4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
53 #include <linux/swapops.h>
54 #include <linux/balloon_compaction.h>
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/vmscan.h>
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
65 /* This context's GFP mask */
68 /* Allocation order */
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup *target_mem_cgroup;
83 /* Scan (total_size >> priority) pages at once */
86 unsigned int may_writepage:1;
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
94 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash:1;
97 unsigned int hibernation_mode:1;
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned;
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed;
109 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
111 #ifdef ARCH_HAS_PREFETCH
112 #define prefetch_prev_lru_page(_page, _base, _field) \
114 if ((_page)->lru.prev != _base) { \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetch(&prev->_field); \
122 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125 #ifdef ARCH_HAS_PREFETCHW
126 #define prefetchw_prev_lru_page(_page, _base, _field) \
128 if ((_page)->lru.prev != _base) { \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetchw(&prev->_field); \
136 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
140 * From 0 .. 100. Higher means more swappy.
142 int vm_swappiness = 60;
144 * The total number of pages which are beyond the high watermark within all
147 unsigned long vm_total_pages;
149 static LIST_HEAD(shrinker_list);
150 static DECLARE_RWSEM(shrinker_rwsem);
153 static bool global_reclaim(struct scan_control *sc)
155 return !sc->target_mem_cgroup;
158 static bool global_reclaim(struct scan_control *sc)
164 static unsigned long zone_reclaimable_pages(struct zone *zone)
168 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
169 zone_page_state(zone, NR_INACTIVE_FILE);
171 if (get_nr_swap_pages() > 0)
172 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
173 zone_page_state(zone, NR_INACTIVE_ANON);
178 bool zone_reclaimable(struct zone *zone)
180 return zone_page_state(zone, NR_PAGES_SCANNED) <
181 zone_reclaimable_pages(zone) * 6;
184 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
186 if (!mem_cgroup_disabled())
187 return mem_cgroup_get_lru_size(lruvec, lru);
189 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
193 * Add a shrinker callback to be called from the vm.
195 int register_shrinker(struct shrinker *shrinker)
197 size_t size = sizeof(*shrinker->nr_deferred);
200 * If we only have one possible node in the system anyway, save
201 * ourselves the trouble and disable NUMA aware behavior. This way we
202 * will save memory and some small loop time later.
204 if (nr_node_ids == 1)
205 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
207 if (shrinker->flags & SHRINKER_NUMA_AWARE)
210 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
211 if (!shrinker->nr_deferred)
214 down_write(&shrinker_rwsem);
215 list_add_tail(&shrinker->list, &shrinker_list);
216 up_write(&shrinker_rwsem);
219 EXPORT_SYMBOL(register_shrinker);
224 void unregister_shrinker(struct shrinker *shrinker)
226 down_write(&shrinker_rwsem);
227 list_del(&shrinker->list);
228 up_write(&shrinker_rwsem);
229 kfree(shrinker->nr_deferred);
231 EXPORT_SYMBOL(unregister_shrinker);
233 #define SHRINK_BATCH 128
235 static unsigned long shrink_slabs(struct shrink_control *shrinkctl,
236 struct shrinker *shrinker,
237 unsigned long nr_scanned,
238 unsigned long nr_eligible)
240 unsigned long freed = 0;
241 unsigned long long delta;
246 int nid = shrinkctl->nid;
247 long batch_size = shrinker->batch ? shrinker->batch
250 freeable = shrinker->count_objects(shrinker, shrinkctl);
255 * copy the current shrinker scan count into a local variable
256 * and zero it so that other concurrent shrinker invocations
257 * don't also do this scanning work.
259 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
262 delta = (4 * nr_scanned) / shrinker->seeks;
264 do_div(delta, nr_eligible + 1);
266 if (total_scan < 0) {
267 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
268 shrinker->scan_objects, total_scan);
269 total_scan = freeable;
273 * We need to avoid excessive windup on filesystem shrinkers
274 * due to large numbers of GFP_NOFS allocations causing the
275 * shrinkers to return -1 all the time. This results in a large
276 * nr being built up so when a shrink that can do some work
277 * comes along it empties the entire cache due to nr >>>
278 * freeable. This is bad for sustaining a working set in
281 * Hence only allow the shrinker to scan the entire cache when
282 * a large delta change is calculated directly.
284 if (delta < freeable / 4)
285 total_scan = min(total_scan, freeable / 2);
288 * Avoid risking looping forever due to too large nr value:
289 * never try to free more than twice the estimate number of
292 if (total_scan > freeable * 2)
293 total_scan = freeable * 2;
295 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
296 nr_scanned, nr_eligible,
297 freeable, delta, total_scan);
300 * Normally, we should not scan less than batch_size objects in one
301 * pass to avoid too frequent shrinker calls, but if the slab has less
302 * than batch_size objects in total and we are really tight on memory,
303 * we will try to reclaim all available objects, otherwise we can end
304 * up failing allocations although there are plenty of reclaimable
305 * objects spread over several slabs with usage less than the
308 * We detect the "tight on memory" situations by looking at the total
309 * number of objects we want to scan (total_scan). If it is greater
310 * than the total number of objects on slab (freeable), we must be
311 * scanning at high prio and therefore should try to reclaim as much as
314 while (total_scan >= batch_size ||
315 total_scan >= freeable) {
317 unsigned long nr_to_scan = min(batch_size, total_scan);
319 shrinkctl->nr_to_scan = nr_to_scan;
320 ret = shrinker->scan_objects(shrinker, shrinkctl);
321 if (ret == SHRINK_STOP)
325 count_vm_events(SLABS_SCANNED, nr_to_scan);
326 total_scan -= nr_to_scan;
332 * move the unused scan count back into the shrinker in a
333 * manner that handles concurrent updates. If we exhausted the
334 * scan, there is no need to do an update.
337 new_nr = atomic_long_add_return(total_scan,
338 &shrinker->nr_deferred[nid]);
340 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
342 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
347 * shrink_node_slabs - shrink slab caches of a given node
348 * @gfp_mask: allocation context
349 * @nid: node whose slab caches to target
350 * @nr_scanned: pressure numerator
351 * @nr_eligible: pressure denominator
353 * Call the shrink functions to age shrinkable caches.
355 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
356 * unaware shrinkers will receive a node id of 0 instead.
358 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
359 * the available objects should be scanned. Page reclaim for example
360 * passes the number of pages scanned and the number of pages on the
361 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
362 * when it encountered mapped pages. The ratio is further biased by
363 * the ->seeks setting of the shrink function, which indicates the
364 * cost to recreate an object relative to that of an LRU page.
366 * Returns the number of reclaimed slab objects.
368 unsigned long shrink_node_slabs(gfp_t gfp_mask, int nid,
369 unsigned long nr_scanned,
370 unsigned long nr_eligible)
372 struct shrinker *shrinker;
373 unsigned long freed = 0;
376 nr_scanned = SWAP_CLUSTER_MAX;
378 if (!down_read_trylock(&shrinker_rwsem)) {
380 * If we would return 0, our callers would understand that we
381 * have nothing else to shrink and give up trying. By returning
382 * 1 we keep it going and assume we'll be able to shrink next
389 list_for_each_entry(shrinker, &shrinker_list, list) {
390 struct shrink_control sc = {
391 .gfp_mask = gfp_mask,
395 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
398 freed += shrink_slabs(&sc, shrinker, nr_scanned, nr_eligible);
401 up_read(&shrinker_rwsem);
407 static inline int is_page_cache_freeable(struct page *page)
410 * A freeable page cache page is referenced only by the caller
411 * that isolated the page, the page cache radix tree and
412 * optional buffer heads at page->private.
414 return page_count(page) - page_has_private(page) == 2;
417 static int may_write_to_queue(struct backing_dev_info *bdi,
418 struct scan_control *sc)
420 if (current->flags & PF_SWAPWRITE)
422 if (!bdi_write_congested(bdi))
424 if (bdi == current->backing_dev_info)
430 * We detected a synchronous write error writing a page out. Probably
431 * -ENOSPC. We need to propagate that into the address_space for a subsequent
432 * fsync(), msync() or close().
434 * The tricky part is that after writepage we cannot touch the mapping: nothing
435 * prevents it from being freed up. But we have a ref on the page and once
436 * that page is locked, the mapping is pinned.
438 * We're allowed to run sleeping lock_page() here because we know the caller has
441 static void handle_write_error(struct address_space *mapping,
442 struct page *page, int error)
445 if (page_mapping(page) == mapping)
446 mapping_set_error(mapping, error);
450 /* possible outcome of pageout() */
452 /* failed to write page out, page is locked */
454 /* move page to the active list, page is locked */
456 /* page has been sent to the disk successfully, page is unlocked */
458 /* page is clean and locked */
463 * pageout is called by shrink_page_list() for each dirty page.
464 * Calls ->writepage().
466 static pageout_t pageout(struct page *page, struct address_space *mapping,
467 struct scan_control *sc)
470 * If the page is dirty, only perform writeback if that write
471 * will be non-blocking. To prevent this allocation from being
472 * stalled by pagecache activity. But note that there may be
473 * stalls if we need to run get_block(). We could test
474 * PagePrivate for that.
476 * If this process is currently in __generic_file_write_iter() against
477 * this page's queue, we can perform writeback even if that
480 * If the page is swapcache, write it back even if that would
481 * block, for some throttling. This happens by accident, because
482 * swap_backing_dev_info is bust: it doesn't reflect the
483 * congestion state of the swapdevs. Easy to fix, if needed.
485 if (!is_page_cache_freeable(page))
489 * Some data journaling orphaned pages can have
490 * page->mapping == NULL while being dirty with clean buffers.
492 if (page_has_private(page)) {
493 if (try_to_free_buffers(page)) {
494 ClearPageDirty(page);
495 pr_info("%s: orphaned page\n", __func__);
501 if (mapping->a_ops->writepage == NULL)
502 return PAGE_ACTIVATE;
503 if (!may_write_to_queue(mapping->backing_dev_info, sc))
506 if (clear_page_dirty_for_io(page)) {
508 struct writeback_control wbc = {
509 .sync_mode = WB_SYNC_NONE,
510 .nr_to_write = SWAP_CLUSTER_MAX,
512 .range_end = LLONG_MAX,
516 SetPageReclaim(page);
517 res = mapping->a_ops->writepage(page, &wbc);
519 handle_write_error(mapping, page, res);
520 if (res == AOP_WRITEPAGE_ACTIVATE) {
521 ClearPageReclaim(page);
522 return PAGE_ACTIVATE;
525 if (!PageWriteback(page)) {
526 /* synchronous write or broken a_ops? */
527 ClearPageReclaim(page);
529 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
530 inc_zone_page_state(page, NR_VMSCAN_WRITE);
538 * Same as remove_mapping, but if the page is removed from the mapping, it
539 * gets returned with a refcount of 0.
541 static int __remove_mapping(struct address_space *mapping, struct page *page,
544 BUG_ON(!PageLocked(page));
545 BUG_ON(mapping != page_mapping(page));
547 spin_lock_irq(&mapping->tree_lock);
549 * The non racy check for a busy page.
551 * Must be careful with the order of the tests. When someone has
552 * a ref to the page, it may be possible that they dirty it then
553 * drop the reference. So if PageDirty is tested before page_count
554 * here, then the following race may occur:
556 * get_user_pages(&page);
557 * [user mapping goes away]
559 * !PageDirty(page) [good]
560 * SetPageDirty(page);
562 * !page_count(page) [good, discard it]
564 * [oops, our write_to data is lost]
566 * Reversing the order of the tests ensures such a situation cannot
567 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
568 * load is not satisfied before that of page->_count.
570 * Note that if SetPageDirty is always performed via set_page_dirty,
571 * and thus under tree_lock, then this ordering is not required.
573 if (!page_freeze_refs(page, 2))
575 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
576 if (unlikely(PageDirty(page))) {
577 page_unfreeze_refs(page, 2);
581 if (PageSwapCache(page)) {
582 swp_entry_t swap = { .val = page_private(page) };
583 mem_cgroup_swapout(page, swap);
584 __delete_from_swap_cache(page);
585 spin_unlock_irq(&mapping->tree_lock);
586 swapcache_free(swap);
588 void (*freepage)(struct page *);
591 freepage = mapping->a_ops->freepage;
593 * Remember a shadow entry for reclaimed file cache in
594 * order to detect refaults, thus thrashing, later on.
596 * But don't store shadows in an address space that is
597 * already exiting. This is not just an optizimation,
598 * inode reclaim needs to empty out the radix tree or
599 * the nodes are lost. Don't plant shadows behind its
602 if (reclaimed && page_is_file_cache(page) &&
603 !mapping_exiting(mapping))
604 shadow = workingset_eviction(mapping, page);
605 __delete_from_page_cache(page, shadow);
606 spin_unlock_irq(&mapping->tree_lock);
608 if (freepage != NULL)
615 spin_unlock_irq(&mapping->tree_lock);
620 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
621 * someone else has a ref on the page, abort and return 0. If it was
622 * successfully detached, return 1. Assumes the caller has a single ref on
625 int remove_mapping(struct address_space *mapping, struct page *page)
627 if (__remove_mapping(mapping, page, false)) {
629 * Unfreezing the refcount with 1 rather than 2 effectively
630 * drops the pagecache ref for us without requiring another
633 page_unfreeze_refs(page, 1);
640 * putback_lru_page - put previously isolated page onto appropriate LRU list
641 * @page: page to be put back to appropriate lru list
643 * Add previously isolated @page to appropriate LRU list.
644 * Page may still be unevictable for other reasons.
646 * lru_lock must not be held, interrupts must be enabled.
648 void putback_lru_page(struct page *page)
651 int was_unevictable = PageUnevictable(page);
653 VM_BUG_ON_PAGE(PageLRU(page), page);
656 ClearPageUnevictable(page);
658 if (page_evictable(page)) {
660 * For evictable pages, we can use the cache.
661 * In event of a race, worst case is we end up with an
662 * unevictable page on [in]active list.
663 * We know how to handle that.
665 is_unevictable = false;
669 * Put unevictable pages directly on zone's unevictable
672 is_unevictable = true;
673 add_page_to_unevictable_list(page);
675 * When racing with an mlock or AS_UNEVICTABLE clearing
676 * (page is unlocked) make sure that if the other thread
677 * does not observe our setting of PG_lru and fails
678 * isolation/check_move_unevictable_pages,
679 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
680 * the page back to the evictable list.
682 * The other side is TestClearPageMlocked() or shmem_lock().
688 * page's status can change while we move it among lru. If an evictable
689 * page is on unevictable list, it never be freed. To avoid that,
690 * check after we added it to the list, again.
692 if (is_unevictable && page_evictable(page)) {
693 if (!isolate_lru_page(page)) {
697 /* This means someone else dropped this page from LRU
698 * So, it will be freed or putback to LRU again. There is
699 * nothing to do here.
703 if (was_unevictable && !is_unevictable)
704 count_vm_event(UNEVICTABLE_PGRESCUED);
705 else if (!was_unevictable && is_unevictable)
706 count_vm_event(UNEVICTABLE_PGCULLED);
708 put_page(page); /* drop ref from isolate */
711 enum page_references {
713 PAGEREF_RECLAIM_CLEAN,
718 static enum page_references page_check_references(struct page *page,
719 struct scan_control *sc)
721 int referenced_ptes, referenced_page;
722 unsigned long vm_flags;
724 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
726 referenced_page = TestClearPageReferenced(page);
729 * Mlock lost the isolation race with us. Let try_to_unmap()
730 * move the page to the unevictable list.
732 if (vm_flags & VM_LOCKED)
733 return PAGEREF_RECLAIM;
735 if (referenced_ptes) {
736 if (PageSwapBacked(page))
737 return PAGEREF_ACTIVATE;
739 * All mapped pages start out with page table
740 * references from the instantiating fault, so we need
741 * to look twice if a mapped file page is used more
744 * Mark it and spare it for another trip around the
745 * inactive list. Another page table reference will
746 * lead to its activation.
748 * Note: the mark is set for activated pages as well
749 * so that recently deactivated but used pages are
752 SetPageReferenced(page);
754 if (referenced_page || referenced_ptes > 1)
755 return PAGEREF_ACTIVATE;
758 * Activate file-backed executable pages after first usage.
760 if (vm_flags & VM_EXEC)
761 return PAGEREF_ACTIVATE;
766 /* Reclaim if clean, defer dirty pages to writeback */
767 if (referenced_page && !PageSwapBacked(page))
768 return PAGEREF_RECLAIM_CLEAN;
770 return PAGEREF_RECLAIM;
773 /* Check if a page is dirty or under writeback */
774 static void page_check_dirty_writeback(struct page *page,
775 bool *dirty, bool *writeback)
777 struct address_space *mapping;
780 * Anonymous pages are not handled by flushers and must be written
781 * from reclaim context. Do not stall reclaim based on them
783 if (!page_is_file_cache(page)) {
789 /* By default assume that the page flags are accurate */
790 *dirty = PageDirty(page);
791 *writeback = PageWriteback(page);
793 /* Verify dirty/writeback state if the filesystem supports it */
794 if (!page_has_private(page))
797 mapping = page_mapping(page);
798 if (mapping && mapping->a_ops->is_dirty_writeback)
799 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
803 * shrink_page_list() returns the number of reclaimed pages
805 static unsigned long shrink_page_list(struct list_head *page_list,
807 struct scan_control *sc,
808 enum ttu_flags ttu_flags,
809 unsigned long *ret_nr_dirty,
810 unsigned long *ret_nr_unqueued_dirty,
811 unsigned long *ret_nr_congested,
812 unsigned long *ret_nr_writeback,
813 unsigned long *ret_nr_immediate,
816 LIST_HEAD(ret_pages);
817 LIST_HEAD(free_pages);
819 unsigned long nr_unqueued_dirty = 0;
820 unsigned long nr_dirty = 0;
821 unsigned long nr_congested = 0;
822 unsigned long nr_reclaimed = 0;
823 unsigned long nr_writeback = 0;
824 unsigned long nr_immediate = 0;
828 while (!list_empty(page_list)) {
829 struct address_space *mapping;
832 enum page_references references = PAGEREF_RECLAIM_CLEAN;
833 bool dirty, writeback;
837 page = lru_to_page(page_list);
838 list_del(&page->lru);
840 if (!trylock_page(page))
843 VM_BUG_ON_PAGE(PageActive(page), page);
844 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
848 if (unlikely(!page_evictable(page)))
851 if (!sc->may_unmap && page_mapped(page))
854 /* Double the slab pressure for mapped and swapcache pages */
855 if (page_mapped(page) || PageSwapCache(page))
858 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
859 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
862 * The number of dirty pages determines if a zone is marked
863 * reclaim_congested which affects wait_iff_congested. kswapd
864 * will stall and start writing pages if the tail of the LRU
865 * is all dirty unqueued pages.
867 page_check_dirty_writeback(page, &dirty, &writeback);
868 if (dirty || writeback)
871 if (dirty && !writeback)
875 * Treat this page as congested if the underlying BDI is or if
876 * pages are cycling through the LRU so quickly that the
877 * pages marked for immediate reclaim are making it to the
878 * end of the LRU a second time.
880 mapping = page_mapping(page);
881 if (((dirty || writeback) && mapping &&
882 bdi_write_congested(mapping->backing_dev_info)) ||
883 (writeback && PageReclaim(page)))
887 * If a page at the tail of the LRU is under writeback, there
888 * are three cases to consider.
890 * 1) If reclaim is encountering an excessive number of pages
891 * under writeback and this page is both under writeback and
892 * PageReclaim then it indicates that pages are being queued
893 * for IO but are being recycled through the LRU before the
894 * IO can complete. Waiting on the page itself risks an
895 * indefinite stall if it is impossible to writeback the
896 * page due to IO error or disconnected storage so instead
897 * note that the LRU is being scanned too quickly and the
898 * caller can stall after page list has been processed.
900 * 2) Global reclaim encounters a page, memcg encounters a
901 * page that is not marked for immediate reclaim or
902 * the caller does not have __GFP_IO. In this case mark
903 * the page for immediate reclaim and continue scanning.
905 * __GFP_IO is checked because a loop driver thread might
906 * enter reclaim, and deadlock if it waits on a page for
907 * which it is needed to do the write (loop masks off
908 * __GFP_IO|__GFP_FS for this reason); but more thought
909 * would probably show more reasons.
911 * Don't require __GFP_FS, since we're not going into the
912 * FS, just waiting on its writeback completion. Worryingly,
913 * ext4 gfs2 and xfs allocate pages with
914 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
915 * may_enter_fs here is liable to OOM on them.
917 * 3) memcg encounters a page that is not already marked
918 * PageReclaim. memcg does not have any dirty pages
919 * throttling so we could easily OOM just because too many
920 * pages are in writeback and there is nothing else to
921 * reclaim. Wait for the writeback to complete.
923 if (PageWriteback(page)) {
925 if (current_is_kswapd() &&
927 test_bit(ZONE_WRITEBACK, &zone->flags)) {
932 } else if (global_reclaim(sc) ||
933 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
935 * This is slightly racy - end_page_writeback()
936 * might have just cleared PageReclaim, then
937 * setting PageReclaim here end up interpreted
938 * as PageReadahead - but that does not matter
939 * enough to care. What we do want is for this
940 * page to have PageReclaim set next time memcg
941 * reclaim reaches the tests above, so it will
942 * then wait_on_page_writeback() to avoid OOM;
943 * and it's also appropriate in global reclaim.
945 SetPageReclaim(page);
952 wait_on_page_writeback(page);
957 references = page_check_references(page, sc);
959 switch (references) {
960 case PAGEREF_ACTIVATE:
961 goto activate_locked;
964 case PAGEREF_RECLAIM:
965 case PAGEREF_RECLAIM_CLEAN:
966 ; /* try to reclaim the page below */
970 * Anonymous process memory has backing store?
971 * Try to allocate it some swap space here.
973 if (PageAnon(page) && !PageSwapCache(page)) {
974 if (!(sc->gfp_mask & __GFP_IO))
976 if (!add_to_swap(page, page_list))
977 goto activate_locked;
980 /* Adding to swap updated mapping */
981 mapping = page_mapping(page);
985 * The page is mapped into the page tables of one or more
986 * processes. Try to unmap it here.
988 if (page_mapped(page) && mapping) {
989 switch (try_to_unmap(page, ttu_flags)) {
991 goto activate_locked;
997 ; /* try to free the page below */
1001 if (PageDirty(page)) {
1003 * Only kswapd can writeback filesystem pages to
1004 * avoid risk of stack overflow but only writeback
1005 * if many dirty pages have been encountered.
1007 if (page_is_file_cache(page) &&
1008 (!current_is_kswapd() ||
1009 !test_bit(ZONE_DIRTY, &zone->flags))) {
1011 * Immediately reclaim when written back.
1012 * Similar in principal to deactivate_page()
1013 * except we already have the page isolated
1014 * and know it's dirty
1016 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1017 SetPageReclaim(page);
1022 if (references == PAGEREF_RECLAIM_CLEAN)
1026 if (!sc->may_writepage)
1029 /* Page is dirty, try to write it out here */
1030 switch (pageout(page, mapping, sc)) {
1034 goto activate_locked;
1036 if (PageWriteback(page))
1038 if (PageDirty(page))
1042 * A synchronous write - probably a ramdisk. Go
1043 * ahead and try to reclaim the page.
1045 if (!trylock_page(page))
1047 if (PageDirty(page) || PageWriteback(page))
1049 mapping = page_mapping(page);
1051 ; /* try to free the page below */
1056 * If the page has buffers, try to free the buffer mappings
1057 * associated with this page. If we succeed we try to free
1060 * We do this even if the page is PageDirty().
1061 * try_to_release_page() does not perform I/O, but it is
1062 * possible for a page to have PageDirty set, but it is actually
1063 * clean (all its buffers are clean). This happens if the
1064 * buffers were written out directly, with submit_bh(). ext3
1065 * will do this, as well as the blockdev mapping.
1066 * try_to_release_page() will discover that cleanness and will
1067 * drop the buffers and mark the page clean - it can be freed.
1069 * Rarely, pages can have buffers and no ->mapping. These are
1070 * the pages which were not successfully invalidated in
1071 * truncate_complete_page(). We try to drop those buffers here
1072 * and if that worked, and the page is no longer mapped into
1073 * process address space (page_count == 1) it can be freed.
1074 * Otherwise, leave the page on the LRU so it is swappable.
1076 if (page_has_private(page)) {
1077 if (!try_to_release_page(page, sc->gfp_mask))
1078 goto activate_locked;
1079 if (!mapping && page_count(page) == 1) {
1081 if (put_page_testzero(page))
1085 * rare race with speculative reference.
1086 * the speculative reference will free
1087 * this page shortly, so we may
1088 * increment nr_reclaimed here (and
1089 * leave it off the LRU).
1097 if (!mapping || !__remove_mapping(mapping, page, true))
1101 * At this point, we have no other references and there is
1102 * no way to pick any more up (removed from LRU, removed
1103 * from pagecache). Can use non-atomic bitops now (and
1104 * we obviously don't have to worry about waking up a process
1105 * waiting on the page lock, because there are no references.
1107 __clear_page_locked(page);
1112 * Is there need to periodically free_page_list? It would
1113 * appear not as the counts should be low
1115 list_add(&page->lru, &free_pages);
1119 if (PageSwapCache(page))
1120 try_to_free_swap(page);
1122 putback_lru_page(page);
1126 /* Not a candidate for swapping, so reclaim swap space. */
1127 if (PageSwapCache(page) && vm_swap_full())
1128 try_to_free_swap(page);
1129 VM_BUG_ON_PAGE(PageActive(page), page);
1130 SetPageActive(page);
1135 list_add(&page->lru, &ret_pages);
1136 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1139 mem_cgroup_uncharge_list(&free_pages);
1140 free_hot_cold_page_list(&free_pages, true);
1142 list_splice(&ret_pages, page_list);
1143 count_vm_events(PGACTIVATE, pgactivate);
1145 *ret_nr_dirty += nr_dirty;
1146 *ret_nr_congested += nr_congested;
1147 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1148 *ret_nr_writeback += nr_writeback;
1149 *ret_nr_immediate += nr_immediate;
1150 return nr_reclaimed;
1153 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1154 struct list_head *page_list)
1156 struct scan_control sc = {
1157 .gfp_mask = GFP_KERNEL,
1158 .priority = DEF_PRIORITY,
1161 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1162 struct page *page, *next;
1163 LIST_HEAD(clean_pages);
1165 list_for_each_entry_safe(page, next, page_list, lru) {
1166 if (page_is_file_cache(page) && !PageDirty(page) &&
1167 !isolated_balloon_page(page)) {
1168 ClearPageActive(page);
1169 list_move(&page->lru, &clean_pages);
1173 ret = shrink_page_list(&clean_pages, zone, &sc,
1174 TTU_UNMAP|TTU_IGNORE_ACCESS,
1175 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1176 list_splice(&clean_pages, page_list);
1177 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1182 * Attempt to remove the specified page from its LRU. Only take this page
1183 * if it is of the appropriate PageActive status. Pages which are being
1184 * freed elsewhere are also ignored.
1186 * page: page to consider
1187 * mode: one of the LRU isolation modes defined above
1189 * returns 0 on success, -ve errno on failure.
1191 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1195 /* Only take pages on the LRU. */
1199 /* Compaction should not handle unevictable pages but CMA can do so */
1200 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1206 * To minimise LRU disruption, the caller can indicate that it only
1207 * wants to isolate pages it will be able to operate on without
1208 * blocking - clean pages for the most part.
1210 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1211 * is used by reclaim when it is cannot write to backing storage
1213 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1214 * that it is possible to migrate without blocking
1216 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1217 /* All the caller can do on PageWriteback is block */
1218 if (PageWriteback(page))
1221 if (PageDirty(page)) {
1222 struct address_space *mapping;
1224 /* ISOLATE_CLEAN means only clean pages */
1225 if (mode & ISOLATE_CLEAN)
1229 * Only pages without mappings or that have a
1230 * ->migratepage callback are possible to migrate
1233 mapping = page_mapping(page);
1234 if (mapping && !mapping->a_ops->migratepage)
1239 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1242 if (likely(get_page_unless_zero(page))) {
1244 * Be careful not to clear PageLRU until after we're
1245 * sure the page is not being freed elsewhere -- the
1246 * page release code relies on it.
1256 * zone->lru_lock is heavily contended. Some of the functions that
1257 * shrink the lists perform better by taking out a batch of pages
1258 * and working on them outside the LRU lock.
1260 * For pagecache intensive workloads, this function is the hottest
1261 * spot in the kernel (apart from copy_*_user functions).
1263 * Appropriate locks must be held before calling this function.
1265 * @nr_to_scan: The number of pages to look through on the list.
1266 * @lruvec: The LRU vector to pull pages from.
1267 * @dst: The temp list to put pages on to.
1268 * @nr_scanned: The number of pages that were scanned.
1269 * @sc: The scan_control struct for this reclaim session
1270 * @mode: One of the LRU isolation modes
1271 * @lru: LRU list id for isolating
1273 * returns how many pages were moved onto *@dst.
1275 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1276 struct lruvec *lruvec, struct list_head *dst,
1277 unsigned long *nr_scanned, struct scan_control *sc,
1278 isolate_mode_t mode, enum lru_list lru)
1280 struct list_head *src = &lruvec->lists[lru];
1281 unsigned long nr_taken = 0;
1284 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1288 page = lru_to_page(src);
1289 prefetchw_prev_lru_page(page, src, flags);
1291 VM_BUG_ON_PAGE(!PageLRU(page), page);
1293 switch (__isolate_lru_page(page, mode)) {
1295 nr_pages = hpage_nr_pages(page);
1296 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1297 list_move(&page->lru, dst);
1298 nr_taken += nr_pages;
1302 /* else it is being freed elsewhere */
1303 list_move(&page->lru, src);
1312 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1313 nr_taken, mode, is_file_lru(lru));
1318 * isolate_lru_page - tries to isolate a page from its LRU list
1319 * @page: page to isolate from its LRU list
1321 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1322 * vmstat statistic corresponding to whatever LRU list the page was on.
1324 * Returns 0 if the page was removed from an LRU list.
1325 * Returns -EBUSY if the page was not on an LRU list.
1327 * The returned page will have PageLRU() cleared. If it was found on
1328 * the active list, it will have PageActive set. If it was found on
1329 * the unevictable list, it will have the PageUnevictable bit set. That flag
1330 * may need to be cleared by the caller before letting the page go.
1332 * The vmstat statistic corresponding to the list on which the page was
1333 * found will be decremented.
1336 * (1) Must be called with an elevated refcount on the page. This is a
1337 * fundamentnal difference from isolate_lru_pages (which is called
1338 * without a stable reference).
1339 * (2) the lru_lock must not be held.
1340 * (3) interrupts must be enabled.
1342 int isolate_lru_page(struct page *page)
1346 VM_BUG_ON_PAGE(!page_count(page), page);
1348 if (PageLRU(page)) {
1349 struct zone *zone = page_zone(page);
1350 struct lruvec *lruvec;
1352 spin_lock_irq(&zone->lru_lock);
1353 lruvec = mem_cgroup_page_lruvec(page, zone);
1354 if (PageLRU(page)) {
1355 int lru = page_lru(page);
1358 del_page_from_lru_list(page, lruvec, lru);
1361 spin_unlock_irq(&zone->lru_lock);
1367 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1368 * then get resheduled. When there are massive number of tasks doing page
1369 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1370 * the LRU list will go small and be scanned faster than necessary, leading to
1371 * unnecessary swapping, thrashing and OOM.
1373 static int too_many_isolated(struct zone *zone, int file,
1374 struct scan_control *sc)
1376 unsigned long inactive, isolated;
1378 if (current_is_kswapd())
1381 if (!global_reclaim(sc))
1385 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1386 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1388 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1389 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1393 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1394 * won't get blocked by normal direct-reclaimers, forming a circular
1397 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1400 return isolated > inactive;
1403 static noinline_for_stack void
1404 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1406 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1407 struct zone *zone = lruvec_zone(lruvec);
1408 LIST_HEAD(pages_to_free);
1411 * Put back any unfreeable pages.
1413 while (!list_empty(page_list)) {
1414 struct page *page = lru_to_page(page_list);
1417 VM_BUG_ON_PAGE(PageLRU(page), page);
1418 list_del(&page->lru);
1419 if (unlikely(!page_evictable(page))) {
1420 spin_unlock_irq(&zone->lru_lock);
1421 putback_lru_page(page);
1422 spin_lock_irq(&zone->lru_lock);
1426 lruvec = mem_cgroup_page_lruvec(page, zone);
1429 lru = page_lru(page);
1430 add_page_to_lru_list(page, lruvec, lru);
1432 if (is_active_lru(lru)) {
1433 int file = is_file_lru(lru);
1434 int numpages = hpage_nr_pages(page);
1435 reclaim_stat->recent_rotated[file] += numpages;
1437 if (put_page_testzero(page)) {
1438 __ClearPageLRU(page);
1439 __ClearPageActive(page);
1440 del_page_from_lru_list(page, lruvec, lru);
1442 if (unlikely(PageCompound(page))) {
1443 spin_unlock_irq(&zone->lru_lock);
1444 mem_cgroup_uncharge(page);
1445 (*get_compound_page_dtor(page))(page);
1446 spin_lock_irq(&zone->lru_lock);
1448 list_add(&page->lru, &pages_to_free);
1453 * To save our caller's stack, now use input list for pages to free.
1455 list_splice(&pages_to_free, page_list);
1459 * If a kernel thread (such as nfsd for loop-back mounts) services
1460 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1461 * In that case we should only throttle if the backing device it is
1462 * writing to is congested. In other cases it is safe to throttle.
1464 static int current_may_throttle(void)
1466 return !(current->flags & PF_LESS_THROTTLE) ||
1467 current->backing_dev_info == NULL ||
1468 bdi_write_congested(current->backing_dev_info);
1472 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1473 * of reclaimed pages
1475 static noinline_for_stack unsigned long
1476 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1477 struct scan_control *sc, enum lru_list lru)
1479 LIST_HEAD(page_list);
1480 unsigned long nr_scanned;
1481 unsigned long nr_reclaimed = 0;
1482 unsigned long nr_taken;
1483 unsigned long nr_dirty = 0;
1484 unsigned long nr_congested = 0;
1485 unsigned long nr_unqueued_dirty = 0;
1486 unsigned long nr_writeback = 0;
1487 unsigned long nr_immediate = 0;
1488 isolate_mode_t isolate_mode = 0;
1489 int file = is_file_lru(lru);
1490 struct zone *zone = lruvec_zone(lruvec);
1491 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1493 while (unlikely(too_many_isolated(zone, file, sc))) {
1494 congestion_wait(BLK_RW_ASYNC, HZ/10);
1496 /* We are about to die and free our memory. Return now. */
1497 if (fatal_signal_pending(current))
1498 return SWAP_CLUSTER_MAX;
1504 isolate_mode |= ISOLATE_UNMAPPED;
1505 if (!sc->may_writepage)
1506 isolate_mode |= ISOLATE_CLEAN;
1508 spin_lock_irq(&zone->lru_lock);
1510 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1511 &nr_scanned, sc, isolate_mode, lru);
1513 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1514 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1516 if (global_reclaim(sc)) {
1517 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1518 if (current_is_kswapd())
1519 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1521 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1523 spin_unlock_irq(&zone->lru_lock);
1528 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1529 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1530 &nr_writeback, &nr_immediate,
1533 spin_lock_irq(&zone->lru_lock);
1535 reclaim_stat->recent_scanned[file] += nr_taken;
1537 if (global_reclaim(sc)) {
1538 if (current_is_kswapd())
1539 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1542 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1546 putback_inactive_pages(lruvec, &page_list);
1548 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1550 spin_unlock_irq(&zone->lru_lock);
1552 mem_cgroup_uncharge_list(&page_list);
1553 free_hot_cold_page_list(&page_list, true);
1556 * If reclaim is isolating dirty pages under writeback, it implies
1557 * that the long-lived page allocation rate is exceeding the page
1558 * laundering rate. Either the global limits are not being effective
1559 * at throttling processes due to the page distribution throughout
1560 * zones or there is heavy usage of a slow backing device. The
1561 * only option is to throttle from reclaim context which is not ideal
1562 * as there is no guarantee the dirtying process is throttled in the
1563 * same way balance_dirty_pages() manages.
1565 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1566 * of pages under pages flagged for immediate reclaim and stall if any
1567 * are encountered in the nr_immediate check below.
1569 if (nr_writeback && nr_writeback == nr_taken)
1570 set_bit(ZONE_WRITEBACK, &zone->flags);
1573 * memcg will stall in page writeback so only consider forcibly
1574 * stalling for global reclaim
1576 if (global_reclaim(sc)) {
1578 * Tag a zone as congested if all the dirty pages scanned were
1579 * backed by a congested BDI and wait_iff_congested will stall.
1581 if (nr_dirty && nr_dirty == nr_congested)
1582 set_bit(ZONE_CONGESTED, &zone->flags);
1585 * If dirty pages are scanned that are not queued for IO, it
1586 * implies that flushers are not keeping up. In this case, flag
1587 * the zone ZONE_DIRTY and kswapd will start writing pages from
1590 if (nr_unqueued_dirty == nr_taken)
1591 set_bit(ZONE_DIRTY, &zone->flags);
1594 * If kswapd scans pages marked marked for immediate
1595 * reclaim and under writeback (nr_immediate), it implies
1596 * that pages are cycling through the LRU faster than
1597 * they are written so also forcibly stall.
1599 if (nr_immediate && current_may_throttle())
1600 congestion_wait(BLK_RW_ASYNC, HZ/10);
1604 * Stall direct reclaim for IO completions if underlying BDIs or zone
1605 * is congested. Allow kswapd to continue until it starts encountering
1606 * unqueued dirty pages or cycling through the LRU too quickly.
1608 if (!sc->hibernation_mode && !current_is_kswapd() &&
1609 current_may_throttle())
1610 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1612 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1614 nr_scanned, nr_reclaimed,
1616 trace_shrink_flags(file));
1617 return nr_reclaimed;
1621 * This moves pages from the active list to the inactive list.
1623 * We move them the other way if the page is referenced by one or more
1624 * processes, from rmap.
1626 * If the pages are mostly unmapped, the processing is fast and it is
1627 * appropriate to hold zone->lru_lock across the whole operation. But if
1628 * the pages are mapped, the processing is slow (page_referenced()) so we
1629 * should drop zone->lru_lock around each page. It's impossible to balance
1630 * this, so instead we remove the pages from the LRU while processing them.
1631 * It is safe to rely on PG_active against the non-LRU pages in here because
1632 * nobody will play with that bit on a non-LRU page.
1634 * The downside is that we have to touch page->_count against each page.
1635 * But we had to alter page->flags anyway.
1638 static void move_active_pages_to_lru(struct lruvec *lruvec,
1639 struct list_head *list,
1640 struct list_head *pages_to_free,
1643 struct zone *zone = lruvec_zone(lruvec);
1644 unsigned long pgmoved = 0;
1648 while (!list_empty(list)) {
1649 page = lru_to_page(list);
1650 lruvec = mem_cgroup_page_lruvec(page, zone);
1652 VM_BUG_ON_PAGE(PageLRU(page), page);
1655 nr_pages = hpage_nr_pages(page);
1656 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1657 list_move(&page->lru, &lruvec->lists[lru]);
1658 pgmoved += nr_pages;
1660 if (put_page_testzero(page)) {
1661 __ClearPageLRU(page);
1662 __ClearPageActive(page);
1663 del_page_from_lru_list(page, lruvec, lru);
1665 if (unlikely(PageCompound(page))) {
1666 spin_unlock_irq(&zone->lru_lock);
1667 mem_cgroup_uncharge(page);
1668 (*get_compound_page_dtor(page))(page);
1669 spin_lock_irq(&zone->lru_lock);
1671 list_add(&page->lru, pages_to_free);
1674 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1675 if (!is_active_lru(lru))
1676 __count_vm_events(PGDEACTIVATE, pgmoved);
1679 static void shrink_active_list(unsigned long nr_to_scan,
1680 struct lruvec *lruvec,
1681 struct scan_control *sc,
1684 unsigned long nr_taken;
1685 unsigned long nr_scanned;
1686 unsigned long vm_flags;
1687 LIST_HEAD(l_hold); /* The pages which were snipped off */
1688 LIST_HEAD(l_active);
1689 LIST_HEAD(l_inactive);
1691 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1692 unsigned long nr_rotated = 0;
1693 isolate_mode_t isolate_mode = 0;
1694 int file = is_file_lru(lru);
1695 struct zone *zone = lruvec_zone(lruvec);
1700 isolate_mode |= ISOLATE_UNMAPPED;
1701 if (!sc->may_writepage)
1702 isolate_mode |= ISOLATE_CLEAN;
1704 spin_lock_irq(&zone->lru_lock);
1706 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1707 &nr_scanned, sc, isolate_mode, lru);
1708 if (global_reclaim(sc))
1709 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1711 reclaim_stat->recent_scanned[file] += nr_taken;
1713 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1714 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1715 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1716 spin_unlock_irq(&zone->lru_lock);
1718 while (!list_empty(&l_hold)) {
1720 page = lru_to_page(&l_hold);
1721 list_del(&page->lru);
1723 if (unlikely(!page_evictable(page))) {
1724 putback_lru_page(page);
1728 if (unlikely(buffer_heads_over_limit)) {
1729 if (page_has_private(page) && trylock_page(page)) {
1730 if (page_has_private(page))
1731 try_to_release_page(page, 0);
1736 if (page_referenced(page, 0, sc->target_mem_cgroup,
1738 nr_rotated += hpage_nr_pages(page);
1740 * Identify referenced, file-backed active pages and
1741 * give them one more trip around the active list. So
1742 * that executable code get better chances to stay in
1743 * memory under moderate memory pressure. Anon pages
1744 * are not likely to be evicted by use-once streaming
1745 * IO, plus JVM can create lots of anon VM_EXEC pages,
1746 * so we ignore them here.
1748 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1749 list_add(&page->lru, &l_active);
1754 ClearPageActive(page); /* we are de-activating */
1755 list_add(&page->lru, &l_inactive);
1759 * Move pages back to the lru list.
1761 spin_lock_irq(&zone->lru_lock);
1763 * Count referenced pages from currently used mappings as rotated,
1764 * even though only some of them are actually re-activated. This
1765 * helps balance scan pressure between file and anonymous pages in
1768 reclaim_stat->recent_rotated[file] += nr_rotated;
1770 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1771 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1772 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1773 spin_unlock_irq(&zone->lru_lock);
1775 mem_cgroup_uncharge_list(&l_hold);
1776 free_hot_cold_page_list(&l_hold, true);
1780 static int inactive_anon_is_low_global(struct zone *zone)
1782 unsigned long active, inactive;
1784 active = zone_page_state(zone, NR_ACTIVE_ANON);
1785 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1787 if (inactive * zone->inactive_ratio < active)
1794 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1795 * @lruvec: LRU vector to check
1797 * Returns true if the zone does not have enough inactive anon pages,
1798 * meaning some active anon pages need to be deactivated.
1800 static int inactive_anon_is_low(struct lruvec *lruvec)
1803 * If we don't have swap space, anonymous page deactivation
1806 if (!total_swap_pages)
1809 if (!mem_cgroup_disabled())
1810 return mem_cgroup_inactive_anon_is_low(lruvec);
1812 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1815 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1822 * inactive_file_is_low - check if file pages need to be deactivated
1823 * @lruvec: LRU vector to check
1825 * When the system is doing streaming IO, memory pressure here
1826 * ensures that active file pages get deactivated, until more
1827 * than half of the file pages are on the inactive list.
1829 * Once we get to that situation, protect the system's working
1830 * set from being evicted by disabling active file page aging.
1832 * This uses a different ratio than the anonymous pages, because
1833 * the page cache uses a use-once replacement algorithm.
1835 static int inactive_file_is_low(struct lruvec *lruvec)
1837 unsigned long inactive;
1838 unsigned long active;
1840 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1841 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1843 return active > inactive;
1846 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1848 if (is_file_lru(lru))
1849 return inactive_file_is_low(lruvec);
1851 return inactive_anon_is_low(lruvec);
1854 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1855 struct lruvec *lruvec, struct scan_control *sc)
1857 if (is_active_lru(lru)) {
1858 if (inactive_list_is_low(lruvec, lru))
1859 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1863 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1874 * Determine how aggressively the anon and file LRU lists should be
1875 * scanned. The relative value of each set of LRU lists is determined
1876 * by looking at the fraction of the pages scanned we did rotate back
1877 * onto the active list instead of evict.
1879 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1880 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1882 static void get_scan_count(struct lruvec *lruvec, int swappiness,
1883 struct scan_control *sc, unsigned long *nr,
1884 unsigned long *lru_pages)
1886 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1888 u64 denominator = 0; /* gcc */
1889 struct zone *zone = lruvec_zone(lruvec);
1890 unsigned long anon_prio, file_prio;
1891 enum scan_balance scan_balance;
1892 unsigned long anon, file;
1893 bool force_scan = false;
1894 unsigned long ap, fp;
1900 * If the zone or memcg is small, nr[l] can be 0. This
1901 * results in no scanning on this priority and a potential
1902 * priority drop. Global direct reclaim can go to the next
1903 * zone and tends to have no problems. Global kswapd is for
1904 * zone balancing and it needs to scan a minimum amount. When
1905 * reclaiming for a memcg, a priority drop can cause high
1906 * latencies, so it's better to scan a minimum amount there as
1909 if (current_is_kswapd()) {
1910 if (!zone_reclaimable(zone))
1912 if (!mem_cgroup_lruvec_online(lruvec))
1915 if (!global_reclaim(sc))
1918 /* If we have no swap space, do not bother scanning anon pages. */
1919 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1920 scan_balance = SCAN_FILE;
1925 * Global reclaim will swap to prevent OOM even with no
1926 * swappiness, but memcg users want to use this knob to
1927 * disable swapping for individual groups completely when
1928 * using the memory controller's swap limit feature would be
1931 if (!global_reclaim(sc) && !swappiness) {
1932 scan_balance = SCAN_FILE;
1937 * Do not apply any pressure balancing cleverness when the
1938 * system is close to OOM, scan both anon and file equally
1939 * (unless the swappiness setting disagrees with swapping).
1941 if (!sc->priority && swappiness) {
1942 scan_balance = SCAN_EQUAL;
1947 * Prevent the reclaimer from falling into the cache trap: as
1948 * cache pages start out inactive, every cache fault will tip
1949 * the scan balance towards the file LRU. And as the file LRU
1950 * shrinks, so does the window for rotation from references.
1951 * This means we have a runaway feedback loop where a tiny
1952 * thrashing file LRU becomes infinitely more attractive than
1953 * anon pages. Try to detect this based on file LRU size.
1955 if (global_reclaim(sc)) {
1956 unsigned long zonefile;
1957 unsigned long zonefree;
1959 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1960 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1961 zone_page_state(zone, NR_INACTIVE_FILE);
1963 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
1964 scan_balance = SCAN_ANON;
1970 * There is enough inactive page cache, do not reclaim
1971 * anything from the anonymous working set right now.
1973 if (!inactive_file_is_low(lruvec)) {
1974 scan_balance = SCAN_FILE;
1978 scan_balance = SCAN_FRACT;
1981 * With swappiness at 100, anonymous and file have the same priority.
1982 * This scanning priority is essentially the inverse of IO cost.
1984 anon_prio = swappiness;
1985 file_prio = 200 - anon_prio;
1988 * OK, so we have swap space and a fair amount of page cache
1989 * pages. We use the recently rotated / recently scanned
1990 * ratios to determine how valuable each cache is.
1992 * Because workloads change over time (and to avoid overflow)
1993 * we keep these statistics as a floating average, which ends
1994 * up weighing recent references more than old ones.
1996 * anon in [0], file in [1]
1999 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2000 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2001 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2002 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2004 spin_lock_irq(&zone->lru_lock);
2005 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2006 reclaim_stat->recent_scanned[0] /= 2;
2007 reclaim_stat->recent_rotated[0] /= 2;
2010 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2011 reclaim_stat->recent_scanned[1] /= 2;
2012 reclaim_stat->recent_rotated[1] /= 2;
2016 * The amount of pressure on anon vs file pages is inversely
2017 * proportional to the fraction of recently scanned pages on
2018 * each list that were recently referenced and in active use.
2020 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2021 ap /= reclaim_stat->recent_rotated[0] + 1;
2023 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2024 fp /= reclaim_stat->recent_rotated[1] + 1;
2025 spin_unlock_irq(&zone->lru_lock);
2029 denominator = ap + fp + 1;
2031 some_scanned = false;
2032 /* Only use force_scan on second pass. */
2033 for (pass = 0; !some_scanned && pass < 2; pass++) {
2035 for_each_evictable_lru(lru) {
2036 int file = is_file_lru(lru);
2040 size = get_lru_size(lruvec, lru);
2041 scan = size >> sc->priority;
2043 if (!scan && pass && force_scan)
2044 scan = min(size, SWAP_CLUSTER_MAX);
2046 switch (scan_balance) {
2048 /* Scan lists relative to size */
2052 * Scan types proportional to swappiness and
2053 * their relative recent reclaim efficiency.
2055 scan = div64_u64(scan * fraction[file],
2060 /* Scan one type exclusively */
2061 if ((scan_balance == SCAN_FILE) != file) {
2067 /* Look ma, no brain */
2075 * Skip the second pass and don't force_scan,
2076 * if we found something to scan.
2078 some_scanned |= !!scan;
2084 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2086 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2087 struct scan_control *sc, unsigned long *lru_pages)
2089 unsigned long nr[NR_LRU_LISTS];
2090 unsigned long targets[NR_LRU_LISTS];
2091 unsigned long nr_to_scan;
2093 unsigned long nr_reclaimed = 0;
2094 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2095 struct blk_plug plug;
2098 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2100 /* Record the original scan target for proportional adjustments later */
2101 memcpy(targets, nr, sizeof(nr));
2104 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2105 * event that can occur when there is little memory pressure e.g.
2106 * multiple streaming readers/writers. Hence, we do not abort scanning
2107 * when the requested number of pages are reclaimed when scanning at
2108 * DEF_PRIORITY on the assumption that the fact we are direct
2109 * reclaiming implies that kswapd is not keeping up and it is best to
2110 * do a batch of work at once. For memcg reclaim one check is made to
2111 * abort proportional reclaim if either the file or anon lru has already
2112 * dropped to zero at the first pass.
2114 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2115 sc->priority == DEF_PRIORITY);
2117 blk_start_plug(&plug);
2118 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2119 nr[LRU_INACTIVE_FILE]) {
2120 unsigned long nr_anon, nr_file, percentage;
2121 unsigned long nr_scanned;
2123 for_each_evictable_lru(lru) {
2125 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2126 nr[lru] -= nr_to_scan;
2128 nr_reclaimed += shrink_list(lru, nr_to_scan,
2133 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2137 * For kswapd and memcg, reclaim at least the number of pages
2138 * requested. Ensure that the anon and file LRUs are scanned
2139 * proportionally what was requested by get_scan_count(). We
2140 * stop reclaiming one LRU and reduce the amount scanning
2141 * proportional to the original scan target.
2143 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2144 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2147 * It's just vindictive to attack the larger once the smaller
2148 * has gone to zero. And given the way we stop scanning the
2149 * smaller below, this makes sure that we only make one nudge
2150 * towards proportionality once we've got nr_to_reclaim.
2152 if (!nr_file || !nr_anon)
2155 if (nr_file > nr_anon) {
2156 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2157 targets[LRU_ACTIVE_ANON] + 1;
2159 percentage = nr_anon * 100 / scan_target;
2161 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2162 targets[LRU_ACTIVE_FILE] + 1;
2164 percentage = nr_file * 100 / scan_target;
2167 /* Stop scanning the smaller of the LRU */
2169 nr[lru + LRU_ACTIVE] = 0;
2172 * Recalculate the other LRU scan count based on its original
2173 * scan target and the percentage scanning already complete
2175 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2176 nr_scanned = targets[lru] - nr[lru];
2177 nr[lru] = targets[lru] * (100 - percentage) / 100;
2178 nr[lru] -= min(nr[lru], nr_scanned);
2181 nr_scanned = targets[lru] - nr[lru];
2182 nr[lru] = targets[lru] * (100 - percentage) / 100;
2183 nr[lru] -= min(nr[lru], nr_scanned);
2185 scan_adjusted = true;
2187 blk_finish_plug(&plug);
2188 sc->nr_reclaimed += nr_reclaimed;
2191 * Even if we did not try to evict anon pages at all, we want to
2192 * rebalance the anon lru active/inactive ratio.
2194 if (inactive_anon_is_low(lruvec))
2195 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2196 sc, LRU_ACTIVE_ANON);
2198 throttle_vm_writeout(sc->gfp_mask);
2201 /* Use reclaim/compaction for costly allocs or under memory pressure */
2202 static bool in_reclaim_compaction(struct scan_control *sc)
2204 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2205 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2206 sc->priority < DEF_PRIORITY - 2))
2213 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2214 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2215 * true if more pages should be reclaimed such that when the page allocator
2216 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2217 * It will give up earlier than that if there is difficulty reclaiming pages.
2219 static inline bool should_continue_reclaim(struct zone *zone,
2220 unsigned long nr_reclaimed,
2221 unsigned long nr_scanned,
2222 struct scan_control *sc)
2224 unsigned long pages_for_compaction;
2225 unsigned long inactive_lru_pages;
2227 /* If not in reclaim/compaction mode, stop */
2228 if (!in_reclaim_compaction(sc))
2231 /* Consider stopping depending on scan and reclaim activity */
2232 if (sc->gfp_mask & __GFP_REPEAT) {
2234 * For __GFP_REPEAT allocations, stop reclaiming if the
2235 * full LRU list has been scanned and we are still failing
2236 * to reclaim pages. This full LRU scan is potentially
2237 * expensive but a __GFP_REPEAT caller really wants to succeed
2239 if (!nr_reclaimed && !nr_scanned)
2243 * For non-__GFP_REPEAT allocations which can presumably
2244 * fail without consequence, stop if we failed to reclaim
2245 * any pages from the last SWAP_CLUSTER_MAX number of
2246 * pages that were scanned. This will return to the
2247 * caller faster at the risk reclaim/compaction and
2248 * the resulting allocation attempt fails
2255 * If we have not reclaimed enough pages for compaction and the
2256 * inactive lists are large enough, continue reclaiming
2258 pages_for_compaction = (2UL << sc->order);
2259 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2260 if (get_nr_swap_pages() > 0)
2261 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2262 if (sc->nr_reclaimed < pages_for_compaction &&
2263 inactive_lru_pages > pages_for_compaction)
2266 /* If compaction would go ahead or the allocation would succeed, stop */
2267 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2268 case COMPACT_PARTIAL:
2269 case COMPACT_CONTINUE:
2276 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2279 unsigned long nr_reclaimed, nr_scanned;
2280 bool reclaimable = false;
2283 struct mem_cgroup *root = sc->target_mem_cgroup;
2284 struct mem_cgroup_reclaim_cookie reclaim = {
2286 .priority = sc->priority,
2288 unsigned long zone_lru_pages = 0;
2289 struct mem_cgroup *memcg;
2291 nr_reclaimed = sc->nr_reclaimed;
2292 nr_scanned = sc->nr_scanned;
2294 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2296 unsigned long lru_pages;
2297 struct lruvec *lruvec;
2300 if (mem_cgroup_low(root, memcg)) {
2301 if (!sc->may_thrash)
2303 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2306 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2307 swappiness = mem_cgroup_swappiness(memcg);
2309 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2310 zone_lru_pages += lru_pages;
2313 * Direct reclaim and kswapd have to scan all memory
2314 * cgroups to fulfill the overall scan target for the
2317 * Limit reclaim, on the other hand, only cares about
2318 * nr_to_reclaim pages to be reclaimed and it will
2319 * retry with decreasing priority if one round over the
2320 * whole hierarchy is not sufficient.
2322 if (!global_reclaim(sc) &&
2323 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2324 mem_cgroup_iter_break(root, memcg);
2327 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2330 * Shrink the slab caches in the same proportion that
2331 * the eligible LRU pages were scanned.
2333 if (global_reclaim(sc) && is_classzone) {
2334 struct reclaim_state *reclaim_state;
2336 shrink_node_slabs(sc->gfp_mask, zone_to_nid(zone),
2337 sc->nr_scanned - nr_scanned,
2340 reclaim_state = current->reclaim_state;
2341 if (reclaim_state) {
2343 reclaim_state->reclaimed_slab;
2344 reclaim_state->reclaimed_slab = 0;
2348 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2349 sc->nr_scanned - nr_scanned,
2350 sc->nr_reclaimed - nr_reclaimed);
2352 if (sc->nr_reclaimed - nr_reclaimed)
2355 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2356 sc->nr_scanned - nr_scanned, sc));
2362 * Returns true if compaction should go ahead for a high-order request, or
2363 * the high-order allocation would succeed without compaction.
2365 static inline bool compaction_ready(struct zone *zone, int order)
2367 unsigned long balance_gap, watermark;
2371 * Compaction takes time to run and there are potentially other
2372 * callers using the pages just freed. Continue reclaiming until
2373 * there is a buffer of free pages available to give compaction
2374 * a reasonable chance of completing and allocating the page
2376 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2377 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2378 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2379 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2382 * If compaction is deferred, reclaim up to a point where
2383 * compaction will have a chance of success when re-enabled
2385 if (compaction_deferred(zone, order))
2386 return watermark_ok;
2389 * If compaction is not ready to start and allocation is not likely
2390 * to succeed without it, then keep reclaiming.
2392 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2395 return watermark_ok;
2399 * This is the direct reclaim path, for page-allocating processes. We only
2400 * try to reclaim pages from zones which will satisfy the caller's allocation
2403 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2405 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2407 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2408 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2409 * zone defense algorithm.
2411 * If a zone is deemed to be full of pinned pages then just give it a light
2412 * scan then give up on it.
2414 * Returns true if a zone was reclaimable.
2416 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2420 unsigned long nr_soft_reclaimed;
2421 unsigned long nr_soft_scanned;
2423 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2424 bool reclaimable = false;
2427 * If the number of buffer_heads in the machine exceeds the maximum
2428 * allowed level, force direct reclaim to scan the highmem zone as
2429 * highmem pages could be pinning lowmem pages storing buffer_heads
2431 orig_mask = sc->gfp_mask;
2432 if (buffer_heads_over_limit)
2433 sc->gfp_mask |= __GFP_HIGHMEM;
2435 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2436 requested_highidx, sc->nodemask) {
2437 enum zone_type classzone_idx;
2439 if (!populated_zone(zone))
2442 classzone_idx = requested_highidx;
2443 while (!populated_zone(zone->zone_pgdat->node_zones +
2448 * Take care memory controller reclaiming has small influence
2451 if (global_reclaim(sc)) {
2452 if (!cpuset_zone_allowed(zone,
2453 GFP_KERNEL | __GFP_HARDWALL))
2456 if (sc->priority != DEF_PRIORITY &&
2457 !zone_reclaimable(zone))
2458 continue; /* Let kswapd poll it */
2461 * If we already have plenty of memory free for
2462 * compaction in this zone, don't free any more.
2463 * Even though compaction is invoked for any
2464 * non-zero order, only frequent costly order
2465 * reclamation is disruptive enough to become a
2466 * noticeable problem, like transparent huge
2469 if (IS_ENABLED(CONFIG_COMPACTION) &&
2470 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2471 zonelist_zone_idx(z) <= requested_highidx &&
2472 compaction_ready(zone, sc->order)) {
2473 sc->compaction_ready = true;
2478 * This steals pages from memory cgroups over softlimit
2479 * and returns the number of reclaimed pages and
2480 * scanned pages. This works for global memory pressure
2481 * and balancing, not for a memcg's limit.
2483 nr_soft_scanned = 0;
2484 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2485 sc->order, sc->gfp_mask,
2487 sc->nr_reclaimed += nr_soft_reclaimed;
2488 sc->nr_scanned += nr_soft_scanned;
2489 if (nr_soft_reclaimed)
2491 /* need some check for avoid more shrink_zone() */
2494 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2497 if (global_reclaim(sc) &&
2498 !reclaimable && zone_reclaimable(zone))
2503 * Restore to original mask to avoid the impact on the caller if we
2504 * promoted it to __GFP_HIGHMEM.
2506 sc->gfp_mask = orig_mask;
2512 * This is the main entry point to direct page reclaim.
2514 * If a full scan of the inactive list fails to free enough memory then we
2515 * are "out of memory" and something needs to be killed.
2517 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2518 * high - the zone may be full of dirty or under-writeback pages, which this
2519 * caller can't do much about. We kick the writeback threads and take explicit
2520 * naps in the hope that some of these pages can be written. But if the
2521 * allocating task holds filesystem locks which prevent writeout this might not
2522 * work, and the allocation attempt will fail.
2524 * returns: 0, if no pages reclaimed
2525 * else, the number of pages reclaimed
2527 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2528 struct scan_control *sc)
2530 int initial_priority = sc->priority;
2531 unsigned long total_scanned = 0;
2532 unsigned long writeback_threshold;
2533 bool zones_reclaimable;
2535 delayacct_freepages_start();
2537 if (global_reclaim(sc))
2538 count_vm_event(ALLOCSTALL);
2541 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2544 zones_reclaimable = shrink_zones(zonelist, sc);
2546 total_scanned += sc->nr_scanned;
2547 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2550 if (sc->compaction_ready)
2554 * If we're getting trouble reclaiming, start doing
2555 * writepage even in laptop mode.
2557 if (sc->priority < DEF_PRIORITY - 2)
2558 sc->may_writepage = 1;
2561 * Try to write back as many pages as we just scanned. This
2562 * tends to cause slow streaming writers to write data to the
2563 * disk smoothly, at the dirtying rate, which is nice. But
2564 * that's undesirable in laptop mode, where we *want* lumpy
2565 * writeout. So in laptop mode, write out the whole world.
2567 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2568 if (total_scanned > writeback_threshold) {
2569 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2570 WB_REASON_TRY_TO_FREE_PAGES);
2571 sc->may_writepage = 1;
2573 } while (--sc->priority >= 0);
2575 delayacct_freepages_end();
2577 if (sc->nr_reclaimed)
2578 return sc->nr_reclaimed;
2580 /* Aborted reclaim to try compaction? don't OOM, then */
2581 if (sc->compaction_ready)
2584 /* Untapped cgroup reserves? Don't OOM, retry. */
2585 if (!sc->may_thrash) {
2586 sc->priority = initial_priority;
2591 /* Any of the zones still reclaimable? Don't OOM. */
2592 if (zones_reclaimable)
2598 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2601 unsigned long pfmemalloc_reserve = 0;
2602 unsigned long free_pages = 0;
2606 for (i = 0; i <= ZONE_NORMAL; i++) {
2607 zone = &pgdat->node_zones[i];
2608 if (!populated_zone(zone))
2611 pfmemalloc_reserve += min_wmark_pages(zone);
2612 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2615 /* If there are no reserves (unexpected config) then do not throttle */
2616 if (!pfmemalloc_reserve)
2619 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2621 /* kswapd must be awake if processes are being throttled */
2622 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2623 pgdat->classzone_idx = min(pgdat->classzone_idx,
2624 (enum zone_type)ZONE_NORMAL);
2625 wake_up_interruptible(&pgdat->kswapd_wait);
2632 * Throttle direct reclaimers if backing storage is backed by the network
2633 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2634 * depleted. kswapd will continue to make progress and wake the processes
2635 * when the low watermark is reached.
2637 * Returns true if a fatal signal was delivered during throttling. If this
2638 * happens, the page allocator should not consider triggering the OOM killer.
2640 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2641 nodemask_t *nodemask)
2645 pg_data_t *pgdat = NULL;
2648 * Kernel threads should not be throttled as they may be indirectly
2649 * responsible for cleaning pages necessary for reclaim to make forward
2650 * progress. kjournald for example may enter direct reclaim while
2651 * committing a transaction where throttling it could forcing other
2652 * processes to block on log_wait_commit().
2654 if (current->flags & PF_KTHREAD)
2658 * If a fatal signal is pending, this process should not throttle.
2659 * It should return quickly so it can exit and free its memory
2661 if (fatal_signal_pending(current))
2665 * Check if the pfmemalloc reserves are ok by finding the first node
2666 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2667 * GFP_KERNEL will be required for allocating network buffers when
2668 * swapping over the network so ZONE_HIGHMEM is unusable.
2670 * Throttling is based on the first usable node and throttled processes
2671 * wait on a queue until kswapd makes progress and wakes them. There
2672 * is an affinity then between processes waking up and where reclaim
2673 * progress has been made assuming the process wakes on the same node.
2674 * More importantly, processes running on remote nodes will not compete
2675 * for remote pfmemalloc reserves and processes on different nodes
2676 * should make reasonable progress.
2678 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2679 gfp_zone(gfp_mask), nodemask) {
2680 if (zone_idx(zone) > ZONE_NORMAL)
2683 /* Throttle based on the first usable node */
2684 pgdat = zone->zone_pgdat;
2685 if (pfmemalloc_watermark_ok(pgdat))
2690 /* If no zone was usable by the allocation flags then do not throttle */
2694 /* Account for the throttling */
2695 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2698 * If the caller cannot enter the filesystem, it's possible that it
2699 * is due to the caller holding an FS lock or performing a journal
2700 * transaction in the case of a filesystem like ext[3|4]. In this case,
2701 * it is not safe to block on pfmemalloc_wait as kswapd could be
2702 * blocked waiting on the same lock. Instead, throttle for up to a
2703 * second before continuing.
2705 if (!(gfp_mask & __GFP_FS)) {
2706 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2707 pfmemalloc_watermark_ok(pgdat), HZ);
2712 /* Throttle until kswapd wakes the process */
2713 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2714 pfmemalloc_watermark_ok(pgdat));
2717 if (fatal_signal_pending(current))
2724 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2725 gfp_t gfp_mask, nodemask_t *nodemask)
2727 unsigned long nr_reclaimed;
2728 struct scan_control sc = {
2729 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2730 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2732 .nodemask = nodemask,
2733 .priority = DEF_PRIORITY,
2734 .may_writepage = !laptop_mode,
2740 * Do not enter reclaim if fatal signal was delivered while throttled.
2741 * 1 is returned so that the page allocator does not OOM kill at this
2744 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2747 trace_mm_vmscan_direct_reclaim_begin(order,
2751 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2753 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2755 return nr_reclaimed;
2760 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2761 gfp_t gfp_mask, bool noswap,
2763 unsigned long *nr_scanned)
2765 struct scan_control sc = {
2766 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2767 .target_mem_cgroup = memcg,
2768 .may_writepage = !laptop_mode,
2770 .may_swap = !noswap,
2772 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2773 int swappiness = mem_cgroup_swappiness(memcg);
2774 unsigned long lru_pages;
2776 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2777 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2779 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2784 * NOTE: Although we can get the priority field, using it
2785 * here is not a good idea, since it limits the pages we can scan.
2786 * if we don't reclaim here, the shrink_zone from balance_pgdat
2787 * will pick up pages from other mem cgroup's as well. We hack
2788 * the priority and make it zero.
2790 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
2792 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2794 *nr_scanned = sc.nr_scanned;
2795 return sc.nr_reclaimed;
2798 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2799 unsigned long nr_pages,
2803 struct zonelist *zonelist;
2804 unsigned long nr_reclaimed;
2806 struct scan_control sc = {
2807 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2808 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2809 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2810 .target_mem_cgroup = memcg,
2811 .priority = DEF_PRIORITY,
2812 .may_writepage = !laptop_mode,
2814 .may_swap = may_swap,
2818 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2819 * take care of from where we get pages. So the node where we start the
2820 * scan does not need to be the current node.
2822 nid = mem_cgroup_select_victim_node(memcg);
2824 zonelist = NODE_DATA(nid)->node_zonelists;
2826 trace_mm_vmscan_memcg_reclaim_begin(0,
2830 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2832 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2834 return nr_reclaimed;
2838 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2840 struct mem_cgroup *memcg;
2842 if (!total_swap_pages)
2845 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2847 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2849 if (inactive_anon_is_low(lruvec))
2850 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2851 sc, LRU_ACTIVE_ANON);
2853 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2857 static bool zone_balanced(struct zone *zone, int order,
2858 unsigned long balance_gap, int classzone_idx)
2860 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2861 balance_gap, classzone_idx, 0))
2864 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2865 order, 0, classzone_idx) == COMPACT_SKIPPED)
2872 * pgdat_balanced() is used when checking if a node is balanced.
2874 * For order-0, all zones must be balanced!
2876 * For high-order allocations only zones that meet watermarks and are in a
2877 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2878 * total of balanced pages must be at least 25% of the zones allowed by
2879 * classzone_idx for the node to be considered balanced. Forcing all zones to
2880 * be balanced for high orders can cause excessive reclaim when there are
2882 * The choice of 25% is due to
2883 * o a 16M DMA zone that is balanced will not balance a zone on any
2884 * reasonable sized machine
2885 * o On all other machines, the top zone must be at least a reasonable
2886 * percentage of the middle zones. For example, on 32-bit x86, highmem
2887 * would need to be at least 256M for it to be balance a whole node.
2888 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2889 * to balance a node on its own. These seemed like reasonable ratios.
2891 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2893 unsigned long managed_pages = 0;
2894 unsigned long balanced_pages = 0;
2897 /* Check the watermark levels */
2898 for (i = 0; i <= classzone_idx; i++) {
2899 struct zone *zone = pgdat->node_zones + i;
2901 if (!populated_zone(zone))
2904 managed_pages += zone->managed_pages;
2907 * A special case here:
2909 * balance_pgdat() skips over all_unreclaimable after
2910 * DEF_PRIORITY. Effectively, it considers them balanced so
2911 * they must be considered balanced here as well!
2913 if (!zone_reclaimable(zone)) {
2914 balanced_pages += zone->managed_pages;
2918 if (zone_balanced(zone, order, 0, i))
2919 balanced_pages += zone->managed_pages;
2925 return balanced_pages >= (managed_pages >> 2);
2931 * Prepare kswapd for sleeping. This verifies that there are no processes
2932 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2934 * Returns true if kswapd is ready to sleep
2936 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2939 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2944 * The throttled processes are normally woken up in balance_pgdat() as
2945 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2946 * race between when kswapd checks the watermarks and a process gets
2947 * throttled. There is also a potential race if processes get
2948 * throttled, kswapd wakes, a large process exits thereby balancing the
2949 * zones, which causes kswapd to exit balance_pgdat() before reaching
2950 * the wake up checks. If kswapd is going to sleep, no process should
2951 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2952 * the wake up is premature, processes will wake kswapd and get
2953 * throttled again. The difference from wake ups in balance_pgdat() is
2954 * that here we are under prepare_to_wait().
2956 if (waitqueue_active(&pgdat->pfmemalloc_wait))
2957 wake_up_all(&pgdat->pfmemalloc_wait);
2959 return pgdat_balanced(pgdat, order, classzone_idx);
2963 * kswapd shrinks the zone by the number of pages required to reach
2964 * the high watermark.
2966 * Returns true if kswapd scanned at least the requested number of pages to
2967 * reclaim or if the lack of progress was due to pages under writeback.
2968 * This is used to determine if the scanning priority needs to be raised.
2970 static bool kswapd_shrink_zone(struct zone *zone,
2972 struct scan_control *sc,
2973 unsigned long *nr_attempted)
2975 int testorder = sc->order;
2976 unsigned long balance_gap;
2977 bool lowmem_pressure;
2979 /* Reclaim above the high watermark. */
2980 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2983 * Kswapd reclaims only single pages with compaction enabled. Trying
2984 * too hard to reclaim until contiguous free pages have become
2985 * available can hurt performance by evicting too much useful data
2986 * from memory. Do not reclaim more than needed for compaction.
2988 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2989 compaction_suitable(zone, sc->order, 0, classzone_idx)
2994 * We put equal pressure on every zone, unless one zone has way too
2995 * many pages free already. The "too many pages" is defined as the
2996 * high wmark plus a "gap" where the gap is either the low
2997 * watermark or 1% of the zone, whichever is smaller.
2999 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3000 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3003 * If there is no low memory pressure or the zone is balanced then no
3004 * reclaim is necessary
3006 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3007 if (!lowmem_pressure && zone_balanced(zone, testorder,
3008 balance_gap, classzone_idx))
3011 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3013 /* Account for the number of pages attempted to reclaim */
3014 *nr_attempted += sc->nr_to_reclaim;
3016 clear_bit(ZONE_WRITEBACK, &zone->flags);
3019 * If a zone reaches its high watermark, consider it to be no longer
3020 * congested. It's possible there are dirty pages backed by congested
3021 * BDIs but as pressure is relieved, speculatively avoid congestion
3024 if (zone_reclaimable(zone) &&
3025 zone_balanced(zone, testorder, 0, classzone_idx)) {
3026 clear_bit(ZONE_CONGESTED, &zone->flags);
3027 clear_bit(ZONE_DIRTY, &zone->flags);
3030 return sc->nr_scanned >= sc->nr_to_reclaim;
3034 * For kswapd, balance_pgdat() will work across all this node's zones until
3035 * they are all at high_wmark_pages(zone).
3037 * Returns the final order kswapd was reclaiming at
3039 * There is special handling here for zones which are full of pinned pages.
3040 * This can happen if the pages are all mlocked, or if they are all used by
3041 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3042 * What we do is to detect the case where all pages in the zone have been
3043 * scanned twice and there has been zero successful reclaim. Mark the zone as
3044 * dead and from now on, only perform a short scan. Basically we're polling
3045 * the zone for when the problem goes away.
3047 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3048 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3049 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3050 * lower zones regardless of the number of free pages in the lower zones. This
3051 * interoperates with the page allocator fallback scheme to ensure that aging
3052 * of pages is balanced across the zones.
3054 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3058 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3059 unsigned long nr_soft_reclaimed;
3060 unsigned long nr_soft_scanned;
3061 struct scan_control sc = {
3062 .gfp_mask = GFP_KERNEL,
3064 .priority = DEF_PRIORITY,
3065 .may_writepage = !laptop_mode,
3069 count_vm_event(PAGEOUTRUN);
3072 unsigned long nr_attempted = 0;
3073 bool raise_priority = true;
3074 bool pgdat_needs_compaction = (order > 0);
3076 sc.nr_reclaimed = 0;
3079 * Scan in the highmem->dma direction for the highest
3080 * zone which needs scanning
3082 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3083 struct zone *zone = pgdat->node_zones + i;
3085 if (!populated_zone(zone))
3088 if (sc.priority != DEF_PRIORITY &&
3089 !zone_reclaimable(zone))
3093 * Do some background aging of the anon list, to give
3094 * pages a chance to be referenced before reclaiming.
3096 age_active_anon(zone, &sc);
3099 * If the number of buffer_heads in the machine
3100 * exceeds the maximum allowed level and this node
3101 * has a highmem zone, force kswapd to reclaim from
3102 * it to relieve lowmem pressure.
3104 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3109 if (!zone_balanced(zone, order, 0, 0)) {
3114 * If balanced, clear the dirty and congested
3117 clear_bit(ZONE_CONGESTED, &zone->flags);
3118 clear_bit(ZONE_DIRTY, &zone->flags);
3125 for (i = 0; i <= end_zone; i++) {
3126 struct zone *zone = pgdat->node_zones + i;
3128 if (!populated_zone(zone))
3132 * If any zone is currently balanced then kswapd will
3133 * not call compaction as it is expected that the
3134 * necessary pages are already available.
3136 if (pgdat_needs_compaction &&
3137 zone_watermark_ok(zone, order,
3138 low_wmark_pages(zone),
3140 pgdat_needs_compaction = false;
3144 * If we're getting trouble reclaiming, start doing writepage
3145 * even in laptop mode.
3147 if (sc.priority < DEF_PRIORITY - 2)
3148 sc.may_writepage = 1;
3151 * Now scan the zone in the dma->highmem direction, stopping
3152 * at the last zone which needs scanning.
3154 * We do this because the page allocator works in the opposite
3155 * direction. This prevents the page allocator from allocating
3156 * pages behind kswapd's direction of progress, which would
3157 * cause too much scanning of the lower zones.
3159 for (i = 0; i <= end_zone; i++) {
3160 struct zone *zone = pgdat->node_zones + i;
3162 if (!populated_zone(zone))
3165 if (sc.priority != DEF_PRIORITY &&
3166 !zone_reclaimable(zone))
3171 nr_soft_scanned = 0;
3173 * Call soft limit reclaim before calling shrink_zone.
3175 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3178 sc.nr_reclaimed += nr_soft_reclaimed;
3181 * There should be no need to raise the scanning
3182 * priority if enough pages are already being scanned
3183 * that that high watermark would be met at 100%
3186 if (kswapd_shrink_zone(zone, end_zone,
3187 &sc, &nr_attempted))
3188 raise_priority = false;
3192 * If the low watermark is met there is no need for processes
3193 * to be throttled on pfmemalloc_wait as they should not be
3194 * able to safely make forward progress. Wake them
3196 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3197 pfmemalloc_watermark_ok(pgdat))
3198 wake_up_all(&pgdat->pfmemalloc_wait);
3201 * Fragmentation may mean that the system cannot be rebalanced
3202 * for high-order allocations in all zones. If twice the
3203 * allocation size has been reclaimed and the zones are still
3204 * not balanced then recheck the watermarks at order-0 to
3205 * prevent kswapd reclaiming excessively. Assume that a
3206 * process requested a high-order can direct reclaim/compact.
3208 if (order && sc.nr_reclaimed >= 2UL << order)
3209 order = sc.order = 0;
3211 /* Check if kswapd should be suspending */
3212 if (try_to_freeze() || kthread_should_stop())
3216 * Compact if necessary and kswapd is reclaiming at least the
3217 * high watermark number of pages as requsted
3219 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3220 compact_pgdat(pgdat, order);
3223 * Raise priority if scanning rate is too low or there was no
3224 * progress in reclaiming pages
3226 if (raise_priority || !sc.nr_reclaimed)
3228 } while (sc.priority >= 1 &&
3229 !pgdat_balanced(pgdat, order, *classzone_idx));
3233 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3234 * makes a decision on the order we were last reclaiming at. However,
3235 * if another caller entered the allocator slow path while kswapd
3236 * was awake, order will remain at the higher level
3238 *classzone_idx = end_zone;
3242 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3247 if (freezing(current) || kthread_should_stop())
3250 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3252 /* Try to sleep for a short interval */
3253 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3254 remaining = schedule_timeout(HZ/10);
3255 finish_wait(&pgdat->kswapd_wait, &wait);
3256 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3260 * After a short sleep, check if it was a premature sleep. If not, then
3261 * go fully to sleep until explicitly woken up.
3263 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3264 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3267 * vmstat counters are not perfectly accurate and the estimated
3268 * value for counters such as NR_FREE_PAGES can deviate from the
3269 * true value by nr_online_cpus * threshold. To avoid the zone
3270 * watermarks being breached while under pressure, we reduce the
3271 * per-cpu vmstat threshold while kswapd is awake and restore
3272 * them before going back to sleep.
3274 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3277 * Compaction records what page blocks it recently failed to
3278 * isolate pages from and skips them in the future scanning.
3279 * When kswapd is going to sleep, it is reasonable to assume
3280 * that pages and compaction may succeed so reset the cache.
3282 reset_isolation_suitable(pgdat);
3284 if (!kthread_should_stop())
3287 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3290 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3292 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3294 finish_wait(&pgdat->kswapd_wait, &wait);
3298 * The background pageout daemon, started as a kernel thread
3299 * from the init process.
3301 * This basically trickles out pages so that we have _some_
3302 * free memory available even if there is no other activity
3303 * that frees anything up. This is needed for things like routing
3304 * etc, where we otherwise might have all activity going on in
3305 * asynchronous contexts that cannot page things out.
3307 * If there are applications that are active memory-allocators
3308 * (most normal use), this basically shouldn't matter.
3310 static int kswapd(void *p)
3312 unsigned long order, new_order;
3313 unsigned balanced_order;
3314 int classzone_idx, new_classzone_idx;
3315 int balanced_classzone_idx;
3316 pg_data_t *pgdat = (pg_data_t*)p;
3317 struct task_struct *tsk = current;
3319 struct reclaim_state reclaim_state = {
3320 .reclaimed_slab = 0,
3322 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3324 lockdep_set_current_reclaim_state(GFP_KERNEL);
3326 if (!cpumask_empty(cpumask))
3327 set_cpus_allowed_ptr(tsk, cpumask);
3328 current->reclaim_state = &reclaim_state;
3331 * Tell the memory management that we're a "memory allocator",
3332 * and that if we need more memory we should get access to it
3333 * regardless (see "__alloc_pages()"). "kswapd" should
3334 * never get caught in the normal page freeing logic.
3336 * (Kswapd normally doesn't need memory anyway, but sometimes
3337 * you need a small amount of memory in order to be able to
3338 * page out something else, and this flag essentially protects
3339 * us from recursively trying to free more memory as we're
3340 * trying to free the first piece of memory in the first place).
3342 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3345 order = new_order = 0;
3347 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3348 balanced_classzone_idx = classzone_idx;
3353 * If the last balance_pgdat was unsuccessful it's unlikely a
3354 * new request of a similar or harder type will succeed soon
3355 * so consider going to sleep on the basis we reclaimed at
3357 if (balanced_classzone_idx >= new_classzone_idx &&
3358 balanced_order == new_order) {
3359 new_order = pgdat->kswapd_max_order;
3360 new_classzone_idx = pgdat->classzone_idx;
3361 pgdat->kswapd_max_order = 0;
3362 pgdat->classzone_idx = pgdat->nr_zones - 1;
3365 if (order < new_order || classzone_idx > new_classzone_idx) {
3367 * Don't sleep if someone wants a larger 'order'
3368 * allocation or has tigher zone constraints
3371 classzone_idx = new_classzone_idx;
3373 kswapd_try_to_sleep(pgdat, balanced_order,
3374 balanced_classzone_idx);
3375 order = pgdat->kswapd_max_order;
3376 classzone_idx = pgdat->classzone_idx;
3378 new_classzone_idx = classzone_idx;
3379 pgdat->kswapd_max_order = 0;
3380 pgdat->classzone_idx = pgdat->nr_zones - 1;
3383 ret = try_to_freeze();
3384 if (kthread_should_stop())
3388 * We can speed up thawing tasks if we don't call balance_pgdat
3389 * after returning from the refrigerator
3392 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3393 balanced_classzone_idx = classzone_idx;
3394 balanced_order = balance_pgdat(pgdat, order,
3395 &balanced_classzone_idx);
3399 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3400 current->reclaim_state = NULL;
3401 lockdep_clear_current_reclaim_state();
3407 * A zone is low on free memory, so wake its kswapd task to service it.
3409 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3413 if (!populated_zone(zone))
3416 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3418 pgdat = zone->zone_pgdat;
3419 if (pgdat->kswapd_max_order < order) {
3420 pgdat->kswapd_max_order = order;
3421 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3423 if (!waitqueue_active(&pgdat->kswapd_wait))
3425 if (zone_balanced(zone, order, 0, 0))
3428 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3429 wake_up_interruptible(&pgdat->kswapd_wait);
3432 #ifdef CONFIG_HIBERNATION
3434 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3437 * Rather than trying to age LRUs the aim is to preserve the overall
3438 * LRU order by reclaiming preferentially
3439 * inactive > active > active referenced > active mapped
3441 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3443 struct reclaim_state reclaim_state;
3444 struct scan_control sc = {
3445 .nr_to_reclaim = nr_to_reclaim,
3446 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3447 .priority = DEF_PRIORITY,
3451 .hibernation_mode = 1,
3453 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3454 struct task_struct *p = current;
3455 unsigned long nr_reclaimed;
3457 p->flags |= PF_MEMALLOC;
3458 lockdep_set_current_reclaim_state(sc.gfp_mask);
3459 reclaim_state.reclaimed_slab = 0;
3460 p->reclaim_state = &reclaim_state;
3462 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3464 p->reclaim_state = NULL;
3465 lockdep_clear_current_reclaim_state();
3466 p->flags &= ~PF_MEMALLOC;
3468 return nr_reclaimed;
3470 #endif /* CONFIG_HIBERNATION */
3472 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3473 not required for correctness. So if the last cpu in a node goes
3474 away, we get changed to run anywhere: as the first one comes back,
3475 restore their cpu bindings. */
3476 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3481 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3482 for_each_node_state(nid, N_MEMORY) {
3483 pg_data_t *pgdat = NODE_DATA(nid);
3484 const struct cpumask *mask;
3486 mask = cpumask_of_node(pgdat->node_id);
3488 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3489 /* One of our CPUs online: restore mask */
3490 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3497 * This kswapd start function will be called by init and node-hot-add.
3498 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3500 int kswapd_run(int nid)
3502 pg_data_t *pgdat = NODE_DATA(nid);
3508 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3509 if (IS_ERR(pgdat->kswapd)) {
3510 /* failure at boot is fatal */
3511 BUG_ON(system_state == SYSTEM_BOOTING);
3512 pr_err("Failed to start kswapd on node %d\n", nid);
3513 ret = PTR_ERR(pgdat->kswapd);
3514 pgdat->kswapd = NULL;
3520 * Called by memory hotplug when all memory in a node is offlined. Caller must
3521 * hold mem_hotplug_begin/end().
3523 void kswapd_stop(int nid)
3525 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3528 kthread_stop(kswapd);
3529 NODE_DATA(nid)->kswapd = NULL;
3533 static int __init kswapd_init(void)
3538 for_each_node_state(nid, N_MEMORY)
3540 hotcpu_notifier(cpu_callback, 0);
3544 module_init(kswapd_init)
3550 * If non-zero call zone_reclaim when the number of free pages falls below
3553 int zone_reclaim_mode __read_mostly;
3555 #define RECLAIM_OFF 0
3556 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3557 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3558 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3561 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3562 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3565 #define ZONE_RECLAIM_PRIORITY 4
3568 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3571 int sysctl_min_unmapped_ratio = 1;
3574 * If the number of slab pages in a zone grows beyond this percentage then
3575 * slab reclaim needs to occur.
3577 int sysctl_min_slab_ratio = 5;
3579 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3581 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3582 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3583 zone_page_state(zone, NR_ACTIVE_FILE);
3586 * It's possible for there to be more file mapped pages than
3587 * accounted for by the pages on the file LRU lists because
3588 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3590 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3593 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3594 static long zone_pagecache_reclaimable(struct zone *zone)
3596 long nr_pagecache_reclaimable;
3600 * If RECLAIM_SWAP is set, then all file pages are considered
3601 * potentially reclaimable. Otherwise, we have to worry about
3602 * pages like swapcache and zone_unmapped_file_pages() provides
3605 if (zone_reclaim_mode & RECLAIM_SWAP)
3606 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3608 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3610 /* If we can't clean pages, remove dirty pages from consideration */
3611 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3612 delta += zone_page_state(zone, NR_FILE_DIRTY);
3614 /* Watch for any possible underflows due to delta */
3615 if (unlikely(delta > nr_pagecache_reclaimable))
3616 delta = nr_pagecache_reclaimable;
3618 return nr_pagecache_reclaimable - delta;
3622 * Try to free up some pages from this zone through reclaim.
3624 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3626 /* Minimum pages needed in order to stay on node */
3627 const unsigned long nr_pages = 1 << order;
3628 struct task_struct *p = current;
3629 struct reclaim_state reclaim_state;
3630 struct scan_control sc = {
3631 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3632 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3634 .priority = ZONE_RECLAIM_PRIORITY,
3635 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3636 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3642 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3643 * and we also need to be able to write out pages for RECLAIM_WRITE
3646 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3647 lockdep_set_current_reclaim_state(gfp_mask);
3648 reclaim_state.reclaimed_slab = 0;
3649 p->reclaim_state = &reclaim_state;
3651 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3653 * Free memory by calling shrink zone with increasing
3654 * priorities until we have enough memory freed.
3657 shrink_zone(zone, &sc, true);
3658 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3661 p->reclaim_state = NULL;
3662 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3663 lockdep_clear_current_reclaim_state();
3664 return sc.nr_reclaimed >= nr_pages;
3667 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3673 * Zone reclaim reclaims unmapped file backed pages and
3674 * slab pages if we are over the defined limits.
3676 * A small portion of unmapped file backed pages is needed for
3677 * file I/O otherwise pages read by file I/O will be immediately
3678 * thrown out if the zone is overallocated. So we do not reclaim
3679 * if less than a specified percentage of the zone is used by
3680 * unmapped file backed pages.
3682 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3683 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3684 return ZONE_RECLAIM_FULL;
3686 if (!zone_reclaimable(zone))
3687 return ZONE_RECLAIM_FULL;
3690 * Do not scan if the allocation should not be delayed.
3692 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3693 return ZONE_RECLAIM_NOSCAN;
3696 * Only run zone reclaim on the local zone or on zones that do not
3697 * have associated processors. This will favor the local processor
3698 * over remote processors and spread off node memory allocations
3699 * as wide as possible.
3701 node_id = zone_to_nid(zone);
3702 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3703 return ZONE_RECLAIM_NOSCAN;
3705 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3706 return ZONE_RECLAIM_NOSCAN;
3708 ret = __zone_reclaim(zone, gfp_mask, order);
3709 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3712 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3719 * page_evictable - test whether a page is evictable
3720 * @page: the page to test
3722 * Test whether page is evictable--i.e., should be placed on active/inactive
3723 * lists vs unevictable list.
3725 * Reasons page might not be evictable:
3726 * (1) page's mapping marked unevictable
3727 * (2) page is part of an mlocked VMA
3730 int page_evictable(struct page *page)
3732 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3737 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3738 * @pages: array of pages to check
3739 * @nr_pages: number of pages to check
3741 * Checks pages for evictability and moves them to the appropriate lru list.
3743 * This function is only used for SysV IPC SHM_UNLOCK.
3745 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3747 struct lruvec *lruvec;
3748 struct zone *zone = NULL;
3753 for (i = 0; i < nr_pages; i++) {
3754 struct page *page = pages[i];
3755 struct zone *pagezone;
3758 pagezone = page_zone(page);
3759 if (pagezone != zone) {
3761 spin_unlock_irq(&zone->lru_lock);
3763 spin_lock_irq(&zone->lru_lock);
3765 lruvec = mem_cgroup_page_lruvec(page, zone);
3767 if (!PageLRU(page) || !PageUnevictable(page))
3770 if (page_evictable(page)) {
3771 enum lru_list lru = page_lru_base_type(page);
3773 VM_BUG_ON_PAGE(PageActive(page), page);
3774 ClearPageUnevictable(page);
3775 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3776 add_page_to_lru_list(page, lruvec, lru);
3782 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3783 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3784 spin_unlock_irq(&zone->lru_lock);
3787 #endif /* CONFIG_SHMEM */