 3.15.0-rc5 3.15.0-rc5 vanilla shrinker-v1r1 Ops/sec Transactions 21.00 ( 0.00%) 24.00 ( 14.29%) Ops/sec FilesCreate 39.00 ( 0.00%) 44.00 ( 12.82%) Ops/sec CreateTransact 10.00 ( 0.00%) 12.00 ( 20.00%) Ops/sec FilesDeleted 6202.00 ( 0.00%) 6202.00 ( 0.00%) Ops/sec DeleteTransact 11.00 ( 0.00%) 12.00 ( 9.09%) Ops/sec DataRead/MB 25.97 ( 0.00%) 29.10 ( 12.05%) Ops/sec DataWrite/MB 49.99 ( 0.00%) 56.02 ( 12.06%) ffsb running in a configuration that is meant to simulate a mail server showed 3.15.0-rc5 3.15.0-rc5 vanilla shrinker-v1r1 Ops/sec readall 9402.63 ( 0.00%) 9567.97 ( 1.76%) Ops/sec create 4695.45 ( 0.00%) 4735.00 ( 0.84%) Ops/sec delete 173.72 ( 0.00%) 179.83 ( 3.52%) Ops/sec Transactions 14271.80 ( 0.00%) 14482.81 ( 1.48%) Ops/sec Read 37.00 ( 0.00%) 37.60 ( 1.62%) Ops/sec Write 18.20 ( 0.00%) 18.30 ( 0.55%) Signed-off-by:
Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by:
-
Dave Chinner authored
commit 28f2cd4f upstream. This series is aimed at regressions noticed during reclaim activity. The first two patches are shrinker patches that were posted ages ago but never merged for reasons that are unclear to me. I'm posting them again to see if there was a reason they were dropped or if they just got lost. Dave? Time? The last patch adjusts proportional reclaim. Yuanhan Liu, can you retest the vm scalability test cases on a larger machine? Hugh, does this work for you on the memcg test cases? Based on ext4, I get the following results but unfortunately my larger test machines are all unavailable so this is based on a relatively small machine. postmark 3.15.0-rc5 3.15.0-rc5 vanilla proportion-v1r4 Ops/sec Transactions 21.00 ( 0.00%) 25.00 ( 19.05%) Ops/sec FilesCreate 39.00 ( 0.00%) 45.00 ( 15.38%) Ops/sec CreateTransact 10.00 ( 0.00%) 12.00 ( 20.00%) Ops/sec FilesDeleted 6202.00 ( 0.00%) 6202.00 ( 0.00%) Ops/sec DeleteTransact 11.00 ( 0.00%) 12.00 ( 9.09%) Ops/sec DataRead/MB 25.97 ( 0.00%) 30.02 ( 15.59%) Ops/sec DataWrite/MB 49.99 ( 0.00%) 57.78 ( 15.58%) ffsb (mail server simulator) 3.15.0-rc5 3.15.0-rc5 vanilla proportion-v1r4 Ops/sec readall 9402.63 ( 0.00%) 9805.74 ( 4.29%) Ops/sec create 4695.45 ( 0.00%) 4781.39 ( 1.83%) Ops/sec delete 173.72 ( 0.00%) 177.23 ( 2.02%) Ops/sec Transactions 14271.80 ( 0.00%) 14764.37 ( 3.45%) Ops/sec Read 37.00 ( 0.00%) 38.50 ( 4.05%) Ops/sec Write 18.20 ( 0.00%) 18.50 ( 1.65%) dd of a large file 3.15.0-rc5 3.15.0-rc5 vanilla proportion-v1r4 WallTime DownloadTar 75.00 ( 0.00%) 61.00 ( 18.67%) WallTime DD 423.00 ( 0.00%) 401.00 ( 5.20%) WallTime Delete 2.00 ( 0.00%) 5.00 (-150.00%) stutter (times mmap latency during large amounts of IO) 3.15.0-rc5 3.15.0-rc5 vanilla proportion-v1r4 Unit >5ms Delays 80252.0000 ( 0.00%) 81523.0000 ( -1.58%) Unit Mmap min 8.2118 ( 0.00%) 8.3206 ( -1.33%) Unit Mmap mean 17.4614 ( 0.00%) 17.2868 ( 1.00%) Unit Mmap stddev 24.9059 ( 0.00%) 34.6771 (-39.23%) Unit Mmap max 2811.6433 ( 0.00%) 2645.1398 ( 5.92%) Unit Mmap 90% 20.5098 ( 0.00%) 18.3105 ( 10.72%) Unit Mmap 93% 22.9180 ( 0.00%) 20.1751 ( 11.97%) Unit Mmap 95% 25.2114 ( 0.00%) 22.4988 ( 10.76%) Unit Mmap 99% 46.1430 ( 0.00%) 43.5952 ( 5.52%) Unit Ideal Tput 85.2623 ( 0.00%) 78.8906 ( 7.47%) Unit Tput min 44.0666 ( 0.00%) 43.9609 ( 0.24%) Unit Tput mean 45.5646 ( 0.00%) 45.2009 ( 0.80%) Unit Tput stddev 0.9318 ( 0.00%) 1.1084 (-18.95%) Unit Tput max 46.7375 ( 0.00%) 46.7539 ( -0.04%) This patch (of 3): We will like to unregister the sb shrinker before ->kill_sb(). This will allow cached objects to be counted without call to grab_super_passive() to update ref count on sb. We want to avoid locking during memory reclamation especially when we are skipping the memory reclaim when we are out of cached objects. This is safe because grab_super_passive does a try-lock on the sb->s_umount now, and so if we are in the unmount process, it won't ever block. That means what used to be a deadlock and races we were avoiding by using grab_super_passive() is now: shrinker umount down_read(shrinker_rwsem) down_write(sb->s_umount) shrinker_unregister down_write(shrinker_rwsem) <blocks> grab_super_passive(sb) down_read_trylock(sb->s_umount) <fails> <shrinker aborts> .... <shrinkers finish running> up_read(shrinker_rwsem) <unblocks> <removes shrinker> up_write(shrinker_rwsem) ->kill_sb() .... So it is safe to deregister the shrinker before ->kill_sb(). Signed-off-by:
-
Hugh Dickins authored
commit 8bdd6380 upstream. Shortly before 3.16-rc1, Dave Jones reported: WARNING: CPU: 3 PID: 19721 at fs/xfs/xfs_aops.c:971 xfs_vm_writepage+0x5ce/0x630 [xfs]() CPU: 3 PID: 19721 Comm: trinity-c61 Not tainted 3.15.0+ #3 Call Trace: xfs_vm_writepage+0x5ce/0x630 [xfs] shrink_page_list+0x8f9/0xb90 shrink_inactive_list+0x253/0x510 shrink_lruvec+0x563/0x6c0 shrink_zone+0x3b/0x100 shrink_zones+0x1f1/0x3c0 try_to_free_pages+0x164/0x380 __alloc_pages_nodemask+0x822/0xc90 alloc_pages_vma+0xaf/0x1c0 handle_mm_fault+0xa31/0xc50 etc. 970 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) == 971 PF_MEMALLOC)) I did not respond at the time, because a glance at the PageDirty block in shrink_page_list() quickly shows that this is impossible: we don't do writeback on file pages (other than tmpfs) from direct reclaim nowadays. Dave was hallucinating, but it would have been disrespectful to say so. However, my own /var/log/messages now shows similar complaints WARNING: CPU: 1 PID: 28814 at fs/ext4/inode.c:1881 ext4_writepage+0xa7/0x38b() WARNING: CPU: 0 PID: 27347 at fs/ext4/inode.c:1764 ext4_writepage+0xa7/0x38b() from stressing some mmotm trees during July. Could a dirty xfs or ext4 file page somehow get marked PageSwapBacked, so fail shrink_page_list()'s page_is_file_cache() test, and so proceed to mapping->a_ops->writepage()? Yes, 3.16-rc1's commit 68711a74 ("mm, migration: add destination page freeing callback") has provided such a way to compaction: if migrating a SwapBacked page fails, its newpage may be put back on the list for later use with PageSwapBacked still set, and nothing will clear it. Whether that can do anything worse than issue WARN_ON_ONCEs, and get some statistics wrong, is unclear: easier to fix than to think through the consequences. Fixing it here, before the put_new_page(), addresses the bug directly, but is probably the worst place to fix it. Page migration is doing too many parts of the job on too many levels: fixing it in move_to_new_page() to complement its SetPageSwapBacked would be preferable, except why is it (and newpage->mapping and newpage->index) done there, rather than down in migrate_page_move_mapping(), once we are sure of success? Not a cleanup to get into right now, especially not with memcg cleanups coming in 3.17. Reported-by:
Dave Jones <davej@redhat.com> Signed-off-by:
-
Shaohua Li authored
commit b13b1d2d upstream. We use the accessed bit to age a page at page reclaim time, and currently we also flush the TLB when doing so. But in some workloads TLB flush overhead is very heavy. In my simple multithreaded app with a lot of swap to several pcie SSDs, removing the tlb flush gives about 20% ~ 30% swapout speedup. Fortunately just removing the TLB flush is a valid optimization: on x86 CPUs, clearing the accessed bit without a TLB flush doesn't cause data corruption. It could cause incorrect page aging and the (mistaken) reclaim of hot pages, but the chance of that should be relatively low. So as a performance optimization don't flush the TLB when clearing the accessed bit, it will eventually be flushed by a context switch or a VM operation anyway. [ In the rare event of it not getting flushed for a long time the delay shouldn't really matter because there's no real memory pressure for swapout to react to. ] Suggested-by:
Shaohua Li <shli@fusionio.com> Acked-by:
Ingo Molnar <mingo@kernel.org> Signed-off-by:
-
Vlastimil Babka authored
commit be976572 upstream. Compaction uses compact_checklock_irqsave() function to periodically check for lock contention and need_resched() to either abort async compaction, or to free the lock, schedule and retake the lock. When aborting, cc->contended is set to signal the contended state to the caller. Two problems have been identified in this mechanism. First, compaction also calls directly cond_resched() in both scanners when no lock is yet taken. This call either does not abort async compaction, or set cc->contended appropriately. This patch introduces a new compact_should_abort() function to achieve both. In isolate_freepages(), the check frequency is reduced to once by SWAP_CLUSTER_MAX pageblocks to match what the migration scanner does in the preliminary page checks. In case a pageblock is found suitable for calling isolate_freepages_block(), the checks within there are done on higher frequency. Second, isolate_freepages() does not check if isolate_freepages_block() aborted due to contention, and advances to the next pageblock. This violates the principle of aborting on contention, and might result in pageblocks not being scanned completely, since the scanning cursor is advanced. This problem has been noticed in the code by Joonsoo Kim when reviewing related patches. This patch makes isolate_freepages_block() check the cc->contended flag and abort. In case isolate_freepages() has already isolated some pages before aborting due to contention, page migration will proceed, which is OK since we do not want to waste the work that has been done, and page migration has own checks for contention. However, we do not want another isolation attempt by either of the scanners, so cc->contended flag check is added also to compaction_alloc() and compact_finished() to make sure compaction is aborted right after the migration. The outcome of the patch should be reduced lock contention by async compaction and lower latencies for higher-order allocations where direct compaction is involved. [akpm@linux-foundation.org: fix typo in comment] Reported-by:
Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Michal Nazarewicz <mina86@mina86.com> Cc: Christoph Lameter <cl@linux.com> Cc: Rik van Riel <riel@redhat.com> Acked-by:
Michal Nazarewicz <mina86@mina86.com> Tested-by:
Shawn Guo <shawn.guo@linaro.org> Tested-by:
Kevin Hilman <khilman@linaro.org> Tested-by:
Stephen Warren <swarren@nvidia.com> Tested-by:
Fabio Estevam <fabio.estevam@freescale.com> Cc: David Rientjes <rientjes@google.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by:
-
Vlastimil Babka authored
commit e9ade569 upstream. The compaction free scanner in isolate_freepages() currently remembers PFN of the highest pageblock where it successfully isolates, to be used as the starting pageblock for the next invocation. The rationale behind this is that page migration might return free pages to the allocator when migration fails and we don't want to skip them if the compaction continues. Since migration now returns free pages back to compaction code where they can be reused, this is no longer a concern. This patch changes isolate_freepages() so that the PFN for restarting is updated with each pageblock where isolation is attempted. Using stress-highalloc from mmtests, this resulted in 10% reduction of the pages scanned by the free scanner. Note that the somewhat similar functionality that records highest successful pageblock in zone->compact_cached_free_pfn, remains unchanged. This cache is used when the whole compaction is restarted, not for multiple invocations of the free scanner during single compaction. Signed-off-by:
-
Vlastimil Babka authored
commit f8c9301f upstream. During compaction, update_nr_listpages() has been used to count remaining non-migrated and free pages after a call to migrage_pages(). The freepages counting has become unneccessary, and it turns out that migratepages counting is also unnecessary in most cases. The only situation when it's needed to count cc->migratepages is when migrate_pages() returns with a negative error code. Otherwise, the non-negative return value is the number of pages that were not migrated, which is exactly the count of remaining pages in the cc->migratepages list. Furthermore, any non-zero count is only interesting for the tracepoint of mm_compaction_migratepages events, because after that all remaining unmigrated pages are put back and their count is set to 0. This patch therefore removes update_nr_listpages() completely, and changes the tracepoint definition so that the manual counting is done only when the tracepoint is enabled, and only when migrate_pages() returns a negative error code. Furthermore, migrate_pages() and the tracepoints won't be called when there's nothing to migrate. This potentially avoids some wasted cycles and reduces the volume of uninteresting mm_compaction_migratepages events where "nr_migrated=0 nr_failed=0". In the stress-highalloc mmtest, this was about 75% of the events. The mm_compaction_isolate_migratepages event is better for determining that nothing was isolated for migration, and this one was just duplicating the info. Signed-off-by:
-
David Rientjes authored
commit aeef4b83 upstream. Async compaction terminates prematurely when need_resched(), see compact_checklock_irqsave(). This can never trigger, however, if the cond_resched() in isolate_migratepages_range() always takes care of the scheduling. If the cond_resched() actually triggers, then terminate this pageblock scan for async compaction as well. Signed-off-by:
-
