- 26 Sep, 2014 40 commits
-
-
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:
Vlastimil Babka <vbabka@suse.cz> Reviewed-by:
Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Acked-by:
Michal Nazarewicz <mina86@mina86.com> Cc: Christoph Lameter <cl@linux.com> Cc: Rik van Riel <riel@redhat.com> Acked-by:
David Rientjes <rientjes@google.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by:
Mel Gorman <mgorman@suse.de> Signed-off-by:
Jiri Slaby <jslaby@suse.cz>
-
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:
-
David Rientjes authored
commit e0b9daeb upstream. We're going to want to manipulate the migration mode for compaction in the page allocator, and currently compact_control's sync field is only a bool. Currently, we only do MIGRATE_ASYNC or MIGRATE_SYNC_LIGHT compaction depending on the value of this bool. Convert the bool to enum migrate_mode and pass the migration mode in directly. Later, we'll want to avoid MIGRATE_SYNC_LIGHT for thp allocations in the pagefault patch to avoid unnecessary latency. This also alters compaction triggered from sysfs, either for the entire system or for a node, to force MIGRATE_SYNC. [akpm@linux-foundation.org: fix build] [iamjoonsoo.kim@lge.com: use MIGRATE_SYNC in alloc_contig_range()] Signed-off-by:
Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
-
David Rientjes authored
commit 35979ef3 upstream. Each zone has a cached migration scanner pfn for memory compaction so that subsequent calls to memory compaction can start where the previous call left off. Currently, the compaction migration scanner only updates the per-zone cached pfn when pageblocks were not skipped for async compaction. This creates a dependency on calling sync compaction to avoid having subsequent calls to async compaction from scanning an enormous amount of non-MOVABLE pageblocks each time it is called. On large machines, this could be potentially very expensive. This patch adds a per-zone cached migration scanner pfn only for async compaction. It is updated everytime a pageblock has been scanned in its entirety and when no pages from it were successfully isolated. The cached migration scanner pfn for sync compaction is updated only when called for sync compaction. Signed-off-by:
-
David Rientjes authored
commit d53aea3d upstream. Greg reported that he found isolated free pages were returned back to the VM rather than the compaction freelist. This will cause holes behind the free scanner and cause it to reallocate additional memory if necessary later. He detected the problem at runtime seeing that ext4 metadata pages (esp the ones read by "sbi->s_group_desc[i] = sb_bread(sb, block)") were constantly visited by compaction calls of migrate_pages(). These pages had a non-zero b_count which caused fallback_migrate_page() -> try_to_release_page() -> try_to_free_buffers() to fail. Memory compaction works by having a "freeing scanner" scan from one end of a zone which isolates pages as migration targets while another "migrating scanner" scans from the other end of the same zone which isolates pages for migration. When page migration fails for an isolated page, the target page is returned to the system rather than the freelist built by the freeing scanner. This may require the freeing scanner to continue scanning memory after suitable migration targets have already been returned to the system needlessly. This patch returns destination pages to the freeing scanner freelist when page migration fails. This prevents unnecessary work done by the freeing scanner but also encourages memory to be as compacted as possible at the end of the zone. Signed-off-by:
Greg Thelen <gthelen@google.com> Acked-by:
-
David Rientjes authored
commit 68711a74 upstream. Memory migration uses a callback defined by the caller to determine how to allocate destination pages. When migration fails for a source page, however, it frees the destination page back to the system. This patch adds a memory migration callback defined by the caller to determine how to free destination pages. If a caller, such as memory compaction, builds its own freelist for migration targets, this can reuse already freed memory instead of scanning additional memory. If the caller provides a function to handle freeing of destination pages, it is called when page migration fails. If the caller passes NULL then freeing back to the system will be handled as usual. This patch introduces no functional change. Signed-off-by:
-
Vlastimil Babka authored
commit c96b9e50 upstream. isolate_freepages() is currently somewhat hard to follow thanks to many looks like it is related to the 'low_pfn' variable, but in fact it is not. This patch renames the 'high_pfn' variable to a hopefully less confusing name, and slightly changes its handling without a functional change. A comment made obsolete by recent changes is also updated. [akpm@linux-foundation.org: comment fixes, per Minchan] [iamjoonsoo.kim@lge.com: cleanups] Signed-off-by:
-
Heesub Shin authored
commit 13fb44e4 upstream. Remove code lines currently not in use or never called. Signed-off-by:
Heesub Shin <heesub.shin@samsung.com> Acked-by:
-
Fabian Frederick authored
commit 29f175d1 upstream. Commit f9acc8c7 ("readahead: sanify file_ra_state names") left ra_submit with a single function call. Move ra_submit to internal.h and inline it to save some stack. Thanks to Andrew Morton for commenting different versions. Signed-off-by:
Fabian Frederick <fabf@skynet.be> Suggested-by:
-
Al Viro authored
commit 9e8c2af9 upstream. ... it does that itself (via kmap_atomic()) Signed-off-by:
Al Viro <viro@zeniv.linux.org.uk> Signed-off-by:
-
Sasha Levin authored
commit 67f9fd91 upstream. This patch removes read_cache_page_async() which wasn't really needed anywhere and simplifies the code around it a bit. read_cache_page_async() is useful when we want to read a page into the cache without waiting for it to complete. This happens when the appropriate callback 'filler' doesn't complete its read operation and releases the page lock immediately, and instead queues a different completion routine to do that. This never actually happened anywhere in the code. read_cache_page_async() had 3 different callers: - read_cache_page() which is the sync version, it would just wait for the requested read to complete using wait_on_page_read(). - JFFS2 would call it from jffs2_gc_fetch_page(), but the filler function it supplied doesn't do any async reads, and would complete before the filler function returns - making it actually a sync read. - CRAMFS would call it using the read_mapping_page_async() wrapper, with a similar story to JFFS2 - the filler function doesn't do anything that reminds async reads and would always complete before the filler function returns. To sum it up, the code in mm/filemap.c never took advantage of having read_cache_page_async(). While there are filler callbacks that do async reads (such as the block one), we always called it with the read_cache_page(). This patch adds a mandatory wait for read to complete when adding a new page to the cache, and removes read_cache_page_async() and its wrappers. Signed-off-by:
Sasha Levin <sasha.levin@oracle.com> Signed-off-by:
-
Johannes Weiner authored
commit 55231e5c upstream. MADV_WILLNEED currently does not read swapped out shmem pages back in. Commit 0cd6144a ("mm + fs: prepare for non-page entries in page cache radix trees") made find_get_page() filter exceptional radix tree entries but failed to convert all find_get_page() callers that WANT exceptional entries over to find_get_entry(). One of them is shmem swap readahead in madvise, which now skips over any swap-out records. Convert it to find_get_entry(). Fixes: 0cd6144a ("mm + fs: prepare for non-page entries in page cache radix trees") Signed-off-by:
Johannes Weiner <hannes@cmpxchg.org> Reported-by:
Hugh Dickins <hughd@google.com> Signed-off-by:
-
Johannes Weiner authored
commit 0cd6144a upstream. shmem mappings already contain exceptional entries where swap slot information is remembered. To be able to store eviction information for regular page cache, prepare every site dealing with the radix trees directly to handle entries other than pages. The common lookup functions will filter out non-page entries and return NULL for page cache holes, just as before. But provide a raw version of the API which returns non-page entries as well, and switch shmem over to use it. Signed-off-by:
Rik van Riel <riel@redhat.com> Reviewed-by:
Minchan Kim <minchan@kernel.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Bob Liu <bob.liu@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg Thelen <gthelen@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jan Kara <jack@suse.cz> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Luigi Semenzato <semenzato@google.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Metin Doslu <metin@citusdata.com> Cc: Michel Lespinasse <walken@google.com> Cc: Ozgun Erdogan <ozgun@citusdata.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Roman Gushchin <klamm@yandex-team.ru> Cc: Ryan Mallon <rmallon@gmail.com> Cc: Tejun Heo <tj@kernel.org> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by:
-
Johannes Weiner authored
commit e7b563bb upstream. The radix tree hole searching code is only used for page cache, for example the readahead code trying to get a a picture of the area surrounding a fault. It sufficed to rely on the radix tree definition of holes, which is "empty tree slot". But this is about to change, though, as shadow page descriptors will be stored in the page cache after the actual pages get evicted from memory. Move the functions over to mm/filemap.c and make them native page cache operations, where they can later be adapted to handle the new definition of "page cache hole". Signed-off-by:
-
Johannes Weiner authored
commit 6dbaf22c upstream. Page cache radix tree slots are usually stabilized by the page lock, but shmem's swap cookies have no such thing. Because the overall truncation loop is lockless, the swap entry is currently confirmed by a tree lookup and then deleted by another tree lookup under the same tree lock region. Use radix_tree_delete_item() instead, which does the verification and deletion with only one lookup. This also allows removing the delete-only special case from shmem_radix_tree_replace(). Signed-off-by:
-
Johannes Weiner authored
commit 53c59f26 upstream. Provide a function that does not just delete an entry at a given index, but also allows passing in an expected item. Delete only if that item is still located at the specified index. This is handy when lockless tree traversals want to delete entries as well because they don't have to do an second, locked lookup to verify the slot has not changed under them before deleting the entry. Signed-off-by:
-
Linus Torvalds authored
commit 50f5aa8a upstream. BUG_ON() is a big hammer, and should be used _only_ if there is some major corruption that you cannot possibly recover from, making it imperative that the current process (and possibly the whole machine) be terminated with extreme prejudice. The trivial sanity check in the vmacache code is *not* such a fatal error. Recovering from it is absolutely trivial, and using BUG_ON() just makes it harder to debug for no actual advantage. To make matters worse, the placement of the BUG_ON() (only if the range check matched) actually makes it harder to hit the sanity check to begin with, so _if_ there is a bug (and we just got a report from Srivatsa Bhat that this can indeed trigger), it is harder to debug not just because the machine is possibly dead, but because we don't have better coverage. BUG_ON() must *die*. Maybe we should add a checkpatch warning for it, because it is simply just about the worst thing you can ever do if you hit some "this cannot happen" situation. Reported-by:
Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
-
Davidlohr Bueso authored
commit 615d6e87 upstream. This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by:
Davidlohr Bueso <davidlohr@hp.com> Reviewed-by:
Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by:
-
Christoph Lameter authored
commit 83da7510 upstream. Seems to be called with preemption enabled. Therefore it must use mod_zone_page_state instead. Signed-off-by:
Christoph Lameter <cl@linux.com> Reported-by:
Grygorii Strashko <grygorii.strashko@ti.com> Tested-by:
-
Vladimir Davydov authored
commit d5bc5fd3 upstream. The name `max_pass' is misleading, because this variable actually keeps the estimate number of freeable objects, not the maximal number of objects we can scan in this pass, which can be twice that. Rename it to reflect its actual meaning. Signed-off-by:
Vladimir Davydov <vdavydov@parallels.com> Acked-by:
-
Vladimir Davydov authored
commit 99120b77 upstream. When direct reclaim is executed by a process bound to a set of NUMA nodes, we should scan only those nodes when possible, but currently we will scan kmem from all online nodes even if the kmem shrinker is NUMA aware. That said, binding a process to a particular NUMA node won't prevent it from shrinking inode/dentry caches from other nodes, which is not good. Fix this. Signed-off-by:
-
Jens Axboe authored
commit 7fcbbaf1 upstream. In some testing I ran today (some fio jobs that spread over two nodes), we end up spending 40% of the time in filemap_check_errors(). That smells fishy. Looking further, this is basically what happens: blkdev_aio_read() generic_file_aio_read() filemap_write_and_wait_range() if (!mapping->nr_pages) filemap_check_errors() and filemap_check_errors() always attempts two test_and_clear_bit() on the mapping flags, thus dirtying it for every single invocation. The patch below tests each of these bits before clearing them, avoiding this issue. In my test case (4-socket box), performance went from 1.7M IOPS to 4.0M IOPS. Signed-off-by:
Jens Axboe <axboe@fb.com> Acked-by:
Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by:
-
Mel Gorman authored
commit d26914d1 upstream. Since put_mems_allowed() is strictly optional, its a seqcount retry, we don't need to evaluate the function if the allocation was in fact successful, saving a smp_rmb some loads and comparisons on some relative fast-paths. Since the naming, get/put_mems_allowed() does suggest a mandatory pairing, rename the interface, as suggested by Mel, to resemble the seqcount interface. This gives us: read_mems_allowed_begin() and read_mems_allowed_retry(), where it is important to note that the return value of the latter call is inverted from its previous incarnation. Signed-off-by:
Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by:
-
Raghavendra K T authored
commit 6d2be915 upstream. Currently max_sane_readahead() returns zero on the cpu whose NUMA node has no local memory which leads to readahead failure. Fix this readahead failure by returning minimum of (requested pages, 512). Users running applications on a memory-less cpu which needs readahead such as streaming application see considerable boost in the performance. Result: fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU with 1GB testfile (12 iterations) yielded around 46.66% improvement. fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile 32GB* 4G RAM numa machine (12 iterations) showed no impact on the normal NUMA cases w/ patch. Kernel Avg Stddev base 7.4975 3.92% patched 7.4174 3.26% [Andrew: making return value PAGE_SIZE independent] Suggested-by:
Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by:
Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: David Rientjes <rientjes@google.com> Signed-off-by:
-
David Rientjes authored
commit 91ca9186 upstream. The cached pageblock hint should be ignored when triggering compaction through /proc/sys/vm/compact_memory so all eligible memory is isolated. Manually invoking compaction is known to be expensive, there's no need to skip pageblocks based on heuristics (mainly for debugging). Signed-off-by:
-
David Rientjes authored
commit da1c67a7 upstream. The conditions that control the isolation mode in isolate_migratepages_range() do not change during the iteration, so extract them out and only define the value once. This actually does have an effect, gcc doesn't optimize it itself because of cc->sync. Signed-off-by:
-
Joonsoo Kim authored
commit b6c75016 upstream. It is just for clean-up to reduce code size and improve readability. There is no functional change. Signed-off-by:
-
Joonsoo Kim authored
commit c122b208 upstream. isolation_suitable() and migrate_async_suitable() is used to be sure that this pageblock range is fine to be migragted. It isn't needed to call it on every page. Current code do well if not suitable, but, don't do well when suitable. 1) It re-checks isolation_suitable() on each page of a pageblock that was already estabilished as suitable. 2) It re-checks migrate_async_suitable() on each page of a pageblock that was not entered through the next_pageblock: label, because last_pageblock_nr is not otherwise updated. This patch fixes situation by 1) calling isolation_suitable() only once per pageblock and 2) always updating last_pageblock_nr to the pageblock that was just checked. Additionally, move PageBuddy() check after pageblock unit check, since pageblock check is the first thing we should do and makes things more simple. [vbabka@suse.cz: rephrase commit description] Signed-off-by:
-
Joonsoo Kim authored
commit be1aa03b upstream. It is odd to drop the spinlock when we scan (SWAP_CLUSTER_MAX - 1) th pfn page. This may results in below situation while isolating migratepage. 1. try isolate 0x0 ~ 0x200 pfn pages. 2. When low_pfn is 0x1ff, ((low_pfn+1) % SWAP_CLUSTER_MAX) == 0, so drop the spinlock. 3. Then, to complete isolating, retry to aquire the lock. I think that it is better to use SWAP_CLUSTER_MAX th pfn for checking the criteria about dropping the lock. This has no harm 0x0 pfn, because, at this time, locked variable would be false. Signed-off-by:
-
Joonsoo Kim authored
commit 01ead534 upstream. suitable_migration_target() checks that pageblock is suitable for migration target. In isolate_freepages_block(), it is called on every page and this is inefficient. So make it called once per pageblock. suitable_migration_target() also checks if page is highorder or not, but it's criteria for highorder is pageblock order. So calling it once within pageblock range has no problem. Signed-off-by:
-
Joonsoo Kim authored
commit 7d348b9e upstream. Purpose of compaction is to get a high order page. Currently, if we find high-order page while searching migration target page, we break it to order-0 pages and use them as migration target. It is contrary to purpose of compaction, so disallow high-order page to be used for migration target. Additionally, clean-up logic in suitable_migration_target() to simplify the code. There is no functional changes from this clean-up. Signed-off-by:
-
David Rientjes authored
commit 119d6d59 upstream. Page migration will fail for memory that is pinned in memory with, for example, get_user_pages(). In this case, it is unnecessary to take zone->lru_lock or isolating the page and passing it to page migration which will ultimately fail. This is a racy check, the page can still change from under us, but in that case we'll just fail later when attempting to move the page. This avoids very expensive memory compaction when faulting transparent hugepages after pinning a lot of memory with a Mellanox driver. On a 128GB machine and pinning ~120GB of memory, before this patch we see the enormous disparity in the number of page migration failures because of the pinning (from /proc/vmstat): compact_pages_moved 8450 compact_pagemigrate_failed 15614415 0.05% of pages isolated are successfully migrated and explicitly triggering memory compaction takes 102 seconds. After the patch: compact_pages_moved 9197 compact_pagemigrate_failed 7 99.9% of pages isolated are now successfully migrated in this configuration and memory compaction takes less than one second. Signed-off-by:
-
Yasuaki Ishimatsu authored
commit 943dca1a upstream. Yasuaki Ishimatsu reported memory hot-add spent more than 5 _hours_ on 9TB memory machine since onlining memory sections is too slow. And we found out setup_zone_migrate_reserve spent >90% of the time. The problem is, setup_zone_migrate_reserve scans all pageblocks unconditionally, but it is only necessary if the number of reserved block was reduced (i.e. memory hot remove). Moreover, maximum MIGRATE_RESERVE per zone is currently 2. It means that the number of reserved pageblocks is almost always unchanged. This patch adds zone->nr_migrate_reserve_block to maintain the number of MIGRATE_RESERVE pageblocks and it reduces the overhead of setup_zone_migrate_reserve dramatically. The following table shows time of onlining a memory section. Amount of memory | 128GB | 192GB | 256GB| --------------------------------------------- linux-3.12 | 23.9 | 31.4 | 44.5 | This patch | 8.3 | 8.3 | 8.6 | Mel's proposal patch | 10.9 | 19.2 | 31.3 | --------------------------------------------- (millisecond) 128GB : 4 nodes and each node has 32GB of memory 192GB : 6 nodes and each node has 32GB of memory 256GB : 8 nodes and each node has 32GB of memory (*1) Mel proposed his idea by the following threads. https://lkml.org/lkml/2013/10/30/272 [akpm@linux-foundation.org: tweak comment] Signed-off-by:
KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by:
Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by:
-
Vladimir Davydov authored
commit ec97097b upstream. If a shrinker is not NUMA-aware, shrink_slab() should call it exactly once with nid=0, but currently it is not true: if node 0 is not set in the nodemask or if it is not online, we will not call such shrinkers at all. As a result some slabs will be left untouched under some circumstances. Let us fix it. Signed-off-by:
Dave Chinner <dchinner@redhat.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.cz> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Rik van Riel <riel@redhat.com> Cc: Glauber Costa <glommer@gmail.com> Signed-off-by:
-
Vladimir Davydov authored
commit 0b1fb40a upstream. When reclaiming kmem, we currently don't scan slabs that have less than batch_size objects (see shrink_slab_node()): while (total_scan >= batch_size) { shrinkctl->nr_to_scan = batch_size; shrinker->scan_objects(shrinker, shrinkctl); total_scan -= batch_size; } If there are only a few shrinkers available, such a behavior won't cause any problems, because the batch_size is usually small, but if we have a lot of slab shrinkers, which is perfectly possible since FS shrinkers are now per-superblock, we can end up with hundreds of megabytes of practically unreclaimable kmem objects. For instance, mounting a thousand of ext2 FS images with a hundred of files in each and iterating over all the files using du(1) will result in about 200 Mb of FS caches that cannot be dropped even with the aid of the vm.drop_caches sysctl! This problem was initially pointed out by Glauber Costa [*]. Glauber proposed to fix it by making the shrink_slab() always take at least one pass, to put it simply, turning the scan loop above to a do{}while() loop. However, this proposal was rejected, because it could result in more aggressive and frequent slab shrinking even under low memory pressure when total_scan is naturally very small. This patch is a slightly modified version of Glauber's approach. Similarly to Glauber's patch, it makes shrink_slab() scan less than batch_size objects, but only if the total number of objects we want to scan (total_scan) is greater than the total number of objects available (max_pass). Since total_scan is biased as half max_pass if the current delta change is small: if (delta < max_pass / 4) total_scan = min(total_scan, max_pass / 2); this is only possible if we are scanning at high prio. That said, this patch shouldn't change the vmscan behaviour if the memory pressure is low, but if we are tight on memory, we will do our best by trying to reclaim all available objects, which sounds reasonable. [*] http://www.spinics.net/lists/cgroups/msg06913.htmlSigned-off-by:
-
Shaohua Li authored
commit 579f8290 upstream. This is a patch to improve swap readahead algorithm. It's from Hugh and I slightly changed it. Hugh's original changelog: swapin readahead does a blind readahead, whether or not the swapin is sequential. This may be ok on harddisk, because large reads have relatively small costs, and if the readahead pages are unneeded they can be reclaimed easily - though, what if their allocation forced reclaim of useful pages? But on SSD devices large reads are more expensive than small ones: if the readahead pages are unneeded, reading them in caused significant overhead. This patch adds very simplistic random read detection. Stealing the PageReadahead technique from Konstantin Khlebnikov's patch, avoiding the vma/anon_vma sophistications of Shaohua Li's patch, swapin_nr_pages() simply looks at readahead's current success rate, and narrows or widens its readahead window accordingly. There is little science to its heuristic: it's about as stupid as can be whilst remaining effective. The table below shows elapsed times (in centiseconds) when running a single repetitive swapping load across a 1000MB mapping in 900MB ram with 1GB swap (the harddisk tests had taken painfully too long when I used mem=500M, but SSD shows similar results for that). Vanilla is the 3.6-rc7 kernel on which I started; Shaohua denotes his Sep 3 patch in mmotm and linux-next; HughOld denotes my Oct 1 patch which Shaohua showed to be defective; HughNew this Nov 14 patch, with page_cluster as usual at default of 3 (8-page reads); HughPC4 this same patch with page_cluster 4 (16-page reads); HughPC0 with page_cluster 0 (1-page reads: no readahead). HDD for swapping to harddisk, SSD for swapping to VertexII SSD. Seq for sequential access to the mapping, cycling five times around; Rand for the same number of random touches. Anon for a MAP_PRIVATE anon mapping; Shmem for a MAP_SHARED anon mapping, equivalent to tmpfs. One weakness of Shaohua's vma/anon_vma approach was that it did not optimize Shmem: seen below. Konstantin's approach was perhaps mistuned, 50% slower on Seq: did not compete and is not shown below. HDD Vanilla Shaohua HughOld HughNew HughPC4 HughPC0 Seq Anon 73921 76210 75611 76904 78191 121542 Seq Shmem 73601 73176 73855 72947 74543 118322 Rand Anon 895392 831243 871569 845197 846496 841680 Rand Shmem 1058375 1053486 827935 764955 764376 756489 SSD Vanilla Shaohua HughOld HughNew HughPC4 HughPC0 Seq Anon 24634 24198 24673 25107 21614 70018 Seq Shmem 24959 24932 25052 25703 22030 69678 Rand Anon 43014 26146 28075 25989 26935 25901 Rand Shmem 45349 45215 28249 24268 24138 24332 These tests are, of course, two extremes of a very simple case: under heavier mixed loads I've not yet observed any consistent improvement or degradation, and wider testing would be welcome. Shaohua Li: Test shows Vanilla is slightly better in sequential workload than Hugh's patch. I observed with Hugh's patch sometimes the readahead size is shrinked too fast (from 8 to 1 immediately) in sequential workload if there is no hit. And in such case, continuing doing readahead is good actually. I don't prepare a sophisticated algorithm for the sequential workload because so far we can't guarantee sequential accessed pages are swap out sequentially. So I slightly change Hugh's heuristic - don't shrink readahead size too fast. Here is my test result (unit second, 3 runs average): Vanilla Hugh New Seq 356 370 360 Random 4525 2447 2444 Attached graph is the swapin/swapout throughput I collected with 'vmstat 2'. The first part is running a random workload (till around 1200 of the x-axis) and the second part is running a sequential workload. swapin and swapout throughput are almost identical in steady state in both workloads. These are expected behavior. while in Vanilla, swapin is much bigger than swapout especially in random workload (because wrong readahead). Original patches by: Shaohua Li and Konstantin Khlebnikov. [fengguang.wu@intel.com: swapin_nr_pages() can be static] Signed-off-by:
Shaohua Li <shli@fusionio.com> Signed-off-by:
Fengguang Wu <fengguang.wu@intel.com> Cc: Rik van Riel <riel@redhat.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Signed-off-by:
-
Vlastimil Babka authored
commit 55b7c4c9 upstream. Compaction used to start its migrate and free page scaners at the zone's lowest and highest pfn, respectively. Later, caching was introduced to remember the scanners' progress across compaction attempts so that pageblocks are not re-scanned uselessly. Additionally, pageblocks where isolation failed are marked to be quickly skipped when encountered again in future compactions. Currently, both the reset of cached pfn's and clearing of the pageblock skip information for a zone is done in __reset_isolation_suitable(). This function gets called when: - compaction is restarting after being deferred - compact_blockskip_flush flag is set in compact_finished() when the scanners meet (and not again cleared when direct compaction succeeds in allocation) and kswapd acts upon this flag before going to sleep This behavior is suboptimal for several reasons: - when direct sync compaction is called after async compaction fails (in the allocation slowpath), it will effectively do nothing, unless kswapd happens to process the compact_blockskip_flush flag meanwhile. This is racy and goes against the purpose of sync compaction to more thoroughly retry the compaction of a zone where async compaction has failed. The restart-after-deferring path cannot help here as deferring happens only after the sync compaction fails. It is also done only for the preferred zone, while the compaction might be done for a fallback zone. - the mechanism of marking pageblock to be skipped has little value since the cached pfn's are reset only together with the pageblock skip flags. This effectively limits pageblock skip usage to parallel compactions. This patch changes compact_finished() so that cached pfn's are reset immediately when the scanners meet. Clearing pageblock skip flags is unchanged, as well as the other situations where cached pfn's are reset. This allows the sync-after-async compaction to retry pageblocks not marked as skipped, such as blocks !MIGRATE_MOVABLE blocks that async compactions now skips without marking them. Signed-off-by:
-
Vlastimil Babka authored
commit 50b5b094 upstream. Compaction temporarily marks pageblocks where it fails to isolate pages as to-be-skipped in further compactions, in order to improve efficiency. One of the reasons to fail isolating pages is that isolation is not attempted in pageblocks that are not of MIGRATE_MOVABLE (or CMA) type. The problem is that blocks skipped due to not being MIGRATE_MOVABLE in async compaction become skipped due to the temporary mark also in future sync compaction. Moreover, this may follow quite soon during __alloc_page_slowpath, without much time for kswapd to clear the pageblock skip marks. This goes against the idea that sync compaction should try to scan these blocks more thoroughly than the async compaction. The fix is to ensure in async compaction that these !MIGRATE_MOVABLE blocks are not marked to be skipped. Note this should not affect performance or locking impact of further async compactions, as skipping a block due to being !MIGRATE_MOVABLE is done soon after skipping a block marked to be skipped, both without locking. Signed-off-by:
-
Vlastimil Babka authored
commit de6c60a6 upstream. Currently there are several functions to manipulate the deferred compaction state variables. The remaining case where the variables are touched directly is when a successful allocation occurs in direct compaction, or is expected to be successful in the future by kswapd. Here, the lowest order that is expected to fail is updated, and in the case of successful allocation, the deferred status and counter is reset completely. Create a new function compaction_defer_reset() to encapsulate this functionality and make it easier to understand the code. No functional change. Signed-off-by:
-
Mel Gorman authored
commit 0eb927c0 upstream. The broad goal of the series is to improve allocation success rates for huge pages through memory compaction, while trying not to increase the compaction overhead. The original objective was to reintroduce capturing of high-order pages freed by the compaction, before they are split by concurrent activity. However, several bugs and opportunities for simple improvements were found in the current implementation, mostly through extra tracepoints (which are however too ugly for now to be considered for sending). The patches mostly deal with two mechanisms that reduce compaction overhead, which is caching the progress of migrate and free scanners, and marking pageblocks where isolation failed to be skipped during further scans. Patch 1 (from mgorman) adds tracepoints that allow calculate time spent in compaction and potentially debug scanner pfn values. Patch 2 encapsulates the some functionality for handling deferred compactions for better maintainability, without a functional change type is not determined without being actually needed. Patch 3 fixes a bug where cached scanner pfn's are sometimes reset only after they have been read to initialize a compaction run. Patch 4 fixes a bug where scanners meeting is sometimes not properly detected and can lead to multiple compaction attempts quitting early without doing any work. Patch 5 improves the chances of sync compaction to process pageblocks that async compaction has skipped due to being !MIGRATE_MOVABLE. Patch 6 improves the chances of sync direct compaction to actually do anything when called after async compaction fails during allocation slowpath. The impact of patches were validated using mmtests's stress-highalloc benchmark with mmtests's stress-highalloc benchmark on a x86_64 machine with 4GB memory. Due to instability of the results (mostly related to the bugs fixed by patches 2 and 3), 10 iterations were performed, taking min,mean,max values for success rates and mean values for time and vmstat-based metrics. First, the default GFP_HIGHUSER_MOVABLE allocations were tested with the patches stacked on top of v3.13-rc2. Patch 2 is OK to serve as baseline due to no functional changes in 1 and 2. Comments below. stress-highalloc 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 2-nothp 3-nothp 4-nothp 5-nothp 6-nothp Success 1 Min 9.00 ( 0.00%) 10.00 (-11.11%) 43.00 (-377.78%) 43.00 (-377.78%) 33.00 (-266.67%) Success 1 Mean 27.50 ( 0.00%) 25.30 ( 8.00%) 45.50 (-65.45%) 45.90 (-66.91%) 46.30 (-68.36%) Success 1 Max 36.00 ( 0.00%) 36.00 ( 0.00%) 47.00 (-30.56%) 48.00 (-33.33%) 52.00 (-44.44%) Success 2 Min 10.00 ( 0.00%) 8.00 ( 20.00%) 46.00 (-360.00%) 45.00 (-350.00%) 35.00 (-250.00%) Success 2 Mean 26.40 ( 0.00%) 23.50 ( 10.98%) 47.30 (-79.17%) 47.60 (-80.30%) 48.10 (-82.20%) Success 2 Max 34.00 ( 0.00%) 33.00 ( 2.94%) 48.00 (-41.18%) 50.00 (-47.06%) 54.00 (-58.82%) Success 3 Min 65.00 ( 0.00%) 63.00 ( 3.08%) 85.00 (-30.77%) 84.00 (-29.23%) 85.00 (-30.77%) Success 3 Mean 76.70 ( 0.00%) 70.50 ( 8.08%) 86.20 (-12.39%) 85.50 (-11.47%) 86.00 (-12.13%) Success 3 Max 87.00 ( 0.00%) 86.00 ( 1.15%) 88.00 ( -1.15%) 87.00 ( 0.00%) 87.00 ( 0.00%) 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 2-nothp 3-nothp 4-nothp 5-nothp 6-nothp User 6437.72 6459.76 5960.32 5974.55 6019.67 System 1049.65 1049.09 1029.32 1031.47 1032.31 Elapsed 1856.77 1874.48 1949.97 1994.22 1983.15 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 2-nothp 3-nothp 4-nothp 5-nothp 6-nothp Minor Faults 253952267 254581900 250030122 250507333 250157829 Major Faults 420 407 506 530 530 Swap Ins 4 9 9 6 6 Swap Outs 398 375 345 346 333 Direct pages scanned 197538 189017 298574 287019 299063 Kswapd pages scanned 1809843 1801308 1846674 1873184 1861089 Kswapd pages reclaimed 1806972 1798684 1844219 1870509 1858622 Direct pages reclaimed 197227 188829 298380 286822 298835 Kswapd efficiency 99% 99% 99% 99% 99% Kswapd velocity 953.382 970.449 952.243 934.569 922.286 Direct efficiency 99% 99% 99% 99% 99% Direct velocity 104.058 101.832 153.961 143.200 148.205 Percentage direct scans 9% 9% 13% 13% 13% Zone normal velocity 347.289 359.676 348.063 339.933 332.983 Zone dma32 velocity 710.151 712.605 758.140 737.835 737.507 Zone dma velocity 0.000 0.000 0.000 0.000 0.000 Page writes by reclaim 557.600 429.000 353.600 426.400 381.800 Page writes file 159 53 7 79 48 Page writes anon 398 375 345 346 333 Page reclaim immediate 825 644 411 575 420 Sector Reads 2781750 2769780 2878547 2939128 2910483 Sector Writes 12080843 12083351 12012892 12002132 12010745 Page rescued immediate 0 0 0 0 0 Slabs scanned 1575654 1545344 1778406 1786700 1794073 Direct inode steals 9657 10037 15795 14104 14645 Kswapd inode steals 46857 46335 50543 50716 51796 Kswapd skipped wait 0 0 0 0 0 THP fault alloc 97 91 81 71 77 THP collapse alloc 456 506 546 544 565 THP splits 6 5 5 4 4 THP fault fallback 0 1 0 0 0 THP collapse fail 14 14 12 13 12 Compaction stalls 1006 980 1537 1536 1548 Compaction success 303 284 562 559 578 Compaction failures 702 696 974 976 969 Page migrate success 1177325 1070077 3927538 3781870 3877057 Page migrate failure 0 0 0 0 0 Compaction pages isolated 2547248 2306457 8301218 8008500 8200674 Compaction migrate scanned 42290478 38832618 153961130 154143900 159141197 Compaction free scanned 89199429 79189151 356529027 351943166 356326727 Compaction cost 1566 1426 5312 5156 5294 NUMA PTE updates 0 0 0 0 0 NUMA hint faults 0 0 0 0 0 NUMA hint local faults 0 0 0 0 0 NUMA hint local percent 100 100 100 100 100 NUMA pages migrated 0 0 0 0 0 AutoNUMA cost 0 0 0 0 0 Observations: - The "Success 3" line is allocation success rate with system idle (phases 1 and 2 are with background interference). I used to get stable values around 85% with vanilla 3.11. The lower min and mean values came with 3.12. This was bisected to commit 81c0a2bb ("mm: page_alloc: fair zone allocator policy") As explained in comment for patch 3, I don't think the commit is wrong, but that it makes the effect of compaction bugs worse. From patch 3 onwards, the results are OK and match the 3.11 results. - Patch 4 also clearly helps phases 1 and 2, and exceeds any results I've seen with 3.11 (I didn't measure it that thoroughly then, but it was never above 40%). - Compaction cost and number of scanned pages is higher, especially due to patch 4. However, keep in mind that patches 3 and 4 fix existing bugs in the current design of compaction overhead mitigation, they do not change it. If overhead is found unacceptable, then it should be decreased differently (and consistently, not due to random conditions) than the current implementation does. In contrast, patches 5 and 6 (which are not strictly bug fixes) do not increase the overhead (but also not success rates). This might be a limitation of the stress-highalloc benchmark as it's quite uniform. Another set of results is when configuring stress-highalloc t allocate with similar flags as THP uses: (GFP_HIGHUSER_MOVABLE|__GFP_NOMEMALLOC|__GFP_NORETRY|__GFP_NO_KSWAPD) stress-highalloc 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 2-thp 3-thp 4-thp 5-thp 6-thp Success 1 Min 2.00 ( 0.00%) 7.00 (-250.00%) 18.00 (-800.00%) 19.00 (-850.00%) 26.00 (-1200.00%) Success 1 Mean 19.20 ( 0.00%) 17.80 ( 7.29%) 29.20 (-52.08%) 29.90 (-55.73%) 32.80 (-70.83%) Success 1 Max 27.00 ( 0.00%) 29.00 ( -7.41%) 35.00 (-29.63%) 36.00 (-33.33%) 37.00 (-37.04%) Success 2 Min 3.00 ( 0.00%) 8.00 (-166.67%) 21.00 (-600.00%) 21.00 (-600.00%) 32.00 (-966.67%) Success 2 Mean 19.30 ( 0.00%) 17.90 ( 7.25%) 32.20 (-66.84%) 32.60 (-68.91%) 35.70 (-84.97%) Success 2 Max 27.00 ( 0.00%) 30.00 (-11.11%) 36.00 (-33.33%) 37.00 (-37.04%) 39.00 (-44.44%) Success 3 Min 62.00 ( 0.00%) 62.00 ( 0.00%) 85.00 (-37.10%) 75.00 (-20.97%) 64.00 ( -3.23%) Success 3 Mean 66.30 ( 0.00%) 65.50 ( 1.21%) 85.60 (-29.11%) 83.40 (-25.79%) 83.50 (-25.94%) Success 3 Max 70.00 ( 0.00%) 69.00 ( 1.43%) 87.00 (-24.29%) 86.00 (-22.86%) 87.00 (-24.29%) 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 2-thp 3-thp 4-thp 5-thp 6-thp User 6547.93 6475.85 6265.54 6289.46 6189.96 System 1053.42 1047.28 1043.23 1042.73 1038.73 Elapsed 1835.43 1821.96 1908.67 1912.74 1956.38 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 2-thp 3-thp 4-thp 5-thp 6-thp Minor Faults 256805673 253106328 253222299 249830289 251184418 Major Faults 395 375 423 434 448 Swap Ins 12 10 10 12 9 Swap Outs 530 537 487 455 415 Direct pages scanned 71859 86046 153244 152764 190713 Kswapd pages scanned 1900994 1870240 1898012 1892864 1880520 Kswapd pages reclaimed 1897814 1867428 1894939 1890125 1877924 Direct pages reclaimed 71766 85908 153167 152643 190600 Kswapd efficiency 99% 99% 99% 99% 99% Kswapd velocity 1029.000 1067.782 1000.091 991.049 951.218 Direct efficiency 99% 99% 99% 99% 99% Direct velocity 38.897 49.127 80.747 79.983 96.468 Percentage direct scans 3% 4% 7% 7% 9% Zone normal velocity 351.377 372.494 348.910 341.689 335.310 Zone dma32 velocity 716.520 744.414 731.928 729.343 712.377 Zone dma velocity 0.000 0.000 0.000 0.000 0.000 Page writes by reclaim 669.300 604.000 545.700 538.900 429.900 Page writes file 138 66 58 83 14 Page writes anon 530 537 487 455 415 Page reclaim immediate 806 655 772 548 517 Sector Reads 2711956 2703239 2811602 2818248 2839459 Sector Writes 12163238 12018662 12038248 11954736 11994892 Page rescued immediate 0 0 0 0 0 Slabs scanned 1385088 1388364 1507968 1513292 1558656 Direct inode steals 1739 2564 4622 5496 6007 Kswapd inode steals 47461 46406 47804 48013 48466 Kswapd skipped wait 0 0 0 0 0 THP fault alloc 110 82 84 69 70 THP collapse alloc 445 482 467 462 539 THP splits 6 5 4 5 3 THP fault fallback 3 0 0 0 0 THP collapse fail 15 14 14 14 13 Compaction stalls 659 685 1033 1073 1111 Compaction success 222 225 410 427 456 Compaction failures 436 460 622 646 655 Page migrate success 446594 439978 1085640 1095062 1131716 Page migrate failure 0 0 0 0 0 Compaction pages isolated 1029475 1013490 2453074 2482698 2565400 Compaction migrate scanned 9955461 11344259 24375202 27978356 30494204 Compaction free scanned 27715272 28544654 80150615 82898631 85756132 Compaction cost 552 555 1344 1379 1436 NUMA PTE updates 0 0 0 0 0 NUMA hint faults 0 0 0 0 0 NUMA hint local faults 0 0 0 0 0 NUMA hint local percent 100 100 100 100 100 NUMA pages migrated 0 0 0 0 0 AutoNUMA cost 0 0 0 0 0 There are some differences from the previous results for THP-like allocations: - Here, the bad result for unpatched kernel in phase 3 is much more consistent to be between 65-70% and not related to the "regression" in 3.12. Still there is the improvement from patch 4 onwards, which brings it on par with simple GFP_HIGHUSER_MOVABLE allocations. - Compaction costs have increased, but nowhere near as much as the non-THP case. Again, the patches should be worth the gained determininsm. - Patches 5 and 6 somewhat increase the number of migrate-scanned pages. This is most likely due to __GFP_NO_KSWAPD flag, which means the cached pfn's and pageblock skip bits are not reset by kswapd that often (at least in phase 3 where no concurrent activity would wake up kswapd) and the patches thus help the sync-after-async compaction. It doesn't however show that the sync compaction would help so much with success rates, which can be again seen as a limitation of the benchmark scenario. This patch (of 6): Add two tracepoints for compaction begin and end of a zone. Using this it is possible to calculate how much time a workload is spending within compaction and potentially debug problems related to cached pfns for scanning. In combination with the direct reclaim and slab trace points it should be possible to estimate most allocation-related overhead for a workload. Signed-off-by:
-
