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.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
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;
159 * sane_reclaim - is the usual dirty throttling mechanism operational?
160 * @sc: scan_control in question
162 * The normal page dirty throttling mechanism in balance_dirty_pages() is
163 * completely broken with the legacy memcg and direct stalling in
164 * shrink_page_list() is used for throttling instead, which lacks all the
165 * niceties such as fairness, adaptive pausing, bandwidth proportional
166 * allocation and configurability.
168 * This function tests whether the vmscan currently in progress can assume
169 * that the normal dirty throttling mechanism is operational.
171 static bool sane_reclaim(struct scan_control *sc)
173 struct mem_cgroup *memcg = sc->target_mem_cgroup;
177 #ifdef CONFIG_CGROUP_WRITEBACK
178 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
184 static bool global_reclaim(struct scan_control *sc)
189 static bool sane_reclaim(struct scan_control *sc)
195 static unsigned long zone_reclaimable_pages(struct zone *zone)
199 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
200 zone_page_state(zone, NR_INACTIVE_FILE);
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
204 zone_page_state(zone, NR_INACTIVE_ANON);
209 bool zone_reclaimable(struct zone *zone)
211 return zone_page_state(zone, NR_PAGES_SCANNED) <
212 zone_reclaimable_pages(zone) * 6;
215 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
217 if (!mem_cgroup_disabled())
218 return mem_cgroup_get_lru_size(lruvec, lru);
220 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
224 * Add a shrinker callback to be called from the vm.
226 int register_shrinker(struct shrinker *shrinker)
228 size_t size = sizeof(*shrinker->nr_deferred);
231 * If we only have one possible node in the system anyway, save
232 * ourselves the trouble and disable NUMA aware behavior. This way we
233 * will save memory and some small loop time later.
235 if (nr_node_ids == 1)
236 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
238 if (shrinker->flags & SHRINKER_NUMA_AWARE)
241 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
242 if (!shrinker->nr_deferred)
245 down_write(&shrinker_rwsem);
246 list_add_tail(&shrinker->list, &shrinker_list);
247 up_write(&shrinker_rwsem);
250 EXPORT_SYMBOL(register_shrinker);
255 void unregister_shrinker(struct shrinker *shrinker)
257 down_write(&shrinker_rwsem);
258 list_del(&shrinker->list);
259 up_write(&shrinker_rwsem);
260 kfree(shrinker->nr_deferred);
262 EXPORT_SYMBOL(unregister_shrinker);
264 #define SHRINK_BATCH 128
266 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
267 struct shrinker *shrinker,
268 unsigned long nr_scanned,
269 unsigned long nr_eligible)
271 unsigned long freed = 0;
272 unsigned long long delta;
277 int nid = shrinkctl->nid;
278 long batch_size = shrinker->batch ? shrinker->batch
281 freeable = shrinker->count_objects(shrinker, shrinkctl);
286 * copy the current shrinker scan count into a local variable
287 * and zero it so that other concurrent shrinker invocations
288 * don't also do this scanning work.
290 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
293 delta = (4 * nr_scanned) / shrinker->seeks;
295 do_div(delta, nr_eligible + 1);
297 if (total_scan < 0) {
298 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
299 shrinker->scan_objects, total_scan);
300 total_scan = freeable;
304 * We need to avoid excessive windup on filesystem shrinkers
305 * due to large numbers of GFP_NOFS allocations causing the
306 * shrinkers to return -1 all the time. This results in a large
307 * nr being built up so when a shrink that can do some work
308 * comes along it empties the entire cache due to nr >>>
309 * freeable. This is bad for sustaining a working set in
312 * Hence only allow the shrinker to scan the entire cache when
313 * a large delta change is calculated directly.
315 if (delta < freeable / 4)
316 total_scan = min(total_scan, freeable / 2);
319 * Avoid risking looping forever due to too large nr value:
320 * never try to free more than twice the estimate number of
323 if (total_scan > freeable * 2)
324 total_scan = freeable * 2;
326 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
327 nr_scanned, nr_eligible,
328 freeable, delta, total_scan);
331 * Normally, we should not scan less than batch_size objects in one
332 * pass to avoid too frequent shrinker calls, but if the slab has less
333 * than batch_size objects in total and we are really tight on memory,
334 * we will try to reclaim all available objects, otherwise we can end
335 * up failing allocations although there are plenty of reclaimable
336 * objects spread over several slabs with usage less than the
339 * We detect the "tight on memory" situations by looking at the total
340 * number of objects we want to scan (total_scan). If it is greater
341 * than the total number of objects on slab (freeable), we must be
342 * scanning at high prio and therefore should try to reclaim as much as
345 while (total_scan >= batch_size ||
346 total_scan >= freeable) {
348 unsigned long nr_to_scan = min(batch_size, total_scan);
350 shrinkctl->nr_to_scan = nr_to_scan;
351 ret = shrinker->scan_objects(shrinker, shrinkctl);
352 if (ret == SHRINK_STOP)
356 count_vm_events(SLABS_SCANNED, nr_to_scan);
357 total_scan -= nr_to_scan;
363 * move the unused scan count back into the shrinker in a
364 * manner that handles concurrent updates. If we exhausted the
365 * scan, there is no need to do an update.
368 new_nr = atomic_long_add_return(total_scan,
369 &shrinker->nr_deferred[nid]);
371 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
373 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
378 * shrink_slab - shrink slab caches
379 * @gfp_mask: allocation context
380 * @nid: node whose slab caches to target
381 * @memcg: memory cgroup whose slab caches to target
382 * @nr_scanned: pressure numerator
383 * @nr_eligible: pressure denominator
385 * Call the shrink functions to age shrinkable caches.
387 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
388 * unaware shrinkers will receive a node id of 0 instead.
390 * @memcg specifies the memory cgroup to target. If it is not NULL,
391 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
392 * objects from the memory cgroup specified. Otherwise all shrinkers
393 * are called, and memcg aware shrinkers are supposed to scan the
396 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
397 * the available objects should be scanned. Page reclaim for example
398 * passes the number of pages scanned and the number of pages on the
399 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
400 * when it encountered mapped pages. The ratio is further biased by
401 * the ->seeks setting of the shrink function, which indicates the
402 * cost to recreate an object relative to that of an LRU page.
404 * Returns the number of reclaimed slab objects.
406 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
407 struct mem_cgroup *memcg,
408 unsigned long nr_scanned,
409 unsigned long nr_eligible)
411 struct shrinker *shrinker;
412 unsigned long freed = 0;
414 if (memcg && !memcg_kmem_is_active(memcg))
418 nr_scanned = SWAP_CLUSTER_MAX;
420 if (!down_read_trylock(&shrinker_rwsem)) {
422 * If we would return 0, our callers would understand that we
423 * have nothing else to shrink and give up trying. By returning
424 * 1 we keep it going and assume we'll be able to shrink next
431 list_for_each_entry(shrinker, &shrinker_list, list) {
432 struct shrink_control sc = {
433 .gfp_mask = gfp_mask,
438 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
441 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
444 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
447 up_read(&shrinker_rwsem);
453 void drop_slab_node(int nid)
458 struct mem_cgroup *memcg = NULL;
462 freed += shrink_slab(GFP_KERNEL, nid, memcg,
464 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
465 } while (freed > 10);
472 for_each_online_node(nid)
476 static inline int is_page_cache_freeable(struct page *page)
479 * A freeable page cache page is referenced only by the caller
480 * that isolated the page, the page cache radix tree and
481 * optional buffer heads at page->private.
483 return page_count(page) - page_has_private(page) == 2;
486 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
488 if (current->flags & PF_SWAPWRITE)
490 if (!inode_write_congested(inode))
492 if (inode_to_bdi(inode) == current->backing_dev_info)
498 * We detected a synchronous write error writing a page out. Probably
499 * -ENOSPC. We need to propagate that into the address_space for a subsequent
500 * fsync(), msync() or close().
502 * The tricky part is that after writepage we cannot touch the mapping: nothing
503 * prevents it from being freed up. But we have a ref on the page and once
504 * that page is locked, the mapping is pinned.
506 * We're allowed to run sleeping lock_page() here because we know the caller has
509 static void handle_write_error(struct address_space *mapping,
510 struct page *page, int error)
513 if (page_mapping(page) == mapping)
514 mapping_set_error(mapping, error);
518 /* possible outcome of pageout() */
520 /* failed to write page out, page is locked */
522 /* move page to the active list, page is locked */
524 /* page has been sent to the disk successfully, page is unlocked */
526 /* page is clean and locked */
531 * pageout is called by shrink_page_list() for each dirty page.
532 * Calls ->writepage().
534 static pageout_t pageout(struct page *page, struct address_space *mapping,
535 struct scan_control *sc)
538 * If the page is dirty, only perform writeback if that write
539 * will be non-blocking. To prevent this allocation from being
540 * stalled by pagecache activity. But note that there may be
541 * stalls if we need to run get_block(). We could test
542 * PagePrivate for that.
544 * If this process is currently in __generic_file_write_iter() against
545 * this page's queue, we can perform writeback even if that
548 * If the page is swapcache, write it back even if that would
549 * block, for some throttling. This happens by accident, because
550 * swap_backing_dev_info is bust: it doesn't reflect the
551 * congestion state of the swapdevs. Easy to fix, if needed.
553 if (!is_page_cache_freeable(page))
557 * Some data journaling orphaned pages can have
558 * page->mapping == NULL while being dirty with clean buffers.
560 if (page_has_private(page)) {
561 if (try_to_free_buffers(page)) {
562 ClearPageDirty(page);
563 pr_info("%s: orphaned page\n", __func__);
569 if (mapping->a_ops->writepage == NULL)
570 return PAGE_ACTIVATE;
571 if (!may_write_to_inode(mapping->host, sc))
574 if (clear_page_dirty_for_io(page)) {
576 struct writeback_control wbc = {
577 .sync_mode = WB_SYNC_NONE,
578 .nr_to_write = SWAP_CLUSTER_MAX,
580 .range_end = LLONG_MAX,
584 SetPageReclaim(page);
585 res = mapping->a_ops->writepage(page, &wbc);
587 handle_write_error(mapping, page, res);
588 if (res == AOP_WRITEPAGE_ACTIVATE) {
589 ClearPageReclaim(page);
590 return PAGE_ACTIVATE;
593 if (!PageWriteback(page)) {
594 /* synchronous write or broken a_ops? */
595 ClearPageReclaim(page);
597 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
598 inc_zone_page_state(page, NR_VMSCAN_WRITE);
606 * Same as remove_mapping, but if the page is removed from the mapping, it
607 * gets returned with a refcount of 0.
609 static int __remove_mapping(struct address_space *mapping, struct page *page,
613 struct mem_cgroup *memcg;
615 BUG_ON(!PageLocked(page));
616 BUG_ON(mapping != page_mapping(page));
618 memcg = mem_cgroup_begin_page_stat(page);
619 spin_lock_irqsave(&mapping->tree_lock, flags);
621 * The non racy check for a busy page.
623 * Must be careful with the order of the tests. When someone has
624 * a ref to the page, it may be possible that they dirty it then
625 * drop the reference. So if PageDirty is tested before page_count
626 * here, then the following race may occur:
628 * get_user_pages(&page);
629 * [user mapping goes away]
631 * !PageDirty(page) [good]
632 * SetPageDirty(page);
634 * !page_count(page) [good, discard it]
636 * [oops, our write_to data is lost]
638 * Reversing the order of the tests ensures such a situation cannot
639 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
640 * load is not satisfied before that of page->_count.
642 * Note that if SetPageDirty is always performed via set_page_dirty,
643 * and thus under tree_lock, then this ordering is not required.
645 if (!page_freeze_refs(page, 2))
647 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
648 if (unlikely(PageDirty(page))) {
649 page_unfreeze_refs(page, 2);
653 if (PageSwapCache(page)) {
654 swp_entry_t swap = { .val = page_private(page) };
655 mem_cgroup_swapout(page, swap);
656 __delete_from_swap_cache(page);
657 spin_unlock_irqrestore(&mapping->tree_lock, flags);
658 mem_cgroup_end_page_stat(memcg);
659 swapcache_free(swap);
661 void (*freepage)(struct page *);
664 freepage = mapping->a_ops->freepage;
666 * Remember a shadow entry for reclaimed file cache in
667 * order to detect refaults, thus thrashing, later on.
669 * But don't store shadows in an address space that is
670 * already exiting. This is not just an optizimation,
671 * inode reclaim needs to empty out the radix tree or
672 * the nodes are lost. Don't plant shadows behind its
675 if (reclaimed && page_is_file_cache(page) &&
676 !mapping_exiting(mapping))
677 shadow = workingset_eviction(mapping, page);
678 __delete_from_page_cache(page, shadow, memcg);
679 spin_unlock_irqrestore(&mapping->tree_lock, flags);
680 mem_cgroup_end_page_stat(memcg);
682 if (freepage != NULL)
689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
690 mem_cgroup_end_page_stat(memcg);
695 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
696 * someone else has a ref on the page, abort and return 0. If it was
697 * successfully detached, return 1. Assumes the caller has a single ref on
700 int remove_mapping(struct address_space *mapping, struct page *page)
702 if (__remove_mapping(mapping, page, false)) {
704 * Unfreezing the refcount with 1 rather than 2 effectively
705 * drops the pagecache ref for us without requiring another
708 page_unfreeze_refs(page, 1);
715 * putback_lru_page - put previously isolated page onto appropriate LRU list
716 * @page: page to be put back to appropriate lru list
718 * Add previously isolated @page to appropriate LRU list.
719 * Page may still be unevictable for other reasons.
721 * lru_lock must not be held, interrupts must be enabled.
723 void putback_lru_page(struct page *page)
726 int was_unevictable = PageUnevictable(page);
728 VM_BUG_ON_PAGE(PageLRU(page), page);
731 ClearPageUnevictable(page);
733 if (page_evictable(page)) {
735 * For evictable pages, we can use the cache.
736 * In event of a race, worst case is we end up with an
737 * unevictable page on [in]active list.
738 * We know how to handle that.
740 is_unevictable = false;
744 * Put unevictable pages directly on zone's unevictable
747 is_unevictable = true;
748 add_page_to_unevictable_list(page);
750 * When racing with an mlock or AS_UNEVICTABLE clearing
751 * (page is unlocked) make sure that if the other thread
752 * does not observe our setting of PG_lru and fails
753 * isolation/check_move_unevictable_pages,
754 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
755 * the page back to the evictable list.
757 * The other side is TestClearPageMlocked() or shmem_lock().
763 * page's status can change while we move it among lru. If an evictable
764 * page is on unevictable list, it never be freed. To avoid that,
765 * check after we added it to the list, again.
767 if (is_unevictable && page_evictable(page)) {
768 if (!isolate_lru_page(page)) {
772 /* This means someone else dropped this page from LRU
773 * So, it will be freed or putback to LRU again. There is
774 * nothing to do here.
778 if (was_unevictable && !is_unevictable)
779 count_vm_event(UNEVICTABLE_PGRESCUED);
780 else if (!was_unevictable && is_unevictable)
781 count_vm_event(UNEVICTABLE_PGCULLED);
783 put_page(page); /* drop ref from isolate */
786 enum page_references {
788 PAGEREF_RECLAIM_CLEAN,
793 static enum page_references page_check_references(struct page *page,
794 struct scan_control *sc)
796 int referenced_ptes, referenced_page;
797 unsigned long vm_flags;
799 VM_BUG_ON_PAGE(!PageLocked(page), page);
801 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
803 referenced_page = TestClearPageReferenced(page);
806 * Mlock lost the isolation race with us. Let try_to_unmap()
807 * move the page to the unevictable list.
809 if (vm_flags & VM_LOCKED)
810 return PAGEREF_RECLAIM;
812 if (referenced_ptes) {
813 if (PageSwapBacked(page))
814 return PAGEREF_ACTIVATE;
816 * All mapped pages start out with page table
817 * references from the instantiating fault, so we need
818 * to look twice if a mapped file page is used more
821 * Mark it and spare it for another trip around the
822 * inactive list. Another page table reference will
823 * lead to its activation.
825 * Note: the mark is set for activated pages as well
826 * so that recently deactivated but used pages are
829 SetPageReferenced(page);
831 if (referenced_page || referenced_ptes > 1)
832 return PAGEREF_ACTIVATE;
835 * Activate file-backed executable pages after first usage.
837 if (vm_flags & VM_EXEC)
838 return PAGEREF_ACTIVATE;
843 /* Reclaim if clean, defer dirty pages to writeback */
844 if (referenced_page && !PageSwapBacked(page))
845 return PAGEREF_RECLAIM_CLEAN;
847 return PAGEREF_RECLAIM;
850 /* Check if a page is dirty or under writeback */
851 static void page_check_dirty_writeback(struct page *page,
852 bool *dirty, bool *writeback)
854 struct address_space *mapping;
857 * Anonymous pages are not handled by flushers and must be written
858 * from reclaim context. Do not stall reclaim based on them
860 if (!page_is_file_cache(page)) {
866 /* By default assume that the page flags are accurate */
867 *dirty = PageDirty(page);
868 *writeback = PageWriteback(page);
870 /* Verify dirty/writeback state if the filesystem supports it */
871 if (!page_has_private(page))
874 mapping = page_mapping(page);
875 if (mapping && mapping->a_ops->is_dirty_writeback)
876 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
880 * shrink_page_list() returns the number of reclaimed pages
882 static unsigned long shrink_page_list(struct list_head *page_list,
884 struct scan_control *sc,
885 enum ttu_flags ttu_flags,
886 unsigned long *ret_nr_dirty,
887 unsigned long *ret_nr_unqueued_dirty,
888 unsigned long *ret_nr_congested,
889 unsigned long *ret_nr_writeback,
890 unsigned long *ret_nr_immediate,
893 LIST_HEAD(ret_pages);
894 LIST_HEAD(free_pages);
896 unsigned long nr_unqueued_dirty = 0;
897 unsigned long nr_dirty = 0;
898 unsigned long nr_congested = 0;
899 unsigned long nr_reclaimed = 0;
900 unsigned long nr_writeback = 0;
901 unsigned long nr_immediate = 0;
905 while (!list_empty(page_list)) {
906 struct address_space *mapping;
909 enum page_references references = PAGEREF_RECLAIM_CLEAN;
910 bool dirty, writeback;
911 bool freeable = false;
915 page = lru_to_page(page_list);
916 list_del(&page->lru);
918 if (!trylock_page(page))
921 VM_BUG_ON_PAGE(PageActive(page), page);
922 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
926 if (unlikely(!page_evictable(page)))
929 if (!sc->may_unmap && page_mapped(page))
932 /* Double the slab pressure for mapped and swapcache pages */
933 if (page_mapped(page) || PageSwapCache(page))
936 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
937 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
940 * The number of dirty pages determines if a zone is marked
941 * reclaim_congested which affects wait_iff_congested. kswapd
942 * will stall and start writing pages if the tail of the LRU
943 * is all dirty unqueued pages.
945 page_check_dirty_writeback(page, &dirty, &writeback);
946 if (dirty || writeback)
949 if (dirty && !writeback)
953 * Treat this page as congested if the underlying BDI is or if
954 * pages are cycling through the LRU so quickly that the
955 * pages marked for immediate reclaim are making it to the
956 * end of the LRU a second time.
958 mapping = page_mapping(page);
959 if (((dirty || writeback) && mapping &&
960 inode_write_congested(mapping->host)) ||
961 (writeback && PageReclaim(page)))
965 * If a page at the tail of the LRU is under writeback, there
966 * are three cases to consider.
968 * 1) If reclaim is encountering an excessive number of pages
969 * under writeback and this page is both under writeback and
970 * PageReclaim then it indicates that pages are being queued
971 * for IO but are being recycled through the LRU before the
972 * IO can complete. Waiting on the page itself risks an
973 * indefinite stall if it is impossible to writeback the
974 * page due to IO error or disconnected storage so instead
975 * note that the LRU is being scanned too quickly and the
976 * caller can stall after page list has been processed.
978 * 2) Global or new memcg reclaim encounters a page that is
979 * not marked for immediate reclaim, or the caller does not
980 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
981 * not to fs). In this case mark the page for immediate
982 * reclaim and continue scanning.
984 * Require may_enter_fs because we would wait on fs, which
985 * may not have submitted IO yet. And the loop driver might
986 * enter reclaim, and deadlock if it waits on a page for
987 * which it is needed to do the write (loop masks off
988 * __GFP_IO|__GFP_FS for this reason); but more thought
989 * would probably show more reasons.
991 * 3) Legacy memcg encounters a page that is already marked
992 * PageReclaim. memcg does not have any dirty pages
993 * throttling so we could easily OOM just because too many
994 * pages are in writeback and there is nothing else to
995 * reclaim. Wait for the writeback to complete.
997 if (PageWriteback(page)) {
999 if (current_is_kswapd() &&
1000 PageReclaim(page) &&
1001 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1006 } else if (sane_reclaim(sc) ||
1007 !PageReclaim(page) || !may_enter_fs) {
1009 * This is slightly racy - end_page_writeback()
1010 * might have just cleared PageReclaim, then
1011 * setting PageReclaim here end up interpreted
1012 * as PageReadahead - but that does not matter
1013 * enough to care. What we do want is for this
1014 * page to have PageReclaim set next time memcg
1015 * reclaim reaches the tests above, so it will
1016 * then wait_on_page_writeback() to avoid OOM;
1017 * and it's also appropriate in global reclaim.
1019 SetPageReclaim(page);
1026 wait_on_page_writeback(page);
1027 /* then go back and try same page again */
1028 list_add_tail(&page->lru, page_list);
1034 references = page_check_references(page, sc);
1036 switch (references) {
1037 case PAGEREF_ACTIVATE:
1038 goto activate_locked;
1041 case PAGEREF_RECLAIM:
1042 case PAGEREF_RECLAIM_CLEAN:
1043 ; /* try to reclaim the page below */
1047 * Anonymous process memory has backing store?
1048 * Try to allocate it some swap space here.
1050 if (PageAnon(page) && !PageSwapCache(page)) {
1051 if (!(sc->gfp_mask & __GFP_IO))
1053 if (!add_to_swap(page, page_list))
1054 goto activate_locked;
1057 /* Adding to swap updated mapping */
1058 mapping = page_mapping(page);
1062 * The page is mapped into the page tables of one or more
1063 * processes. Try to unmap it here.
1065 if (page_mapped(page) && mapping) {
1066 switch (try_to_unmap(page, freeable ?
1067 ttu_flags | TTU_BATCH_FLUSH | TTU_FREE :
1068 ttu_flags | TTU_BATCH_FLUSH)) {
1070 goto activate_locked;
1076 ; /* try to free the page below */
1080 if (PageDirty(page)) {
1082 * Only kswapd can writeback filesystem pages to
1083 * avoid risk of stack overflow but only writeback
1084 * if many dirty pages have been encountered.
1086 if (page_is_file_cache(page) &&
1087 (!current_is_kswapd() ||
1088 !test_bit(ZONE_DIRTY, &zone->flags))) {
1090 * Immediately reclaim when written back.
1091 * Similar in principal to deactivate_page()
1092 * except we already have the page isolated
1093 * and know it's dirty
1095 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1096 SetPageReclaim(page);
1101 if (references == PAGEREF_RECLAIM_CLEAN)
1105 if (!sc->may_writepage)
1109 * Page is dirty. Flush the TLB if a writable entry
1110 * potentially exists to avoid CPU writes after IO
1111 * starts and then write it out here.
1113 try_to_unmap_flush_dirty();
1114 switch (pageout(page, mapping, sc)) {
1118 goto activate_locked;
1120 if (PageWriteback(page))
1122 if (PageDirty(page))
1126 * A synchronous write - probably a ramdisk. Go
1127 * ahead and try to reclaim the page.
1129 if (!trylock_page(page))
1131 if (PageDirty(page) || PageWriteback(page))
1133 mapping = page_mapping(page);
1135 ; /* try to free the page below */
1140 * If the page has buffers, try to free the buffer mappings
1141 * associated with this page. If we succeed we try to free
1144 * We do this even if the page is PageDirty().
1145 * try_to_release_page() does not perform I/O, but it is
1146 * possible for a page to have PageDirty set, but it is actually
1147 * clean (all its buffers are clean). This happens if the
1148 * buffers were written out directly, with submit_bh(). ext3
1149 * will do this, as well as the blockdev mapping.
1150 * try_to_release_page() will discover that cleanness and will
1151 * drop the buffers and mark the page clean - it can be freed.
1153 * Rarely, pages can have buffers and no ->mapping. These are
1154 * the pages which were not successfully invalidated in
1155 * truncate_complete_page(). We try to drop those buffers here
1156 * and if that worked, and the page is no longer mapped into
1157 * process address space (page_count == 1) it can be freed.
1158 * Otherwise, leave the page on the LRU so it is swappable.
1160 if (page_has_private(page)) {
1161 if (!try_to_release_page(page, sc->gfp_mask))
1162 goto activate_locked;
1163 if (!mapping && page_count(page) == 1) {
1165 if (put_page_testzero(page))
1169 * rare race with speculative reference.
1170 * the speculative reference will free
1171 * this page shortly, so we may
1172 * increment nr_reclaimed here (and
1173 * leave it off the LRU).
1181 if (!mapping || !__remove_mapping(mapping, page, true))
1185 * At this point, we have no other references and there is
1186 * no way to pick any more up (removed from LRU, removed
1187 * from pagecache). Can use non-atomic bitops now (and
1188 * we obviously don't have to worry about waking up a process
1189 * waiting on the page lock, because there are no references.
1191 __ClearPageLocked(page);
1193 if (freeable && !PageDirty(page))
1194 count_vm_event(PGLAZYFREED);
1199 * Is there need to periodically free_page_list? It would
1200 * appear not as the counts should be low
1202 list_add(&page->lru, &free_pages);
1206 if (PageSwapCache(page))
1207 try_to_free_swap(page);
1209 list_add(&page->lru, &ret_pages);
1213 /* Not a candidate for swapping, so reclaim swap space. */
1214 if (PageSwapCache(page) && vm_swap_full())
1215 try_to_free_swap(page);
1216 VM_BUG_ON_PAGE(PageActive(page), page);
1217 SetPageActive(page);
1222 list_add(&page->lru, &ret_pages);
1223 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1226 mem_cgroup_uncharge_list(&free_pages);
1227 try_to_unmap_flush();
1228 free_hot_cold_page_list(&free_pages, true);
1230 list_splice(&ret_pages, page_list);
1231 count_vm_events(PGACTIVATE, pgactivate);
1233 *ret_nr_dirty += nr_dirty;
1234 *ret_nr_congested += nr_congested;
1235 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1236 *ret_nr_writeback += nr_writeback;
1237 *ret_nr_immediate += nr_immediate;
1238 return nr_reclaimed;
1241 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1242 struct list_head *page_list)
1244 struct scan_control sc = {
1245 .gfp_mask = GFP_KERNEL,
1246 .priority = DEF_PRIORITY,
1249 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1250 struct page *page, *next;
1251 LIST_HEAD(clean_pages);
1253 list_for_each_entry_safe(page, next, page_list, lru) {
1254 if (page_is_file_cache(page) && !PageDirty(page) &&
1255 !isolated_balloon_page(page)) {
1256 ClearPageActive(page);
1257 list_move(&page->lru, &clean_pages);
1261 ret = shrink_page_list(&clean_pages, zone, &sc,
1262 TTU_UNMAP|TTU_IGNORE_ACCESS,
1263 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1264 list_splice(&clean_pages, page_list);
1265 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1270 * Attempt to remove the specified page from its LRU. Only take this page
1271 * if it is of the appropriate PageActive status. Pages which are being
1272 * freed elsewhere are also ignored.
1274 * page: page to consider
1275 * mode: one of the LRU isolation modes defined above
1277 * returns 0 on success, -ve errno on failure.
1279 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1283 /* Only take pages on the LRU. */
1287 /* Compaction should not handle unevictable pages but CMA can do so */
1288 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1294 * To minimise LRU disruption, the caller can indicate that it only
1295 * wants to isolate pages it will be able to operate on without
1296 * blocking - clean pages for the most part.
1298 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1299 * is used by reclaim when it is cannot write to backing storage
1301 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1302 * that it is possible to migrate without blocking
1304 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1305 /* All the caller can do on PageWriteback is block */
1306 if (PageWriteback(page))
1309 if (PageDirty(page)) {
1310 struct address_space *mapping;
1312 /* ISOLATE_CLEAN means only clean pages */
1313 if (mode & ISOLATE_CLEAN)
1317 * Only pages without mappings or that have a
1318 * ->migratepage callback are possible to migrate
1321 mapping = page_mapping(page);
1322 if (mapping && !mapping->a_ops->migratepage)
1327 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1330 if (likely(get_page_unless_zero(page))) {
1332 * Be careful not to clear PageLRU until after we're
1333 * sure the page is not being freed elsewhere -- the
1334 * page release code relies on it.
1344 * zone->lru_lock is heavily contended. Some of the functions that
1345 * shrink the lists perform better by taking out a batch of pages
1346 * and working on them outside the LRU lock.
1348 * For pagecache intensive workloads, this function is the hottest
1349 * spot in the kernel (apart from copy_*_user functions).
1351 * Appropriate locks must be held before calling this function.
1353 * @nr_to_scan: The number of pages to look through on the list.
1354 * @lruvec: The LRU vector to pull pages from.
1355 * @dst: The temp list to put pages on to.
1356 * @nr_scanned: The number of pages that were scanned.
1357 * @sc: The scan_control struct for this reclaim session
1358 * @mode: One of the LRU isolation modes
1359 * @lru: LRU list id for isolating
1361 * returns how many pages were moved onto *@dst.
1363 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1364 struct lruvec *lruvec, struct list_head *dst,
1365 unsigned long *nr_scanned, struct scan_control *sc,
1366 isolate_mode_t mode, enum lru_list lru)
1368 struct list_head *src = &lruvec->lists[lru];
1369 unsigned long nr_taken = 0;
1372 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1373 !list_empty(src); scan++) {
1377 page = lru_to_page(src);
1378 prefetchw_prev_lru_page(page, src, flags);
1380 VM_BUG_ON_PAGE(!PageLRU(page), page);
1382 switch (__isolate_lru_page(page, mode)) {
1384 nr_pages = hpage_nr_pages(page);
1385 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1386 list_move(&page->lru, dst);
1387 nr_taken += nr_pages;
1391 /* else it is being freed elsewhere */
1392 list_move(&page->lru, src);
1401 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1402 nr_taken, mode, is_file_lru(lru));
1407 * isolate_lru_page - tries to isolate a page from its LRU list
1408 * @page: page to isolate from its LRU list
1410 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1411 * vmstat statistic corresponding to whatever LRU list the page was on.
1413 * Returns 0 if the page was removed from an LRU list.
1414 * Returns -EBUSY if the page was not on an LRU list.
1416 * The returned page will have PageLRU() cleared. If it was found on
1417 * the active list, it will have PageActive set. If it was found on
1418 * the unevictable list, it will have the PageUnevictable bit set. That flag
1419 * may need to be cleared by the caller before letting the page go.
1421 * The vmstat statistic corresponding to the list on which the page was
1422 * found will be decremented.
1425 * (1) Must be called with an elevated refcount on the page. This is a
1426 * fundamentnal difference from isolate_lru_pages (which is called
1427 * without a stable reference).
1428 * (2) the lru_lock must not be held.
1429 * (3) interrupts must be enabled.
1431 int isolate_lru_page(struct page *page)
1435 VM_BUG_ON_PAGE(!page_count(page), page);
1437 if (PageLRU(page)) {
1438 struct zone *zone = page_zone(page);
1439 struct lruvec *lruvec;
1441 spin_lock_irq(&zone->lru_lock);
1442 lruvec = mem_cgroup_page_lruvec(page, zone);
1443 if (PageLRU(page)) {
1444 int lru = page_lru(page);
1447 del_page_from_lru_list(page, lruvec, lru);
1450 spin_unlock_irq(&zone->lru_lock);
1456 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1457 * then get resheduled. When there are massive number of tasks doing page
1458 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1459 * the LRU list will go small and be scanned faster than necessary, leading to
1460 * unnecessary swapping, thrashing and OOM.
1462 static int too_many_isolated(struct zone *zone, int file,
1463 struct scan_control *sc)
1465 unsigned long inactive, isolated;
1467 if (current_is_kswapd())
1470 if (!sane_reclaim(sc))
1474 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1475 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1477 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1478 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1482 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1483 * won't get blocked by normal direct-reclaimers, forming a circular
1486 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1489 return isolated > inactive;
1492 static noinline_for_stack void
1493 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1495 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1496 struct zone *zone = lruvec_zone(lruvec);
1497 LIST_HEAD(pages_to_free);
1500 * Put back any unfreeable pages.
1502 while (!list_empty(page_list)) {
1503 struct page *page = lru_to_page(page_list);
1506 VM_BUG_ON_PAGE(PageLRU(page), page);
1507 list_del(&page->lru);
1508 if (unlikely(!page_evictable(page))) {
1509 spin_unlock_irq(&zone->lru_lock);
1510 putback_lru_page(page);
1511 spin_lock_irq(&zone->lru_lock);
1515 lruvec = mem_cgroup_page_lruvec(page, zone);
1518 lru = page_lru(page);
1519 add_page_to_lru_list(page, lruvec, lru);
1521 if (is_active_lru(lru)) {
1522 int file = is_file_lru(lru);
1523 int numpages = hpage_nr_pages(page);
1524 reclaim_stat->recent_rotated[file] += numpages;
1526 if (put_page_testzero(page)) {
1527 __ClearPageLRU(page);
1528 __ClearPageActive(page);
1529 del_page_from_lru_list(page, lruvec, lru);
1531 if (unlikely(PageCompound(page))) {
1532 spin_unlock_irq(&zone->lru_lock);
1533 mem_cgroup_uncharge(page);
1534 (*get_compound_page_dtor(page))(page);
1535 spin_lock_irq(&zone->lru_lock);
1537 list_add(&page->lru, &pages_to_free);
1542 * To save our caller's stack, now use input list for pages to free.
1544 list_splice(&pages_to_free, page_list);
1548 * If a kernel thread (such as nfsd for loop-back mounts) services
1549 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1550 * In that case we should only throttle if the backing device it is
1551 * writing to is congested. In other cases it is safe to throttle.
1553 static int current_may_throttle(void)
1555 return !(current->flags & PF_LESS_THROTTLE) ||
1556 current->backing_dev_info == NULL ||
1557 bdi_write_congested(current->backing_dev_info);
1561 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1562 * of reclaimed pages
1564 static noinline_for_stack unsigned long
1565 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1566 struct scan_control *sc, enum lru_list lru)
1568 LIST_HEAD(page_list);
1569 unsigned long nr_scanned;
1570 unsigned long nr_reclaimed = 0;
1571 unsigned long nr_taken;
1572 unsigned long nr_dirty = 0;
1573 unsigned long nr_congested = 0;
1574 unsigned long nr_unqueued_dirty = 0;
1575 unsigned long nr_writeback = 0;
1576 unsigned long nr_immediate = 0;
1577 isolate_mode_t isolate_mode = 0;
1578 int file = is_file_lru(lru);
1579 struct zone *zone = lruvec_zone(lruvec);
1580 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1582 while (unlikely(too_many_isolated(zone, file, sc))) {
1583 congestion_wait(BLK_RW_ASYNC, HZ/10);
1585 /* We are about to die and free our memory. Return now. */
1586 if (fatal_signal_pending(current))
1587 return SWAP_CLUSTER_MAX;
1593 isolate_mode |= ISOLATE_UNMAPPED;
1594 if (!sc->may_writepage)
1595 isolate_mode |= ISOLATE_CLEAN;
1597 spin_lock_irq(&zone->lru_lock);
1599 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1600 &nr_scanned, sc, isolate_mode, lru);
1602 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1603 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1605 if (global_reclaim(sc)) {
1606 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1607 if (current_is_kswapd())
1608 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1610 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1612 spin_unlock_irq(&zone->lru_lock);
1617 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1618 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1619 &nr_writeback, &nr_immediate,
1622 spin_lock_irq(&zone->lru_lock);
1624 reclaim_stat->recent_scanned[file] += nr_taken;
1626 if (global_reclaim(sc)) {
1627 if (current_is_kswapd())
1628 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1631 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1635 putback_inactive_pages(lruvec, &page_list);
1637 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1639 spin_unlock_irq(&zone->lru_lock);
1641 mem_cgroup_uncharge_list(&page_list);
1642 free_hot_cold_page_list(&page_list, true);
1645 * If reclaim is isolating dirty pages under writeback, it implies
1646 * that the long-lived page allocation rate is exceeding the page
1647 * laundering rate. Either the global limits are not being effective
1648 * at throttling processes due to the page distribution throughout
1649 * zones or there is heavy usage of a slow backing device. The
1650 * only option is to throttle from reclaim context which is not ideal
1651 * as there is no guarantee the dirtying process is throttled in the
1652 * same way balance_dirty_pages() manages.
1654 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1655 * of pages under pages flagged for immediate reclaim and stall if any
1656 * are encountered in the nr_immediate check below.
1658 if (nr_writeback && nr_writeback == nr_taken)
1659 set_bit(ZONE_WRITEBACK, &zone->flags);
1662 * Legacy memcg will stall in page writeback so avoid forcibly
1665 if (sane_reclaim(sc)) {
1667 * Tag a zone as congested if all the dirty pages scanned were
1668 * backed by a congested BDI and wait_iff_congested will stall.
1670 if (nr_dirty && nr_dirty == nr_congested)
1671 set_bit(ZONE_CONGESTED, &zone->flags);
1674 * If dirty pages are scanned that are not queued for IO, it
1675 * implies that flushers are not keeping up. In this case, flag
1676 * the zone ZONE_DIRTY and kswapd will start writing pages from
1679 if (nr_unqueued_dirty == nr_taken)
1680 set_bit(ZONE_DIRTY, &zone->flags);
1683 * If kswapd scans pages marked marked for immediate
1684 * reclaim and under writeback (nr_immediate), it implies
1685 * that pages are cycling through the LRU faster than
1686 * they are written so also forcibly stall.
1688 if (nr_immediate && current_may_throttle())
1689 congestion_wait(BLK_RW_ASYNC, HZ/10);
1693 * Stall direct reclaim for IO completions if underlying BDIs or zone
1694 * is congested. Allow kswapd to continue until it starts encountering
1695 * unqueued dirty pages or cycling through the LRU too quickly.
1697 if (!sc->hibernation_mode && !current_is_kswapd() &&
1698 current_may_throttle())
1699 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1701 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1703 nr_scanned, nr_reclaimed,
1705 trace_shrink_flags(file));
1706 return nr_reclaimed;
1710 * This moves pages from the active list to the inactive list.
1712 * We move them the other way if the page is referenced by one or more
1713 * processes, from rmap.
1715 * If the pages are mostly unmapped, the processing is fast and it is
1716 * appropriate to hold zone->lru_lock across the whole operation. But if
1717 * the pages are mapped, the processing is slow (page_referenced()) so we
1718 * should drop zone->lru_lock around each page. It's impossible to balance
1719 * this, so instead we remove the pages from the LRU while processing them.
1720 * It is safe to rely on PG_active against the non-LRU pages in here because
1721 * nobody will play with that bit on a non-LRU page.
1723 * The downside is that we have to touch page->_count against each page.
1724 * But we had to alter page->flags anyway.
1727 static void move_active_pages_to_lru(struct lruvec *lruvec,
1728 struct list_head *list,
1729 struct list_head *pages_to_free,
1732 struct zone *zone = lruvec_zone(lruvec);
1733 unsigned long pgmoved = 0;
1737 while (!list_empty(list)) {
1738 page = lru_to_page(list);
1739 lruvec = mem_cgroup_page_lruvec(page, zone);
1741 VM_BUG_ON_PAGE(PageLRU(page), page);
1744 nr_pages = hpage_nr_pages(page);
1745 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1746 list_move(&page->lru, &lruvec->lists[lru]);
1747 pgmoved += nr_pages;
1749 if (put_page_testzero(page)) {
1750 __ClearPageLRU(page);
1751 __ClearPageActive(page);
1752 del_page_from_lru_list(page, lruvec, lru);
1754 if (unlikely(PageCompound(page))) {
1755 spin_unlock_irq(&zone->lru_lock);
1756 mem_cgroup_uncharge(page);
1757 (*get_compound_page_dtor(page))(page);
1758 spin_lock_irq(&zone->lru_lock);
1760 list_add(&page->lru, pages_to_free);
1763 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1764 if (!is_active_lru(lru))
1765 __count_vm_events(PGDEACTIVATE, pgmoved);
1768 static void shrink_active_list(unsigned long nr_to_scan,
1769 struct lruvec *lruvec,
1770 struct scan_control *sc,
1773 unsigned long nr_taken;
1774 unsigned long nr_scanned;
1775 unsigned long vm_flags;
1776 LIST_HEAD(l_hold); /* The pages which were snipped off */
1777 LIST_HEAD(l_active);
1778 LIST_HEAD(l_inactive);
1780 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1781 unsigned long nr_rotated = 0;
1782 isolate_mode_t isolate_mode = 0;
1783 int file = is_file_lru(lru);
1784 struct zone *zone = lruvec_zone(lruvec);
1789 isolate_mode |= ISOLATE_UNMAPPED;
1790 if (!sc->may_writepage)
1791 isolate_mode |= ISOLATE_CLEAN;
1793 spin_lock_irq(&zone->lru_lock);
1795 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1796 &nr_scanned, sc, isolate_mode, lru);
1797 if (global_reclaim(sc))
1798 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1800 reclaim_stat->recent_scanned[file] += nr_taken;
1802 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1803 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1804 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1805 spin_unlock_irq(&zone->lru_lock);
1807 while (!list_empty(&l_hold)) {
1809 page = lru_to_page(&l_hold);
1810 list_del(&page->lru);
1812 if (unlikely(!page_evictable(page))) {
1813 putback_lru_page(page);
1817 if (unlikely(buffer_heads_over_limit)) {
1818 if (page_has_private(page) && trylock_page(page)) {
1819 if (page_has_private(page))
1820 try_to_release_page(page, 0);
1825 if (page_referenced(page, 0, sc->target_mem_cgroup,
1827 nr_rotated += hpage_nr_pages(page);
1829 * Identify referenced, file-backed active pages and
1830 * give them one more trip around the active list. So
1831 * that executable code get better chances to stay in
1832 * memory under moderate memory pressure. Anon pages
1833 * are not likely to be evicted by use-once streaming
1834 * IO, plus JVM can create lots of anon VM_EXEC pages,
1835 * so we ignore them here.
1837 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1838 list_add(&page->lru, &l_active);
1843 ClearPageActive(page); /* we are de-activating */
1844 list_add(&page->lru, &l_inactive);
1848 * Move pages back to the lru list.
1850 spin_lock_irq(&zone->lru_lock);
1852 * Count referenced pages from currently used mappings as rotated,
1853 * even though only some of them are actually re-activated. This
1854 * helps balance scan pressure between file and anonymous pages in
1857 reclaim_stat->recent_rotated[file] += nr_rotated;
1859 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1860 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1861 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1862 spin_unlock_irq(&zone->lru_lock);
1864 mem_cgroup_uncharge_list(&l_hold);
1865 free_hot_cold_page_list(&l_hold, true);
1869 static bool inactive_anon_is_low_global(struct zone *zone)
1871 unsigned long active, inactive;
1873 active = zone_page_state(zone, NR_ACTIVE_ANON);
1874 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1876 return inactive * zone->inactive_ratio < active;
1880 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1881 * @lruvec: LRU vector to check
1883 * Returns true if the zone does not have enough inactive anon pages,
1884 * meaning some active anon pages need to be deactivated.
1886 static bool inactive_anon_is_low(struct lruvec *lruvec)
1889 * If we don't have swap space, anonymous page deactivation
1892 if (!total_swap_pages)
1895 if (!mem_cgroup_disabled())
1896 return mem_cgroup_inactive_anon_is_low(lruvec);
1898 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1901 static inline bool inactive_anon_is_low(struct lruvec *lruvec)
1908 * inactive_file_is_low - check if file pages need to be deactivated
1909 * @lruvec: LRU vector to check
1911 * When the system is doing streaming IO, memory pressure here
1912 * ensures that active file pages get deactivated, until more
1913 * than half of the file pages are on the inactive list.
1915 * Once we get to that situation, protect the system's working
1916 * set from being evicted by disabling active file page aging.
1918 * This uses a different ratio than the anonymous pages, because
1919 * the page cache uses a use-once replacement algorithm.
1921 static bool inactive_file_is_low(struct lruvec *lruvec)
1923 unsigned long inactive;
1924 unsigned long active;
1926 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1927 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1929 return active > inactive;
1932 static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1934 if (is_file_lru(lru))
1935 return inactive_file_is_low(lruvec);
1937 return inactive_anon_is_low(lruvec);
1940 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1941 struct lruvec *lruvec, struct scan_control *sc)
1943 if (is_active_lru(lru)) {
1944 if (inactive_list_is_low(lruvec, lru))
1945 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1949 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1960 * Determine how aggressively the anon and file LRU lists should be
1961 * scanned. The relative value of each set of LRU lists is determined
1962 * by looking at the fraction of the pages scanned we did rotate back
1963 * onto the active list instead of evict.
1965 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1966 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1968 static void get_scan_count(struct lruvec *lruvec, int swappiness,
1969 struct scan_control *sc, unsigned long *nr,
1970 unsigned long *lru_pages)
1972 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1974 u64 denominator = 0; /* gcc */
1975 struct zone *zone = lruvec_zone(lruvec);
1976 unsigned long anon_prio, file_prio;
1977 enum scan_balance scan_balance;
1978 unsigned long anon, file;
1979 bool force_scan = false;
1980 unsigned long ap, fp;
1986 * If the zone or memcg is small, nr[l] can be 0. This
1987 * results in no scanning on this priority and a potential
1988 * priority drop. Global direct reclaim can go to the next
1989 * zone and tends to have no problems. Global kswapd is for
1990 * zone balancing and it needs to scan a minimum amount. When
1991 * reclaiming for a memcg, a priority drop can cause high
1992 * latencies, so it's better to scan a minimum amount there as
1995 if (current_is_kswapd()) {
1996 if (!zone_reclaimable(zone))
1998 if (!mem_cgroup_lruvec_online(lruvec))
2001 if (!global_reclaim(sc))
2004 /* If we have no swap space, do not bother scanning anon pages. */
2005 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
2006 scan_balance = SCAN_FILE;
2011 * Global reclaim will swap to prevent OOM even with no
2012 * swappiness, but memcg users want to use this knob to
2013 * disable swapping for individual groups completely when
2014 * using the memory controller's swap limit feature would be
2017 if (!global_reclaim(sc) && !swappiness) {
2018 scan_balance = SCAN_FILE;
2023 * Do not apply any pressure balancing cleverness when the
2024 * system is close to OOM, scan both anon and file equally
2025 * (unless the swappiness setting disagrees with swapping).
2027 if (!sc->priority && swappiness) {
2028 scan_balance = SCAN_EQUAL;
2033 * Prevent the reclaimer from falling into the cache trap: as
2034 * cache pages start out inactive, every cache fault will tip
2035 * the scan balance towards the file LRU. And as the file LRU
2036 * shrinks, so does the window for rotation from references.
2037 * This means we have a runaway feedback loop where a tiny
2038 * thrashing file LRU becomes infinitely more attractive than
2039 * anon pages. Try to detect this based on file LRU size.
2041 if (global_reclaim(sc)) {
2042 unsigned long zonefile;
2043 unsigned long zonefree;
2045 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2046 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2047 zone_page_state(zone, NR_INACTIVE_FILE);
2049 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2050 scan_balance = SCAN_ANON;
2056 * There is enough inactive page cache, do not reclaim
2057 * anything from the anonymous working set right now.
2059 if (!inactive_file_is_low(lruvec)) {
2060 scan_balance = SCAN_FILE;
2064 scan_balance = SCAN_FRACT;
2067 * With swappiness at 100, anonymous and file have the same priority.
2068 * This scanning priority is essentially the inverse of IO cost.
2070 anon_prio = swappiness;
2071 file_prio = 200 - anon_prio;
2074 * OK, so we have swap space and a fair amount of page cache
2075 * pages. We use the recently rotated / recently scanned
2076 * ratios to determine how valuable each cache is.
2078 * Because workloads change over time (and to avoid overflow)
2079 * we keep these statistics as a floating average, which ends
2080 * up weighing recent references more than old ones.
2082 * anon in [0], file in [1]
2085 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2086 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2087 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2088 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2090 spin_lock_irq(&zone->lru_lock);
2091 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2092 reclaim_stat->recent_scanned[0] /= 2;
2093 reclaim_stat->recent_rotated[0] /= 2;
2096 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2097 reclaim_stat->recent_scanned[1] /= 2;
2098 reclaim_stat->recent_rotated[1] /= 2;
2102 * The amount of pressure on anon vs file pages is inversely
2103 * proportional to the fraction of recently scanned pages on
2104 * each list that were recently referenced and in active use.
2106 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2107 ap /= reclaim_stat->recent_rotated[0] + 1;
2109 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2110 fp /= reclaim_stat->recent_rotated[1] + 1;
2111 spin_unlock_irq(&zone->lru_lock);
2115 denominator = ap + fp + 1;
2117 some_scanned = false;
2118 /* Only use force_scan on second pass. */
2119 for (pass = 0; !some_scanned && pass < 2; pass++) {
2121 for_each_evictable_lru(lru) {
2122 int file = is_file_lru(lru);
2126 size = get_lru_size(lruvec, lru);
2127 scan = size >> sc->priority;
2129 if (!scan && pass && force_scan)
2130 scan = min(size, SWAP_CLUSTER_MAX);
2132 switch (scan_balance) {
2134 /* Scan lists relative to size */
2138 * Scan types proportional to swappiness and
2139 * their relative recent reclaim efficiency.
2141 scan = div64_u64(scan * fraction[file],
2146 /* Scan one type exclusively */
2147 if ((scan_balance == SCAN_FILE) != file) {
2153 /* Look ma, no brain */
2161 * Skip the second pass and don't force_scan,
2162 * if we found something to scan.
2164 some_scanned |= !!scan;
2169 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2170 static void init_tlb_ubc(void)
2173 * This deliberately does not clear the cpumask as it's expensive
2174 * and unnecessary. If there happens to be data in there then the
2175 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2176 * then will be cleared.
2178 current->tlb_ubc.flush_required = false;
2181 static inline void init_tlb_ubc(void)
2184 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2187 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2189 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2190 struct scan_control *sc, unsigned long *lru_pages)
2192 unsigned long nr[NR_LRU_LISTS];
2193 unsigned long targets[NR_LRU_LISTS];
2194 unsigned long nr_to_scan;
2196 unsigned long nr_reclaimed = 0;
2197 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2198 struct blk_plug plug;
2201 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2203 /* Record the original scan target for proportional adjustments later */
2204 memcpy(targets, nr, sizeof(nr));
2207 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2208 * event that can occur when there is little memory pressure e.g.
2209 * multiple streaming readers/writers. Hence, we do not abort scanning
2210 * when the requested number of pages are reclaimed when scanning at
2211 * DEF_PRIORITY on the assumption that the fact we are direct
2212 * reclaiming implies that kswapd is not keeping up and it is best to
2213 * do a batch of work at once. For memcg reclaim one check is made to
2214 * abort proportional reclaim if either the file or anon lru has already
2215 * dropped to zero at the first pass.
2217 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2218 sc->priority == DEF_PRIORITY);
2222 blk_start_plug(&plug);
2223 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2224 nr[LRU_INACTIVE_FILE]) {
2225 unsigned long nr_anon, nr_file, percentage;
2226 unsigned long nr_scanned;
2228 for_each_evictable_lru(lru) {
2230 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2231 nr[lru] -= nr_to_scan;
2233 nr_reclaimed += shrink_list(lru, nr_to_scan,
2238 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2242 * For kswapd and memcg, reclaim at least the number of pages
2243 * requested. Ensure that the anon and file LRUs are scanned
2244 * proportionally what was requested by get_scan_count(). We
2245 * stop reclaiming one LRU and reduce the amount scanning
2246 * proportional to the original scan target.
2248 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2249 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2252 * It's just vindictive to attack the larger once the smaller
2253 * has gone to zero. And given the way we stop scanning the
2254 * smaller below, this makes sure that we only make one nudge
2255 * towards proportionality once we've got nr_to_reclaim.
2257 if (!nr_file || !nr_anon)
2260 if (nr_file > nr_anon) {
2261 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2262 targets[LRU_ACTIVE_ANON] + 1;
2264 percentage = nr_anon * 100 / scan_target;
2266 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2267 targets[LRU_ACTIVE_FILE] + 1;
2269 percentage = nr_file * 100 / scan_target;
2272 /* Stop scanning the smaller of the LRU */
2274 nr[lru + LRU_ACTIVE] = 0;
2277 * Recalculate the other LRU scan count based on its original
2278 * scan target and the percentage scanning already complete
2280 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2281 nr_scanned = targets[lru] - nr[lru];
2282 nr[lru] = targets[lru] * (100 - percentage) / 100;
2283 nr[lru] -= min(nr[lru], nr_scanned);
2286 nr_scanned = targets[lru] - nr[lru];
2287 nr[lru] = targets[lru] * (100 - percentage) / 100;
2288 nr[lru] -= min(nr[lru], nr_scanned);
2290 scan_adjusted = true;
2292 blk_finish_plug(&plug);
2293 sc->nr_reclaimed += nr_reclaimed;
2296 * Even if we did not try to evict anon pages at all, we want to
2297 * rebalance the anon lru active/inactive ratio.
2299 if (inactive_anon_is_low(lruvec))
2300 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2301 sc, LRU_ACTIVE_ANON);
2303 throttle_vm_writeout(sc->gfp_mask);
2306 /* Use reclaim/compaction for costly allocs or under memory pressure */
2307 static bool in_reclaim_compaction(struct scan_control *sc)
2309 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2310 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2311 sc->priority < DEF_PRIORITY - 2))
2318 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2319 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2320 * true if more pages should be reclaimed such that when the page allocator
2321 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2322 * It will give up earlier than that if there is difficulty reclaiming pages.
2324 static inline bool should_continue_reclaim(struct zone *zone,
2325 unsigned long nr_reclaimed,
2326 unsigned long nr_scanned,
2327 struct scan_control *sc)
2329 unsigned long pages_for_compaction;
2330 unsigned long inactive_lru_pages;
2332 /* If not in reclaim/compaction mode, stop */
2333 if (!in_reclaim_compaction(sc))
2336 /* Consider stopping depending on scan and reclaim activity */
2337 if (sc->gfp_mask & __GFP_REPEAT) {
2339 * For __GFP_REPEAT allocations, stop reclaiming if the
2340 * full LRU list has been scanned and we are still failing
2341 * to reclaim pages. This full LRU scan is potentially
2342 * expensive but a __GFP_REPEAT caller really wants to succeed
2344 if (!nr_reclaimed && !nr_scanned)
2348 * For non-__GFP_REPEAT allocations which can presumably
2349 * fail without consequence, stop if we failed to reclaim
2350 * any pages from the last SWAP_CLUSTER_MAX number of
2351 * pages that were scanned. This will return to the
2352 * caller faster at the risk reclaim/compaction and
2353 * the resulting allocation attempt fails
2360 * If we have not reclaimed enough pages for compaction and the
2361 * inactive lists are large enough, continue reclaiming
2363 pages_for_compaction = (2UL << sc->order);
2364 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2365 if (get_nr_swap_pages() > 0)
2366 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2367 if (sc->nr_reclaimed < pages_for_compaction &&
2368 inactive_lru_pages > pages_for_compaction)
2371 /* If compaction would go ahead or the allocation would succeed, stop */
2372 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2373 case COMPACT_PARTIAL:
2374 case COMPACT_CONTINUE:
2381 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2384 struct reclaim_state *reclaim_state = current->reclaim_state;
2385 unsigned long nr_reclaimed, nr_scanned;
2386 bool reclaimable = false;
2389 struct mem_cgroup *root = sc->target_mem_cgroup;
2390 struct mem_cgroup_reclaim_cookie reclaim = {
2392 .priority = sc->priority,
2394 unsigned long zone_lru_pages = 0;
2395 struct mem_cgroup *memcg;
2397 nr_reclaimed = sc->nr_reclaimed;
2398 nr_scanned = sc->nr_scanned;
2400 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2402 unsigned long lru_pages;
2403 unsigned long scanned;
2404 struct lruvec *lruvec;
2407 if (mem_cgroup_low(root, memcg)) {
2408 if (!sc->may_thrash)
2410 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2413 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2414 swappiness = mem_cgroup_swappiness(memcg);
2415 scanned = sc->nr_scanned;
2417 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2418 zone_lru_pages += lru_pages;
2420 if (memcg && is_classzone)
2421 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2422 memcg, sc->nr_scanned - scanned,
2426 * Direct reclaim and kswapd have to scan all memory
2427 * cgroups to fulfill the overall scan target for the
2430 * Limit reclaim, on the other hand, only cares about
2431 * nr_to_reclaim pages to be reclaimed and it will
2432 * retry with decreasing priority if one round over the
2433 * whole hierarchy is not sufficient.
2435 if (!global_reclaim(sc) &&
2436 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2437 mem_cgroup_iter_break(root, memcg);
2440 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2443 * Shrink the slab caches in the same proportion that
2444 * the eligible LRU pages were scanned.
2446 if (global_reclaim(sc) && is_classzone)
2447 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2448 sc->nr_scanned - nr_scanned,
2451 if (reclaim_state) {
2452 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2453 reclaim_state->reclaimed_slab = 0;
2456 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2457 sc->nr_scanned - nr_scanned,
2458 sc->nr_reclaimed - nr_reclaimed);
2460 if (sc->nr_reclaimed - nr_reclaimed)
2463 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2464 sc->nr_scanned - nr_scanned, sc));
2470 * Returns true if compaction should go ahead for a high-order request, or
2471 * the high-order allocation would succeed without compaction.
2473 static inline bool compaction_ready(struct zone *zone, int order)
2475 unsigned long balance_gap, watermark;
2479 * Compaction takes time to run and there are potentially other
2480 * callers using the pages just freed. Continue reclaiming until
2481 * there is a buffer of free pages available to give compaction
2482 * a reasonable chance of completing and allocating the page
2484 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2485 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2486 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2487 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2490 * If compaction is deferred, reclaim up to a point where
2491 * compaction will have a chance of success when re-enabled
2493 if (compaction_deferred(zone, order))
2494 return watermark_ok;
2497 * If compaction is not ready to start and allocation is not likely
2498 * to succeed without it, then keep reclaiming.
2500 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2503 return watermark_ok;
2507 * This is the direct reclaim path, for page-allocating processes. We only
2508 * try to reclaim pages from zones which will satisfy the caller's allocation
2511 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2513 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2515 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2516 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2517 * zone defense algorithm.
2519 * If a zone is deemed to be full of pinned pages then just give it a light
2520 * scan then give up on it.
2522 * Returns true if a zone was reclaimable.
2524 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2528 unsigned long nr_soft_reclaimed;
2529 unsigned long nr_soft_scanned;
2531 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2532 bool reclaimable = false;
2535 * If the number of buffer_heads in the machine exceeds the maximum
2536 * allowed level, force direct reclaim to scan the highmem zone as
2537 * highmem pages could be pinning lowmem pages storing buffer_heads
2539 orig_mask = sc->gfp_mask;
2540 if (buffer_heads_over_limit)
2541 sc->gfp_mask |= __GFP_HIGHMEM;
2543 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2544 requested_highidx, sc->nodemask) {
2545 enum zone_type classzone_idx;
2547 if (!populated_zone(zone))
2550 classzone_idx = requested_highidx;
2551 while (!populated_zone(zone->zone_pgdat->node_zones +
2556 * Take care memory controller reclaiming has small influence
2559 if (global_reclaim(sc)) {
2560 if (!cpuset_zone_allowed(zone,
2561 GFP_KERNEL | __GFP_HARDWALL))
2564 if (sc->priority != DEF_PRIORITY &&
2565 !zone_reclaimable(zone))
2566 continue; /* Let kswapd poll it */
2569 * If we already have plenty of memory free for
2570 * compaction in this zone, don't free any more.
2571 * Even though compaction is invoked for any
2572 * non-zero order, only frequent costly order
2573 * reclamation is disruptive enough to become a
2574 * noticeable problem, like transparent huge
2577 if (IS_ENABLED(CONFIG_COMPACTION) &&
2578 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2579 zonelist_zone_idx(z) <= requested_highidx &&
2580 compaction_ready(zone, sc->order)) {
2581 sc->compaction_ready = true;
2586 * This steals pages from memory cgroups over softlimit
2587 * and returns the number of reclaimed pages and
2588 * scanned pages. This works for global memory pressure
2589 * and balancing, not for a memcg's limit.
2591 nr_soft_scanned = 0;
2592 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2593 sc->order, sc->gfp_mask,
2595 sc->nr_reclaimed += nr_soft_reclaimed;
2596 sc->nr_scanned += nr_soft_scanned;
2597 if (nr_soft_reclaimed)
2599 /* need some check for avoid more shrink_zone() */
2602 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2605 if (global_reclaim(sc) &&
2606 !reclaimable && zone_reclaimable(zone))
2611 * Restore to original mask to avoid the impact on the caller if we
2612 * promoted it to __GFP_HIGHMEM.
2614 sc->gfp_mask = orig_mask;
2620 * This is the main entry point to direct page reclaim.
2622 * If a full scan of the inactive list fails to free enough memory then we
2623 * are "out of memory" and something needs to be killed.
2625 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2626 * high - the zone may be full of dirty or under-writeback pages, which this
2627 * caller can't do much about. We kick the writeback threads and take explicit
2628 * naps in the hope that some of these pages can be written. But if the
2629 * allocating task holds filesystem locks which prevent writeout this might not
2630 * work, and the allocation attempt will fail.
2632 * returns: 0, if no pages reclaimed
2633 * else, the number of pages reclaimed
2635 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2636 struct scan_control *sc)
2638 int initial_priority = sc->priority;
2639 unsigned long total_scanned = 0;
2640 unsigned long writeback_threshold;
2641 bool zones_reclaimable;
2643 delayacct_freepages_start();
2645 if (global_reclaim(sc))
2646 count_vm_event(ALLOCSTALL);
2649 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2652 zones_reclaimable = shrink_zones(zonelist, sc);
2654 total_scanned += sc->nr_scanned;
2655 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2658 if (sc->compaction_ready)
2662 * If we're getting trouble reclaiming, start doing
2663 * writepage even in laptop mode.
2665 if (sc->priority < DEF_PRIORITY - 2)
2666 sc->may_writepage = 1;
2669 * Try to write back as many pages as we just scanned. This
2670 * tends to cause slow streaming writers to write data to the
2671 * disk smoothly, at the dirtying rate, which is nice. But
2672 * that's undesirable in laptop mode, where we *want* lumpy
2673 * writeout. So in laptop mode, write out the whole world.
2675 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2676 if (total_scanned > writeback_threshold) {
2677 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2678 WB_REASON_TRY_TO_FREE_PAGES);
2679 sc->may_writepage = 1;
2681 } while (--sc->priority >= 0);
2683 delayacct_freepages_end();
2685 if (sc->nr_reclaimed)
2686 return sc->nr_reclaimed;
2688 /* Aborted reclaim to try compaction? don't OOM, then */
2689 if (sc->compaction_ready)
2692 /* Untapped cgroup reserves? Don't OOM, retry. */
2693 if (!sc->may_thrash) {
2694 sc->priority = initial_priority;
2699 /* Any of the zones still reclaimable? Don't OOM. */
2700 if (zones_reclaimable)
2706 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2709 unsigned long pfmemalloc_reserve = 0;
2710 unsigned long free_pages = 0;
2714 for (i = 0; i <= ZONE_NORMAL; i++) {
2715 zone = &pgdat->node_zones[i];
2716 if (!populated_zone(zone) ||
2717 zone_reclaimable_pages(zone) == 0)
2720 pfmemalloc_reserve += min_wmark_pages(zone);
2721 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2724 /* If there are no reserves (unexpected config) then do not throttle */
2725 if (!pfmemalloc_reserve)
2728 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2730 /* kswapd must be awake if processes are being throttled */
2731 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2732 pgdat->classzone_idx = min(pgdat->classzone_idx,
2733 (enum zone_type)ZONE_NORMAL);
2734 wake_up_interruptible(&pgdat->kswapd_wait);
2741 * Throttle direct reclaimers if backing storage is backed by the network
2742 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2743 * depleted. kswapd will continue to make progress and wake the processes
2744 * when the low watermark is reached.
2746 * Returns true if a fatal signal was delivered during throttling. If this
2747 * happens, the page allocator should not consider triggering the OOM killer.
2749 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2750 nodemask_t *nodemask)
2754 pg_data_t *pgdat = NULL;
2757 * Kernel threads should not be throttled as they may be indirectly
2758 * responsible for cleaning pages necessary for reclaim to make forward
2759 * progress. kjournald for example may enter direct reclaim while
2760 * committing a transaction where throttling it could forcing other
2761 * processes to block on log_wait_commit().
2763 if (current->flags & PF_KTHREAD)
2767 * If a fatal signal is pending, this process should not throttle.
2768 * It should return quickly so it can exit and free its memory
2770 if (fatal_signal_pending(current))
2774 * Check if the pfmemalloc reserves are ok by finding the first node
2775 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2776 * GFP_KERNEL will be required for allocating network buffers when
2777 * swapping over the network so ZONE_HIGHMEM is unusable.
2779 * Throttling is based on the first usable node and throttled processes
2780 * wait on a queue until kswapd makes progress and wakes them. There
2781 * is an affinity then between processes waking up and where reclaim
2782 * progress has been made assuming the process wakes on the same node.
2783 * More importantly, processes running on remote nodes will not compete
2784 * for remote pfmemalloc reserves and processes on different nodes
2785 * should make reasonable progress.
2787 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2788 gfp_zone(gfp_mask), nodemask) {
2789 if (zone_idx(zone) > ZONE_NORMAL)
2792 /* Throttle based on the first usable node */
2793 pgdat = zone->zone_pgdat;
2794 if (pfmemalloc_watermark_ok(pgdat))
2799 /* If no zone was usable by the allocation flags then do not throttle */
2803 /* Account for the throttling */
2804 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2807 * If the caller cannot enter the filesystem, it's possible that it
2808 * is due to the caller holding an FS lock or performing a journal
2809 * transaction in the case of a filesystem like ext[3|4]. In this case,
2810 * it is not safe to block on pfmemalloc_wait as kswapd could be
2811 * blocked waiting on the same lock. Instead, throttle for up to a
2812 * second before continuing.
2814 if (!(gfp_mask & __GFP_FS)) {
2815 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2816 pfmemalloc_watermark_ok(pgdat), HZ);
2821 /* Throttle until kswapd wakes the process */
2822 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2823 pfmemalloc_watermark_ok(pgdat));
2826 if (fatal_signal_pending(current))
2833 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2834 gfp_t gfp_mask, nodemask_t *nodemask)
2836 unsigned long nr_reclaimed;
2837 struct scan_control sc = {
2838 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2839 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2841 .nodemask = nodemask,
2842 .priority = DEF_PRIORITY,
2843 .may_writepage = !laptop_mode,
2849 * Do not enter reclaim if fatal signal was delivered while throttled.
2850 * 1 is returned so that the page allocator does not OOM kill at this
2853 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2856 trace_mm_vmscan_direct_reclaim_begin(order,
2860 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2862 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2864 return nr_reclaimed;
2869 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2870 gfp_t gfp_mask, bool noswap,
2872 unsigned long *nr_scanned)
2874 struct scan_control sc = {
2875 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2876 .target_mem_cgroup = memcg,
2877 .may_writepage = !laptop_mode,
2879 .may_swap = !noswap,
2881 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2882 int swappiness = mem_cgroup_swappiness(memcg);
2883 unsigned long lru_pages;
2885 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2886 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2888 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2893 * NOTE: Although we can get the priority field, using it
2894 * here is not a good idea, since it limits the pages we can scan.
2895 * if we don't reclaim here, the shrink_zone from balance_pgdat
2896 * will pick up pages from other mem cgroup's as well. We hack
2897 * the priority and make it zero.
2899 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
2901 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2903 *nr_scanned = sc.nr_scanned;
2904 return sc.nr_reclaimed;
2907 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2908 unsigned long nr_pages,
2912 struct zonelist *zonelist;
2913 unsigned long nr_reclaimed;
2915 struct scan_control sc = {
2916 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2917 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2918 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2919 .target_mem_cgroup = memcg,
2920 .priority = DEF_PRIORITY,
2921 .may_writepage = !laptop_mode,
2923 .may_swap = may_swap,
2927 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2928 * take care of from where we get pages. So the node where we start the
2929 * scan does not need to be the current node.
2931 nid = mem_cgroup_select_victim_node(memcg);
2933 zonelist = NODE_DATA(nid)->node_zonelists;
2935 trace_mm_vmscan_memcg_reclaim_begin(0,
2939 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2941 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2943 return nr_reclaimed;
2947 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2949 struct mem_cgroup *memcg;
2951 if (!total_swap_pages)
2954 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2956 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2958 if (inactive_anon_is_low(lruvec))
2959 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2960 sc, LRU_ACTIVE_ANON);
2962 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2966 static bool zone_balanced(struct zone *zone, int order,
2967 unsigned long balance_gap, int classzone_idx)
2969 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2970 balance_gap, classzone_idx))
2973 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2974 order, 0, classzone_idx) == COMPACT_SKIPPED)
2981 * pgdat_balanced() is used when checking if a node is balanced.
2983 * For order-0, all zones must be balanced!
2985 * For high-order allocations only zones that meet watermarks and are in a
2986 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2987 * total of balanced pages must be at least 25% of the zones allowed by
2988 * classzone_idx for the node to be considered balanced. Forcing all zones to
2989 * be balanced for high orders can cause excessive reclaim when there are
2991 * The choice of 25% is due to
2992 * o a 16M DMA zone that is balanced will not balance a zone on any
2993 * reasonable sized machine
2994 * o On all other machines, the top zone must be at least a reasonable
2995 * percentage of the middle zones. For example, on 32-bit x86, highmem
2996 * would need to be at least 256M for it to be balance a whole node.
2997 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2998 * to balance a node on its own. These seemed like reasonable ratios.
3000 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3002 unsigned long managed_pages = 0;
3003 unsigned long balanced_pages = 0;
3006 /* Check the watermark levels */
3007 for (i = 0; i <= classzone_idx; i++) {
3008 struct zone *zone = pgdat->node_zones + i;
3010 if (!populated_zone(zone))
3013 managed_pages += zone->managed_pages;
3016 * A special case here:
3018 * balance_pgdat() skips over all_unreclaimable after
3019 * DEF_PRIORITY. Effectively, it considers them balanced so
3020 * they must be considered balanced here as well!
3022 if (!zone_reclaimable(zone)) {
3023 balanced_pages += zone->managed_pages;
3027 if (zone_balanced(zone, order, 0, i))
3028 balanced_pages += zone->managed_pages;
3034 return balanced_pages >= (managed_pages >> 2);
3040 * Prepare kswapd for sleeping. This verifies that there are no processes
3041 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3043 * Returns true if kswapd is ready to sleep
3045 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3048 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3053 * The throttled processes are normally woken up in balance_pgdat() as
3054 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3055 * race between when kswapd checks the watermarks and a process gets
3056 * throttled. There is also a potential race if processes get
3057 * throttled, kswapd wakes, a large process exits thereby balancing the
3058 * zones, which causes kswapd to exit balance_pgdat() before reaching
3059 * the wake up checks. If kswapd is going to sleep, no process should
3060 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3061 * the wake up is premature, processes will wake kswapd and get
3062 * throttled again. The difference from wake ups in balance_pgdat() is
3063 * that here we are under prepare_to_wait().
3065 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3066 wake_up_all(&pgdat->pfmemalloc_wait);
3068 return pgdat_balanced(pgdat, order, classzone_idx);
3072 * kswapd shrinks the zone by the number of pages required to reach
3073 * the high watermark.
3075 * Returns true if kswapd scanned at least the requested number of pages to
3076 * reclaim or if the lack of progress was due to pages under writeback.
3077 * This is used to determine if the scanning priority needs to be raised.
3079 static bool kswapd_shrink_zone(struct zone *zone,
3081 struct scan_control *sc,
3082 unsigned long *nr_attempted)
3084 int testorder = sc->order;
3085 unsigned long balance_gap;
3086 bool lowmem_pressure;
3088 /* Reclaim above the high watermark. */
3089 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3092 * Kswapd reclaims only single pages with compaction enabled. Trying
3093 * too hard to reclaim until contiguous free pages have become
3094 * available can hurt performance by evicting too much useful data
3095 * from memory. Do not reclaim more than needed for compaction.
3097 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3098 compaction_suitable(zone, sc->order, 0, classzone_idx)
3103 * We put equal pressure on every zone, unless one zone has way too
3104 * many pages free already. The "too many pages" is defined as the
3105 * high wmark plus a "gap" where the gap is either the low
3106 * watermark or 1% of the zone, whichever is smaller.
3108 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3109 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3112 * If there is no low memory pressure or the zone is balanced then no
3113 * reclaim is necessary
3115 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3116 if (!lowmem_pressure && zone_balanced(zone, testorder,
3117 balance_gap, classzone_idx))
3120 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3122 /* Account for the number of pages attempted to reclaim */
3123 *nr_attempted += sc->nr_to_reclaim;
3125 clear_bit(ZONE_WRITEBACK, &zone->flags);
3128 * If a zone reaches its high watermark, consider it to be no longer
3129 * congested. It's possible there are dirty pages backed by congested
3130 * BDIs but as pressure is relieved, speculatively avoid congestion
3133 if (zone_reclaimable(zone) &&
3134 zone_balanced(zone, testorder, 0, classzone_idx)) {
3135 clear_bit(ZONE_CONGESTED, &zone->flags);
3136 clear_bit(ZONE_DIRTY, &zone->flags);
3139 return sc->nr_scanned >= sc->nr_to_reclaim;
3143 * For kswapd, balance_pgdat() will work across all this node's zones until
3144 * they are all at high_wmark_pages(zone).
3146 * Returns the final order kswapd was reclaiming at
3148 * There is special handling here for zones which are full of pinned pages.
3149 * This can happen if the pages are all mlocked, or if they are all used by
3150 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3151 * What we do is to detect the case where all pages in the zone have been
3152 * scanned twice and there has been zero successful reclaim. Mark the zone as
3153 * dead and from now on, only perform a short scan. Basically we're polling
3154 * the zone for when the problem goes away.
3156 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3157 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3158 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3159 * lower zones regardless of the number of free pages in the lower zones. This
3160 * interoperates with the page allocator fallback scheme to ensure that aging
3161 * of pages is balanced across the zones.
3163 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3167 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3168 unsigned long nr_soft_reclaimed;
3169 unsigned long nr_soft_scanned;
3170 struct scan_control sc = {
3171 .gfp_mask = GFP_KERNEL,
3173 .priority = DEF_PRIORITY,
3174 .may_writepage = !laptop_mode,
3178 count_vm_event(PAGEOUTRUN);
3181 unsigned long nr_attempted = 0;
3182 bool raise_priority = true;
3183 bool pgdat_needs_compaction = (order > 0);
3185 sc.nr_reclaimed = 0;
3188 * Scan in the highmem->dma direction for the highest
3189 * zone which needs scanning
3191 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3192 struct zone *zone = pgdat->node_zones + i;
3194 if (!populated_zone(zone))
3197 if (sc.priority != DEF_PRIORITY &&
3198 !zone_reclaimable(zone))
3202 * Do some background aging of the anon list, to give
3203 * pages a chance to be referenced before reclaiming.
3205 age_active_anon(zone, &sc);
3208 * If the number of buffer_heads in the machine
3209 * exceeds the maximum allowed level and this node
3210 * has a highmem zone, force kswapd to reclaim from
3211 * it to relieve lowmem pressure.
3213 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3218 if (!zone_balanced(zone, order, 0, 0)) {
3223 * If balanced, clear the dirty and congested
3226 clear_bit(ZONE_CONGESTED, &zone->flags);
3227 clear_bit(ZONE_DIRTY, &zone->flags);
3234 for (i = 0; i <= end_zone; i++) {
3235 struct zone *zone = pgdat->node_zones + i;
3237 if (!populated_zone(zone))
3241 * If any zone is currently balanced then kswapd will
3242 * not call compaction as it is expected that the
3243 * necessary pages are already available.
3245 if (pgdat_needs_compaction &&
3246 zone_watermark_ok(zone, order,
3247 low_wmark_pages(zone),
3249 pgdat_needs_compaction = false;
3253 * If we're getting trouble reclaiming, start doing writepage
3254 * even in laptop mode.
3256 if (sc.priority < DEF_PRIORITY - 2)
3257 sc.may_writepage = 1;
3260 * Now scan the zone in the dma->highmem direction, stopping
3261 * at the last zone which needs scanning.
3263 * We do this because the page allocator works in the opposite
3264 * direction. This prevents the page allocator from allocating
3265 * pages behind kswapd's direction of progress, which would
3266 * cause too much scanning of the lower zones.
3268 for (i = 0; i <= end_zone; i++) {
3269 struct zone *zone = pgdat->node_zones + i;
3271 if (!populated_zone(zone))
3274 if (sc.priority != DEF_PRIORITY &&
3275 !zone_reclaimable(zone))
3280 nr_soft_scanned = 0;
3282 * Call soft limit reclaim before calling shrink_zone.
3284 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3287 sc.nr_reclaimed += nr_soft_reclaimed;
3290 * There should be no need to raise the scanning
3291 * priority if enough pages are already being scanned
3292 * that that high watermark would be met at 100%
3295 if (kswapd_shrink_zone(zone, end_zone,
3296 &sc, &nr_attempted))
3297 raise_priority = false;
3301 * If the low watermark is met there is no need for processes
3302 * to be throttled on pfmemalloc_wait as they should not be
3303 * able to safely make forward progress. Wake them
3305 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3306 pfmemalloc_watermark_ok(pgdat))
3307 wake_up_all(&pgdat->pfmemalloc_wait);
3310 * Fragmentation may mean that the system cannot be rebalanced
3311 * for high-order allocations in all zones. If twice the
3312 * allocation size has been reclaimed and the zones are still
3313 * not balanced then recheck the watermarks at order-0 to
3314 * prevent kswapd reclaiming excessively. Assume that a
3315 * process requested a high-order can direct reclaim/compact.
3317 if (order && sc.nr_reclaimed >= 2UL << order)
3318 order = sc.order = 0;
3320 /* Check if kswapd should be suspending */
3321 if (try_to_freeze() || kthread_should_stop())
3325 * Compact if necessary and kswapd is reclaiming at least the
3326 * high watermark number of pages as requsted
3328 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3329 compact_pgdat(pgdat, order);
3332 * Raise priority if scanning rate is too low or there was no
3333 * progress in reclaiming pages
3335 if (raise_priority || !sc.nr_reclaimed)
3337 } while (sc.priority >= 1 &&
3338 !pgdat_balanced(pgdat, order, *classzone_idx));
3342 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3343 * makes a decision on the order we were last reclaiming at. However,
3344 * if another caller entered the allocator slow path while kswapd
3345 * was awake, order will remain at the higher level
3347 *classzone_idx = end_zone;
3351 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3356 if (freezing(current) || kthread_should_stop())
3359 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3361 /* Try to sleep for a short interval */
3362 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3363 remaining = schedule_timeout(HZ/10);
3364 finish_wait(&pgdat->kswapd_wait, &wait);
3365 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3369 * After a short sleep, check if it was a premature sleep. If not, then
3370 * go fully to sleep until explicitly woken up.
3372 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3373 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3376 * vmstat counters are not perfectly accurate and the estimated
3377 * value for counters such as NR_FREE_PAGES can deviate from the
3378 * true value by nr_online_cpus * threshold. To avoid the zone
3379 * watermarks being breached while under pressure, we reduce the
3380 * per-cpu vmstat threshold while kswapd is awake and restore
3381 * them before going back to sleep.
3383 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3386 * Compaction records what page blocks it recently failed to
3387 * isolate pages from and skips them in the future scanning.
3388 * When kswapd is going to sleep, it is reasonable to assume
3389 * that pages and compaction may succeed so reset the cache.
3391 reset_isolation_suitable(pgdat);
3393 if (!kthread_should_stop())
3396 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3399 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3401 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3403 finish_wait(&pgdat->kswapd_wait, &wait);
3407 * The background pageout daemon, started as a kernel thread
3408 * from the init process.
3410 * This basically trickles out pages so that we have _some_
3411 * free memory available even if there is no other activity
3412 * that frees anything up. This is needed for things like routing
3413 * etc, where we otherwise might have all activity going on in
3414 * asynchronous contexts that cannot page things out.
3416 * If there are applications that are active memory-allocators
3417 * (most normal use), this basically shouldn't matter.
3419 static int kswapd(void *p)
3421 unsigned long order, new_order;
3422 unsigned balanced_order;
3423 int classzone_idx, new_classzone_idx;
3424 int balanced_classzone_idx;
3425 pg_data_t *pgdat = (pg_data_t*)p;
3426 struct task_struct *tsk = current;
3428 struct reclaim_state reclaim_state = {
3429 .reclaimed_slab = 0,
3431 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3433 lockdep_set_current_reclaim_state(GFP_KERNEL);
3435 if (!cpumask_empty(cpumask))
3436 set_cpus_allowed_ptr(tsk, cpumask);
3437 current->reclaim_state = &reclaim_state;
3440 * Tell the memory management that we're a "memory allocator",
3441 * and that if we need more memory we should get access to it
3442 * regardless (see "__alloc_pages()"). "kswapd" should
3443 * never get caught in the normal page freeing logic.
3445 * (Kswapd normally doesn't need memory anyway, but sometimes
3446 * you need a small amount of memory in order to be able to
3447 * page out something else, and this flag essentially protects
3448 * us from recursively trying to free more memory as we're
3449 * trying to free the first piece of memory in the first place).
3451 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3454 order = new_order = 0;
3456 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3457 balanced_classzone_idx = classzone_idx;
3462 * If the last balance_pgdat was unsuccessful it's unlikely a
3463 * new request of a similar or harder type will succeed soon
3464 * so consider going to sleep on the basis we reclaimed at
3466 if (balanced_classzone_idx >= new_classzone_idx &&
3467 balanced_order == new_order) {
3468 new_order = pgdat->kswapd_max_order;
3469 new_classzone_idx = pgdat->classzone_idx;
3470 pgdat->kswapd_max_order = 0;
3471 pgdat->classzone_idx = pgdat->nr_zones - 1;
3474 if (order < new_order || classzone_idx > new_classzone_idx) {
3476 * Don't sleep if someone wants a larger 'order'
3477 * allocation or has tigher zone constraints
3480 classzone_idx = new_classzone_idx;
3482 kswapd_try_to_sleep(pgdat, balanced_order,
3483 balanced_classzone_idx);
3484 order = pgdat->kswapd_max_order;
3485 classzone_idx = pgdat->classzone_idx;
3487 new_classzone_idx = classzone_idx;
3488 pgdat->kswapd_max_order = 0;
3489 pgdat->classzone_idx = pgdat->nr_zones - 1;
3492 ret = try_to_freeze();
3493 if (kthread_should_stop())
3497 * We can speed up thawing tasks if we don't call balance_pgdat
3498 * after returning from the refrigerator
3501 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3502 balanced_classzone_idx = classzone_idx;
3503 balanced_order = balance_pgdat(pgdat, order,
3504 &balanced_classzone_idx);
3508 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3509 current->reclaim_state = NULL;
3510 lockdep_clear_current_reclaim_state();
3516 * A zone is low on free memory, so wake its kswapd task to service it.
3518 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3522 if (!populated_zone(zone))
3525 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3527 pgdat = zone->zone_pgdat;
3528 if (pgdat->kswapd_max_order < order) {
3529 pgdat->kswapd_max_order = order;
3530 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3532 if (!waitqueue_active(&pgdat->kswapd_wait))
3534 if (zone_balanced(zone, order, 0, 0))
3537 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3538 wake_up_interruptible(&pgdat->kswapd_wait);
3541 #ifdef CONFIG_HIBERNATION
3543 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3546 * Rather than trying to age LRUs the aim is to preserve the overall
3547 * LRU order by reclaiming preferentially
3548 * inactive > active > active referenced > active mapped
3550 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3552 struct reclaim_state reclaim_state;
3553 struct scan_control sc = {
3554 .nr_to_reclaim = nr_to_reclaim,
3555 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3556 .priority = DEF_PRIORITY,
3560 .hibernation_mode = 1,
3562 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3563 struct task_struct *p = current;
3564 unsigned long nr_reclaimed;
3566 p->flags |= PF_MEMALLOC;
3567 lockdep_set_current_reclaim_state(sc.gfp_mask);
3568 reclaim_state.reclaimed_slab = 0;
3569 p->reclaim_state = &reclaim_state;
3571 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3573 p->reclaim_state = NULL;
3574 lockdep_clear_current_reclaim_state();
3575 p->flags &= ~PF_MEMALLOC;
3577 return nr_reclaimed;
3579 #endif /* CONFIG_HIBERNATION */
3581 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3582 not required for correctness. So if the last cpu in a node goes
3583 away, we get changed to run anywhere: as the first one comes back,
3584 restore their cpu bindings. */
3585 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3590 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3591 for_each_node_state(nid, N_MEMORY) {
3592 pg_data_t *pgdat = NODE_DATA(nid);
3593 const struct cpumask *mask;
3595 mask = cpumask_of_node(pgdat->node_id);
3597 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3598 /* One of our CPUs online: restore mask */
3599 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3606 * This kswapd start function will be called by init and node-hot-add.
3607 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3609 int kswapd_run(int nid)
3611 pg_data_t *pgdat = NODE_DATA(nid);
3617 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3618 if (IS_ERR(pgdat->kswapd)) {
3619 /* failure at boot is fatal */
3620 BUG_ON(system_state == SYSTEM_BOOTING);
3621 pr_err("Failed to start kswapd on node %d\n", nid);
3622 ret = PTR_ERR(pgdat->kswapd);
3623 pgdat->kswapd = NULL;
3629 * Called by memory hotplug when all memory in a node is offlined. Caller must
3630 * hold mem_hotplug_begin/end().
3632 void kswapd_stop(int nid)
3634 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3637 kthread_stop(kswapd);
3638 NODE_DATA(nid)->kswapd = NULL;
3642 static int __init kswapd_init(void)
3647 for_each_node_state(nid, N_MEMORY)
3649 hotcpu_notifier(cpu_callback, 0);
3653 module_init(kswapd_init)
3659 * If non-zero call zone_reclaim when the number of free pages falls below
3662 int zone_reclaim_mode __read_mostly;
3664 #define RECLAIM_OFF 0
3665 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3666 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3667 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3670 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3671 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3674 #define ZONE_RECLAIM_PRIORITY 4
3677 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3680 int sysctl_min_unmapped_ratio = 1;
3683 * If the number of slab pages in a zone grows beyond this percentage then
3684 * slab reclaim needs to occur.
3686 int sysctl_min_slab_ratio = 5;
3688 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3690 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3691 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3692 zone_page_state(zone, NR_ACTIVE_FILE);
3695 * It's possible for there to be more file mapped pages than
3696 * accounted for by the pages on the file LRU lists because
3697 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3699 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3702 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3703 static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3705 unsigned long nr_pagecache_reclaimable;
3706 unsigned long delta = 0;
3709 * If RECLAIM_UNMAP is set, then all file pages are considered
3710 * potentially reclaimable. Otherwise, we have to worry about
3711 * pages like swapcache and zone_unmapped_file_pages() provides
3714 if (zone_reclaim_mode & RECLAIM_UNMAP)
3715 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3717 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3719 /* If we can't clean pages, remove dirty pages from consideration */
3720 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3721 delta += zone_page_state(zone, NR_FILE_DIRTY);
3723 /* Watch for any possible underflows due to delta */
3724 if (unlikely(delta > nr_pagecache_reclaimable))
3725 delta = nr_pagecache_reclaimable;
3727 return nr_pagecache_reclaimable - delta;
3731 * Try to free up some pages from this zone through reclaim.
3733 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3735 /* Minimum pages needed in order to stay on node */
3736 const unsigned long nr_pages = 1 << order;
3737 struct task_struct *p = current;
3738 struct reclaim_state reclaim_state;
3739 struct scan_control sc = {
3740 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3741 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3743 .priority = ZONE_RECLAIM_PRIORITY,
3744 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3745 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3751 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3752 * and we also need to be able to write out pages for RECLAIM_WRITE
3753 * and RECLAIM_UNMAP.
3755 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3756 lockdep_set_current_reclaim_state(gfp_mask);
3757 reclaim_state.reclaimed_slab = 0;
3758 p->reclaim_state = &reclaim_state;
3760 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3762 * Free memory by calling shrink zone with increasing
3763 * priorities until we have enough memory freed.
3766 shrink_zone(zone, &sc, true);
3767 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3770 p->reclaim_state = NULL;
3771 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3772 lockdep_clear_current_reclaim_state();
3773 return sc.nr_reclaimed >= nr_pages;
3776 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3782 * Zone reclaim reclaims unmapped file backed pages and
3783 * slab pages if we are over the defined limits.
3785 * A small portion of unmapped file backed pages is needed for
3786 * file I/O otherwise pages read by file I/O will be immediately
3787 * thrown out if the zone is overallocated. So we do not reclaim
3788 * if less than a specified percentage of the zone is used by
3789 * unmapped file backed pages.
3791 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3792 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3793 return ZONE_RECLAIM_FULL;
3795 if (!zone_reclaimable(zone))
3796 return ZONE_RECLAIM_FULL;
3799 * Do not scan if the allocation should not be delayed.
3801 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3802 return ZONE_RECLAIM_NOSCAN;
3805 * Only run zone reclaim on the local zone or on zones that do not
3806 * have associated processors. This will favor the local processor
3807 * over remote processors and spread off node memory allocations
3808 * as wide as possible.
3810 node_id = zone_to_nid(zone);
3811 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3812 return ZONE_RECLAIM_NOSCAN;
3814 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3815 return ZONE_RECLAIM_NOSCAN;
3817 ret = __zone_reclaim(zone, gfp_mask, order);
3818 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3821 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3828 * page_evictable - test whether a page is evictable
3829 * @page: the page to test
3831 * Test whether page is evictable--i.e., should be placed on active/inactive
3832 * lists vs unevictable list.
3834 * Reasons page might not be evictable:
3835 * (1) page's mapping marked unevictable
3836 * (2) page is part of an mlocked VMA
3839 int page_evictable(struct page *page)
3841 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3846 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3847 * @pages: array of pages to check
3848 * @nr_pages: number of pages to check
3850 * Checks pages for evictability and moves them to the appropriate lru list.
3852 * This function is only used for SysV IPC SHM_UNLOCK.
3854 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3856 struct lruvec *lruvec;
3857 struct zone *zone = NULL;
3862 for (i = 0; i < nr_pages; i++) {
3863 struct page *page = pages[i];
3864 struct zone *pagezone;
3867 pagezone = page_zone(page);
3868 if (pagezone != zone) {
3870 spin_unlock_irq(&zone->lru_lock);
3872 spin_lock_irq(&zone->lru_lock);
3874 lruvec = mem_cgroup_page_lruvec(page, zone);
3876 if (!PageLRU(page) || !PageUnevictable(page))
3879 if (page_evictable(page)) {
3880 enum lru_list lru = page_lru_base_type(page);
3882 VM_BUG_ON_PAGE(PageActive(page), page);
3883 ClearPageUnevictable(page);
3884 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3885 add_page_to_lru_list(page, lruvec, lru);
3891 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3892 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3893 spin_unlock_irq(&zone->lru_lock);
3896 #endif /* CONFIG_SHMEM */