1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup *target_mem_cgroup;
83 * Scan pressure balancing between anon and file LRUs
85 unsigned long anon_cost;
86 unsigned long file_cost;
88 /* Can active pages be deactivated as part of reclaim? */
89 #define DEACTIVATE_ANON 1
90 #define DEACTIVATE_FILE 2
91 unsigned int may_deactivate:2;
92 unsigned int force_deactivate:1;
93 unsigned int skipped_deactivate:1;
95 /* Writepage batching in laptop mode; RECLAIM_WRITE */
96 unsigned int may_writepage:1;
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap:1;
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap:1;
105 * Cgroups are not reclaimed below their configured memory.low,
106 * unless we threaten to OOM. If any cgroups are skipped due to
107 * memory.low and nothing was reclaimed, go back for memory.low.
109 unsigned int memcg_low_reclaim:1;
110 unsigned int memcg_low_skipped:1;
112 unsigned int hibernation_mode:1;
114 /* One of the zones is ready for compaction */
115 unsigned int compaction_ready:1;
117 /* There is easily reclaimable cold cache in the current node */
118 unsigned int cache_trim_mode:1;
120 /* The file pages on the current node are dangerously low */
121 unsigned int file_is_tiny:1;
123 /* Allocation order */
126 /* Scan (total_size >> priority) pages at once */
129 /* The highest zone to isolate pages for reclaim from */
132 /* This context's GFP mask */
135 /* Incremented by the number of inactive pages that were scanned */
136 unsigned long nr_scanned;
138 /* Number of pages freed so far during a call to shrink_zones() */
139 unsigned long nr_reclaimed;
143 unsigned int unqueued_dirty;
144 unsigned int congested;
145 unsigned int writeback;
146 unsigned int immediate;
147 unsigned int file_taken;
151 /* for recording the reclaimed slab by now */
152 struct reclaim_state reclaim_state;
155 #ifdef ARCH_HAS_PREFETCHW
156 #define prefetchw_prev_lru_page(_page, _base, _field) \
158 if ((_page)->lru.prev != _base) { \
161 prev = lru_to_page(&(_page->lru)); \
162 prefetchw(&prev->_field); \
166 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
170 * From 0 .. 200. Higher means more swappy.
172 int vm_swappiness = 60;
174 static void set_task_reclaim_state(struct task_struct *task,
175 struct reclaim_state *rs)
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs && task->reclaim_state);
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs && !task->reclaim_state);
183 task->reclaim_state = rs;
186 static LIST_HEAD(shrinker_list);
187 static DECLARE_RWSEM(shrinker_rwsem);
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
201 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
203 static DEFINE_IDR(shrinker_idr);
204 static int shrinker_nr_max;
206 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
208 int id, ret = -ENOMEM;
210 down_write(&shrinker_rwsem);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
216 if (id >= shrinker_nr_max) {
217 if (memcg_expand_shrinker_maps(id)) {
218 idr_remove(&shrinker_idr, id);
222 shrinker_nr_max = id + 1;
227 up_write(&shrinker_rwsem);
231 static void unregister_memcg_shrinker(struct shrinker *shrinker)
233 int id = shrinker->id;
237 down_write(&shrinker_rwsem);
238 idr_remove(&shrinker_idr, id);
239 up_write(&shrinker_rwsem);
242 static bool cgroup_reclaim(struct scan_control *sc)
244 return sc->target_mem_cgroup;
248 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
249 * @sc: scan_control in question
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
260 static bool writeback_throttling_sane(struct scan_control *sc)
262 if (!cgroup_reclaim(sc))
264 #ifdef CONFIG_CGROUP_WRITEBACK
265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
271 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
276 static void unregister_memcg_shrinker(struct shrinker *shrinker)
280 static bool cgroup_reclaim(struct scan_control *sc)
285 static bool writeback_throttling_sane(struct scan_control *sc)
292 * This misses isolated pages which are not accounted for to save counters.
293 * As the data only determines if reclaim or compaction continues, it is
294 * not expected that isolated pages will be a dominating factor.
296 unsigned long zone_reclaimable_pages(struct zone *zone)
300 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
301 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
302 if (get_nr_swap_pages() > 0)
303 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
304 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
310 * lruvec_lru_size - Returns the number of pages on the given LRU list.
311 * @lruvec: lru vector
313 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
315 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
317 unsigned long size = 0;
320 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
321 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
323 if (!managed_zone(zone))
326 if (!mem_cgroup_disabled())
327 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
329 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
335 * Add a shrinker callback to be called from the vm.
337 int prealloc_shrinker(struct shrinker *shrinker)
339 unsigned int size = sizeof(*shrinker->nr_deferred);
341 if (shrinker->flags & SHRINKER_NUMA_AWARE)
344 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
345 if (!shrinker->nr_deferred)
348 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
349 if (prealloc_memcg_shrinker(shrinker))
356 kfree(shrinker->nr_deferred);
357 shrinker->nr_deferred = NULL;
361 void free_prealloced_shrinker(struct shrinker *shrinker)
363 if (!shrinker->nr_deferred)
366 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
367 unregister_memcg_shrinker(shrinker);
369 kfree(shrinker->nr_deferred);
370 shrinker->nr_deferred = NULL;
373 void register_shrinker_prepared(struct shrinker *shrinker)
375 down_write(&shrinker_rwsem);
376 list_add_tail(&shrinker->list, &shrinker_list);
378 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
379 idr_replace(&shrinker_idr, shrinker, shrinker->id);
381 up_write(&shrinker_rwsem);
384 int register_shrinker(struct shrinker *shrinker)
386 int err = prealloc_shrinker(shrinker);
390 register_shrinker_prepared(shrinker);
393 EXPORT_SYMBOL(register_shrinker);
398 void unregister_shrinker(struct shrinker *shrinker)
400 if (!shrinker->nr_deferred)
402 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
403 unregister_memcg_shrinker(shrinker);
404 down_write(&shrinker_rwsem);
405 list_del(&shrinker->list);
406 up_write(&shrinker_rwsem);
407 kfree(shrinker->nr_deferred);
408 shrinker->nr_deferred = NULL;
410 EXPORT_SYMBOL(unregister_shrinker);
412 #define SHRINK_BATCH 128
414 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
415 struct shrinker *shrinker, int priority)
417 unsigned long freed = 0;
418 unsigned long long delta;
423 int nid = shrinkctl->nid;
424 long batch_size = shrinker->batch ? shrinker->batch
426 long scanned = 0, next_deferred;
428 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
431 freeable = shrinker->count_objects(shrinker, shrinkctl);
432 if (freeable == 0 || freeable == SHRINK_EMPTY)
436 * copy the current shrinker scan count into a local variable
437 * and zero it so that other concurrent shrinker invocations
438 * don't also do this scanning work.
440 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
443 if (shrinker->seeks) {
444 delta = freeable >> priority;
446 do_div(delta, shrinker->seeks);
449 * These objects don't require any IO to create. Trim
450 * them aggressively under memory pressure to keep
451 * them from causing refetches in the IO caches.
453 delta = freeable / 2;
457 if (total_scan < 0) {
458 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
459 shrinker->scan_objects, total_scan);
460 total_scan = freeable;
463 next_deferred = total_scan;
466 * We need to avoid excessive windup on filesystem shrinkers
467 * due to large numbers of GFP_NOFS allocations causing the
468 * shrinkers to return -1 all the time. This results in a large
469 * nr being built up so when a shrink that can do some work
470 * comes along it empties the entire cache due to nr >>>
471 * freeable. This is bad for sustaining a working set in
474 * Hence only allow the shrinker to scan the entire cache when
475 * a large delta change is calculated directly.
477 if (delta < freeable / 4)
478 total_scan = min(total_scan, freeable / 2);
481 * Avoid risking looping forever due to too large nr value:
482 * never try to free more than twice the estimate number of
485 if (total_scan > freeable * 2)
486 total_scan = freeable * 2;
488 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
489 freeable, delta, total_scan, priority);
492 * Normally, we should not scan less than batch_size objects in one
493 * pass to avoid too frequent shrinker calls, but if the slab has less
494 * than batch_size objects in total and we are really tight on memory,
495 * we will try to reclaim all available objects, otherwise we can end
496 * up failing allocations although there are plenty of reclaimable
497 * objects spread over several slabs with usage less than the
500 * We detect the "tight on memory" situations by looking at the total
501 * number of objects we want to scan (total_scan). If it is greater
502 * than the total number of objects on slab (freeable), we must be
503 * scanning at high prio and therefore should try to reclaim as much as
506 while (total_scan >= batch_size ||
507 total_scan >= freeable) {
509 unsigned long nr_to_scan = min(batch_size, total_scan);
511 shrinkctl->nr_to_scan = nr_to_scan;
512 shrinkctl->nr_scanned = nr_to_scan;
513 ret = shrinker->scan_objects(shrinker, shrinkctl);
514 if (ret == SHRINK_STOP)
518 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
519 total_scan -= shrinkctl->nr_scanned;
520 scanned += shrinkctl->nr_scanned;
525 if (next_deferred >= scanned)
526 next_deferred -= scanned;
530 * move the unused scan count back into the shrinker in a
531 * manner that handles concurrent updates. If we exhausted the
532 * scan, there is no need to do an update.
534 if (next_deferred > 0)
535 new_nr = atomic_long_add_return(next_deferred,
536 &shrinker->nr_deferred[nid]);
538 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
540 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
545 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
546 struct mem_cgroup *memcg, int priority)
548 struct memcg_shrinker_map *map;
549 unsigned long ret, freed = 0;
552 if (!mem_cgroup_online(memcg))
555 if (!down_read_trylock(&shrinker_rwsem))
558 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
563 for_each_set_bit(i, map->map, shrinker_nr_max) {
564 struct shrink_control sc = {
565 .gfp_mask = gfp_mask,
569 struct shrinker *shrinker;
571 shrinker = idr_find(&shrinker_idr, i);
572 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
574 clear_bit(i, map->map);
578 /* Call non-slab shrinkers even though kmem is disabled */
579 if (!memcg_kmem_enabled() &&
580 !(shrinker->flags & SHRINKER_NONSLAB))
583 ret = do_shrink_slab(&sc, shrinker, priority);
584 if (ret == SHRINK_EMPTY) {
585 clear_bit(i, map->map);
587 * After the shrinker reported that it had no objects to
588 * free, but before we cleared the corresponding bit in
589 * the memcg shrinker map, a new object might have been
590 * added. To make sure, we have the bit set in this
591 * case, we invoke the shrinker one more time and reset
592 * the bit if it reports that it is not empty anymore.
593 * The memory barrier here pairs with the barrier in
594 * memcg_set_shrinker_bit():
596 * list_lru_add() shrink_slab_memcg()
597 * list_add_tail() clear_bit()
599 * set_bit() do_shrink_slab()
601 smp_mb__after_atomic();
602 ret = do_shrink_slab(&sc, shrinker, priority);
603 if (ret == SHRINK_EMPTY)
606 memcg_set_shrinker_bit(memcg, nid, i);
610 if (rwsem_is_contended(&shrinker_rwsem)) {
616 up_read(&shrinker_rwsem);
619 #else /* CONFIG_MEMCG */
620 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
621 struct mem_cgroup *memcg, int priority)
625 #endif /* CONFIG_MEMCG */
628 * shrink_slab - shrink slab caches
629 * @gfp_mask: allocation context
630 * @nid: node whose slab caches to target
631 * @memcg: memory cgroup whose slab caches to target
632 * @priority: the reclaim priority
634 * Call the shrink functions to age shrinkable caches.
636 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
637 * unaware shrinkers will receive a node id of 0 instead.
639 * @memcg specifies the memory cgroup to target. Unaware shrinkers
640 * are called only if it is the root cgroup.
642 * @priority is sc->priority, we take the number of objects and >> by priority
643 * in order to get the scan target.
645 * Returns the number of reclaimed slab objects.
647 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
648 struct mem_cgroup *memcg,
651 unsigned long ret, freed = 0;
652 struct shrinker *shrinker;
655 * The root memcg might be allocated even though memcg is disabled
656 * via "cgroup_disable=memory" boot parameter. This could make
657 * mem_cgroup_is_root() return false, then just run memcg slab
658 * shrink, but skip global shrink. This may result in premature
661 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
662 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
664 if (!down_read_trylock(&shrinker_rwsem))
667 list_for_each_entry(shrinker, &shrinker_list, list) {
668 struct shrink_control sc = {
669 .gfp_mask = gfp_mask,
674 ret = do_shrink_slab(&sc, shrinker, priority);
675 if (ret == SHRINK_EMPTY)
679 * Bail out if someone want to register a new shrinker to
680 * prevent the registration from being stalled for long periods
681 * by parallel ongoing shrinking.
683 if (rwsem_is_contended(&shrinker_rwsem)) {
689 up_read(&shrinker_rwsem);
695 void drop_slab_node(int nid)
700 struct mem_cgroup *memcg = NULL;
702 if (fatal_signal_pending(current))
706 memcg = mem_cgroup_iter(NULL, NULL, NULL);
708 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
709 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
710 } while (freed > 10);
717 for_each_online_node(nid)
721 static inline int is_page_cache_freeable(struct page *page)
724 * A freeable page cache page is referenced only by the caller
725 * that isolated the page, the page cache and optional buffer
726 * heads at page->private.
728 int page_cache_pins = thp_nr_pages(page);
729 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
732 static int may_write_to_inode(struct inode *inode)
734 if (current->flags & PF_SWAPWRITE)
736 if (!inode_write_congested(inode))
738 if (inode_to_bdi(inode) == current->backing_dev_info)
744 * We detected a synchronous write error writing a page out. Probably
745 * -ENOSPC. We need to propagate that into the address_space for a subsequent
746 * fsync(), msync() or close().
748 * The tricky part is that after writepage we cannot touch the mapping: nothing
749 * prevents it from being freed up. But we have a ref on the page and once
750 * that page is locked, the mapping is pinned.
752 * We're allowed to run sleeping lock_page() here because we know the caller has
755 static void handle_write_error(struct address_space *mapping,
756 struct page *page, int error)
759 if (page_mapping(page) == mapping)
760 mapping_set_error(mapping, error);
764 /* possible outcome of pageout() */
766 /* failed to write page out, page is locked */
768 /* move page to the active list, page is locked */
770 /* page has been sent to the disk successfully, page is unlocked */
772 /* page is clean and locked */
777 * pageout is called by shrink_page_list() for each dirty page.
778 * Calls ->writepage().
780 static pageout_t pageout(struct page *page, struct address_space *mapping)
783 * If the page is dirty, only perform writeback if that write
784 * will be non-blocking. To prevent this allocation from being
785 * stalled by pagecache activity. But note that there may be
786 * stalls if we need to run get_block(). We could test
787 * PagePrivate for that.
789 * If this process is currently in __generic_file_write_iter() against
790 * this page's queue, we can perform writeback even if that
793 * If the page is swapcache, write it back even if that would
794 * block, for some throttling. This happens by accident, because
795 * swap_backing_dev_info is bust: it doesn't reflect the
796 * congestion state of the swapdevs. Easy to fix, if needed.
798 if (!is_page_cache_freeable(page))
802 * Some data journaling orphaned pages can have
803 * page->mapping == NULL while being dirty with clean buffers.
805 if (page_has_private(page)) {
806 if (try_to_free_buffers(page)) {
807 ClearPageDirty(page);
808 pr_info("%s: orphaned page\n", __func__);
814 if (mapping->a_ops->writepage == NULL)
815 return PAGE_ACTIVATE;
816 if (!may_write_to_inode(mapping->host))
819 if (clear_page_dirty_for_io(page)) {
821 struct writeback_control wbc = {
822 .sync_mode = WB_SYNC_NONE,
823 .nr_to_write = SWAP_CLUSTER_MAX,
825 .range_end = LLONG_MAX,
829 SetPageReclaim(page);
830 res = mapping->a_ops->writepage(page, &wbc);
832 handle_write_error(mapping, page, res);
833 if (res == AOP_WRITEPAGE_ACTIVATE) {
834 ClearPageReclaim(page);
835 return PAGE_ACTIVATE;
838 if (!PageWriteback(page)) {
839 /* synchronous write or broken a_ops? */
840 ClearPageReclaim(page);
842 trace_mm_vmscan_writepage(page);
843 inc_node_page_state(page, NR_VMSCAN_WRITE);
851 * Same as remove_mapping, but if the page is removed from the mapping, it
852 * gets returned with a refcount of 0.
854 static int __remove_mapping(struct address_space *mapping, struct page *page,
855 bool reclaimed, struct mem_cgroup *target_memcg)
861 BUG_ON(!PageLocked(page));
862 BUG_ON(mapping != page_mapping(page));
864 xa_lock_irqsave(&mapping->i_pages, flags);
866 * The non racy check for a busy page.
868 * Must be careful with the order of the tests. When someone has
869 * a ref to the page, it may be possible that they dirty it then
870 * drop the reference. So if PageDirty is tested before page_count
871 * here, then the following race may occur:
873 * get_user_pages(&page);
874 * [user mapping goes away]
876 * !PageDirty(page) [good]
877 * SetPageDirty(page);
879 * !page_count(page) [good, discard it]
881 * [oops, our write_to data is lost]
883 * Reversing the order of the tests ensures such a situation cannot
884 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
885 * load is not satisfied before that of page->_refcount.
887 * Note that if SetPageDirty is always performed via set_page_dirty,
888 * and thus under the i_pages lock, then this ordering is not required.
890 refcount = 1 + compound_nr(page);
891 if (!page_ref_freeze(page, refcount))
893 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
894 if (unlikely(PageDirty(page))) {
895 page_ref_unfreeze(page, refcount);
899 if (PageSwapCache(page)) {
900 swp_entry_t swap = { .val = page_private(page) };
901 mem_cgroup_swapout(page, swap);
902 if (reclaimed && !mapping_exiting(mapping))
903 shadow = workingset_eviction(page, target_memcg);
904 __delete_from_swap_cache(page, swap, shadow);
905 xa_unlock_irqrestore(&mapping->i_pages, flags);
906 put_swap_page(page, swap);
908 void (*freepage)(struct page *);
910 freepage = mapping->a_ops->freepage;
912 * Remember a shadow entry for reclaimed file cache in
913 * order to detect refaults, thus thrashing, later on.
915 * But don't store shadows in an address space that is
916 * already exiting. This is not just an optimization,
917 * inode reclaim needs to empty out the radix tree or
918 * the nodes are lost. Don't plant shadows behind its
921 * We also don't store shadows for DAX mappings because the
922 * only page cache pages found in these are zero pages
923 * covering holes, and because we don't want to mix DAX
924 * exceptional entries and shadow exceptional entries in the
925 * same address_space.
927 if (reclaimed && page_is_file_lru(page) &&
928 !mapping_exiting(mapping) && !dax_mapping(mapping))
929 shadow = workingset_eviction(page, target_memcg);
930 __delete_from_page_cache(page, shadow);
931 xa_unlock_irqrestore(&mapping->i_pages, flags);
933 if (freepage != NULL)
940 xa_unlock_irqrestore(&mapping->i_pages, flags);
945 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
946 * someone else has a ref on the page, abort and return 0. If it was
947 * successfully detached, return 1. Assumes the caller has a single ref on
950 int remove_mapping(struct address_space *mapping, struct page *page)
952 if (__remove_mapping(mapping, page, false, NULL)) {
954 * Unfreezing the refcount with 1 rather than 2 effectively
955 * drops the pagecache ref for us without requiring another
958 page_ref_unfreeze(page, 1);
965 * putback_lru_page - put previously isolated page onto appropriate LRU list
966 * @page: page to be put back to appropriate lru list
968 * Add previously isolated @page to appropriate LRU list.
969 * Page may still be unevictable for other reasons.
971 * lru_lock must not be held, interrupts must be enabled.
973 void putback_lru_page(struct page *page)
976 put_page(page); /* drop ref from isolate */
979 enum page_references {
981 PAGEREF_RECLAIM_CLEAN,
986 static enum page_references page_check_references(struct page *page,
987 struct scan_control *sc)
989 int referenced_ptes, referenced_page;
990 unsigned long vm_flags;
992 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
994 referenced_page = TestClearPageReferenced(page);
997 * Mlock lost the isolation race with us. Let try_to_unmap()
998 * move the page to the unevictable list.
1000 if (vm_flags & VM_LOCKED)
1001 return PAGEREF_RECLAIM;
1003 if (referenced_ptes) {
1005 * All mapped pages start out with page table
1006 * references from the instantiating fault, so we need
1007 * to look twice if a mapped file page is used more
1010 * Mark it and spare it for another trip around the
1011 * inactive list. Another page table reference will
1012 * lead to its activation.
1014 * Note: the mark is set for activated pages as well
1015 * so that recently deactivated but used pages are
1016 * quickly recovered.
1018 SetPageReferenced(page);
1020 if (referenced_page || referenced_ptes > 1)
1021 return PAGEREF_ACTIVATE;
1024 * Activate file-backed executable pages after first usage.
1026 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1027 return PAGEREF_ACTIVATE;
1029 return PAGEREF_KEEP;
1032 /* Reclaim if clean, defer dirty pages to writeback */
1033 if (referenced_page && !PageSwapBacked(page))
1034 return PAGEREF_RECLAIM_CLEAN;
1036 return PAGEREF_RECLAIM;
1039 /* Check if a page is dirty or under writeback */
1040 static void page_check_dirty_writeback(struct page *page,
1041 bool *dirty, bool *writeback)
1043 struct address_space *mapping;
1046 * Anonymous pages are not handled by flushers and must be written
1047 * from reclaim context. Do not stall reclaim based on them
1049 if (!page_is_file_lru(page) ||
1050 (PageAnon(page) && !PageSwapBacked(page))) {
1056 /* By default assume that the page flags are accurate */
1057 *dirty = PageDirty(page);
1058 *writeback = PageWriteback(page);
1060 /* Verify dirty/writeback state if the filesystem supports it */
1061 if (!page_has_private(page))
1064 mapping = page_mapping(page);
1065 if (mapping && mapping->a_ops->is_dirty_writeback)
1066 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1070 * shrink_page_list() returns the number of reclaimed pages
1072 static unsigned int shrink_page_list(struct list_head *page_list,
1073 struct pglist_data *pgdat,
1074 struct scan_control *sc,
1075 struct reclaim_stat *stat,
1076 bool ignore_references)
1078 LIST_HEAD(ret_pages);
1079 LIST_HEAD(free_pages);
1080 unsigned int nr_reclaimed = 0;
1081 unsigned int pgactivate = 0;
1083 memset(stat, 0, sizeof(*stat));
1086 while (!list_empty(page_list)) {
1087 struct address_space *mapping;
1089 enum page_references references = PAGEREF_RECLAIM;
1090 bool dirty, writeback, may_enter_fs;
1091 unsigned int nr_pages;
1095 page = lru_to_page(page_list);
1096 list_del(&page->lru);
1098 if (!trylock_page(page))
1101 VM_BUG_ON_PAGE(PageActive(page), page);
1103 nr_pages = compound_nr(page);
1105 /* Account the number of base pages even though THP */
1106 sc->nr_scanned += nr_pages;
1108 if (unlikely(!page_evictable(page)))
1109 goto activate_locked;
1111 if (!sc->may_unmap && page_mapped(page))
1114 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1115 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1118 * The number of dirty pages determines if a node is marked
1119 * reclaim_congested which affects wait_iff_congested. kswapd
1120 * will stall and start writing pages if the tail of the LRU
1121 * is all dirty unqueued pages.
1123 page_check_dirty_writeback(page, &dirty, &writeback);
1124 if (dirty || writeback)
1127 if (dirty && !writeback)
1128 stat->nr_unqueued_dirty++;
1131 * Treat this page as congested if the underlying BDI is or if
1132 * pages are cycling through the LRU so quickly that the
1133 * pages marked for immediate reclaim are making it to the
1134 * end of the LRU a second time.
1136 mapping = page_mapping(page);
1137 if (((dirty || writeback) && mapping &&
1138 inode_write_congested(mapping->host)) ||
1139 (writeback && PageReclaim(page)))
1140 stat->nr_congested++;
1143 * If a page at the tail of the LRU is under writeback, there
1144 * are three cases to consider.
1146 * 1) If reclaim is encountering an excessive number of pages
1147 * under writeback and this page is both under writeback and
1148 * PageReclaim then it indicates that pages are being queued
1149 * for IO but are being recycled through the LRU before the
1150 * IO can complete. Waiting on the page itself risks an
1151 * indefinite stall if it is impossible to writeback the
1152 * page due to IO error or disconnected storage so instead
1153 * note that the LRU is being scanned too quickly and the
1154 * caller can stall after page list has been processed.
1156 * 2) Global or new memcg reclaim encounters a page that is
1157 * not marked for immediate reclaim, or the caller does not
1158 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1159 * not to fs). In this case mark the page for immediate
1160 * reclaim and continue scanning.
1162 * Require may_enter_fs because we would wait on fs, which
1163 * may not have submitted IO yet. And the loop driver might
1164 * enter reclaim, and deadlock if it waits on a page for
1165 * which it is needed to do the write (loop masks off
1166 * __GFP_IO|__GFP_FS for this reason); but more thought
1167 * would probably show more reasons.
1169 * 3) Legacy memcg encounters a page that is already marked
1170 * PageReclaim. memcg does not have any dirty pages
1171 * throttling so we could easily OOM just because too many
1172 * pages are in writeback and there is nothing else to
1173 * reclaim. Wait for the writeback to complete.
1175 * In cases 1) and 2) we activate the pages to get them out of
1176 * the way while we continue scanning for clean pages on the
1177 * inactive list and refilling from the active list. The
1178 * observation here is that waiting for disk writes is more
1179 * expensive than potentially causing reloads down the line.
1180 * Since they're marked for immediate reclaim, they won't put
1181 * memory pressure on the cache working set any longer than it
1182 * takes to write them to disk.
1184 if (PageWriteback(page)) {
1186 if (current_is_kswapd() &&
1187 PageReclaim(page) &&
1188 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1189 stat->nr_immediate++;
1190 goto activate_locked;
1193 } else if (writeback_throttling_sane(sc) ||
1194 !PageReclaim(page) || !may_enter_fs) {
1196 * This is slightly racy - end_page_writeback()
1197 * might have just cleared PageReclaim, then
1198 * setting PageReclaim here end up interpreted
1199 * as PageReadahead - but that does not matter
1200 * enough to care. What we do want is for this
1201 * page to have PageReclaim set next time memcg
1202 * reclaim reaches the tests above, so it will
1203 * then wait_on_page_writeback() to avoid OOM;
1204 * and it's also appropriate in global reclaim.
1206 SetPageReclaim(page);
1207 stat->nr_writeback++;
1208 goto activate_locked;
1213 wait_on_page_writeback(page);
1214 /* then go back and try same page again */
1215 list_add_tail(&page->lru, page_list);
1220 if (!ignore_references)
1221 references = page_check_references(page, sc);
1223 switch (references) {
1224 case PAGEREF_ACTIVATE:
1225 goto activate_locked;
1227 stat->nr_ref_keep += nr_pages;
1229 case PAGEREF_RECLAIM:
1230 case PAGEREF_RECLAIM_CLEAN:
1231 ; /* try to reclaim the page below */
1235 * Anonymous process memory has backing store?
1236 * Try to allocate it some swap space here.
1237 * Lazyfree page could be freed directly
1239 if (PageAnon(page) && PageSwapBacked(page)) {
1240 if (!PageSwapCache(page)) {
1241 if (!(sc->gfp_mask & __GFP_IO))
1243 if (page_maybe_dma_pinned(page))
1245 if (PageTransHuge(page)) {
1246 /* cannot split THP, skip it */
1247 if (!can_split_huge_page(page, NULL))
1248 goto activate_locked;
1250 * Split pages without a PMD map right
1251 * away. Chances are some or all of the
1252 * tail pages can be freed without IO.
1254 if (!compound_mapcount(page) &&
1255 split_huge_page_to_list(page,
1257 goto activate_locked;
1259 if (!add_to_swap(page)) {
1260 if (!PageTransHuge(page))
1261 goto activate_locked_split;
1262 /* Fallback to swap normal pages */
1263 if (split_huge_page_to_list(page,
1265 goto activate_locked;
1266 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1267 count_vm_event(THP_SWPOUT_FALLBACK);
1269 if (!add_to_swap(page))
1270 goto activate_locked_split;
1273 may_enter_fs = true;
1275 /* Adding to swap updated mapping */
1276 mapping = page_mapping(page);
1278 } else if (unlikely(PageTransHuge(page))) {
1279 /* Split file THP */
1280 if (split_huge_page_to_list(page, page_list))
1285 * THP may get split above, need minus tail pages and update
1286 * nr_pages to avoid accounting tail pages twice.
1288 * The tail pages that are added into swap cache successfully
1291 if ((nr_pages > 1) && !PageTransHuge(page)) {
1292 sc->nr_scanned -= (nr_pages - 1);
1297 * The page is mapped into the page tables of one or more
1298 * processes. Try to unmap it here.
1300 if (page_mapped(page)) {
1301 enum ttu_flags flags = TTU_BATCH_FLUSH;
1302 bool was_swapbacked = PageSwapBacked(page);
1304 if (unlikely(PageTransHuge(page)))
1305 flags |= TTU_SPLIT_HUGE_PMD;
1307 if (!try_to_unmap(page, flags)) {
1308 stat->nr_unmap_fail += nr_pages;
1309 if (!was_swapbacked && PageSwapBacked(page))
1310 stat->nr_lazyfree_fail += nr_pages;
1311 goto activate_locked;
1315 if (PageDirty(page)) {
1317 * Only kswapd can writeback filesystem pages
1318 * to avoid risk of stack overflow. But avoid
1319 * injecting inefficient single-page IO into
1320 * flusher writeback as much as possible: only
1321 * write pages when we've encountered many
1322 * dirty pages, and when we've already scanned
1323 * the rest of the LRU for clean pages and see
1324 * the same dirty pages again (PageReclaim).
1326 if (page_is_file_lru(page) &&
1327 (!current_is_kswapd() || !PageReclaim(page) ||
1328 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1330 * Immediately reclaim when written back.
1331 * Similar in principal to deactivate_page()
1332 * except we already have the page isolated
1333 * and know it's dirty
1335 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1336 SetPageReclaim(page);
1338 goto activate_locked;
1341 if (references == PAGEREF_RECLAIM_CLEAN)
1345 if (!sc->may_writepage)
1349 * Page is dirty. Flush the TLB if a writable entry
1350 * potentially exists to avoid CPU writes after IO
1351 * starts and then write it out here.
1353 try_to_unmap_flush_dirty();
1354 switch (pageout(page, mapping)) {
1358 goto activate_locked;
1360 stat->nr_pageout += thp_nr_pages(page);
1362 if (PageWriteback(page))
1364 if (PageDirty(page))
1368 * A synchronous write - probably a ramdisk. Go
1369 * ahead and try to reclaim the page.
1371 if (!trylock_page(page))
1373 if (PageDirty(page) || PageWriteback(page))
1375 mapping = page_mapping(page);
1377 ; /* try to free the page below */
1382 * If the page has buffers, try to free the buffer mappings
1383 * associated with this page. If we succeed we try to free
1386 * We do this even if the page is PageDirty().
1387 * try_to_release_page() does not perform I/O, but it is
1388 * possible for a page to have PageDirty set, but it is actually
1389 * clean (all its buffers are clean). This happens if the
1390 * buffers were written out directly, with submit_bh(). ext3
1391 * will do this, as well as the blockdev mapping.
1392 * try_to_release_page() will discover that cleanness and will
1393 * drop the buffers and mark the page clean - it can be freed.
1395 * Rarely, pages can have buffers and no ->mapping. These are
1396 * the pages which were not successfully invalidated in
1397 * truncate_complete_page(). We try to drop those buffers here
1398 * and if that worked, and the page is no longer mapped into
1399 * process address space (page_count == 1) it can be freed.
1400 * Otherwise, leave the page on the LRU so it is swappable.
1402 if (page_has_private(page)) {
1403 if (!try_to_release_page(page, sc->gfp_mask))
1404 goto activate_locked;
1405 if (!mapping && page_count(page) == 1) {
1407 if (put_page_testzero(page))
1411 * rare race with speculative reference.
1412 * the speculative reference will free
1413 * this page shortly, so we may
1414 * increment nr_reclaimed here (and
1415 * leave it off the LRU).
1423 if (PageAnon(page) && !PageSwapBacked(page)) {
1424 /* follow __remove_mapping for reference */
1425 if (!page_ref_freeze(page, 1))
1427 if (PageDirty(page)) {
1428 page_ref_unfreeze(page, 1);
1432 count_vm_event(PGLAZYFREED);
1433 count_memcg_page_event(page, PGLAZYFREED);
1434 } else if (!mapping || !__remove_mapping(mapping, page, true,
1435 sc->target_mem_cgroup))
1441 * THP may get swapped out in a whole, need account
1444 nr_reclaimed += nr_pages;
1447 * Is there need to periodically free_page_list? It would
1448 * appear not as the counts should be low
1450 if (unlikely(PageTransHuge(page)))
1451 destroy_compound_page(page);
1453 list_add(&page->lru, &free_pages);
1456 activate_locked_split:
1458 * The tail pages that are failed to add into swap cache
1459 * reach here. Fixup nr_scanned and nr_pages.
1462 sc->nr_scanned -= (nr_pages - 1);
1466 /* Not a candidate for swapping, so reclaim swap space. */
1467 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1469 try_to_free_swap(page);
1470 VM_BUG_ON_PAGE(PageActive(page), page);
1471 if (!PageMlocked(page)) {
1472 int type = page_is_file_lru(page);
1473 SetPageActive(page);
1474 stat->nr_activate[type] += nr_pages;
1475 count_memcg_page_event(page, PGACTIVATE);
1480 list_add(&page->lru, &ret_pages);
1481 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1484 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1486 mem_cgroup_uncharge_list(&free_pages);
1487 try_to_unmap_flush();
1488 free_unref_page_list(&free_pages);
1490 list_splice(&ret_pages, page_list);
1491 count_vm_events(PGACTIVATE, pgactivate);
1493 return nr_reclaimed;
1496 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1497 struct list_head *page_list)
1499 struct scan_control sc = {
1500 .gfp_mask = GFP_KERNEL,
1501 .priority = DEF_PRIORITY,
1504 struct reclaim_stat stat;
1505 unsigned int nr_reclaimed;
1506 struct page *page, *next;
1507 LIST_HEAD(clean_pages);
1509 list_for_each_entry_safe(page, next, page_list, lru) {
1510 if (page_is_file_lru(page) && !PageDirty(page) &&
1511 !__PageMovable(page) && !PageUnevictable(page)) {
1512 ClearPageActive(page);
1513 list_move(&page->lru, &clean_pages);
1517 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1519 list_splice(&clean_pages, page_list);
1520 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1521 -(long)nr_reclaimed);
1523 * Since lazyfree pages are isolated from file LRU from the beginning,
1524 * they will rotate back to anonymous LRU in the end if it failed to
1525 * discard so isolated count will be mismatched.
1526 * Compensate the isolated count for both LRU lists.
1528 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1529 stat.nr_lazyfree_fail);
1530 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1531 -(long)stat.nr_lazyfree_fail);
1532 return nr_reclaimed;
1536 * Attempt to remove the specified page from its LRU. Only take this page
1537 * if it is of the appropriate PageActive status. Pages which are being
1538 * freed elsewhere are also ignored.
1540 * page: page to consider
1541 * mode: one of the LRU isolation modes defined above
1543 * returns 0 on success, -ve errno on failure.
1545 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1549 /* Only take pages on the LRU. */
1553 /* Compaction should not handle unevictable pages but CMA can do so */
1554 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1560 * To minimise LRU disruption, the caller can indicate that it only
1561 * wants to isolate pages it will be able to operate on without
1562 * blocking - clean pages for the most part.
1564 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1565 * that it is possible to migrate without blocking
1567 if (mode & ISOLATE_ASYNC_MIGRATE) {
1568 /* All the caller can do on PageWriteback is block */
1569 if (PageWriteback(page))
1572 if (PageDirty(page)) {
1573 struct address_space *mapping;
1577 * Only pages without mappings or that have a
1578 * ->migratepage callback are possible to migrate
1579 * without blocking. However, we can be racing with
1580 * truncation so it's necessary to lock the page
1581 * to stabilise the mapping as truncation holds
1582 * the page lock until after the page is removed
1583 * from the page cache.
1585 if (!trylock_page(page))
1588 mapping = page_mapping(page);
1589 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1596 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1599 if (likely(get_page_unless_zero(page))) {
1601 * Be careful not to clear PageLRU until after we're
1602 * sure the page is not being freed elsewhere -- the
1603 * page release code relies on it.
1614 * Update LRU sizes after isolating pages. The LRU size updates must
1615 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1617 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1618 enum lru_list lru, unsigned long *nr_zone_taken)
1622 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1623 if (!nr_zone_taken[zid])
1626 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1632 * pgdat->lru_lock is heavily contended. Some of the functions that
1633 * shrink the lists perform better by taking out a batch of pages
1634 * and working on them outside the LRU lock.
1636 * For pagecache intensive workloads, this function is the hottest
1637 * spot in the kernel (apart from copy_*_user functions).
1639 * Appropriate locks must be held before calling this function.
1641 * @nr_to_scan: The number of eligible pages to look through on the list.
1642 * @lruvec: The LRU vector to pull pages from.
1643 * @dst: The temp list to put pages on to.
1644 * @nr_scanned: The number of pages that were scanned.
1645 * @sc: The scan_control struct for this reclaim session
1646 * @lru: LRU list id for isolating
1648 * returns how many pages were moved onto *@dst.
1650 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1651 struct lruvec *lruvec, struct list_head *dst,
1652 unsigned long *nr_scanned, struct scan_control *sc,
1655 struct list_head *src = &lruvec->lists[lru];
1656 unsigned long nr_taken = 0;
1657 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1658 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1659 unsigned long skipped = 0;
1660 unsigned long scan, total_scan, nr_pages;
1661 LIST_HEAD(pages_skipped);
1662 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1666 while (scan < nr_to_scan && !list_empty(src)) {
1669 page = lru_to_page(src);
1670 prefetchw_prev_lru_page(page, src, flags);
1672 VM_BUG_ON_PAGE(!PageLRU(page), page);
1674 nr_pages = compound_nr(page);
1675 total_scan += nr_pages;
1677 if (page_zonenum(page) > sc->reclaim_idx) {
1678 list_move(&page->lru, &pages_skipped);
1679 nr_skipped[page_zonenum(page)] += nr_pages;
1684 * Do not count skipped pages because that makes the function
1685 * return with no isolated pages if the LRU mostly contains
1686 * ineligible pages. This causes the VM to not reclaim any
1687 * pages, triggering a premature OOM.
1689 * Account all tail pages of THP. This would not cause
1690 * premature OOM since __isolate_lru_page() returns -EBUSY
1691 * only when the page is being freed somewhere else.
1694 switch (__isolate_lru_page(page, mode)) {
1696 nr_taken += nr_pages;
1697 nr_zone_taken[page_zonenum(page)] += nr_pages;
1698 list_move(&page->lru, dst);
1702 /* else it is being freed elsewhere */
1703 list_move(&page->lru, src);
1712 * Splice any skipped pages to the start of the LRU list. Note that
1713 * this disrupts the LRU order when reclaiming for lower zones but
1714 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1715 * scanning would soon rescan the same pages to skip and put the
1716 * system at risk of premature OOM.
1718 if (!list_empty(&pages_skipped)) {
1721 list_splice(&pages_skipped, src);
1722 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1723 if (!nr_skipped[zid])
1726 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1727 skipped += nr_skipped[zid];
1730 *nr_scanned = total_scan;
1731 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1732 total_scan, skipped, nr_taken, mode, lru);
1733 update_lru_sizes(lruvec, lru, nr_zone_taken);
1738 * isolate_lru_page - tries to isolate a page from its LRU list
1739 * @page: page to isolate from its LRU list
1741 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1742 * vmstat statistic corresponding to whatever LRU list the page was on.
1744 * Returns 0 if the page was removed from an LRU list.
1745 * Returns -EBUSY if the page was not on an LRU list.
1747 * The returned page will have PageLRU() cleared. If it was found on
1748 * the active list, it will have PageActive set. If it was found on
1749 * the unevictable list, it will have the PageUnevictable bit set. That flag
1750 * may need to be cleared by the caller before letting the page go.
1752 * The vmstat statistic corresponding to the list on which the page was
1753 * found will be decremented.
1757 * (1) Must be called with an elevated refcount on the page. This is a
1758 * fundamental difference from isolate_lru_pages (which is called
1759 * without a stable reference).
1760 * (2) the lru_lock must not be held.
1761 * (3) interrupts must be enabled.
1763 int isolate_lru_page(struct page *page)
1767 VM_BUG_ON_PAGE(!page_count(page), page);
1768 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1770 if (PageLRU(page)) {
1771 pg_data_t *pgdat = page_pgdat(page);
1772 struct lruvec *lruvec;
1774 spin_lock_irq(&pgdat->lru_lock);
1775 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1776 if (PageLRU(page)) {
1777 int lru = page_lru(page);
1780 del_page_from_lru_list(page, lruvec, lru);
1783 spin_unlock_irq(&pgdat->lru_lock);
1789 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1790 * then get rescheduled. When there are massive number of tasks doing page
1791 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1792 * the LRU list will go small and be scanned faster than necessary, leading to
1793 * unnecessary swapping, thrashing and OOM.
1795 static int too_many_isolated(struct pglist_data *pgdat, int file,
1796 struct scan_control *sc)
1798 unsigned long inactive, isolated;
1800 if (current_is_kswapd())
1803 if (!writeback_throttling_sane(sc))
1807 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1808 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1810 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1811 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1815 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1816 * won't get blocked by normal direct-reclaimers, forming a circular
1819 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1822 return isolated > inactive;
1826 * This moves pages from @list to corresponding LRU list.
1828 * We move them the other way if the page is referenced by one or more
1829 * processes, from rmap.
1831 * If the pages are mostly unmapped, the processing is fast and it is
1832 * appropriate to hold zone_lru_lock across the whole operation. But if
1833 * the pages are mapped, the processing is slow (page_referenced()) so we
1834 * should drop zone_lru_lock around each page. It's impossible to balance
1835 * this, so instead we remove the pages from the LRU while processing them.
1836 * It is safe to rely on PG_active against the non-LRU pages in here because
1837 * nobody will play with that bit on a non-LRU page.
1839 * The downside is that we have to touch page->_refcount against each page.
1840 * But we had to alter page->flags anyway.
1842 * Returns the number of pages moved to the given lruvec.
1845 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1846 struct list_head *list)
1848 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1849 int nr_pages, nr_moved = 0;
1850 LIST_HEAD(pages_to_free);
1854 while (!list_empty(list)) {
1855 page = lru_to_page(list);
1856 VM_BUG_ON_PAGE(PageLRU(page), page);
1857 if (unlikely(!page_evictable(page))) {
1858 list_del(&page->lru);
1859 spin_unlock_irq(&pgdat->lru_lock);
1860 putback_lru_page(page);
1861 spin_lock_irq(&pgdat->lru_lock);
1864 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1867 lru = page_lru(page);
1869 nr_pages = thp_nr_pages(page);
1870 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1871 list_move(&page->lru, &lruvec->lists[lru]);
1873 if (put_page_testzero(page)) {
1874 __ClearPageLRU(page);
1875 __ClearPageActive(page);
1876 del_page_from_lru_list(page, lruvec, lru);
1878 if (unlikely(PageCompound(page))) {
1879 spin_unlock_irq(&pgdat->lru_lock);
1880 destroy_compound_page(page);
1881 spin_lock_irq(&pgdat->lru_lock);
1883 list_add(&page->lru, &pages_to_free);
1885 nr_moved += nr_pages;
1886 if (PageActive(page))
1887 workingset_age_nonresident(lruvec, nr_pages);
1892 * To save our caller's stack, now use input list for pages to free.
1894 list_splice(&pages_to_free, list);
1900 * If a kernel thread (such as nfsd for loop-back mounts) services
1901 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1902 * In that case we should only throttle if the backing device it is
1903 * writing to is congested. In other cases it is safe to throttle.
1905 static int current_may_throttle(void)
1907 return !(current->flags & PF_LOCAL_THROTTLE) ||
1908 current->backing_dev_info == NULL ||
1909 bdi_write_congested(current->backing_dev_info);
1913 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1914 * of reclaimed pages
1916 static noinline_for_stack unsigned long
1917 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1918 struct scan_control *sc, enum lru_list lru)
1920 LIST_HEAD(page_list);
1921 unsigned long nr_scanned;
1922 unsigned int nr_reclaimed = 0;
1923 unsigned long nr_taken;
1924 struct reclaim_stat stat;
1925 bool file = is_file_lru(lru);
1926 enum vm_event_item item;
1927 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1928 bool stalled = false;
1930 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1934 /* wait a bit for the reclaimer. */
1938 /* We are about to die and free our memory. Return now. */
1939 if (fatal_signal_pending(current))
1940 return SWAP_CLUSTER_MAX;
1945 spin_lock_irq(&pgdat->lru_lock);
1947 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1948 &nr_scanned, sc, lru);
1950 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1951 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1952 if (!cgroup_reclaim(sc))
1953 __count_vm_events(item, nr_scanned);
1954 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1955 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1957 spin_unlock_irq(&pgdat->lru_lock);
1962 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
1964 spin_lock_irq(&pgdat->lru_lock);
1966 move_pages_to_lru(lruvec, &page_list);
1968 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1969 lru_note_cost(lruvec, file, stat.nr_pageout);
1970 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1971 if (!cgroup_reclaim(sc))
1972 __count_vm_events(item, nr_reclaimed);
1973 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1974 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1976 spin_unlock_irq(&pgdat->lru_lock);
1978 mem_cgroup_uncharge_list(&page_list);
1979 free_unref_page_list(&page_list);
1982 * If dirty pages are scanned that are not queued for IO, it
1983 * implies that flushers are not doing their job. This can
1984 * happen when memory pressure pushes dirty pages to the end of
1985 * the LRU before the dirty limits are breached and the dirty
1986 * data has expired. It can also happen when the proportion of
1987 * dirty pages grows not through writes but through memory
1988 * pressure reclaiming all the clean cache. And in some cases,
1989 * the flushers simply cannot keep up with the allocation
1990 * rate. Nudge the flusher threads in case they are asleep.
1992 if (stat.nr_unqueued_dirty == nr_taken)
1993 wakeup_flusher_threads(WB_REASON_VMSCAN);
1995 sc->nr.dirty += stat.nr_dirty;
1996 sc->nr.congested += stat.nr_congested;
1997 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1998 sc->nr.writeback += stat.nr_writeback;
1999 sc->nr.immediate += stat.nr_immediate;
2000 sc->nr.taken += nr_taken;
2002 sc->nr.file_taken += nr_taken;
2004 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2005 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2006 return nr_reclaimed;
2009 static void shrink_active_list(unsigned long nr_to_scan,
2010 struct lruvec *lruvec,
2011 struct scan_control *sc,
2014 unsigned long nr_taken;
2015 unsigned long nr_scanned;
2016 unsigned long vm_flags;
2017 LIST_HEAD(l_hold); /* The pages which were snipped off */
2018 LIST_HEAD(l_active);
2019 LIST_HEAD(l_inactive);
2021 unsigned nr_deactivate, nr_activate;
2022 unsigned nr_rotated = 0;
2023 int file = is_file_lru(lru);
2024 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2028 spin_lock_irq(&pgdat->lru_lock);
2030 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2031 &nr_scanned, sc, lru);
2033 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2035 if (!cgroup_reclaim(sc))
2036 __count_vm_events(PGREFILL, nr_scanned);
2037 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2039 spin_unlock_irq(&pgdat->lru_lock);
2041 while (!list_empty(&l_hold)) {
2043 page = lru_to_page(&l_hold);
2044 list_del(&page->lru);
2046 if (unlikely(!page_evictable(page))) {
2047 putback_lru_page(page);
2051 if (unlikely(buffer_heads_over_limit)) {
2052 if (page_has_private(page) && trylock_page(page)) {
2053 if (page_has_private(page))
2054 try_to_release_page(page, 0);
2059 if (page_referenced(page, 0, sc->target_mem_cgroup,
2062 * Identify referenced, file-backed active pages and
2063 * give them one more trip around the active list. So
2064 * that executable code get better chances to stay in
2065 * memory under moderate memory pressure. Anon pages
2066 * are not likely to be evicted by use-once streaming
2067 * IO, plus JVM can create lots of anon VM_EXEC pages,
2068 * so we ignore them here.
2070 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2071 nr_rotated += thp_nr_pages(page);
2072 list_add(&page->lru, &l_active);
2077 ClearPageActive(page); /* we are de-activating */
2078 SetPageWorkingset(page);
2079 list_add(&page->lru, &l_inactive);
2083 * Move pages back to the lru list.
2085 spin_lock_irq(&pgdat->lru_lock);
2087 nr_activate = move_pages_to_lru(lruvec, &l_active);
2088 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2089 /* Keep all free pages in l_active list */
2090 list_splice(&l_inactive, &l_active);
2092 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2093 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2095 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2096 spin_unlock_irq(&pgdat->lru_lock);
2098 mem_cgroup_uncharge_list(&l_active);
2099 free_unref_page_list(&l_active);
2100 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2101 nr_deactivate, nr_rotated, sc->priority, file);
2104 unsigned long reclaim_pages(struct list_head *page_list)
2106 int nid = NUMA_NO_NODE;
2107 unsigned int nr_reclaimed = 0;
2108 LIST_HEAD(node_page_list);
2109 struct reclaim_stat dummy_stat;
2111 struct scan_control sc = {
2112 .gfp_mask = GFP_KERNEL,
2113 .priority = DEF_PRIORITY,
2119 while (!list_empty(page_list)) {
2120 page = lru_to_page(page_list);
2121 if (nid == NUMA_NO_NODE) {
2122 nid = page_to_nid(page);
2123 INIT_LIST_HEAD(&node_page_list);
2126 if (nid == page_to_nid(page)) {
2127 ClearPageActive(page);
2128 list_move(&page->lru, &node_page_list);
2132 nr_reclaimed += shrink_page_list(&node_page_list,
2134 &sc, &dummy_stat, false);
2135 while (!list_empty(&node_page_list)) {
2136 page = lru_to_page(&node_page_list);
2137 list_del(&page->lru);
2138 putback_lru_page(page);
2144 if (!list_empty(&node_page_list)) {
2145 nr_reclaimed += shrink_page_list(&node_page_list,
2147 &sc, &dummy_stat, false);
2148 while (!list_empty(&node_page_list)) {
2149 page = lru_to_page(&node_page_list);
2150 list_del(&page->lru);
2151 putback_lru_page(page);
2155 return nr_reclaimed;
2158 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2159 struct lruvec *lruvec, struct scan_control *sc)
2161 if (is_active_lru(lru)) {
2162 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2163 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2165 sc->skipped_deactivate = 1;
2169 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2173 * The inactive anon list should be small enough that the VM never has
2174 * to do too much work.
2176 * The inactive file list should be small enough to leave most memory
2177 * to the established workingset on the scan-resistant active list,
2178 * but large enough to avoid thrashing the aggregate readahead window.
2180 * Both inactive lists should also be large enough that each inactive
2181 * page has a chance to be referenced again before it is reclaimed.
2183 * If that fails and refaulting is observed, the inactive list grows.
2185 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2186 * on this LRU, maintained by the pageout code. An inactive_ratio
2187 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2190 * memory ratio inactive
2191 * -------------------------------------
2200 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2202 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2203 unsigned long inactive, active;
2204 unsigned long inactive_ratio;
2207 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2208 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2210 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2212 inactive_ratio = int_sqrt(10 * gb);
2216 return inactive * inactive_ratio < active;
2227 * Determine how aggressively the anon and file LRU lists should be
2228 * scanned. The relative value of each set of LRU lists is determined
2229 * by looking at the fraction of the pages scanned we did rotate back
2230 * onto the active list instead of evict.
2232 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2233 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2235 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2238 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2239 unsigned long anon_cost, file_cost, total_cost;
2240 int swappiness = mem_cgroup_swappiness(memcg);
2241 u64 fraction[ANON_AND_FILE];
2242 u64 denominator = 0; /* gcc */
2243 enum scan_balance scan_balance;
2244 unsigned long ap, fp;
2247 /* If we have no swap space, do not bother scanning anon pages. */
2248 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2249 scan_balance = SCAN_FILE;
2254 * Global reclaim will swap to prevent OOM even with no
2255 * swappiness, but memcg users want to use this knob to
2256 * disable swapping for individual groups completely when
2257 * using the memory controller's swap limit feature would be
2260 if (cgroup_reclaim(sc) && !swappiness) {
2261 scan_balance = SCAN_FILE;
2266 * Do not apply any pressure balancing cleverness when the
2267 * system is close to OOM, scan both anon and file equally
2268 * (unless the swappiness setting disagrees with swapping).
2270 if (!sc->priority && swappiness) {
2271 scan_balance = SCAN_EQUAL;
2276 * If the system is almost out of file pages, force-scan anon.
2278 if (sc->file_is_tiny) {
2279 scan_balance = SCAN_ANON;
2284 * If there is enough inactive page cache, we do not reclaim
2285 * anything from the anonymous working right now.
2287 if (sc->cache_trim_mode) {
2288 scan_balance = SCAN_FILE;
2292 scan_balance = SCAN_FRACT;
2294 * Calculate the pressure balance between anon and file pages.
2296 * The amount of pressure we put on each LRU is inversely
2297 * proportional to the cost of reclaiming each list, as
2298 * determined by the share of pages that are refaulting, times
2299 * the relative IO cost of bringing back a swapped out
2300 * anonymous page vs reloading a filesystem page (swappiness).
2302 * Although we limit that influence to ensure no list gets
2303 * left behind completely: at least a third of the pressure is
2304 * applied, before swappiness.
2306 * With swappiness at 100, anon and file have equal IO cost.
2308 total_cost = sc->anon_cost + sc->file_cost;
2309 anon_cost = total_cost + sc->anon_cost;
2310 file_cost = total_cost + sc->file_cost;
2311 total_cost = anon_cost + file_cost;
2313 ap = swappiness * (total_cost + 1);
2314 ap /= anon_cost + 1;
2316 fp = (200 - swappiness) * (total_cost + 1);
2317 fp /= file_cost + 1;
2321 denominator = ap + fp;
2323 for_each_evictable_lru(lru) {
2324 int file = is_file_lru(lru);
2325 unsigned long lruvec_size;
2327 unsigned long protection;
2329 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2330 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2332 sc->memcg_low_reclaim);
2336 * Scale a cgroup's reclaim pressure by proportioning
2337 * its current usage to its memory.low or memory.min
2340 * This is important, as otherwise scanning aggression
2341 * becomes extremely binary -- from nothing as we
2342 * approach the memory protection threshold, to totally
2343 * nominal as we exceed it. This results in requiring
2344 * setting extremely liberal protection thresholds. It
2345 * also means we simply get no protection at all if we
2346 * set it too low, which is not ideal.
2348 * If there is any protection in place, we reduce scan
2349 * pressure by how much of the total memory used is
2350 * within protection thresholds.
2352 * There is one special case: in the first reclaim pass,
2353 * we skip over all groups that are within their low
2354 * protection. If that fails to reclaim enough pages to
2355 * satisfy the reclaim goal, we come back and override
2356 * the best-effort low protection. However, we still
2357 * ideally want to honor how well-behaved groups are in
2358 * that case instead of simply punishing them all
2359 * equally. As such, we reclaim them based on how much
2360 * memory they are using, reducing the scan pressure
2361 * again by how much of the total memory used is under
2364 unsigned long cgroup_size = mem_cgroup_size(memcg);
2366 /* Avoid TOCTOU with earlier protection check */
2367 cgroup_size = max(cgroup_size, protection);
2369 scan = lruvec_size - lruvec_size * protection /
2373 * Minimally target SWAP_CLUSTER_MAX pages to keep
2374 * reclaim moving forwards, avoiding decrementing
2375 * sc->priority further than desirable.
2377 scan = max(scan, SWAP_CLUSTER_MAX);
2382 scan >>= sc->priority;
2385 * If the cgroup's already been deleted, make sure to
2386 * scrape out the remaining cache.
2388 if (!scan && !mem_cgroup_online(memcg))
2389 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2391 switch (scan_balance) {
2393 /* Scan lists relative to size */
2397 * Scan types proportional to swappiness and
2398 * their relative recent reclaim efficiency.
2399 * Make sure we don't miss the last page on
2400 * the offlined memory cgroups because of a
2403 scan = mem_cgroup_online(memcg) ?
2404 div64_u64(scan * fraction[file], denominator) :
2405 DIV64_U64_ROUND_UP(scan * fraction[file],
2410 /* Scan one type exclusively */
2411 if ((scan_balance == SCAN_FILE) != file)
2415 /* Look ma, no brain */
2423 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2425 unsigned long nr[NR_LRU_LISTS];
2426 unsigned long targets[NR_LRU_LISTS];
2427 unsigned long nr_to_scan;
2429 unsigned long nr_reclaimed = 0;
2430 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2431 struct blk_plug plug;
2434 get_scan_count(lruvec, sc, nr);
2436 /* Record the original scan target for proportional adjustments later */
2437 memcpy(targets, nr, sizeof(nr));
2440 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2441 * event that can occur when there is little memory pressure e.g.
2442 * multiple streaming readers/writers. Hence, we do not abort scanning
2443 * when the requested number of pages are reclaimed when scanning at
2444 * DEF_PRIORITY on the assumption that the fact we are direct
2445 * reclaiming implies that kswapd is not keeping up and it is best to
2446 * do a batch of work at once. For memcg reclaim one check is made to
2447 * abort proportional reclaim if either the file or anon lru has already
2448 * dropped to zero at the first pass.
2450 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2451 sc->priority == DEF_PRIORITY);
2453 blk_start_plug(&plug);
2454 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2455 nr[LRU_INACTIVE_FILE]) {
2456 unsigned long nr_anon, nr_file, percentage;
2457 unsigned long nr_scanned;
2459 for_each_evictable_lru(lru) {
2461 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2462 nr[lru] -= nr_to_scan;
2464 nr_reclaimed += shrink_list(lru, nr_to_scan,
2471 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2475 * For kswapd and memcg, reclaim at least the number of pages
2476 * requested. Ensure that the anon and file LRUs are scanned
2477 * proportionally what was requested by get_scan_count(). We
2478 * stop reclaiming one LRU and reduce the amount scanning
2479 * proportional to the original scan target.
2481 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2482 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2485 * It's just vindictive to attack the larger once the smaller
2486 * has gone to zero. And given the way we stop scanning the
2487 * smaller below, this makes sure that we only make one nudge
2488 * towards proportionality once we've got nr_to_reclaim.
2490 if (!nr_file || !nr_anon)
2493 if (nr_file > nr_anon) {
2494 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2495 targets[LRU_ACTIVE_ANON] + 1;
2497 percentage = nr_anon * 100 / scan_target;
2499 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2500 targets[LRU_ACTIVE_FILE] + 1;
2502 percentage = nr_file * 100 / scan_target;
2505 /* Stop scanning the smaller of the LRU */
2507 nr[lru + LRU_ACTIVE] = 0;
2510 * Recalculate the other LRU scan count based on its original
2511 * scan target and the percentage scanning already complete
2513 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2514 nr_scanned = targets[lru] - nr[lru];
2515 nr[lru] = targets[lru] * (100 - percentage) / 100;
2516 nr[lru] -= min(nr[lru], nr_scanned);
2519 nr_scanned = targets[lru] - nr[lru];
2520 nr[lru] = targets[lru] * (100 - percentage) / 100;
2521 nr[lru] -= min(nr[lru], nr_scanned);
2523 scan_adjusted = true;
2525 blk_finish_plug(&plug);
2526 sc->nr_reclaimed += nr_reclaimed;
2529 * Even if we did not try to evict anon pages at all, we want to
2530 * rebalance the anon lru active/inactive ratio.
2532 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2533 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2534 sc, LRU_ACTIVE_ANON);
2537 /* Use reclaim/compaction for costly allocs or under memory pressure */
2538 static bool in_reclaim_compaction(struct scan_control *sc)
2540 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2541 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2542 sc->priority < DEF_PRIORITY - 2))
2549 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2550 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2551 * true if more pages should be reclaimed such that when the page allocator
2552 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2553 * It will give up earlier than that if there is difficulty reclaiming pages.
2555 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2556 unsigned long nr_reclaimed,
2557 struct scan_control *sc)
2559 unsigned long pages_for_compaction;
2560 unsigned long inactive_lru_pages;
2563 /* If not in reclaim/compaction mode, stop */
2564 if (!in_reclaim_compaction(sc))
2568 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2569 * number of pages that were scanned. This will return to the caller
2570 * with the risk reclaim/compaction and the resulting allocation attempt
2571 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2572 * allocations through requiring that the full LRU list has been scanned
2573 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2574 * scan, but that approximation was wrong, and there were corner cases
2575 * where always a non-zero amount of pages were scanned.
2580 /* If compaction would go ahead or the allocation would succeed, stop */
2581 for (z = 0; z <= sc->reclaim_idx; z++) {
2582 struct zone *zone = &pgdat->node_zones[z];
2583 if (!managed_zone(zone))
2586 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2587 case COMPACT_SUCCESS:
2588 case COMPACT_CONTINUE:
2591 /* check next zone */
2597 * If we have not reclaimed enough pages for compaction and the
2598 * inactive lists are large enough, continue reclaiming
2600 pages_for_compaction = compact_gap(sc->order);
2601 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2602 if (get_nr_swap_pages() > 0)
2603 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2605 return inactive_lru_pages > pages_for_compaction;
2608 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2610 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2611 struct mem_cgroup *memcg;
2613 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2615 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2616 unsigned long reclaimed;
2617 unsigned long scanned;
2620 * This loop can become CPU-bound when target memcgs
2621 * aren't eligible for reclaim - either because they
2622 * don't have any reclaimable pages, or because their
2623 * memory is explicitly protected. Avoid soft lockups.
2627 mem_cgroup_calculate_protection(target_memcg, memcg);
2629 if (mem_cgroup_below_min(memcg)) {
2632 * If there is no reclaimable memory, OOM.
2635 } else if (mem_cgroup_below_low(memcg)) {
2638 * Respect the protection only as long as
2639 * there is an unprotected supply
2640 * of reclaimable memory from other cgroups.
2642 if (!sc->memcg_low_reclaim) {
2643 sc->memcg_low_skipped = 1;
2646 memcg_memory_event(memcg, MEMCG_LOW);
2649 reclaimed = sc->nr_reclaimed;
2650 scanned = sc->nr_scanned;
2652 shrink_lruvec(lruvec, sc);
2654 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2657 /* Record the group's reclaim efficiency */
2658 vmpressure(sc->gfp_mask, memcg, false,
2659 sc->nr_scanned - scanned,
2660 sc->nr_reclaimed - reclaimed);
2662 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2665 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2667 struct reclaim_state *reclaim_state = current->reclaim_state;
2668 unsigned long nr_reclaimed, nr_scanned;
2669 struct lruvec *target_lruvec;
2670 bool reclaimable = false;
2673 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2676 memset(&sc->nr, 0, sizeof(sc->nr));
2678 nr_reclaimed = sc->nr_reclaimed;
2679 nr_scanned = sc->nr_scanned;
2682 * Determine the scan balance between anon and file LRUs.
2684 spin_lock_irq(&pgdat->lru_lock);
2685 sc->anon_cost = target_lruvec->anon_cost;
2686 sc->file_cost = target_lruvec->file_cost;
2687 spin_unlock_irq(&pgdat->lru_lock);
2690 * Target desirable inactive:active list ratios for the anon
2691 * and file LRU lists.
2693 if (!sc->force_deactivate) {
2694 unsigned long refaults;
2696 refaults = lruvec_page_state(target_lruvec,
2697 WORKINGSET_ACTIVATE_ANON);
2698 if (refaults != target_lruvec->refaults[0] ||
2699 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2700 sc->may_deactivate |= DEACTIVATE_ANON;
2702 sc->may_deactivate &= ~DEACTIVATE_ANON;
2705 * When refaults are being observed, it means a new
2706 * workingset is being established. Deactivate to get
2707 * rid of any stale active pages quickly.
2709 refaults = lruvec_page_state(target_lruvec,
2710 WORKINGSET_ACTIVATE_FILE);
2711 if (refaults != target_lruvec->refaults[1] ||
2712 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2713 sc->may_deactivate |= DEACTIVATE_FILE;
2715 sc->may_deactivate &= ~DEACTIVATE_FILE;
2717 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2720 * If we have plenty of inactive file pages that aren't
2721 * thrashing, try to reclaim those first before touching
2724 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2725 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2726 sc->cache_trim_mode = 1;
2728 sc->cache_trim_mode = 0;
2731 * Prevent the reclaimer from falling into the cache trap: as
2732 * cache pages start out inactive, every cache fault will tip
2733 * the scan balance towards the file LRU. And as the file LRU
2734 * shrinks, so does the window for rotation from references.
2735 * This means we have a runaway feedback loop where a tiny
2736 * thrashing file LRU becomes infinitely more attractive than
2737 * anon pages. Try to detect this based on file LRU size.
2739 if (!cgroup_reclaim(sc)) {
2740 unsigned long total_high_wmark = 0;
2741 unsigned long free, anon;
2744 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2745 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2746 node_page_state(pgdat, NR_INACTIVE_FILE);
2748 for (z = 0; z < MAX_NR_ZONES; z++) {
2749 struct zone *zone = &pgdat->node_zones[z];
2750 if (!managed_zone(zone))
2753 total_high_wmark += high_wmark_pages(zone);
2757 * Consider anon: if that's low too, this isn't a
2758 * runaway file reclaim problem, but rather just
2759 * extreme pressure. Reclaim as per usual then.
2761 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2764 file + free <= total_high_wmark &&
2765 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2766 anon >> sc->priority;
2769 shrink_node_memcgs(pgdat, sc);
2771 if (reclaim_state) {
2772 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2773 reclaim_state->reclaimed_slab = 0;
2776 /* Record the subtree's reclaim efficiency */
2777 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2778 sc->nr_scanned - nr_scanned,
2779 sc->nr_reclaimed - nr_reclaimed);
2781 if (sc->nr_reclaimed - nr_reclaimed)
2784 if (current_is_kswapd()) {
2786 * If reclaim is isolating dirty pages under writeback,
2787 * it implies that the long-lived page allocation rate
2788 * is exceeding the page laundering rate. Either the
2789 * global limits are not being effective at throttling
2790 * processes due to the page distribution throughout
2791 * zones or there is heavy usage of a slow backing
2792 * device. The only option is to throttle from reclaim
2793 * context which is not ideal as there is no guarantee
2794 * the dirtying process is throttled in the same way
2795 * balance_dirty_pages() manages.
2797 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2798 * count the number of pages under pages flagged for
2799 * immediate reclaim and stall if any are encountered
2800 * in the nr_immediate check below.
2802 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2803 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2805 /* Allow kswapd to start writing pages during reclaim.*/
2806 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2807 set_bit(PGDAT_DIRTY, &pgdat->flags);
2810 * If kswapd scans pages marked for immediate
2811 * reclaim and under writeback (nr_immediate), it
2812 * implies that pages are cycling through the LRU
2813 * faster than they are written so also forcibly stall.
2815 if (sc->nr.immediate)
2816 congestion_wait(BLK_RW_ASYNC, HZ/10);
2820 * Tag a node/memcg as congested if all the dirty pages
2821 * scanned were backed by a congested BDI and
2822 * wait_iff_congested will stall.
2824 * Legacy memcg will stall in page writeback so avoid forcibly
2825 * stalling in wait_iff_congested().
2827 if ((current_is_kswapd() ||
2828 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2829 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2830 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2833 * Stall direct reclaim for IO completions if underlying BDIs
2834 * and node is congested. Allow kswapd to continue until it
2835 * starts encountering unqueued dirty pages or cycling through
2836 * the LRU too quickly.
2838 if (!current_is_kswapd() && current_may_throttle() &&
2839 !sc->hibernation_mode &&
2840 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2841 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2843 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2848 * Kswapd gives up on balancing particular nodes after too
2849 * many failures to reclaim anything from them and goes to
2850 * sleep. On reclaim progress, reset the failure counter. A
2851 * successful direct reclaim run will revive a dormant kswapd.
2854 pgdat->kswapd_failures = 0;
2858 * Returns true if compaction should go ahead for a costly-order request, or
2859 * the allocation would already succeed without compaction. Return false if we
2860 * should reclaim first.
2862 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2864 unsigned long watermark;
2865 enum compact_result suitable;
2867 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2868 if (suitable == COMPACT_SUCCESS)
2869 /* Allocation should succeed already. Don't reclaim. */
2871 if (suitable == COMPACT_SKIPPED)
2872 /* Compaction cannot yet proceed. Do reclaim. */
2876 * Compaction is already possible, but it takes time to run and there
2877 * are potentially other callers using the pages just freed. So proceed
2878 * with reclaim to make a buffer of free pages available to give
2879 * compaction a reasonable chance of completing and allocating the page.
2880 * Note that we won't actually reclaim the whole buffer in one attempt
2881 * as the target watermark in should_continue_reclaim() is lower. But if
2882 * we are already above the high+gap watermark, don't reclaim at all.
2884 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2886 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2890 * This is the direct reclaim path, for page-allocating processes. We only
2891 * try to reclaim pages from zones which will satisfy the caller's allocation
2894 * If a zone is deemed to be full of pinned pages then just give it a light
2895 * scan then give up on it.
2897 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2901 unsigned long nr_soft_reclaimed;
2902 unsigned long nr_soft_scanned;
2904 pg_data_t *last_pgdat = NULL;
2907 * If the number of buffer_heads in the machine exceeds the maximum
2908 * allowed level, force direct reclaim to scan the highmem zone as
2909 * highmem pages could be pinning lowmem pages storing buffer_heads
2911 orig_mask = sc->gfp_mask;
2912 if (buffer_heads_over_limit) {
2913 sc->gfp_mask |= __GFP_HIGHMEM;
2914 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2917 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2918 sc->reclaim_idx, sc->nodemask) {
2920 * Take care memory controller reclaiming has small influence
2923 if (!cgroup_reclaim(sc)) {
2924 if (!cpuset_zone_allowed(zone,
2925 GFP_KERNEL | __GFP_HARDWALL))
2929 * If we already have plenty of memory free for
2930 * compaction in this zone, don't free any more.
2931 * Even though compaction is invoked for any
2932 * non-zero order, only frequent costly order
2933 * reclamation is disruptive enough to become a
2934 * noticeable problem, like transparent huge
2937 if (IS_ENABLED(CONFIG_COMPACTION) &&
2938 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2939 compaction_ready(zone, sc)) {
2940 sc->compaction_ready = true;
2945 * Shrink each node in the zonelist once. If the
2946 * zonelist is ordered by zone (not the default) then a
2947 * node may be shrunk multiple times but in that case
2948 * the user prefers lower zones being preserved.
2950 if (zone->zone_pgdat == last_pgdat)
2954 * This steals pages from memory cgroups over softlimit
2955 * and returns the number of reclaimed pages and
2956 * scanned pages. This works for global memory pressure
2957 * and balancing, not for a memcg's limit.
2959 nr_soft_scanned = 0;
2960 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2961 sc->order, sc->gfp_mask,
2963 sc->nr_reclaimed += nr_soft_reclaimed;
2964 sc->nr_scanned += nr_soft_scanned;
2965 /* need some check for avoid more shrink_zone() */
2968 /* See comment about same check for global reclaim above */
2969 if (zone->zone_pgdat == last_pgdat)
2971 last_pgdat = zone->zone_pgdat;
2972 shrink_node(zone->zone_pgdat, sc);
2976 * Restore to original mask to avoid the impact on the caller if we
2977 * promoted it to __GFP_HIGHMEM.
2979 sc->gfp_mask = orig_mask;
2982 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2984 struct lruvec *target_lruvec;
2985 unsigned long refaults;
2987 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2988 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
2989 target_lruvec->refaults[0] = refaults;
2990 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
2991 target_lruvec->refaults[1] = refaults;
2995 * This is the main entry point to direct page reclaim.
2997 * If a full scan of the inactive list fails to free enough memory then we
2998 * are "out of memory" and something needs to be killed.
3000 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3001 * high - the zone may be full of dirty or under-writeback pages, which this
3002 * caller can't do much about. We kick the writeback threads and take explicit
3003 * naps in the hope that some of these pages can be written. But if the
3004 * allocating task holds filesystem locks which prevent writeout this might not
3005 * work, and the allocation attempt will fail.
3007 * returns: 0, if no pages reclaimed
3008 * else, the number of pages reclaimed
3010 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3011 struct scan_control *sc)
3013 int initial_priority = sc->priority;
3014 pg_data_t *last_pgdat;
3018 delayacct_freepages_start();
3020 if (!cgroup_reclaim(sc))
3021 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3024 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3027 shrink_zones(zonelist, sc);
3029 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3032 if (sc->compaction_ready)
3036 * If we're getting trouble reclaiming, start doing
3037 * writepage even in laptop mode.
3039 if (sc->priority < DEF_PRIORITY - 2)
3040 sc->may_writepage = 1;
3041 } while (--sc->priority >= 0);
3044 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3046 if (zone->zone_pgdat == last_pgdat)
3048 last_pgdat = zone->zone_pgdat;
3050 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3052 if (cgroup_reclaim(sc)) {
3053 struct lruvec *lruvec;
3055 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3057 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3061 delayacct_freepages_end();
3063 if (sc->nr_reclaimed)
3064 return sc->nr_reclaimed;
3066 /* Aborted reclaim to try compaction? don't OOM, then */
3067 if (sc->compaction_ready)
3071 * We make inactive:active ratio decisions based on the node's
3072 * composition of memory, but a restrictive reclaim_idx or a
3073 * memory.low cgroup setting can exempt large amounts of
3074 * memory from reclaim. Neither of which are very common, so
3075 * instead of doing costly eligibility calculations of the
3076 * entire cgroup subtree up front, we assume the estimates are
3077 * good, and retry with forcible deactivation if that fails.
3079 if (sc->skipped_deactivate) {
3080 sc->priority = initial_priority;
3081 sc->force_deactivate = 1;
3082 sc->skipped_deactivate = 0;
3086 /* Untapped cgroup reserves? Don't OOM, retry. */
3087 if (sc->memcg_low_skipped) {
3088 sc->priority = initial_priority;
3089 sc->force_deactivate = 0;
3090 sc->memcg_low_reclaim = 1;
3091 sc->memcg_low_skipped = 0;
3098 static bool allow_direct_reclaim(pg_data_t *pgdat)
3101 unsigned long pfmemalloc_reserve = 0;
3102 unsigned long free_pages = 0;
3106 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3109 for (i = 0; i <= ZONE_NORMAL; i++) {
3110 zone = &pgdat->node_zones[i];
3111 if (!managed_zone(zone))
3114 if (!zone_reclaimable_pages(zone))
3117 pfmemalloc_reserve += min_wmark_pages(zone);
3118 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3121 /* If there are no reserves (unexpected config) then do not throttle */
3122 if (!pfmemalloc_reserve)
3125 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3127 /* kswapd must be awake if processes are being throttled */
3128 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3129 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3130 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3132 wake_up_interruptible(&pgdat->kswapd_wait);
3139 * Throttle direct reclaimers if backing storage is backed by the network
3140 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3141 * depleted. kswapd will continue to make progress and wake the processes
3142 * when the low watermark is reached.
3144 * Returns true if a fatal signal was delivered during throttling. If this
3145 * happens, the page allocator should not consider triggering the OOM killer.
3147 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3148 nodemask_t *nodemask)
3152 pg_data_t *pgdat = NULL;
3155 * Kernel threads should not be throttled as they may be indirectly
3156 * responsible for cleaning pages necessary for reclaim to make forward
3157 * progress. kjournald for example may enter direct reclaim while
3158 * committing a transaction where throttling it could forcing other
3159 * processes to block on log_wait_commit().
3161 if (current->flags & PF_KTHREAD)
3165 * If a fatal signal is pending, this process should not throttle.
3166 * It should return quickly so it can exit and free its memory
3168 if (fatal_signal_pending(current))
3172 * Check if the pfmemalloc reserves are ok by finding the first node
3173 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3174 * GFP_KERNEL will be required for allocating network buffers when
3175 * swapping over the network so ZONE_HIGHMEM is unusable.
3177 * Throttling is based on the first usable node and throttled processes
3178 * wait on a queue until kswapd makes progress and wakes them. There
3179 * is an affinity then between processes waking up and where reclaim
3180 * progress has been made assuming the process wakes on the same node.
3181 * More importantly, processes running on remote nodes will not compete
3182 * for remote pfmemalloc reserves and processes on different nodes
3183 * should make reasonable progress.
3185 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3186 gfp_zone(gfp_mask), nodemask) {
3187 if (zone_idx(zone) > ZONE_NORMAL)
3190 /* Throttle based on the first usable node */
3191 pgdat = zone->zone_pgdat;
3192 if (allow_direct_reclaim(pgdat))
3197 /* If no zone was usable by the allocation flags then do not throttle */
3201 /* Account for the throttling */
3202 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3205 * If the caller cannot enter the filesystem, it's possible that it
3206 * is due to the caller holding an FS lock or performing a journal
3207 * transaction in the case of a filesystem like ext[3|4]. In this case,
3208 * it is not safe to block on pfmemalloc_wait as kswapd could be
3209 * blocked waiting on the same lock. Instead, throttle for up to a
3210 * second before continuing.
3212 if (!(gfp_mask & __GFP_FS)) {
3213 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3214 allow_direct_reclaim(pgdat), HZ);
3219 /* Throttle until kswapd wakes the process */
3220 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3221 allow_direct_reclaim(pgdat));
3224 if (fatal_signal_pending(current))
3231 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3232 gfp_t gfp_mask, nodemask_t *nodemask)
3234 unsigned long nr_reclaimed;
3235 struct scan_control sc = {
3236 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3237 .gfp_mask = current_gfp_context(gfp_mask),
3238 .reclaim_idx = gfp_zone(gfp_mask),
3240 .nodemask = nodemask,
3241 .priority = DEF_PRIORITY,
3242 .may_writepage = !laptop_mode,
3248 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3249 * Confirm they are large enough for max values.
3251 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3252 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3253 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3256 * Do not enter reclaim if fatal signal was delivered while throttled.
3257 * 1 is returned so that the page allocator does not OOM kill at this
3260 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3263 set_task_reclaim_state(current, &sc.reclaim_state);
3264 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3266 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3268 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3269 set_task_reclaim_state(current, NULL);
3271 return nr_reclaimed;
3276 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3277 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3278 gfp_t gfp_mask, bool noswap,
3280 unsigned long *nr_scanned)
3282 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3283 struct scan_control sc = {
3284 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3285 .target_mem_cgroup = memcg,
3286 .may_writepage = !laptop_mode,
3288 .reclaim_idx = MAX_NR_ZONES - 1,
3289 .may_swap = !noswap,
3292 WARN_ON_ONCE(!current->reclaim_state);
3294 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3295 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3297 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3301 * NOTE: Although we can get the priority field, using it
3302 * here is not a good idea, since it limits the pages we can scan.
3303 * if we don't reclaim here, the shrink_node from balance_pgdat
3304 * will pick up pages from other mem cgroup's as well. We hack
3305 * the priority and make it zero.
3307 shrink_lruvec(lruvec, &sc);
3309 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3311 *nr_scanned = sc.nr_scanned;
3313 return sc.nr_reclaimed;
3316 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3317 unsigned long nr_pages,
3321 unsigned long nr_reclaimed;
3322 unsigned int noreclaim_flag;
3323 struct scan_control sc = {
3324 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3325 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3326 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3327 .reclaim_idx = MAX_NR_ZONES - 1,
3328 .target_mem_cgroup = memcg,
3329 .priority = DEF_PRIORITY,
3330 .may_writepage = !laptop_mode,
3332 .may_swap = may_swap,
3335 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3336 * equal pressure on all the nodes. This is based on the assumption that
3337 * the reclaim does not bail out early.
3339 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3341 set_task_reclaim_state(current, &sc.reclaim_state);
3342 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3343 noreclaim_flag = memalloc_noreclaim_save();
3345 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3347 memalloc_noreclaim_restore(noreclaim_flag);
3348 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3349 set_task_reclaim_state(current, NULL);
3351 return nr_reclaimed;
3355 static void age_active_anon(struct pglist_data *pgdat,
3356 struct scan_control *sc)
3358 struct mem_cgroup *memcg;
3359 struct lruvec *lruvec;
3361 if (!total_swap_pages)
3364 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3365 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3368 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3370 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3371 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3372 sc, LRU_ACTIVE_ANON);
3373 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3377 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3383 * Check for watermark boosts top-down as the higher zones
3384 * are more likely to be boosted. Both watermarks and boosts
3385 * should not be checked at the same time as reclaim would
3386 * start prematurely when there is no boosting and a lower
3389 for (i = highest_zoneidx; i >= 0; i--) {
3390 zone = pgdat->node_zones + i;
3391 if (!managed_zone(zone))
3394 if (zone->watermark_boost)
3402 * Returns true if there is an eligible zone balanced for the request order
3403 * and highest_zoneidx
3405 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3408 unsigned long mark = -1;
3412 * Check watermarks bottom-up as lower zones are more likely to
3415 for (i = 0; i <= highest_zoneidx; i++) {
3416 zone = pgdat->node_zones + i;
3418 if (!managed_zone(zone))
3421 mark = high_wmark_pages(zone);
3422 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3427 * If a node has no populated zone within highest_zoneidx, it does not
3428 * need balancing by definition. This can happen if a zone-restricted
3429 * allocation tries to wake a remote kswapd.
3437 /* Clear pgdat state for congested, dirty or under writeback. */
3438 static void clear_pgdat_congested(pg_data_t *pgdat)
3440 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3442 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3443 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3444 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3448 * Prepare kswapd for sleeping. This verifies that there are no processes
3449 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3451 * Returns true if kswapd is ready to sleep
3453 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3454 int highest_zoneidx)
3457 * The throttled processes are normally woken up in balance_pgdat() as
3458 * soon as allow_direct_reclaim() is true. But there is a potential
3459 * race between when kswapd checks the watermarks and a process gets
3460 * throttled. There is also a potential race if processes get
3461 * throttled, kswapd wakes, a large process exits thereby balancing the
3462 * zones, which causes kswapd to exit balance_pgdat() before reaching
3463 * the wake up checks. If kswapd is going to sleep, no process should
3464 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3465 * the wake up is premature, processes will wake kswapd and get
3466 * throttled again. The difference from wake ups in balance_pgdat() is
3467 * that here we are under prepare_to_wait().
3469 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3470 wake_up_all(&pgdat->pfmemalloc_wait);
3472 /* Hopeless node, leave it to direct reclaim */
3473 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3476 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3477 clear_pgdat_congested(pgdat);
3485 * kswapd shrinks a node of pages that are at or below the highest usable
3486 * zone that is currently unbalanced.
3488 * Returns true if kswapd scanned at least the requested number of pages to
3489 * reclaim or if the lack of progress was due to pages under writeback.
3490 * This is used to determine if the scanning priority needs to be raised.
3492 static bool kswapd_shrink_node(pg_data_t *pgdat,
3493 struct scan_control *sc)
3498 /* Reclaim a number of pages proportional to the number of zones */
3499 sc->nr_to_reclaim = 0;
3500 for (z = 0; z <= sc->reclaim_idx; z++) {
3501 zone = pgdat->node_zones + z;
3502 if (!managed_zone(zone))
3505 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3509 * Historically care was taken to put equal pressure on all zones but
3510 * now pressure is applied based on node LRU order.
3512 shrink_node(pgdat, sc);
3515 * Fragmentation may mean that the system cannot be rebalanced for
3516 * high-order allocations. If twice the allocation size has been
3517 * reclaimed then recheck watermarks only at order-0 to prevent
3518 * excessive reclaim. Assume that a process requested a high-order
3519 * can direct reclaim/compact.
3521 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3524 return sc->nr_scanned >= sc->nr_to_reclaim;
3528 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3529 * that are eligible for use by the caller until at least one zone is
3532 * Returns the order kswapd finished reclaiming at.
3534 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3535 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3536 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3537 * or lower is eligible for reclaim until at least one usable zone is
3540 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3543 unsigned long nr_soft_reclaimed;
3544 unsigned long nr_soft_scanned;
3545 unsigned long pflags;
3546 unsigned long nr_boost_reclaim;
3547 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3550 struct scan_control sc = {
3551 .gfp_mask = GFP_KERNEL,
3556 set_task_reclaim_state(current, &sc.reclaim_state);
3557 psi_memstall_enter(&pflags);
3558 __fs_reclaim_acquire();
3560 count_vm_event(PAGEOUTRUN);
3563 * Account for the reclaim boost. Note that the zone boost is left in
3564 * place so that parallel allocations that are near the watermark will
3565 * stall or direct reclaim until kswapd is finished.
3567 nr_boost_reclaim = 0;
3568 for (i = 0; i <= highest_zoneidx; i++) {
3569 zone = pgdat->node_zones + i;
3570 if (!managed_zone(zone))
3573 nr_boost_reclaim += zone->watermark_boost;
3574 zone_boosts[i] = zone->watermark_boost;
3576 boosted = nr_boost_reclaim;
3579 sc.priority = DEF_PRIORITY;
3581 unsigned long nr_reclaimed = sc.nr_reclaimed;
3582 bool raise_priority = true;
3586 sc.reclaim_idx = highest_zoneidx;
3589 * If the number of buffer_heads exceeds the maximum allowed
3590 * then consider reclaiming from all zones. This has a dual
3591 * purpose -- on 64-bit systems it is expected that
3592 * buffer_heads are stripped during active rotation. On 32-bit
3593 * systems, highmem pages can pin lowmem memory and shrinking
3594 * buffers can relieve lowmem pressure. Reclaim may still not
3595 * go ahead if all eligible zones for the original allocation
3596 * request are balanced to avoid excessive reclaim from kswapd.
3598 if (buffer_heads_over_limit) {
3599 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3600 zone = pgdat->node_zones + i;
3601 if (!managed_zone(zone))
3610 * If the pgdat is imbalanced then ignore boosting and preserve
3611 * the watermarks for a later time and restart. Note that the
3612 * zone watermarks will be still reset at the end of balancing
3613 * on the grounds that the normal reclaim should be enough to
3614 * re-evaluate if boosting is required when kswapd next wakes.
3616 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3617 if (!balanced && nr_boost_reclaim) {
3618 nr_boost_reclaim = 0;
3623 * If boosting is not active then only reclaim if there are no
3624 * eligible zones. Note that sc.reclaim_idx is not used as
3625 * buffer_heads_over_limit may have adjusted it.
3627 if (!nr_boost_reclaim && balanced)
3630 /* Limit the priority of boosting to avoid reclaim writeback */
3631 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3632 raise_priority = false;
3635 * Do not writeback or swap pages for boosted reclaim. The
3636 * intent is to relieve pressure not issue sub-optimal IO
3637 * from reclaim context. If no pages are reclaimed, the
3638 * reclaim will be aborted.
3640 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3641 sc.may_swap = !nr_boost_reclaim;
3644 * Do some background aging of the anon list, to give
3645 * pages a chance to be referenced before reclaiming. All
3646 * pages are rotated regardless of classzone as this is
3647 * about consistent aging.
3649 age_active_anon(pgdat, &sc);
3652 * If we're getting trouble reclaiming, start doing writepage
3653 * even in laptop mode.
3655 if (sc.priority < DEF_PRIORITY - 2)
3656 sc.may_writepage = 1;
3658 /* Call soft limit reclaim before calling shrink_node. */
3660 nr_soft_scanned = 0;
3661 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3662 sc.gfp_mask, &nr_soft_scanned);
3663 sc.nr_reclaimed += nr_soft_reclaimed;
3666 * There should be no need to raise the scanning priority if
3667 * enough pages are already being scanned that that high
3668 * watermark would be met at 100% efficiency.
3670 if (kswapd_shrink_node(pgdat, &sc))
3671 raise_priority = false;
3674 * If the low watermark is met there is no need for processes
3675 * to be throttled on pfmemalloc_wait as they should not be
3676 * able to safely make forward progress. Wake them
3678 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3679 allow_direct_reclaim(pgdat))
3680 wake_up_all(&pgdat->pfmemalloc_wait);
3682 /* Check if kswapd should be suspending */
3683 __fs_reclaim_release();
3684 ret = try_to_freeze();
3685 __fs_reclaim_acquire();
3686 if (ret || kthread_should_stop())
3690 * Raise priority if scanning rate is too low or there was no
3691 * progress in reclaiming pages
3693 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3694 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3697 * If reclaim made no progress for a boost, stop reclaim as
3698 * IO cannot be queued and it could be an infinite loop in
3699 * extreme circumstances.
3701 if (nr_boost_reclaim && !nr_reclaimed)
3704 if (raise_priority || !nr_reclaimed)
3706 } while (sc.priority >= 1);
3708 if (!sc.nr_reclaimed)
3709 pgdat->kswapd_failures++;
3712 /* If reclaim was boosted, account for the reclaim done in this pass */
3714 unsigned long flags;
3716 for (i = 0; i <= highest_zoneidx; i++) {
3717 if (!zone_boosts[i])
3720 /* Increments are under the zone lock */
3721 zone = pgdat->node_zones + i;
3722 spin_lock_irqsave(&zone->lock, flags);
3723 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3724 spin_unlock_irqrestore(&zone->lock, flags);
3728 * As there is now likely space, wakeup kcompact to defragment
3731 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3734 snapshot_refaults(NULL, pgdat);
3735 __fs_reclaim_release();
3736 psi_memstall_leave(&pflags);
3737 set_task_reclaim_state(current, NULL);
3740 * Return the order kswapd stopped reclaiming at as
3741 * prepare_kswapd_sleep() takes it into account. If another caller
3742 * entered the allocator slow path while kswapd was awake, order will
3743 * remain at the higher level.
3749 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3750 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3751 * not a valid index then either kswapd runs for first time or kswapd couldn't
3752 * sleep after previous reclaim attempt (node is still unbalanced). In that
3753 * case return the zone index of the previous kswapd reclaim cycle.
3755 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3756 enum zone_type prev_highest_zoneidx)
3758 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3760 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3763 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3764 unsigned int highest_zoneidx)
3769 if (freezing(current) || kthread_should_stop())
3772 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3775 * Try to sleep for a short interval. Note that kcompactd will only be
3776 * woken if it is possible to sleep for a short interval. This is
3777 * deliberate on the assumption that if reclaim cannot keep an
3778 * eligible zone balanced that it's also unlikely that compaction will
3781 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3783 * Compaction records what page blocks it recently failed to
3784 * isolate pages from and skips them in the future scanning.
3785 * When kswapd is going to sleep, it is reasonable to assume
3786 * that pages and compaction may succeed so reset the cache.
3788 reset_isolation_suitable(pgdat);
3791 * We have freed the memory, now we should compact it to make
3792 * allocation of the requested order possible.
3794 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3796 remaining = schedule_timeout(HZ/10);
3799 * If woken prematurely then reset kswapd_highest_zoneidx and
3800 * order. The values will either be from a wakeup request or
3801 * the previous request that slept prematurely.
3804 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3805 kswapd_highest_zoneidx(pgdat,
3808 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3809 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3812 finish_wait(&pgdat->kswapd_wait, &wait);
3813 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3817 * After a short sleep, check if it was a premature sleep. If not, then
3818 * go fully to sleep until explicitly woken up.
3821 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3822 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3825 * vmstat counters are not perfectly accurate and the estimated
3826 * value for counters such as NR_FREE_PAGES can deviate from the
3827 * true value by nr_online_cpus * threshold. To avoid the zone
3828 * watermarks being breached while under pressure, we reduce the
3829 * per-cpu vmstat threshold while kswapd is awake and restore
3830 * them before going back to sleep.
3832 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3834 if (!kthread_should_stop())
3837 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3840 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3842 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3844 finish_wait(&pgdat->kswapd_wait, &wait);
3848 * The background pageout daemon, started as a kernel thread
3849 * from the init process.
3851 * This basically trickles out pages so that we have _some_
3852 * free memory available even if there is no other activity
3853 * that frees anything up. This is needed for things like routing
3854 * etc, where we otherwise might have all activity going on in
3855 * asynchronous contexts that cannot page things out.
3857 * If there are applications that are active memory-allocators
3858 * (most normal use), this basically shouldn't matter.
3860 static int kswapd(void *p)
3862 unsigned int alloc_order, reclaim_order;
3863 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3864 pg_data_t *pgdat = (pg_data_t*)p;
3865 struct task_struct *tsk = current;
3866 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3868 if (!cpumask_empty(cpumask))
3869 set_cpus_allowed_ptr(tsk, cpumask);
3872 * Tell the memory management that we're a "memory allocator",
3873 * and that if we need more memory we should get access to it
3874 * regardless (see "__alloc_pages()"). "kswapd" should
3875 * never get caught in the normal page freeing logic.
3877 * (Kswapd normally doesn't need memory anyway, but sometimes
3878 * you need a small amount of memory in order to be able to
3879 * page out something else, and this flag essentially protects
3880 * us from recursively trying to free more memory as we're
3881 * trying to free the first piece of memory in the first place).
3883 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3886 WRITE_ONCE(pgdat->kswapd_order, 0);
3887 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3891 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3892 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3896 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3899 /* Read the new order and highest_zoneidx */
3900 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3901 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3903 WRITE_ONCE(pgdat->kswapd_order, 0);
3904 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3906 ret = try_to_freeze();
3907 if (kthread_should_stop())
3911 * We can speed up thawing tasks if we don't call balance_pgdat
3912 * after returning from the refrigerator
3918 * Reclaim begins at the requested order but if a high-order
3919 * reclaim fails then kswapd falls back to reclaiming for
3920 * order-0. If that happens, kswapd will consider sleeping
3921 * for the order it finished reclaiming at (reclaim_order)
3922 * but kcompactd is woken to compact for the original
3923 * request (alloc_order).
3925 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3927 reclaim_order = balance_pgdat(pgdat, alloc_order,
3929 if (reclaim_order < alloc_order)
3930 goto kswapd_try_sleep;
3933 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3939 * A zone is low on free memory or too fragmented for high-order memory. If
3940 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3941 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3942 * has failed or is not needed, still wake up kcompactd if only compaction is
3945 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3946 enum zone_type highest_zoneidx)
3949 enum zone_type curr_idx;
3951 if (!managed_zone(zone))
3954 if (!cpuset_zone_allowed(zone, gfp_flags))
3957 pgdat = zone->zone_pgdat;
3958 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3960 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3961 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3963 if (READ_ONCE(pgdat->kswapd_order) < order)
3964 WRITE_ONCE(pgdat->kswapd_order, order);
3966 if (!waitqueue_active(&pgdat->kswapd_wait))
3969 /* Hopeless node, leave it to direct reclaim if possible */
3970 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3971 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3972 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3974 * There may be plenty of free memory available, but it's too
3975 * fragmented for high-order allocations. Wake up kcompactd
3976 * and rely on compaction_suitable() to determine if it's
3977 * needed. If it fails, it will defer subsequent attempts to
3978 * ratelimit its work.
3980 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3981 wakeup_kcompactd(pgdat, order, highest_zoneidx);
3985 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3987 wake_up_interruptible(&pgdat->kswapd_wait);
3990 #ifdef CONFIG_HIBERNATION
3992 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3995 * Rather than trying to age LRUs the aim is to preserve the overall
3996 * LRU order by reclaiming preferentially
3997 * inactive > active > active referenced > active mapped
3999 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4001 struct scan_control sc = {
4002 .nr_to_reclaim = nr_to_reclaim,
4003 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4004 .reclaim_idx = MAX_NR_ZONES - 1,
4005 .priority = DEF_PRIORITY,
4009 .hibernation_mode = 1,
4011 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4012 unsigned long nr_reclaimed;
4013 unsigned int noreclaim_flag;
4015 fs_reclaim_acquire(sc.gfp_mask);
4016 noreclaim_flag = memalloc_noreclaim_save();
4017 set_task_reclaim_state(current, &sc.reclaim_state);
4019 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4021 set_task_reclaim_state(current, NULL);
4022 memalloc_noreclaim_restore(noreclaim_flag);
4023 fs_reclaim_release(sc.gfp_mask);
4025 return nr_reclaimed;
4027 #endif /* CONFIG_HIBERNATION */
4030 * This kswapd start function will be called by init and node-hot-add.
4031 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4033 int kswapd_run(int nid)
4035 pg_data_t *pgdat = NODE_DATA(nid);
4041 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4042 if (IS_ERR(pgdat->kswapd)) {
4043 /* failure at boot is fatal */
4044 BUG_ON(system_state < SYSTEM_RUNNING);
4045 pr_err("Failed to start kswapd on node %d\n", nid);
4046 ret = PTR_ERR(pgdat->kswapd);
4047 pgdat->kswapd = NULL;
4053 * Called by memory hotplug when all memory in a node is offlined. Caller must
4054 * hold mem_hotplug_begin/end().
4056 void kswapd_stop(int nid)
4058 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4061 kthread_stop(kswapd);
4062 NODE_DATA(nid)->kswapd = NULL;
4066 static int __init kswapd_init(void)
4071 for_each_node_state(nid, N_MEMORY)
4076 module_init(kswapd_init)
4082 * If non-zero call node_reclaim when the number of free pages falls below
4085 int node_reclaim_mode __read_mostly;
4087 #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4088 #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4091 * Priority for NODE_RECLAIM. This determines the fraction of pages
4092 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4095 #define NODE_RECLAIM_PRIORITY 4
4098 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4101 int sysctl_min_unmapped_ratio = 1;
4104 * If the number of slab pages in a zone grows beyond this percentage then
4105 * slab reclaim needs to occur.
4107 int sysctl_min_slab_ratio = 5;
4109 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4111 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4112 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4113 node_page_state(pgdat, NR_ACTIVE_FILE);
4116 * It's possible for there to be more file mapped pages than
4117 * accounted for by the pages on the file LRU lists because
4118 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4120 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4123 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4124 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4126 unsigned long nr_pagecache_reclaimable;
4127 unsigned long delta = 0;
4130 * If RECLAIM_UNMAP is set, then all file pages are considered
4131 * potentially reclaimable. Otherwise, we have to worry about
4132 * pages like swapcache and node_unmapped_file_pages() provides
4135 if (node_reclaim_mode & RECLAIM_UNMAP)
4136 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4138 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4140 /* If we can't clean pages, remove dirty pages from consideration */
4141 if (!(node_reclaim_mode & RECLAIM_WRITE))
4142 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4144 /* Watch for any possible underflows due to delta */
4145 if (unlikely(delta > nr_pagecache_reclaimable))
4146 delta = nr_pagecache_reclaimable;
4148 return nr_pagecache_reclaimable - delta;
4152 * Try to free up some pages from this node through reclaim.
4154 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4156 /* Minimum pages needed in order to stay on node */
4157 const unsigned long nr_pages = 1 << order;
4158 struct task_struct *p = current;
4159 unsigned int noreclaim_flag;
4160 struct scan_control sc = {
4161 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4162 .gfp_mask = current_gfp_context(gfp_mask),
4164 .priority = NODE_RECLAIM_PRIORITY,
4165 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4166 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4168 .reclaim_idx = gfp_zone(gfp_mask),
4171 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4175 fs_reclaim_acquire(sc.gfp_mask);
4177 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4178 * and we also need to be able to write out pages for RECLAIM_WRITE
4179 * and RECLAIM_UNMAP.
4181 noreclaim_flag = memalloc_noreclaim_save();
4182 p->flags |= PF_SWAPWRITE;
4183 set_task_reclaim_state(p, &sc.reclaim_state);
4185 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4187 * Free memory by calling shrink node with increasing
4188 * priorities until we have enough memory freed.
4191 shrink_node(pgdat, &sc);
4192 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4195 set_task_reclaim_state(p, NULL);
4196 current->flags &= ~PF_SWAPWRITE;
4197 memalloc_noreclaim_restore(noreclaim_flag);
4198 fs_reclaim_release(sc.gfp_mask);
4200 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4202 return sc.nr_reclaimed >= nr_pages;
4205 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4210 * Node reclaim reclaims unmapped file backed pages and
4211 * slab pages if we are over the defined limits.
4213 * A small portion of unmapped file backed pages is needed for
4214 * file I/O otherwise pages read by file I/O will be immediately
4215 * thrown out if the node is overallocated. So we do not reclaim
4216 * if less than a specified percentage of the node is used by
4217 * unmapped file backed pages.
4219 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4220 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4221 pgdat->min_slab_pages)
4222 return NODE_RECLAIM_FULL;
4225 * Do not scan if the allocation should not be delayed.
4227 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4228 return NODE_RECLAIM_NOSCAN;
4231 * Only run node reclaim on the local node or on nodes that do not
4232 * have associated processors. This will favor the local processor
4233 * over remote processors and spread off node memory allocations
4234 * as wide as possible.
4236 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4237 return NODE_RECLAIM_NOSCAN;
4239 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4240 return NODE_RECLAIM_NOSCAN;
4242 ret = __node_reclaim(pgdat, gfp_mask, order);
4243 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4246 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4253 * check_move_unevictable_pages - check pages for evictability and move to
4254 * appropriate zone lru list
4255 * @pvec: pagevec with lru pages to check
4257 * Checks pages for evictability, if an evictable page is in the unevictable
4258 * lru list, moves it to the appropriate evictable lru list. This function
4259 * should be only used for lru pages.
4261 void check_move_unevictable_pages(struct pagevec *pvec)
4263 struct lruvec *lruvec;
4264 struct pglist_data *pgdat = NULL;
4269 for (i = 0; i < pvec->nr; i++) {
4270 struct page *page = pvec->pages[i];
4271 struct pglist_data *pagepgdat = page_pgdat(page);
4274 if (PageTransTail(page))
4277 nr_pages = thp_nr_pages(page);
4278 pgscanned += nr_pages;
4280 if (pagepgdat != pgdat) {
4282 spin_unlock_irq(&pgdat->lru_lock);
4284 spin_lock_irq(&pgdat->lru_lock);
4286 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4288 if (!PageLRU(page) || !PageUnevictable(page))
4291 if (page_evictable(page)) {
4292 enum lru_list lru = page_lru_base_type(page);
4294 VM_BUG_ON_PAGE(PageActive(page), page);
4295 ClearPageUnevictable(page);
4296 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4297 add_page_to_lru_list(page, lruvec, lru);
4298 pgrescued += nr_pages;
4303 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4304 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4305 spin_unlock_irq(&pgdat->lru_lock);
4308 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);