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;
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
92 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
96 unsigned int memcg_low_reclaim:1;
97 unsigned int memcg_low_skipped:1;
99 unsigned int hibernation_mode:1;
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready:1;
104 /* Allocation order */
107 /* Scan (total_size >> priority) pages at once */
110 /* The highest zone to isolate pages for reclaim from */
113 /* This context's GFP mask */
116 /* Incremented by the number of inactive pages that were scanned */
117 unsigned long nr_scanned;
119 /* Number of pages freed so far during a call to shrink_zones() */
120 unsigned long nr_reclaimed;
124 unsigned int unqueued_dirty;
125 unsigned int congested;
126 unsigned int writeback;
127 unsigned int immediate;
128 unsigned int file_taken;
132 /* for recording the reclaimed slab by now */
133 struct reclaim_state reclaim_state;
136 #ifdef ARCH_HAS_PREFETCH
137 #define prefetch_prev_lru_page(_page, _base, _field) \
139 if ((_page)->lru.prev != _base) { \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
147 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
150 #ifdef ARCH_HAS_PREFETCHW
151 #define prefetchw_prev_lru_page(_page, _base, _field) \
153 if ((_page)->lru.prev != _base) { \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
161 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
165 * From 0 .. 100. Higher means more swappy.
167 int vm_swappiness = 60;
169 * The total number of pages which are beyond the high watermark within all
172 unsigned long vm_total_pages;
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 global_reclaim(struct scan_control *sc)
244 return !sc->target_mem_cgroup;
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
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 sane_reclaim(struct scan_control *sc)
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
266 #ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
273 static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
277 struct mem_cgroup_per_node *mn;
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
286 static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
289 struct mem_cgroup_per_node *mn;
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
296 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
301 static void unregister_memcg_shrinker(struct shrinker *shrinker)
305 static bool global_reclaim(struct scan_control *sc)
310 static bool sane_reclaim(struct scan_control *sc)
315 static inline void set_memcg_congestion(struct pglist_data *pgdat,
316 struct mem_cgroup *memcg, bool congested)
320 static inline bool memcg_congested(struct pglist_data *pgdat,
321 struct mem_cgroup *memcg)
329 * This misses isolated pages which are not accounted for to save counters.
330 * As the data only determines if reclaim or compaction continues, it is
331 * not expected that isolated pages will be a dominating factor.
333 unsigned long zone_reclaimable_pages(struct zone *zone)
337 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
338 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
339 if (get_nr_swap_pages() > 0)
340 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
341 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
347 * lruvec_lru_size - Returns the number of pages on the given LRU list.
348 * @lruvec: lru vector
350 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
352 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
354 unsigned long size = 0;
357 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
358 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
360 if (!managed_zone(zone))
363 if (!mem_cgroup_disabled())
364 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
366 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
372 * Add a shrinker callback to be called from the vm.
374 int prealloc_shrinker(struct shrinker *shrinker)
376 unsigned int size = sizeof(*shrinker->nr_deferred);
378 if (shrinker->flags & SHRINKER_NUMA_AWARE)
381 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
382 if (!shrinker->nr_deferred)
385 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
386 if (prealloc_memcg_shrinker(shrinker))
393 kfree(shrinker->nr_deferred);
394 shrinker->nr_deferred = NULL;
398 void free_prealloced_shrinker(struct shrinker *shrinker)
400 if (!shrinker->nr_deferred)
403 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
404 unregister_memcg_shrinker(shrinker);
406 kfree(shrinker->nr_deferred);
407 shrinker->nr_deferred = NULL;
410 void register_shrinker_prepared(struct shrinker *shrinker)
412 down_write(&shrinker_rwsem);
413 list_add_tail(&shrinker->list, &shrinker_list);
415 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
416 idr_replace(&shrinker_idr, shrinker, shrinker->id);
418 up_write(&shrinker_rwsem);
421 int register_shrinker(struct shrinker *shrinker)
423 int err = prealloc_shrinker(shrinker);
427 register_shrinker_prepared(shrinker);
430 EXPORT_SYMBOL(register_shrinker);
435 void unregister_shrinker(struct shrinker *shrinker)
437 if (!shrinker->nr_deferred)
439 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
440 unregister_memcg_shrinker(shrinker);
441 down_write(&shrinker_rwsem);
442 list_del(&shrinker->list);
443 up_write(&shrinker_rwsem);
444 kfree(shrinker->nr_deferred);
445 shrinker->nr_deferred = NULL;
447 EXPORT_SYMBOL(unregister_shrinker);
449 #define SHRINK_BATCH 128
451 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
452 struct shrinker *shrinker, int priority)
454 unsigned long freed = 0;
455 unsigned long long delta;
460 int nid = shrinkctl->nid;
461 long batch_size = shrinker->batch ? shrinker->batch
463 long scanned = 0, next_deferred;
465 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
468 freeable = shrinker->count_objects(shrinker, shrinkctl);
469 if (freeable == 0 || freeable == SHRINK_EMPTY)
473 * copy the current shrinker scan count into a local variable
474 * and zero it so that other concurrent shrinker invocations
475 * don't also do this scanning work.
477 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
480 if (shrinker->seeks) {
481 delta = freeable >> priority;
483 do_div(delta, shrinker->seeks);
486 * These objects don't require any IO to create. Trim
487 * them aggressively under memory pressure to keep
488 * them from causing refetches in the IO caches.
490 delta = freeable / 2;
494 if (total_scan < 0) {
495 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
496 shrinker->scan_objects, total_scan);
497 total_scan = freeable;
500 next_deferred = total_scan;
503 * We need to avoid excessive windup on filesystem shrinkers
504 * due to large numbers of GFP_NOFS allocations causing the
505 * shrinkers to return -1 all the time. This results in a large
506 * nr being built up so when a shrink that can do some work
507 * comes along it empties the entire cache due to nr >>>
508 * freeable. This is bad for sustaining a working set in
511 * Hence only allow the shrinker to scan the entire cache when
512 * a large delta change is calculated directly.
514 if (delta < freeable / 4)
515 total_scan = min(total_scan, freeable / 2);
518 * Avoid risking looping forever due to too large nr value:
519 * never try to free more than twice the estimate number of
522 if (total_scan > freeable * 2)
523 total_scan = freeable * 2;
525 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
526 freeable, delta, total_scan, priority);
529 * Normally, we should not scan less than batch_size objects in one
530 * pass to avoid too frequent shrinker calls, but if the slab has less
531 * than batch_size objects in total and we are really tight on memory,
532 * we will try to reclaim all available objects, otherwise we can end
533 * up failing allocations although there are plenty of reclaimable
534 * objects spread over several slabs with usage less than the
537 * We detect the "tight on memory" situations by looking at the total
538 * number of objects we want to scan (total_scan). If it is greater
539 * than the total number of objects on slab (freeable), we must be
540 * scanning at high prio and therefore should try to reclaim as much as
543 while (total_scan >= batch_size ||
544 total_scan >= freeable) {
546 unsigned long nr_to_scan = min(batch_size, total_scan);
548 shrinkctl->nr_to_scan = nr_to_scan;
549 shrinkctl->nr_scanned = nr_to_scan;
550 ret = shrinker->scan_objects(shrinker, shrinkctl);
551 if (ret == SHRINK_STOP)
555 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
556 total_scan -= shrinkctl->nr_scanned;
557 scanned += shrinkctl->nr_scanned;
562 if (next_deferred >= scanned)
563 next_deferred -= scanned;
567 * move the unused scan count back into the shrinker in a
568 * manner that handles concurrent updates. If we exhausted the
569 * scan, there is no need to do an update.
571 if (next_deferred > 0)
572 new_nr = atomic_long_add_return(next_deferred,
573 &shrinker->nr_deferred[nid]);
575 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
577 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
582 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
583 struct mem_cgroup *memcg, int priority)
585 struct memcg_shrinker_map *map;
586 unsigned long ret, freed = 0;
589 if (!mem_cgroup_online(memcg))
592 if (!down_read_trylock(&shrinker_rwsem))
595 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
600 for_each_set_bit(i, map->map, shrinker_nr_max) {
601 struct shrink_control sc = {
602 .gfp_mask = gfp_mask,
606 struct shrinker *shrinker;
608 shrinker = idr_find(&shrinker_idr, i);
609 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
611 clear_bit(i, map->map);
615 /* Call non-slab shrinkers even though kmem is disabled */
616 if (!memcg_kmem_enabled() &&
617 !(shrinker->flags & SHRINKER_NONSLAB))
620 ret = do_shrink_slab(&sc, shrinker, priority);
621 if (ret == SHRINK_EMPTY) {
622 clear_bit(i, map->map);
624 * After the shrinker reported that it had no objects to
625 * free, but before we cleared the corresponding bit in
626 * the memcg shrinker map, a new object might have been
627 * added. To make sure, we have the bit set in this
628 * case, we invoke the shrinker one more time and reset
629 * the bit if it reports that it is not empty anymore.
630 * The memory barrier here pairs with the barrier in
631 * memcg_set_shrinker_bit():
633 * list_lru_add() shrink_slab_memcg()
634 * list_add_tail() clear_bit()
636 * set_bit() do_shrink_slab()
638 smp_mb__after_atomic();
639 ret = do_shrink_slab(&sc, shrinker, priority);
640 if (ret == SHRINK_EMPTY)
643 memcg_set_shrinker_bit(memcg, nid, i);
647 if (rwsem_is_contended(&shrinker_rwsem)) {
653 up_read(&shrinker_rwsem);
656 #else /* CONFIG_MEMCG */
657 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
658 struct mem_cgroup *memcg, int priority)
662 #endif /* CONFIG_MEMCG */
665 * shrink_slab - shrink slab caches
666 * @gfp_mask: allocation context
667 * @nid: node whose slab caches to target
668 * @memcg: memory cgroup whose slab caches to target
669 * @priority: the reclaim priority
671 * Call the shrink functions to age shrinkable caches.
673 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
674 * unaware shrinkers will receive a node id of 0 instead.
676 * @memcg specifies the memory cgroup to target. Unaware shrinkers
677 * are called only if it is the root cgroup.
679 * @priority is sc->priority, we take the number of objects and >> by priority
680 * in order to get the scan target.
682 * Returns the number of reclaimed slab objects.
684 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
685 struct mem_cgroup *memcg,
688 unsigned long ret, freed = 0;
689 struct shrinker *shrinker;
692 * The root memcg might be allocated even though memcg is disabled
693 * via "cgroup_disable=memory" boot parameter. This could make
694 * mem_cgroup_is_root() return false, then just run memcg slab
695 * shrink, but skip global shrink. This may result in premature
698 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
699 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
701 if (!down_read_trylock(&shrinker_rwsem))
704 list_for_each_entry(shrinker, &shrinker_list, list) {
705 struct shrink_control sc = {
706 .gfp_mask = gfp_mask,
711 ret = do_shrink_slab(&sc, shrinker, priority);
712 if (ret == SHRINK_EMPTY)
716 * Bail out if someone want to register a new shrinker to
717 * prevent the regsitration from being stalled for long periods
718 * by parallel ongoing shrinking.
720 if (rwsem_is_contended(&shrinker_rwsem)) {
726 up_read(&shrinker_rwsem);
732 void drop_slab_node(int nid)
737 struct mem_cgroup *memcg = NULL;
740 memcg = mem_cgroup_iter(NULL, NULL, NULL);
742 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
743 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
744 } while (freed > 10);
751 for_each_online_node(nid)
755 static inline int is_page_cache_freeable(struct page *page)
758 * A freeable page cache page is referenced only by the caller
759 * that isolated the page, the page cache and optional buffer
760 * heads at page->private.
762 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
764 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
767 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
769 if (current->flags & PF_SWAPWRITE)
771 if (!inode_write_congested(inode))
773 if (inode_to_bdi(inode) == current->backing_dev_info)
779 * We detected a synchronous write error writing a page out. Probably
780 * -ENOSPC. We need to propagate that into the address_space for a subsequent
781 * fsync(), msync() or close().
783 * The tricky part is that after writepage we cannot touch the mapping: nothing
784 * prevents it from being freed up. But we have a ref on the page and once
785 * that page is locked, the mapping is pinned.
787 * We're allowed to run sleeping lock_page() here because we know the caller has
790 static void handle_write_error(struct address_space *mapping,
791 struct page *page, int error)
794 if (page_mapping(page) == mapping)
795 mapping_set_error(mapping, error);
799 /* possible outcome of pageout() */
801 /* failed to write page out, page is locked */
803 /* move page to the active list, page is locked */
805 /* page has been sent to the disk successfully, page is unlocked */
807 /* page is clean and locked */
812 * pageout is called by shrink_page_list() for each dirty page.
813 * Calls ->writepage().
815 static pageout_t pageout(struct page *page, struct address_space *mapping,
816 struct scan_control *sc)
819 * If the page is dirty, only perform writeback if that write
820 * will be non-blocking. To prevent this allocation from being
821 * stalled by pagecache activity. But note that there may be
822 * stalls if we need to run get_block(). We could test
823 * PagePrivate for that.
825 * If this process is currently in __generic_file_write_iter() against
826 * this page's queue, we can perform writeback even if that
829 * If the page is swapcache, write it back even if that would
830 * block, for some throttling. This happens by accident, because
831 * swap_backing_dev_info is bust: it doesn't reflect the
832 * congestion state of the swapdevs. Easy to fix, if needed.
834 if (!is_page_cache_freeable(page))
838 * Some data journaling orphaned pages can have
839 * page->mapping == NULL while being dirty with clean buffers.
841 if (page_has_private(page)) {
842 if (try_to_free_buffers(page)) {
843 ClearPageDirty(page);
844 pr_info("%s: orphaned page\n", __func__);
850 if (mapping->a_ops->writepage == NULL)
851 return PAGE_ACTIVATE;
852 if (!may_write_to_inode(mapping->host, sc))
855 if (clear_page_dirty_for_io(page)) {
857 struct writeback_control wbc = {
858 .sync_mode = WB_SYNC_NONE,
859 .nr_to_write = SWAP_CLUSTER_MAX,
861 .range_end = LLONG_MAX,
865 SetPageReclaim(page);
866 res = mapping->a_ops->writepage(page, &wbc);
868 handle_write_error(mapping, page, res);
869 if (res == AOP_WRITEPAGE_ACTIVATE) {
870 ClearPageReclaim(page);
871 return PAGE_ACTIVATE;
874 if (!PageWriteback(page)) {
875 /* synchronous write or broken a_ops? */
876 ClearPageReclaim(page);
878 trace_mm_vmscan_writepage(page);
879 inc_node_page_state(page, NR_VMSCAN_WRITE);
887 * Same as remove_mapping, but if the page is removed from the mapping, it
888 * gets returned with a refcount of 0.
890 static int __remove_mapping(struct address_space *mapping, struct page *page,
896 BUG_ON(!PageLocked(page));
897 BUG_ON(mapping != page_mapping(page));
899 xa_lock_irqsave(&mapping->i_pages, flags);
901 * The non racy check for a busy page.
903 * Must be careful with the order of the tests. When someone has
904 * a ref to the page, it may be possible that they dirty it then
905 * drop the reference. So if PageDirty is tested before page_count
906 * here, then the following race may occur:
908 * get_user_pages(&page);
909 * [user mapping goes away]
911 * !PageDirty(page) [good]
912 * SetPageDirty(page);
914 * !page_count(page) [good, discard it]
916 * [oops, our write_to data is lost]
918 * Reversing the order of the tests ensures such a situation cannot
919 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
920 * load is not satisfied before that of page->_refcount.
922 * Note that if SetPageDirty is always performed via set_page_dirty,
923 * and thus under the i_pages lock, then this ordering is not required.
925 refcount = 1 + compound_nr(page);
926 if (!page_ref_freeze(page, refcount))
928 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
929 if (unlikely(PageDirty(page))) {
930 page_ref_unfreeze(page, refcount);
934 if (PageSwapCache(page)) {
935 swp_entry_t swap = { .val = page_private(page) };
936 mem_cgroup_swapout(page, swap);
937 __delete_from_swap_cache(page, swap);
938 xa_unlock_irqrestore(&mapping->i_pages, flags);
939 put_swap_page(page, swap);
941 void (*freepage)(struct page *);
944 freepage = mapping->a_ops->freepage;
946 * Remember a shadow entry for reclaimed file cache in
947 * order to detect refaults, thus thrashing, later on.
949 * But don't store shadows in an address space that is
950 * already exiting. This is not just an optizimation,
951 * inode reclaim needs to empty out the radix tree or
952 * the nodes are lost. Don't plant shadows behind its
955 * We also don't store shadows for DAX mappings because the
956 * only page cache pages found in these are zero pages
957 * covering holes, and because we don't want to mix DAX
958 * exceptional entries and shadow exceptional entries in the
959 * same address_space.
961 if (reclaimed && page_is_file_cache(page) &&
962 !mapping_exiting(mapping) && !dax_mapping(mapping))
963 shadow = workingset_eviction(page);
964 __delete_from_page_cache(page, shadow);
965 xa_unlock_irqrestore(&mapping->i_pages, flags);
967 if (freepage != NULL)
974 xa_unlock_irqrestore(&mapping->i_pages, flags);
979 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
980 * someone else has a ref on the page, abort and return 0. If it was
981 * successfully detached, return 1. Assumes the caller has a single ref on
984 int remove_mapping(struct address_space *mapping, struct page *page)
986 if (__remove_mapping(mapping, page, false)) {
988 * Unfreezing the refcount with 1 rather than 2 effectively
989 * drops the pagecache ref for us without requiring another
992 page_ref_unfreeze(page, 1);
999 * putback_lru_page - put previously isolated page onto appropriate LRU list
1000 * @page: page to be put back to appropriate lru list
1002 * Add previously isolated @page to appropriate LRU list.
1003 * Page may still be unevictable for other reasons.
1005 * lru_lock must not be held, interrupts must be enabled.
1007 void putback_lru_page(struct page *page)
1009 lru_cache_add(page);
1010 put_page(page); /* drop ref from isolate */
1013 enum page_references {
1015 PAGEREF_RECLAIM_CLEAN,
1020 static enum page_references page_check_references(struct page *page,
1021 struct scan_control *sc)
1023 int referenced_ptes, referenced_page;
1024 unsigned long vm_flags;
1026 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1028 referenced_page = TestClearPageReferenced(page);
1031 * Mlock lost the isolation race with us. Let try_to_unmap()
1032 * move the page to the unevictable list.
1034 if (vm_flags & VM_LOCKED)
1035 return PAGEREF_RECLAIM;
1037 if (referenced_ptes) {
1038 if (PageSwapBacked(page))
1039 return PAGEREF_ACTIVATE;
1041 * All mapped pages start out with page table
1042 * references from the instantiating fault, so we need
1043 * to look twice if a mapped file page is used more
1046 * Mark it and spare it for another trip around the
1047 * inactive list. Another page table reference will
1048 * lead to its activation.
1050 * Note: the mark is set for activated pages as well
1051 * so that recently deactivated but used pages are
1052 * quickly recovered.
1054 SetPageReferenced(page);
1056 if (referenced_page || referenced_ptes > 1)
1057 return PAGEREF_ACTIVATE;
1060 * Activate file-backed executable pages after first usage.
1062 if (vm_flags & VM_EXEC)
1063 return PAGEREF_ACTIVATE;
1065 return PAGEREF_KEEP;
1068 /* Reclaim if clean, defer dirty pages to writeback */
1069 if (referenced_page && !PageSwapBacked(page))
1070 return PAGEREF_RECLAIM_CLEAN;
1072 return PAGEREF_RECLAIM;
1075 /* Check if a page is dirty or under writeback */
1076 static void page_check_dirty_writeback(struct page *page,
1077 bool *dirty, bool *writeback)
1079 struct address_space *mapping;
1082 * Anonymous pages are not handled by flushers and must be written
1083 * from reclaim context. Do not stall reclaim based on them
1085 if (!page_is_file_cache(page) ||
1086 (PageAnon(page) && !PageSwapBacked(page))) {
1092 /* By default assume that the page flags are accurate */
1093 *dirty = PageDirty(page);
1094 *writeback = PageWriteback(page);
1096 /* Verify dirty/writeback state if the filesystem supports it */
1097 if (!page_has_private(page))
1100 mapping = page_mapping(page);
1101 if (mapping && mapping->a_ops->is_dirty_writeback)
1102 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1106 * shrink_page_list() returns the number of reclaimed pages
1108 static unsigned long shrink_page_list(struct list_head *page_list,
1109 struct pglist_data *pgdat,
1110 struct scan_control *sc,
1111 enum ttu_flags ttu_flags,
1112 struct reclaim_stat *stat,
1113 bool ignore_references)
1115 LIST_HEAD(ret_pages);
1116 LIST_HEAD(free_pages);
1117 unsigned nr_reclaimed = 0;
1118 unsigned pgactivate = 0;
1120 memset(stat, 0, sizeof(*stat));
1123 while (!list_empty(page_list)) {
1124 struct address_space *mapping;
1127 enum page_references references = PAGEREF_RECLAIM;
1128 bool dirty, writeback;
1129 unsigned int nr_pages;
1133 page = lru_to_page(page_list);
1134 list_del(&page->lru);
1136 if (!trylock_page(page))
1139 VM_BUG_ON_PAGE(PageActive(page), page);
1141 nr_pages = compound_nr(page);
1143 /* Account the number of base pages even though THP */
1144 sc->nr_scanned += nr_pages;
1146 if (unlikely(!page_evictable(page)))
1147 goto activate_locked;
1149 if (!sc->may_unmap && page_mapped(page))
1152 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1153 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1156 * The number of dirty pages determines if a node is marked
1157 * reclaim_congested which affects wait_iff_congested. kswapd
1158 * will stall and start writing pages if the tail of the LRU
1159 * is all dirty unqueued pages.
1161 page_check_dirty_writeback(page, &dirty, &writeback);
1162 if (dirty || writeback)
1165 if (dirty && !writeback)
1166 stat->nr_unqueued_dirty++;
1169 * Treat this page as congested if the underlying BDI is or if
1170 * pages are cycling through the LRU so quickly that the
1171 * pages marked for immediate reclaim are making it to the
1172 * end of the LRU a second time.
1174 mapping = page_mapping(page);
1175 if (((dirty || writeback) && mapping &&
1176 inode_write_congested(mapping->host)) ||
1177 (writeback && PageReclaim(page)))
1178 stat->nr_congested++;
1181 * If a page at the tail of the LRU is under writeback, there
1182 * are three cases to consider.
1184 * 1) If reclaim is encountering an excessive number of pages
1185 * under writeback and this page is both under writeback and
1186 * PageReclaim then it indicates that pages are being queued
1187 * for IO but are being recycled through the LRU before the
1188 * IO can complete. Waiting on the page itself risks an
1189 * indefinite stall if it is impossible to writeback the
1190 * page due to IO error or disconnected storage so instead
1191 * note that the LRU is being scanned too quickly and the
1192 * caller can stall after page list has been processed.
1194 * 2) Global or new memcg reclaim encounters a page that is
1195 * not marked for immediate reclaim, or the caller does not
1196 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1197 * not to fs). In this case mark the page for immediate
1198 * reclaim and continue scanning.
1200 * Require may_enter_fs because we would wait on fs, which
1201 * may not have submitted IO yet. And the loop driver might
1202 * enter reclaim, and deadlock if it waits on a page for
1203 * which it is needed to do the write (loop masks off
1204 * __GFP_IO|__GFP_FS for this reason); but more thought
1205 * would probably show more reasons.
1207 * 3) Legacy memcg encounters a page that is already marked
1208 * PageReclaim. memcg does not have any dirty pages
1209 * throttling so we could easily OOM just because too many
1210 * pages are in writeback and there is nothing else to
1211 * reclaim. Wait for the writeback to complete.
1213 * In cases 1) and 2) we activate the pages to get them out of
1214 * the way while we continue scanning for clean pages on the
1215 * inactive list and refilling from the active list. The
1216 * observation here is that waiting for disk writes is more
1217 * expensive than potentially causing reloads down the line.
1218 * Since they're marked for immediate reclaim, they won't put
1219 * memory pressure on the cache working set any longer than it
1220 * takes to write them to disk.
1222 if (PageWriteback(page)) {
1224 if (current_is_kswapd() &&
1225 PageReclaim(page) &&
1226 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1227 stat->nr_immediate++;
1228 goto activate_locked;
1231 } else if (sane_reclaim(sc) ||
1232 !PageReclaim(page) || !may_enter_fs) {
1234 * This is slightly racy - end_page_writeback()
1235 * might have just cleared PageReclaim, then
1236 * setting PageReclaim here end up interpreted
1237 * as PageReadahead - but that does not matter
1238 * enough to care. What we do want is for this
1239 * page to have PageReclaim set next time memcg
1240 * reclaim reaches the tests above, so it will
1241 * then wait_on_page_writeback() to avoid OOM;
1242 * and it's also appropriate in global reclaim.
1244 SetPageReclaim(page);
1245 stat->nr_writeback++;
1246 goto activate_locked;
1251 wait_on_page_writeback(page);
1252 /* then go back and try same page again */
1253 list_add_tail(&page->lru, page_list);
1258 if (!ignore_references)
1259 references = page_check_references(page, sc);
1261 switch (references) {
1262 case PAGEREF_ACTIVATE:
1263 goto activate_locked;
1265 stat->nr_ref_keep += nr_pages;
1267 case PAGEREF_RECLAIM:
1268 case PAGEREF_RECLAIM_CLEAN:
1269 ; /* try to reclaim the page below */
1273 * Anonymous process memory has backing store?
1274 * Try to allocate it some swap space here.
1275 * Lazyfree page could be freed directly
1277 if (PageAnon(page) && PageSwapBacked(page)) {
1278 if (!PageSwapCache(page)) {
1279 if (!(sc->gfp_mask & __GFP_IO))
1281 if (PageTransHuge(page)) {
1282 /* cannot split THP, skip it */
1283 if (!can_split_huge_page(page, NULL))
1284 goto activate_locked;
1286 * Split pages without a PMD map right
1287 * away. Chances are some or all of the
1288 * tail pages can be freed without IO.
1290 if (!compound_mapcount(page) &&
1291 split_huge_page_to_list(page,
1293 goto activate_locked;
1295 if (!add_to_swap(page)) {
1296 if (!PageTransHuge(page))
1297 goto activate_locked_split;
1298 /* Fallback to swap normal pages */
1299 if (split_huge_page_to_list(page,
1301 goto activate_locked;
1302 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1303 count_vm_event(THP_SWPOUT_FALLBACK);
1305 if (!add_to_swap(page))
1306 goto activate_locked_split;
1311 /* Adding to swap updated mapping */
1312 mapping = page_mapping(page);
1314 } else if (unlikely(PageTransHuge(page))) {
1315 /* Split file THP */
1316 if (split_huge_page_to_list(page, page_list))
1321 * THP may get split above, need minus tail pages and update
1322 * nr_pages to avoid accounting tail pages twice.
1324 * The tail pages that are added into swap cache successfully
1327 if ((nr_pages > 1) && !PageTransHuge(page)) {
1328 sc->nr_scanned -= (nr_pages - 1);
1333 * The page is mapped into the page tables of one or more
1334 * processes. Try to unmap it here.
1336 if (page_mapped(page)) {
1337 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1339 if (unlikely(PageTransHuge(page)))
1340 flags |= TTU_SPLIT_HUGE_PMD;
1341 if (!try_to_unmap(page, flags)) {
1342 stat->nr_unmap_fail += nr_pages;
1343 goto activate_locked;
1347 if (PageDirty(page)) {
1349 * Only kswapd can writeback filesystem pages
1350 * to avoid risk of stack overflow. But avoid
1351 * injecting inefficient single-page IO into
1352 * flusher writeback as much as possible: only
1353 * write pages when we've encountered many
1354 * dirty pages, and when we've already scanned
1355 * the rest of the LRU for clean pages and see
1356 * the same dirty pages again (PageReclaim).
1358 if (page_is_file_cache(page) &&
1359 (!current_is_kswapd() || !PageReclaim(page) ||
1360 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1362 * Immediately reclaim when written back.
1363 * Similar in principal to deactivate_page()
1364 * except we already have the page isolated
1365 * and know it's dirty
1367 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1368 SetPageReclaim(page);
1370 goto activate_locked;
1373 if (references == PAGEREF_RECLAIM_CLEAN)
1377 if (!sc->may_writepage)
1381 * Page is dirty. Flush the TLB if a writable entry
1382 * potentially exists to avoid CPU writes after IO
1383 * starts and then write it out here.
1385 try_to_unmap_flush_dirty();
1386 switch (pageout(page, mapping, sc)) {
1390 goto activate_locked;
1392 if (PageWriteback(page))
1394 if (PageDirty(page))
1398 * A synchronous write - probably a ramdisk. Go
1399 * ahead and try to reclaim the page.
1401 if (!trylock_page(page))
1403 if (PageDirty(page) || PageWriteback(page))
1405 mapping = page_mapping(page);
1407 ; /* try to free the page below */
1412 * If the page has buffers, try to free the buffer mappings
1413 * associated with this page. If we succeed we try to free
1416 * We do this even if the page is PageDirty().
1417 * try_to_release_page() does not perform I/O, but it is
1418 * possible for a page to have PageDirty set, but it is actually
1419 * clean (all its buffers are clean). This happens if the
1420 * buffers were written out directly, with submit_bh(). ext3
1421 * will do this, as well as the blockdev mapping.
1422 * try_to_release_page() will discover that cleanness and will
1423 * drop the buffers and mark the page clean - it can be freed.
1425 * Rarely, pages can have buffers and no ->mapping. These are
1426 * the pages which were not successfully invalidated in
1427 * truncate_complete_page(). We try to drop those buffers here
1428 * and if that worked, and the page is no longer mapped into
1429 * process address space (page_count == 1) it can be freed.
1430 * Otherwise, leave the page on the LRU so it is swappable.
1432 if (page_has_private(page)) {
1433 if (!try_to_release_page(page, sc->gfp_mask))
1434 goto activate_locked;
1435 if (!mapping && page_count(page) == 1) {
1437 if (put_page_testzero(page))
1441 * rare race with speculative reference.
1442 * the speculative reference will free
1443 * this page shortly, so we may
1444 * increment nr_reclaimed here (and
1445 * leave it off the LRU).
1453 if (PageAnon(page) && !PageSwapBacked(page)) {
1454 /* follow __remove_mapping for reference */
1455 if (!page_ref_freeze(page, 1))
1457 if (PageDirty(page)) {
1458 page_ref_unfreeze(page, 1);
1462 count_vm_event(PGLAZYFREED);
1463 count_memcg_page_event(page, PGLAZYFREED);
1464 } else if (!mapping || !__remove_mapping(mapping, page, true))
1470 * THP may get swapped out in a whole, need account
1473 nr_reclaimed += nr_pages;
1476 * Is there need to periodically free_page_list? It would
1477 * appear not as the counts should be low
1479 if (unlikely(PageTransHuge(page)))
1480 (*get_compound_page_dtor(page))(page);
1482 list_add(&page->lru, &free_pages);
1485 activate_locked_split:
1487 * The tail pages that are failed to add into swap cache
1488 * reach here. Fixup nr_scanned and nr_pages.
1491 sc->nr_scanned -= (nr_pages - 1);
1495 /* Not a candidate for swapping, so reclaim swap space. */
1496 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1498 try_to_free_swap(page);
1499 VM_BUG_ON_PAGE(PageActive(page), page);
1500 if (!PageMlocked(page)) {
1501 int type = page_is_file_cache(page);
1502 SetPageActive(page);
1503 stat->nr_activate[type] += nr_pages;
1504 count_memcg_page_event(page, PGACTIVATE);
1509 list_add(&page->lru, &ret_pages);
1510 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1513 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1515 mem_cgroup_uncharge_list(&free_pages);
1516 try_to_unmap_flush();
1517 free_unref_page_list(&free_pages);
1519 list_splice(&ret_pages, page_list);
1520 count_vm_events(PGACTIVATE, pgactivate);
1522 return nr_reclaimed;
1525 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1526 struct list_head *page_list)
1528 struct scan_control sc = {
1529 .gfp_mask = GFP_KERNEL,
1530 .priority = DEF_PRIORITY,
1533 struct reclaim_stat dummy_stat;
1535 struct page *page, *next;
1536 LIST_HEAD(clean_pages);
1538 list_for_each_entry_safe(page, next, page_list, lru) {
1539 if (page_is_file_cache(page) && !PageDirty(page) &&
1540 !__PageMovable(page) && !PageUnevictable(page)) {
1541 ClearPageActive(page);
1542 list_move(&page->lru, &clean_pages);
1546 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1547 TTU_IGNORE_ACCESS, &dummy_stat, true);
1548 list_splice(&clean_pages, page_list);
1549 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1554 * Attempt to remove the specified page from its LRU. Only take this page
1555 * if it is of the appropriate PageActive status. Pages which are being
1556 * freed elsewhere are also ignored.
1558 * page: page to consider
1559 * mode: one of the LRU isolation modes defined above
1561 * returns 0 on success, -ve errno on failure.
1563 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1567 /* Only take pages on the LRU. */
1571 /* Compaction should not handle unevictable pages but CMA can do so */
1572 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1578 * To minimise LRU disruption, the caller can indicate that it only
1579 * wants to isolate pages it will be able to operate on without
1580 * blocking - clean pages for the most part.
1582 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1583 * that it is possible to migrate without blocking
1585 if (mode & ISOLATE_ASYNC_MIGRATE) {
1586 /* All the caller can do on PageWriteback is block */
1587 if (PageWriteback(page))
1590 if (PageDirty(page)) {
1591 struct address_space *mapping;
1595 * Only pages without mappings or that have a
1596 * ->migratepage callback are possible to migrate
1597 * without blocking. However, we can be racing with
1598 * truncation so it's necessary to lock the page
1599 * to stabilise the mapping as truncation holds
1600 * the page lock until after the page is removed
1601 * from the page cache.
1603 if (!trylock_page(page))
1606 mapping = page_mapping(page);
1607 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1614 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1617 if (likely(get_page_unless_zero(page))) {
1619 * Be careful not to clear PageLRU until after we're
1620 * sure the page is not being freed elsewhere -- the
1621 * page release code relies on it.
1632 * Update LRU sizes after isolating pages. The LRU size updates must
1633 * be complete before mem_cgroup_update_lru_size due to a santity check.
1635 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1636 enum lru_list lru, unsigned long *nr_zone_taken)
1640 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1641 if (!nr_zone_taken[zid])
1644 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1646 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1653 * pgdat->lru_lock is heavily contended. Some of the functions that
1654 * shrink the lists perform better by taking out a batch of pages
1655 * and working on them outside the LRU lock.
1657 * For pagecache intensive workloads, this function is the hottest
1658 * spot in the kernel (apart from copy_*_user functions).
1660 * Appropriate locks must be held before calling this function.
1662 * @nr_to_scan: The number of eligible pages to look through on the list.
1663 * @lruvec: The LRU vector to pull pages from.
1664 * @dst: The temp list to put pages on to.
1665 * @nr_scanned: The number of pages that were scanned.
1666 * @sc: The scan_control struct for this reclaim session
1667 * @mode: One of the LRU isolation modes
1668 * @lru: LRU list id for isolating
1670 * returns how many pages were moved onto *@dst.
1672 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1673 struct lruvec *lruvec, struct list_head *dst,
1674 unsigned long *nr_scanned, struct scan_control *sc,
1677 struct list_head *src = &lruvec->lists[lru];
1678 unsigned long nr_taken = 0;
1679 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1680 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1681 unsigned long skipped = 0;
1682 unsigned long scan, total_scan, nr_pages;
1683 LIST_HEAD(pages_skipped);
1684 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1688 while (scan < nr_to_scan && !list_empty(src)) {
1691 page = lru_to_page(src);
1692 prefetchw_prev_lru_page(page, src, flags);
1694 VM_BUG_ON_PAGE(!PageLRU(page), page);
1696 nr_pages = compound_nr(page);
1697 total_scan += nr_pages;
1699 if (page_zonenum(page) > sc->reclaim_idx) {
1700 list_move(&page->lru, &pages_skipped);
1701 nr_skipped[page_zonenum(page)] += nr_pages;
1706 * Do not count skipped pages because that makes the function
1707 * return with no isolated pages if the LRU mostly contains
1708 * ineligible pages. This causes the VM to not reclaim any
1709 * pages, triggering a premature OOM.
1711 * Account all tail pages of THP. This would not cause
1712 * premature OOM since __isolate_lru_page() returns -EBUSY
1713 * only when the page is being freed somewhere else.
1716 switch (__isolate_lru_page(page, mode)) {
1718 nr_taken += nr_pages;
1719 nr_zone_taken[page_zonenum(page)] += nr_pages;
1720 list_move(&page->lru, dst);
1724 /* else it is being freed elsewhere */
1725 list_move(&page->lru, src);
1734 * Splice any skipped pages to the start of the LRU list. Note that
1735 * this disrupts the LRU order when reclaiming for lower zones but
1736 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1737 * scanning would soon rescan the same pages to skip and put the
1738 * system at risk of premature OOM.
1740 if (!list_empty(&pages_skipped)) {
1743 list_splice(&pages_skipped, src);
1744 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1745 if (!nr_skipped[zid])
1748 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1749 skipped += nr_skipped[zid];
1752 *nr_scanned = total_scan;
1753 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1754 total_scan, skipped, nr_taken, mode, lru);
1755 update_lru_sizes(lruvec, lru, nr_zone_taken);
1760 * isolate_lru_page - tries to isolate a page from its LRU list
1761 * @page: page to isolate from its LRU list
1763 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1764 * vmstat statistic corresponding to whatever LRU list the page was on.
1766 * Returns 0 if the page was removed from an LRU list.
1767 * Returns -EBUSY if the page was not on an LRU list.
1769 * The returned page will have PageLRU() cleared. If it was found on
1770 * the active list, it will have PageActive set. If it was found on
1771 * the unevictable list, it will have the PageUnevictable bit set. That flag
1772 * may need to be cleared by the caller before letting the page go.
1774 * The vmstat statistic corresponding to the list on which the page was
1775 * found will be decremented.
1779 * (1) Must be called with an elevated refcount on the page. This is a
1780 * fundamentnal difference from isolate_lru_pages (which is called
1781 * without a stable reference).
1782 * (2) the lru_lock must not be held.
1783 * (3) interrupts must be enabled.
1785 int isolate_lru_page(struct page *page)
1789 VM_BUG_ON_PAGE(!page_count(page), page);
1790 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1792 if (PageLRU(page)) {
1793 pg_data_t *pgdat = page_pgdat(page);
1794 struct lruvec *lruvec;
1796 spin_lock_irq(&pgdat->lru_lock);
1797 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1798 if (PageLRU(page)) {
1799 int lru = page_lru(page);
1802 del_page_from_lru_list(page, lruvec, lru);
1805 spin_unlock_irq(&pgdat->lru_lock);
1811 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1812 * then get resheduled. When there are massive number of tasks doing page
1813 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1814 * the LRU list will go small and be scanned faster than necessary, leading to
1815 * unnecessary swapping, thrashing and OOM.
1817 static int too_many_isolated(struct pglist_data *pgdat, int file,
1818 struct scan_control *sc)
1820 unsigned long inactive, isolated;
1822 if (current_is_kswapd())
1825 if (!sane_reclaim(sc))
1829 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1830 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1832 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1833 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1837 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1838 * won't get blocked by normal direct-reclaimers, forming a circular
1841 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1844 return isolated > inactive;
1848 * This moves pages from @list to corresponding LRU list.
1850 * We move them the other way if the page is referenced by one or more
1851 * processes, from rmap.
1853 * If the pages are mostly unmapped, the processing is fast and it is
1854 * appropriate to hold zone_lru_lock across the whole operation. But if
1855 * the pages are mapped, the processing is slow (page_referenced()) so we
1856 * should drop zone_lru_lock around each page. It's impossible to balance
1857 * this, so instead we remove the pages from the LRU while processing them.
1858 * It is safe to rely on PG_active against the non-LRU pages in here because
1859 * nobody will play with that bit on a non-LRU page.
1861 * The downside is that we have to touch page->_refcount against each page.
1862 * But we had to alter page->flags anyway.
1864 * Returns the number of pages moved to the given lruvec.
1867 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1868 struct list_head *list)
1870 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1871 int nr_pages, nr_moved = 0;
1872 LIST_HEAD(pages_to_free);
1876 while (!list_empty(list)) {
1877 page = lru_to_page(list);
1878 VM_BUG_ON_PAGE(PageLRU(page), page);
1879 if (unlikely(!page_evictable(page))) {
1880 list_del(&page->lru);
1881 spin_unlock_irq(&pgdat->lru_lock);
1882 putback_lru_page(page);
1883 spin_lock_irq(&pgdat->lru_lock);
1886 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1889 lru = page_lru(page);
1891 nr_pages = hpage_nr_pages(page);
1892 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1893 list_move(&page->lru, &lruvec->lists[lru]);
1895 if (put_page_testzero(page)) {
1896 __ClearPageLRU(page);
1897 __ClearPageActive(page);
1898 del_page_from_lru_list(page, lruvec, lru);
1900 if (unlikely(PageCompound(page))) {
1901 spin_unlock_irq(&pgdat->lru_lock);
1902 (*get_compound_page_dtor(page))(page);
1903 spin_lock_irq(&pgdat->lru_lock);
1905 list_add(&page->lru, &pages_to_free);
1907 nr_moved += nr_pages;
1912 * To save our caller's stack, now use input list for pages to free.
1914 list_splice(&pages_to_free, list);
1920 * If a kernel thread (such as nfsd for loop-back mounts) services
1921 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1922 * In that case we should only throttle if the backing device it is
1923 * writing to is congested. In other cases it is safe to throttle.
1925 static int current_may_throttle(void)
1927 return !(current->flags & PF_LESS_THROTTLE) ||
1928 current->backing_dev_info == NULL ||
1929 bdi_write_congested(current->backing_dev_info);
1933 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1934 * of reclaimed pages
1936 static noinline_for_stack unsigned long
1937 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1938 struct scan_control *sc, enum lru_list lru)
1940 LIST_HEAD(page_list);
1941 unsigned long nr_scanned;
1942 unsigned long nr_reclaimed = 0;
1943 unsigned long nr_taken;
1944 struct reclaim_stat stat;
1945 int file = is_file_lru(lru);
1946 enum vm_event_item item;
1947 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1948 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1949 bool stalled = false;
1951 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1955 /* wait a bit for the reclaimer. */
1959 /* We are about to die and free our memory. Return now. */
1960 if (fatal_signal_pending(current))
1961 return SWAP_CLUSTER_MAX;
1966 spin_lock_irq(&pgdat->lru_lock);
1968 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1969 &nr_scanned, sc, lru);
1971 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1972 reclaim_stat->recent_scanned[file] += nr_taken;
1974 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1975 if (global_reclaim(sc))
1976 __count_vm_events(item, nr_scanned);
1977 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1978 spin_unlock_irq(&pgdat->lru_lock);
1983 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1986 spin_lock_irq(&pgdat->lru_lock);
1988 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1989 if (global_reclaim(sc))
1990 __count_vm_events(item, nr_reclaimed);
1991 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1992 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1993 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1995 move_pages_to_lru(lruvec, &page_list);
1997 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1999 spin_unlock_irq(&pgdat->lru_lock);
2001 mem_cgroup_uncharge_list(&page_list);
2002 free_unref_page_list(&page_list);
2005 * If dirty pages are scanned that are not queued for IO, it
2006 * implies that flushers are not doing their job. This can
2007 * happen when memory pressure pushes dirty pages to the end of
2008 * the LRU before the dirty limits are breached and the dirty
2009 * data has expired. It can also happen when the proportion of
2010 * dirty pages grows not through writes but through memory
2011 * pressure reclaiming all the clean cache. And in some cases,
2012 * the flushers simply cannot keep up with the allocation
2013 * rate. Nudge the flusher threads in case they are asleep.
2015 if (stat.nr_unqueued_dirty == nr_taken)
2016 wakeup_flusher_threads(WB_REASON_VMSCAN);
2018 sc->nr.dirty += stat.nr_dirty;
2019 sc->nr.congested += stat.nr_congested;
2020 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2021 sc->nr.writeback += stat.nr_writeback;
2022 sc->nr.immediate += stat.nr_immediate;
2023 sc->nr.taken += nr_taken;
2025 sc->nr.file_taken += nr_taken;
2027 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2028 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2029 return nr_reclaimed;
2032 static void shrink_active_list(unsigned long nr_to_scan,
2033 struct lruvec *lruvec,
2034 struct scan_control *sc,
2037 unsigned long nr_taken;
2038 unsigned long nr_scanned;
2039 unsigned long vm_flags;
2040 LIST_HEAD(l_hold); /* The pages which were snipped off */
2041 LIST_HEAD(l_active);
2042 LIST_HEAD(l_inactive);
2044 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2045 unsigned nr_deactivate, nr_activate;
2046 unsigned nr_rotated = 0;
2047 int file = is_file_lru(lru);
2048 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2052 spin_lock_irq(&pgdat->lru_lock);
2054 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2055 &nr_scanned, sc, lru);
2057 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2058 reclaim_stat->recent_scanned[file] += nr_taken;
2060 __count_vm_events(PGREFILL, nr_scanned);
2061 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2063 spin_unlock_irq(&pgdat->lru_lock);
2065 while (!list_empty(&l_hold)) {
2067 page = lru_to_page(&l_hold);
2068 list_del(&page->lru);
2070 if (unlikely(!page_evictable(page))) {
2071 putback_lru_page(page);
2075 if (unlikely(buffer_heads_over_limit)) {
2076 if (page_has_private(page) && trylock_page(page)) {
2077 if (page_has_private(page))
2078 try_to_release_page(page, 0);
2083 if (page_referenced(page, 0, sc->target_mem_cgroup,
2085 nr_rotated += hpage_nr_pages(page);
2087 * Identify referenced, file-backed active pages and
2088 * give them one more trip around the active list. So
2089 * that executable code get better chances to stay in
2090 * memory under moderate memory pressure. Anon pages
2091 * are not likely to be evicted by use-once streaming
2092 * IO, plus JVM can create lots of anon VM_EXEC pages,
2093 * so we ignore them here.
2095 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2096 list_add(&page->lru, &l_active);
2101 ClearPageActive(page); /* we are de-activating */
2102 SetPageWorkingset(page);
2103 list_add(&page->lru, &l_inactive);
2107 * Move pages back to the lru list.
2109 spin_lock_irq(&pgdat->lru_lock);
2111 * Count referenced pages from currently used mappings as rotated,
2112 * even though only some of them are actually re-activated. This
2113 * helps balance scan pressure between file and anonymous pages in
2116 reclaim_stat->recent_rotated[file] += nr_rotated;
2118 nr_activate = move_pages_to_lru(lruvec, &l_active);
2119 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2120 /* Keep all free pages in l_active list */
2121 list_splice(&l_inactive, &l_active);
2123 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2124 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2126 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2127 spin_unlock_irq(&pgdat->lru_lock);
2129 mem_cgroup_uncharge_list(&l_active);
2130 free_unref_page_list(&l_active);
2131 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2132 nr_deactivate, nr_rotated, sc->priority, file);
2135 unsigned long reclaim_pages(struct list_head *page_list)
2138 unsigned long nr_reclaimed = 0;
2139 LIST_HEAD(node_page_list);
2140 struct reclaim_stat dummy_stat;
2142 struct scan_control sc = {
2143 .gfp_mask = GFP_KERNEL,
2144 .priority = DEF_PRIORITY,
2150 while (!list_empty(page_list)) {
2151 page = lru_to_page(page_list);
2153 nid = page_to_nid(page);
2154 INIT_LIST_HEAD(&node_page_list);
2157 if (nid == page_to_nid(page)) {
2158 ClearPageActive(page);
2159 list_move(&page->lru, &node_page_list);
2163 nr_reclaimed += shrink_page_list(&node_page_list,
2166 &dummy_stat, false);
2167 while (!list_empty(&node_page_list)) {
2168 page = lru_to_page(&node_page_list);
2169 list_del(&page->lru);
2170 putback_lru_page(page);
2176 if (!list_empty(&node_page_list)) {
2177 nr_reclaimed += shrink_page_list(&node_page_list,
2180 &dummy_stat, false);
2181 while (!list_empty(&node_page_list)) {
2182 page = lru_to_page(&node_page_list);
2183 list_del(&page->lru);
2184 putback_lru_page(page);
2188 return nr_reclaimed;
2192 * The inactive anon list should be small enough that the VM never has
2193 * to do too much work.
2195 * The inactive file list should be small enough to leave most memory
2196 * to the established workingset on the scan-resistant active list,
2197 * but large enough to avoid thrashing the aggregate readahead window.
2199 * Both inactive lists should also be large enough that each inactive
2200 * page has a chance to be referenced again before it is reclaimed.
2202 * If that fails and refaulting is observed, the inactive list grows.
2204 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2205 * on this LRU, maintained by the pageout code. An inactive_ratio
2206 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2209 * memory ratio inactive
2210 * -------------------------------------
2219 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2220 struct scan_control *sc, bool trace)
2222 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2223 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2224 enum lru_list inactive_lru = file * LRU_FILE;
2225 unsigned long inactive, active;
2226 unsigned long inactive_ratio;
2227 unsigned long refaults;
2230 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2231 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2234 * When refaults are being observed, it means a new workingset
2235 * is being established. Disable active list protection to get
2236 * rid of the stale workingset quickly.
2238 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2239 if (file && lruvec->refaults != refaults) {
2242 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2244 inactive_ratio = int_sqrt(10 * gb);
2250 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2251 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2252 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2253 inactive_ratio, file);
2255 return inactive * inactive_ratio < active;
2258 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2259 struct lruvec *lruvec, struct scan_control *sc)
2261 if (is_active_lru(lru)) {
2262 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2263 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2267 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2278 * Determine how aggressively the anon and file LRU lists should be
2279 * scanned. The relative value of each set of LRU lists is determined
2280 * by looking at the fraction of the pages scanned we did rotate back
2281 * onto the active list instead of evict.
2283 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2284 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2286 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2287 struct scan_control *sc, unsigned long *nr)
2289 int swappiness = mem_cgroup_swappiness(memcg);
2290 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2292 u64 denominator = 0; /* gcc */
2293 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2294 unsigned long anon_prio, file_prio;
2295 enum scan_balance scan_balance;
2296 unsigned long anon, file;
2297 unsigned long ap, fp;
2300 /* If we have no swap space, do not bother scanning anon pages. */
2301 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2302 scan_balance = SCAN_FILE;
2307 * Global reclaim will swap to prevent OOM even with no
2308 * swappiness, but memcg users want to use this knob to
2309 * disable swapping for individual groups completely when
2310 * using the memory controller's swap limit feature would be
2313 if (!global_reclaim(sc) && !swappiness) {
2314 scan_balance = SCAN_FILE;
2319 * Do not apply any pressure balancing cleverness when the
2320 * system is close to OOM, scan both anon and file equally
2321 * (unless the swappiness setting disagrees with swapping).
2323 if (!sc->priority && swappiness) {
2324 scan_balance = SCAN_EQUAL;
2329 * Prevent the reclaimer from falling into the cache trap: as
2330 * cache pages start out inactive, every cache fault will tip
2331 * the scan balance towards the file LRU. And as the file LRU
2332 * shrinks, so does the window for rotation from references.
2333 * This means we have a runaway feedback loop where a tiny
2334 * thrashing file LRU becomes infinitely more attractive than
2335 * anon pages. Try to detect this based on file LRU size.
2337 if (global_reclaim(sc)) {
2338 unsigned long pgdatfile;
2339 unsigned long pgdatfree;
2341 unsigned long total_high_wmark = 0;
2343 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2344 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2345 node_page_state(pgdat, NR_INACTIVE_FILE);
2347 for (z = 0; z < MAX_NR_ZONES; z++) {
2348 struct zone *zone = &pgdat->node_zones[z];
2349 if (!managed_zone(zone))
2352 total_high_wmark += high_wmark_pages(zone);
2355 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2357 * Force SCAN_ANON if there are enough inactive
2358 * anonymous pages on the LRU in eligible zones.
2359 * Otherwise, the small LRU gets thrashed.
2361 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2362 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2364 scan_balance = SCAN_ANON;
2371 * If there is enough inactive page cache, i.e. if the size of the
2372 * inactive list is greater than that of the active list *and* the
2373 * inactive list actually has some pages to scan on this priority, we
2374 * do not reclaim anything from the anonymous working set right now.
2375 * Without the second condition we could end up never scanning an
2376 * lruvec even if it has plenty of old anonymous pages unless the
2377 * system is under heavy pressure.
2379 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2380 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2381 scan_balance = SCAN_FILE;
2385 scan_balance = SCAN_FRACT;
2388 * With swappiness at 100, anonymous and file have the same priority.
2389 * This scanning priority is essentially the inverse of IO cost.
2391 anon_prio = swappiness;
2392 file_prio = 200 - anon_prio;
2395 * OK, so we have swap space and a fair amount of page cache
2396 * pages. We use the recently rotated / recently scanned
2397 * ratios to determine how valuable each cache is.
2399 * Because workloads change over time (and to avoid overflow)
2400 * we keep these statistics as a floating average, which ends
2401 * up weighing recent references more than old ones.
2403 * anon in [0], file in [1]
2406 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2407 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2408 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2409 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2411 spin_lock_irq(&pgdat->lru_lock);
2412 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2413 reclaim_stat->recent_scanned[0] /= 2;
2414 reclaim_stat->recent_rotated[0] /= 2;
2417 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2418 reclaim_stat->recent_scanned[1] /= 2;
2419 reclaim_stat->recent_rotated[1] /= 2;
2423 * The amount of pressure on anon vs file pages is inversely
2424 * proportional to the fraction of recently scanned pages on
2425 * each list that were recently referenced and in active use.
2427 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2428 ap /= reclaim_stat->recent_rotated[0] + 1;
2430 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2431 fp /= reclaim_stat->recent_rotated[1] + 1;
2432 spin_unlock_irq(&pgdat->lru_lock);
2436 denominator = ap + fp + 1;
2438 for_each_evictable_lru(lru) {
2439 int file = is_file_lru(lru);
2440 unsigned long lruvec_size;
2442 unsigned long protection;
2444 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2445 protection = mem_cgroup_protection(memcg,
2446 sc->memcg_low_reclaim);
2450 * Scale a cgroup's reclaim pressure by proportioning
2451 * its current usage to its memory.low or memory.min
2454 * This is important, as otherwise scanning aggression
2455 * becomes extremely binary -- from nothing as we
2456 * approach the memory protection threshold, to totally
2457 * nominal as we exceed it. This results in requiring
2458 * setting extremely liberal protection thresholds. It
2459 * also means we simply get no protection at all if we
2460 * set it too low, which is not ideal.
2462 * If there is any protection in place, we reduce scan
2463 * pressure by how much of the total memory used is
2464 * within protection thresholds.
2466 * There is one special case: in the first reclaim pass,
2467 * we skip over all groups that are within their low
2468 * protection. If that fails to reclaim enough pages to
2469 * satisfy the reclaim goal, we come back and override
2470 * the best-effort low protection. However, we still
2471 * ideally want to honor how well-behaved groups are in
2472 * that case instead of simply punishing them all
2473 * equally. As such, we reclaim them based on how much
2474 * memory they are using, reducing the scan pressure
2475 * again by how much of the total memory used is under
2478 unsigned long cgroup_size = mem_cgroup_size(memcg);
2480 /* Avoid TOCTOU with earlier protection check */
2481 cgroup_size = max(cgroup_size, protection);
2483 scan = lruvec_size - lruvec_size * protection /
2487 * Minimally target SWAP_CLUSTER_MAX pages to keep
2488 * reclaim moving forwards, avoiding decremeting
2489 * sc->priority further than desirable.
2491 scan = max(scan, SWAP_CLUSTER_MAX);
2496 scan >>= sc->priority;
2499 * If the cgroup's already been deleted, make sure to
2500 * scrape out the remaining cache.
2502 if (!scan && !mem_cgroup_online(memcg))
2503 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2505 switch (scan_balance) {
2507 /* Scan lists relative to size */
2511 * Scan types proportional to swappiness and
2512 * their relative recent reclaim efficiency.
2513 * Make sure we don't miss the last page on
2514 * the offlined memory cgroups because of a
2517 scan = mem_cgroup_online(memcg) ?
2518 div64_u64(scan * fraction[file], denominator) :
2519 DIV64_U64_ROUND_UP(scan * fraction[file],
2524 /* Scan one type exclusively */
2525 if ((scan_balance == SCAN_FILE) != file) {
2531 /* Look ma, no brain */
2540 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2542 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2543 struct scan_control *sc)
2545 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2546 unsigned long nr[NR_LRU_LISTS];
2547 unsigned long targets[NR_LRU_LISTS];
2548 unsigned long nr_to_scan;
2550 unsigned long nr_reclaimed = 0;
2551 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2552 struct blk_plug plug;
2555 get_scan_count(lruvec, memcg, sc, nr);
2557 /* Record the original scan target for proportional adjustments later */
2558 memcpy(targets, nr, sizeof(nr));
2561 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2562 * event that can occur when there is little memory pressure e.g.
2563 * multiple streaming readers/writers. Hence, we do not abort scanning
2564 * when the requested number of pages are reclaimed when scanning at
2565 * DEF_PRIORITY on the assumption that the fact we are direct
2566 * reclaiming implies that kswapd is not keeping up and it is best to
2567 * do a batch of work at once. For memcg reclaim one check is made to
2568 * abort proportional reclaim if either the file or anon lru has already
2569 * dropped to zero at the first pass.
2571 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2572 sc->priority == DEF_PRIORITY);
2574 blk_start_plug(&plug);
2575 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2576 nr[LRU_INACTIVE_FILE]) {
2577 unsigned long nr_anon, nr_file, percentage;
2578 unsigned long nr_scanned;
2580 for_each_evictable_lru(lru) {
2582 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2583 nr[lru] -= nr_to_scan;
2585 nr_reclaimed += shrink_list(lru, nr_to_scan,
2592 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2596 * For kswapd and memcg, reclaim at least the number of pages
2597 * requested. Ensure that the anon and file LRUs are scanned
2598 * proportionally what was requested by get_scan_count(). We
2599 * stop reclaiming one LRU and reduce the amount scanning
2600 * proportional to the original scan target.
2602 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2603 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2606 * It's just vindictive to attack the larger once the smaller
2607 * has gone to zero. And given the way we stop scanning the
2608 * smaller below, this makes sure that we only make one nudge
2609 * towards proportionality once we've got nr_to_reclaim.
2611 if (!nr_file || !nr_anon)
2614 if (nr_file > nr_anon) {
2615 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2616 targets[LRU_ACTIVE_ANON] + 1;
2618 percentage = nr_anon * 100 / scan_target;
2620 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2621 targets[LRU_ACTIVE_FILE] + 1;
2623 percentage = nr_file * 100 / scan_target;
2626 /* Stop scanning the smaller of the LRU */
2628 nr[lru + LRU_ACTIVE] = 0;
2631 * Recalculate the other LRU scan count based on its original
2632 * scan target and the percentage scanning already complete
2634 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2635 nr_scanned = targets[lru] - nr[lru];
2636 nr[lru] = targets[lru] * (100 - percentage) / 100;
2637 nr[lru] -= min(nr[lru], nr_scanned);
2640 nr_scanned = targets[lru] - nr[lru];
2641 nr[lru] = targets[lru] * (100 - percentage) / 100;
2642 nr[lru] -= min(nr[lru], nr_scanned);
2644 scan_adjusted = true;
2646 blk_finish_plug(&plug);
2647 sc->nr_reclaimed += nr_reclaimed;
2650 * Even if we did not try to evict anon pages at all, we want to
2651 * rebalance the anon lru active/inactive ratio.
2653 if (total_swap_pages && inactive_list_is_low(lruvec, false, sc, true))
2654 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2655 sc, LRU_ACTIVE_ANON);
2658 /* Use reclaim/compaction for costly allocs or under memory pressure */
2659 static bool in_reclaim_compaction(struct scan_control *sc)
2661 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2662 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2663 sc->priority < DEF_PRIORITY - 2))
2670 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2671 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2672 * true if more pages should be reclaimed such that when the page allocator
2673 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2674 * It will give up earlier than that if there is difficulty reclaiming pages.
2676 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2677 unsigned long nr_reclaimed,
2678 struct scan_control *sc)
2680 unsigned long pages_for_compaction;
2681 unsigned long inactive_lru_pages;
2684 /* If not in reclaim/compaction mode, stop */
2685 if (!in_reclaim_compaction(sc))
2689 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2690 * number of pages that were scanned. This will return to the caller
2691 * with the risk reclaim/compaction and the resulting allocation attempt
2692 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2693 * allocations through requiring that the full LRU list has been scanned
2694 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2695 * scan, but that approximation was wrong, and there were corner cases
2696 * where always a non-zero amount of pages were scanned.
2701 /* If compaction would go ahead or the allocation would succeed, stop */
2702 for (z = 0; z <= sc->reclaim_idx; z++) {
2703 struct zone *zone = &pgdat->node_zones[z];
2704 if (!managed_zone(zone))
2707 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2708 case COMPACT_SUCCESS:
2709 case COMPACT_CONTINUE:
2712 /* check next zone */
2718 * If we have not reclaimed enough pages for compaction and the
2719 * inactive lists are large enough, continue reclaiming
2721 pages_for_compaction = compact_gap(sc->order);
2722 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2723 if (get_nr_swap_pages() > 0)
2724 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2726 return inactive_lru_pages > pages_for_compaction;
2729 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2731 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2732 (memcg && memcg_congested(pgdat, memcg));
2735 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2737 struct reclaim_state *reclaim_state = current->reclaim_state;
2738 unsigned long nr_reclaimed, nr_scanned;
2739 bool reclaimable = false;
2742 struct mem_cgroup *root = sc->target_mem_cgroup;
2743 struct mem_cgroup *memcg;
2745 memset(&sc->nr, 0, sizeof(sc->nr));
2747 nr_reclaimed = sc->nr_reclaimed;
2748 nr_scanned = sc->nr_scanned;
2750 memcg = mem_cgroup_iter(root, NULL, NULL);
2752 unsigned long reclaimed;
2753 unsigned long scanned;
2755 switch (mem_cgroup_protected(root, memcg)) {
2756 case MEMCG_PROT_MIN:
2759 * If there is no reclaimable memory, OOM.
2762 case MEMCG_PROT_LOW:
2765 * Respect the protection only as long as
2766 * there is an unprotected supply
2767 * of reclaimable memory from other cgroups.
2769 if (!sc->memcg_low_reclaim) {
2770 sc->memcg_low_skipped = 1;
2773 memcg_memory_event(memcg, MEMCG_LOW);
2775 case MEMCG_PROT_NONE:
2777 * All protection thresholds breached. We may
2778 * still choose to vary the scan pressure
2779 * applied based on by how much the cgroup in
2780 * question has exceeded its protection
2781 * thresholds (see get_scan_count).
2786 reclaimed = sc->nr_reclaimed;
2787 scanned = sc->nr_scanned;
2788 shrink_node_memcg(pgdat, memcg, sc);
2790 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2793 /* Record the group's reclaim efficiency */
2794 vmpressure(sc->gfp_mask, memcg, false,
2795 sc->nr_scanned - scanned,
2796 sc->nr_reclaimed - reclaimed);
2798 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2800 if (reclaim_state) {
2801 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2802 reclaim_state->reclaimed_slab = 0;
2805 /* Record the subtree's reclaim efficiency */
2806 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2807 sc->nr_scanned - nr_scanned,
2808 sc->nr_reclaimed - nr_reclaimed);
2810 if (sc->nr_reclaimed - nr_reclaimed)
2813 if (current_is_kswapd()) {
2815 * If reclaim is isolating dirty pages under writeback,
2816 * it implies that the long-lived page allocation rate
2817 * is exceeding the page laundering rate. Either the
2818 * global limits are not being effective at throttling
2819 * processes due to the page distribution throughout
2820 * zones or there is heavy usage of a slow backing
2821 * device. The only option is to throttle from reclaim
2822 * context which is not ideal as there is no guarantee
2823 * the dirtying process is throttled in the same way
2824 * balance_dirty_pages() manages.
2826 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2827 * count the number of pages under pages flagged for
2828 * immediate reclaim and stall if any are encountered
2829 * in the nr_immediate check below.
2831 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2832 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2835 * Tag a node as congested if all the dirty pages
2836 * scanned were backed by a congested BDI and
2837 * wait_iff_congested will stall.
2839 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2840 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2842 /* Allow kswapd to start writing pages during reclaim.*/
2843 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2844 set_bit(PGDAT_DIRTY, &pgdat->flags);
2847 * If kswapd scans pages marked marked for immediate
2848 * reclaim and under writeback (nr_immediate), it
2849 * implies that pages are cycling through the LRU
2850 * faster than they are written so also forcibly stall.
2852 if (sc->nr.immediate)
2853 congestion_wait(BLK_RW_ASYNC, HZ/10);
2857 * Legacy memcg will stall in page writeback so avoid forcibly
2858 * stalling in wait_iff_congested().
2860 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2861 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2862 set_memcg_congestion(pgdat, root, true);
2865 * Stall direct reclaim for IO completions if underlying BDIs
2866 * and node is congested. Allow kswapd to continue until it
2867 * starts encountering unqueued dirty pages or cycling through
2868 * the LRU too quickly.
2870 if (!sc->hibernation_mode && !current_is_kswapd() &&
2871 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2872 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2874 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2878 * Kswapd gives up on balancing particular nodes after too
2879 * many failures to reclaim anything from them and goes to
2880 * sleep. On reclaim progress, reset the failure counter. A
2881 * successful direct reclaim run will revive a dormant kswapd.
2884 pgdat->kswapd_failures = 0;
2890 * Returns true if compaction should go ahead for a costly-order request, or
2891 * the allocation would already succeed without compaction. Return false if we
2892 * should reclaim first.
2894 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2896 unsigned long watermark;
2897 enum compact_result suitable;
2899 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2900 if (suitable == COMPACT_SUCCESS)
2901 /* Allocation should succeed already. Don't reclaim. */
2903 if (suitable == COMPACT_SKIPPED)
2904 /* Compaction cannot yet proceed. Do reclaim. */
2908 * Compaction is already possible, but it takes time to run and there
2909 * are potentially other callers using the pages just freed. So proceed
2910 * with reclaim to make a buffer of free pages available to give
2911 * compaction a reasonable chance of completing and allocating the page.
2912 * Note that we won't actually reclaim the whole buffer in one attempt
2913 * as the target watermark in should_continue_reclaim() is lower. But if
2914 * we are already above the high+gap watermark, don't reclaim at all.
2916 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2918 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2922 * This is the direct reclaim path, for page-allocating processes. We only
2923 * try to reclaim pages from zones which will satisfy the caller's allocation
2926 * If a zone is deemed to be full of pinned pages then just give it a light
2927 * scan then give up on it.
2929 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2933 unsigned long nr_soft_reclaimed;
2934 unsigned long nr_soft_scanned;
2936 pg_data_t *last_pgdat = NULL;
2939 * If the number of buffer_heads in the machine exceeds the maximum
2940 * allowed level, force direct reclaim to scan the highmem zone as
2941 * highmem pages could be pinning lowmem pages storing buffer_heads
2943 orig_mask = sc->gfp_mask;
2944 if (buffer_heads_over_limit) {
2945 sc->gfp_mask |= __GFP_HIGHMEM;
2946 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2949 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2950 sc->reclaim_idx, sc->nodemask) {
2952 * Take care memory controller reclaiming has small influence
2955 if (global_reclaim(sc)) {
2956 if (!cpuset_zone_allowed(zone,
2957 GFP_KERNEL | __GFP_HARDWALL))
2961 * If we already have plenty of memory free for
2962 * compaction in this zone, don't free any more.
2963 * Even though compaction is invoked for any
2964 * non-zero order, only frequent costly order
2965 * reclamation is disruptive enough to become a
2966 * noticeable problem, like transparent huge
2969 if (IS_ENABLED(CONFIG_COMPACTION) &&
2970 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2971 compaction_ready(zone, sc)) {
2972 sc->compaction_ready = true;
2977 * Shrink each node in the zonelist once. If the
2978 * zonelist is ordered by zone (not the default) then a
2979 * node may be shrunk multiple times but in that case
2980 * the user prefers lower zones being preserved.
2982 if (zone->zone_pgdat == last_pgdat)
2986 * This steals pages from memory cgroups over softlimit
2987 * and returns the number of reclaimed pages and
2988 * scanned pages. This works for global memory pressure
2989 * and balancing, not for a memcg's limit.
2991 nr_soft_scanned = 0;
2992 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2993 sc->order, sc->gfp_mask,
2995 sc->nr_reclaimed += nr_soft_reclaimed;
2996 sc->nr_scanned += nr_soft_scanned;
2997 /* need some check for avoid more shrink_zone() */
3000 /* See comment about same check for global reclaim above */
3001 if (zone->zone_pgdat == last_pgdat)
3003 last_pgdat = zone->zone_pgdat;
3004 shrink_node(zone->zone_pgdat, sc);
3008 * Restore to original mask to avoid the impact on the caller if we
3009 * promoted it to __GFP_HIGHMEM.
3011 sc->gfp_mask = orig_mask;
3014 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3016 struct mem_cgroup *memcg;
3018 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3020 unsigned long refaults;
3021 struct lruvec *lruvec;
3023 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3024 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3025 lruvec->refaults = refaults;
3026 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3030 * This is the main entry point to direct page reclaim.
3032 * If a full scan of the inactive list fails to free enough memory then we
3033 * are "out of memory" and something needs to be killed.
3035 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3036 * high - the zone may be full of dirty or under-writeback pages, which this
3037 * caller can't do much about. We kick the writeback threads and take explicit
3038 * naps in the hope that some of these pages can be written. But if the
3039 * allocating task holds filesystem locks which prevent writeout this might not
3040 * work, and the allocation attempt will fail.
3042 * returns: 0, if no pages reclaimed
3043 * else, the number of pages reclaimed
3045 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3046 struct scan_control *sc)
3048 int initial_priority = sc->priority;
3049 pg_data_t *last_pgdat;
3053 delayacct_freepages_start();
3055 if (global_reclaim(sc))
3056 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3059 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3062 shrink_zones(zonelist, sc);
3064 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3067 if (sc->compaction_ready)
3071 * If we're getting trouble reclaiming, start doing
3072 * writepage even in laptop mode.
3074 if (sc->priority < DEF_PRIORITY - 2)
3075 sc->may_writepage = 1;
3076 } while (--sc->priority >= 0);
3079 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3081 if (zone->zone_pgdat == last_pgdat)
3083 last_pgdat = zone->zone_pgdat;
3084 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3085 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3088 delayacct_freepages_end();
3090 if (sc->nr_reclaimed)
3091 return sc->nr_reclaimed;
3093 /* Aborted reclaim to try compaction? don't OOM, then */
3094 if (sc->compaction_ready)
3097 /* Untapped cgroup reserves? Don't OOM, retry. */
3098 if (sc->memcg_low_skipped) {
3099 sc->priority = initial_priority;
3100 sc->memcg_low_reclaim = 1;
3101 sc->memcg_low_skipped = 0;
3108 static bool allow_direct_reclaim(pg_data_t *pgdat)
3111 unsigned long pfmemalloc_reserve = 0;
3112 unsigned long free_pages = 0;
3116 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3119 for (i = 0; i <= ZONE_NORMAL; i++) {
3120 zone = &pgdat->node_zones[i];
3121 if (!managed_zone(zone))
3124 if (!zone_reclaimable_pages(zone))
3127 pfmemalloc_reserve += min_wmark_pages(zone);
3128 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3131 /* If there are no reserves (unexpected config) then do not throttle */
3132 if (!pfmemalloc_reserve)
3135 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3137 /* kswapd must be awake if processes are being throttled */
3138 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3139 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3140 (enum zone_type)ZONE_NORMAL);
3141 wake_up_interruptible(&pgdat->kswapd_wait);
3148 * Throttle direct reclaimers if backing storage is backed by the network
3149 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3150 * depleted. kswapd will continue to make progress and wake the processes
3151 * when the low watermark is reached.
3153 * Returns true if a fatal signal was delivered during throttling. If this
3154 * happens, the page allocator should not consider triggering the OOM killer.
3156 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3157 nodemask_t *nodemask)
3161 pg_data_t *pgdat = NULL;
3164 * Kernel threads should not be throttled as they may be indirectly
3165 * responsible for cleaning pages necessary for reclaim to make forward
3166 * progress. kjournald for example may enter direct reclaim while
3167 * committing a transaction where throttling it could forcing other
3168 * processes to block on log_wait_commit().
3170 if (current->flags & PF_KTHREAD)
3174 * If a fatal signal is pending, this process should not throttle.
3175 * It should return quickly so it can exit and free its memory
3177 if (fatal_signal_pending(current))
3181 * Check if the pfmemalloc reserves are ok by finding the first node
3182 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3183 * GFP_KERNEL will be required for allocating network buffers when
3184 * swapping over the network so ZONE_HIGHMEM is unusable.
3186 * Throttling is based on the first usable node and throttled processes
3187 * wait on a queue until kswapd makes progress and wakes them. There
3188 * is an affinity then between processes waking up and where reclaim
3189 * progress has been made assuming the process wakes on the same node.
3190 * More importantly, processes running on remote nodes will not compete
3191 * for remote pfmemalloc reserves and processes on different nodes
3192 * should make reasonable progress.
3194 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3195 gfp_zone(gfp_mask), nodemask) {
3196 if (zone_idx(zone) > ZONE_NORMAL)
3199 /* Throttle based on the first usable node */
3200 pgdat = zone->zone_pgdat;
3201 if (allow_direct_reclaim(pgdat))
3206 /* If no zone was usable by the allocation flags then do not throttle */
3210 /* Account for the throttling */
3211 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3214 * If the caller cannot enter the filesystem, it's possible that it
3215 * is due to the caller holding an FS lock or performing a journal
3216 * transaction in the case of a filesystem like ext[3|4]. In this case,
3217 * it is not safe to block on pfmemalloc_wait as kswapd could be
3218 * blocked waiting on the same lock. Instead, throttle for up to a
3219 * second before continuing.
3221 if (!(gfp_mask & __GFP_FS)) {
3222 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3223 allow_direct_reclaim(pgdat), HZ);
3228 /* Throttle until kswapd wakes the process */
3229 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3230 allow_direct_reclaim(pgdat));
3233 if (fatal_signal_pending(current))
3240 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3241 gfp_t gfp_mask, nodemask_t *nodemask)
3243 unsigned long nr_reclaimed;
3244 struct scan_control sc = {
3245 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3246 .gfp_mask = current_gfp_context(gfp_mask),
3247 .reclaim_idx = gfp_zone(gfp_mask),
3249 .nodemask = nodemask,
3250 .priority = DEF_PRIORITY,
3251 .may_writepage = !laptop_mode,
3257 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3258 * Confirm they are large enough for max values.
3260 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3261 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3262 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3265 * Do not enter reclaim if fatal signal was delivered while throttled.
3266 * 1 is returned so that the page allocator does not OOM kill at this
3269 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3272 set_task_reclaim_state(current, &sc.reclaim_state);
3273 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3275 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3277 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3278 set_task_reclaim_state(current, NULL);
3280 return nr_reclaimed;
3285 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3286 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3287 gfp_t gfp_mask, bool noswap,
3289 unsigned long *nr_scanned)
3291 struct scan_control sc = {
3292 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3293 .target_mem_cgroup = memcg,
3294 .may_writepage = !laptop_mode,
3296 .reclaim_idx = MAX_NR_ZONES - 1,
3297 .may_swap = !noswap,
3300 WARN_ON_ONCE(!current->reclaim_state);
3302 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3303 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3305 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3309 * NOTE: Although we can get the priority field, using it
3310 * here is not a good idea, since it limits the pages we can scan.
3311 * if we don't reclaim here, the shrink_node from balance_pgdat
3312 * will pick up pages from other mem cgroup's as well. We hack
3313 * the priority and make it zero.
3315 shrink_node_memcg(pgdat, memcg, &sc);
3317 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3319 *nr_scanned = sc.nr_scanned;
3321 return sc.nr_reclaimed;
3324 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3325 unsigned long nr_pages,
3329 struct zonelist *zonelist;
3330 unsigned long nr_reclaimed;
3331 unsigned long pflags;
3333 unsigned int noreclaim_flag;
3334 struct scan_control sc = {
3335 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3336 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3337 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3338 .reclaim_idx = MAX_NR_ZONES - 1,
3339 .target_mem_cgroup = memcg,
3340 .priority = DEF_PRIORITY,
3341 .may_writepage = !laptop_mode,
3343 .may_swap = may_swap,
3346 set_task_reclaim_state(current, &sc.reclaim_state);
3348 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3349 * take care of from where we get pages. So the node where we start the
3350 * scan does not need to be the current node.
3352 nid = mem_cgroup_select_victim_node(memcg);
3354 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3356 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3358 psi_memstall_enter(&pflags);
3359 noreclaim_flag = memalloc_noreclaim_save();
3361 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3363 memalloc_noreclaim_restore(noreclaim_flag);
3364 psi_memstall_leave(&pflags);
3366 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3367 set_task_reclaim_state(current, NULL);
3369 return nr_reclaimed;
3373 static void age_active_anon(struct pglist_data *pgdat,
3374 struct scan_control *sc)
3376 struct mem_cgroup *memcg;
3378 if (!total_swap_pages)
3381 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3383 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3385 if (inactive_list_is_low(lruvec, false, sc, true))
3386 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3387 sc, LRU_ACTIVE_ANON);
3389 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3393 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3399 * Check for watermark boosts top-down as the higher zones
3400 * are more likely to be boosted. Both watermarks and boosts
3401 * should not be checked at the time time as reclaim would
3402 * start prematurely when there is no boosting and a lower
3405 for (i = classzone_idx; i >= 0; i--) {
3406 zone = pgdat->node_zones + i;
3407 if (!managed_zone(zone))
3410 if (zone->watermark_boost)
3418 * Returns true if there is an eligible zone balanced for the request order
3421 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3424 unsigned long mark = -1;
3428 * Check watermarks bottom-up as lower zones are more likely to
3431 for (i = 0; i <= classzone_idx; i++) {
3432 zone = pgdat->node_zones + i;
3434 if (!managed_zone(zone))
3437 mark = high_wmark_pages(zone);
3438 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3443 * If a node has no populated zone within classzone_idx, it does not
3444 * need balancing by definition. This can happen if a zone-restricted
3445 * allocation tries to wake a remote kswapd.
3453 /* Clear pgdat state for congested, dirty or under writeback. */
3454 static void clear_pgdat_congested(pg_data_t *pgdat)
3456 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3457 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3458 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3462 * Prepare kswapd for sleeping. This verifies that there are no processes
3463 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3465 * Returns true if kswapd is ready to sleep
3467 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3470 * The throttled processes are normally woken up in balance_pgdat() as
3471 * soon as allow_direct_reclaim() is true. But there is a potential
3472 * race between when kswapd checks the watermarks and a process gets
3473 * throttled. There is also a potential race if processes get
3474 * throttled, kswapd wakes, a large process exits thereby balancing the
3475 * zones, which causes kswapd to exit balance_pgdat() before reaching
3476 * the wake up checks. If kswapd is going to sleep, no process should
3477 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3478 * the wake up is premature, processes will wake kswapd and get
3479 * throttled again. The difference from wake ups in balance_pgdat() is
3480 * that here we are under prepare_to_wait().
3482 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3483 wake_up_all(&pgdat->pfmemalloc_wait);
3485 /* Hopeless node, leave it to direct reclaim */
3486 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3489 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3490 clear_pgdat_congested(pgdat);
3498 * kswapd shrinks a node of pages that are at or below the highest usable
3499 * zone that is currently unbalanced.
3501 * Returns true if kswapd scanned at least the requested number of pages to
3502 * reclaim or if the lack of progress was due to pages under writeback.
3503 * This is used to determine if the scanning priority needs to be raised.
3505 static bool kswapd_shrink_node(pg_data_t *pgdat,
3506 struct scan_control *sc)
3511 /* Reclaim a number of pages proportional to the number of zones */
3512 sc->nr_to_reclaim = 0;
3513 for (z = 0; z <= sc->reclaim_idx; z++) {
3514 zone = pgdat->node_zones + z;
3515 if (!managed_zone(zone))
3518 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3522 * Historically care was taken to put equal pressure on all zones but
3523 * now pressure is applied based on node LRU order.
3525 shrink_node(pgdat, sc);
3528 * Fragmentation may mean that the system cannot be rebalanced for
3529 * high-order allocations. If twice the allocation size has been
3530 * reclaimed then recheck watermarks only at order-0 to prevent
3531 * excessive reclaim. Assume that a process requested a high-order
3532 * can direct reclaim/compact.
3534 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3537 return sc->nr_scanned >= sc->nr_to_reclaim;
3541 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3542 * that are eligible for use by the caller until at least one zone is
3545 * Returns the order kswapd finished reclaiming at.
3547 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3548 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3549 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3550 * or lower is eligible for reclaim until at least one usable zone is
3553 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3556 unsigned long nr_soft_reclaimed;
3557 unsigned long nr_soft_scanned;
3558 unsigned long pflags;
3559 unsigned long nr_boost_reclaim;
3560 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3563 struct scan_control sc = {
3564 .gfp_mask = GFP_KERNEL,
3569 set_task_reclaim_state(current, &sc.reclaim_state);
3570 psi_memstall_enter(&pflags);
3571 __fs_reclaim_acquire();
3573 count_vm_event(PAGEOUTRUN);
3576 * Account for the reclaim boost. Note that the zone boost is left in
3577 * place so that parallel allocations that are near the watermark will
3578 * stall or direct reclaim until kswapd is finished.
3580 nr_boost_reclaim = 0;
3581 for (i = 0; i <= classzone_idx; i++) {
3582 zone = pgdat->node_zones + i;
3583 if (!managed_zone(zone))
3586 nr_boost_reclaim += zone->watermark_boost;
3587 zone_boosts[i] = zone->watermark_boost;
3589 boosted = nr_boost_reclaim;
3592 sc.priority = DEF_PRIORITY;
3594 unsigned long nr_reclaimed = sc.nr_reclaimed;
3595 bool raise_priority = true;
3599 sc.reclaim_idx = classzone_idx;
3602 * If the number of buffer_heads exceeds the maximum allowed
3603 * then consider reclaiming from all zones. This has a dual
3604 * purpose -- on 64-bit systems it is expected that
3605 * buffer_heads are stripped during active rotation. On 32-bit
3606 * systems, highmem pages can pin lowmem memory and shrinking
3607 * buffers can relieve lowmem pressure. Reclaim may still not
3608 * go ahead if all eligible zones for the original allocation
3609 * request are balanced to avoid excessive reclaim from kswapd.
3611 if (buffer_heads_over_limit) {
3612 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3613 zone = pgdat->node_zones + i;
3614 if (!managed_zone(zone))
3623 * If the pgdat is imbalanced then ignore boosting and preserve
3624 * the watermarks for a later time and restart. Note that the
3625 * zone watermarks will be still reset at the end of balancing
3626 * on the grounds that the normal reclaim should be enough to
3627 * re-evaluate if boosting is required when kswapd next wakes.
3629 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3630 if (!balanced && nr_boost_reclaim) {
3631 nr_boost_reclaim = 0;
3636 * If boosting is not active then only reclaim if there are no
3637 * eligible zones. Note that sc.reclaim_idx is not used as
3638 * buffer_heads_over_limit may have adjusted it.
3640 if (!nr_boost_reclaim && balanced)
3643 /* Limit the priority of boosting to avoid reclaim writeback */
3644 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3645 raise_priority = false;
3648 * Do not writeback or swap pages for boosted reclaim. The
3649 * intent is to relieve pressure not issue sub-optimal IO
3650 * from reclaim context. If no pages are reclaimed, the
3651 * reclaim will be aborted.
3653 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3654 sc.may_swap = !nr_boost_reclaim;
3657 * Do some background aging of the anon list, to give
3658 * pages a chance to be referenced before reclaiming. All
3659 * pages are rotated regardless of classzone as this is
3660 * about consistent aging.
3662 age_active_anon(pgdat, &sc);
3665 * If we're getting trouble reclaiming, start doing writepage
3666 * even in laptop mode.
3668 if (sc.priority < DEF_PRIORITY - 2)
3669 sc.may_writepage = 1;
3671 /* Call soft limit reclaim before calling shrink_node. */
3673 nr_soft_scanned = 0;
3674 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3675 sc.gfp_mask, &nr_soft_scanned);
3676 sc.nr_reclaimed += nr_soft_reclaimed;
3679 * There should be no need to raise the scanning priority if
3680 * enough pages are already being scanned that that high
3681 * watermark would be met at 100% efficiency.
3683 if (kswapd_shrink_node(pgdat, &sc))
3684 raise_priority = false;
3687 * If the low watermark is met there is no need for processes
3688 * to be throttled on pfmemalloc_wait as they should not be
3689 * able to safely make forward progress. Wake them
3691 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3692 allow_direct_reclaim(pgdat))
3693 wake_up_all(&pgdat->pfmemalloc_wait);
3695 /* Check if kswapd should be suspending */
3696 __fs_reclaim_release();
3697 ret = try_to_freeze();
3698 __fs_reclaim_acquire();
3699 if (ret || kthread_should_stop())
3703 * Raise priority if scanning rate is too low or there was no
3704 * progress in reclaiming pages
3706 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3707 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3710 * If reclaim made no progress for a boost, stop reclaim as
3711 * IO cannot be queued and it could be an infinite loop in
3712 * extreme circumstances.
3714 if (nr_boost_reclaim && !nr_reclaimed)
3717 if (raise_priority || !nr_reclaimed)
3719 } while (sc.priority >= 1);
3721 if (!sc.nr_reclaimed)
3722 pgdat->kswapd_failures++;
3725 /* If reclaim was boosted, account for the reclaim done in this pass */
3727 unsigned long flags;
3729 for (i = 0; i <= classzone_idx; i++) {
3730 if (!zone_boosts[i])
3733 /* Increments are under the zone lock */
3734 zone = pgdat->node_zones + i;
3735 spin_lock_irqsave(&zone->lock, flags);
3736 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3737 spin_unlock_irqrestore(&zone->lock, flags);
3741 * As there is now likely space, wakeup kcompact to defragment
3744 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3747 snapshot_refaults(NULL, pgdat);
3748 __fs_reclaim_release();
3749 psi_memstall_leave(&pflags);
3750 set_task_reclaim_state(current, NULL);
3753 * Return the order kswapd stopped reclaiming at as
3754 * prepare_kswapd_sleep() takes it into account. If another caller
3755 * entered the allocator slow path while kswapd was awake, order will
3756 * remain at the higher level.
3762 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3763 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3764 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3765 * after previous reclaim attempt (node is still unbalanced). In that case
3766 * return the zone index of the previous kswapd reclaim cycle.
3768 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3769 enum zone_type prev_classzone_idx)
3771 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3772 return prev_classzone_idx;
3773 return pgdat->kswapd_classzone_idx;
3776 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3777 unsigned int classzone_idx)
3782 if (freezing(current) || kthread_should_stop())
3785 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3788 * Try to sleep for a short interval. Note that kcompactd will only be
3789 * woken if it is possible to sleep for a short interval. This is
3790 * deliberate on the assumption that if reclaim cannot keep an
3791 * eligible zone balanced that it's also unlikely that compaction will
3794 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3796 * Compaction records what page blocks it recently failed to
3797 * isolate pages from and skips them in the future scanning.
3798 * When kswapd is going to sleep, it is reasonable to assume
3799 * that pages and compaction may succeed so reset the cache.
3801 reset_isolation_suitable(pgdat);
3804 * We have freed the memory, now we should compact it to make
3805 * allocation of the requested order possible.
3807 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3809 remaining = schedule_timeout(HZ/10);
3812 * If woken prematurely then reset kswapd_classzone_idx and
3813 * order. The values will either be from a wakeup request or
3814 * the previous request that slept prematurely.
3817 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3818 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3821 finish_wait(&pgdat->kswapd_wait, &wait);
3822 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3826 * After a short sleep, check if it was a premature sleep. If not, then
3827 * go fully to sleep until explicitly woken up.
3830 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3831 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3834 * vmstat counters are not perfectly accurate and the estimated
3835 * value for counters such as NR_FREE_PAGES can deviate from the
3836 * true value by nr_online_cpus * threshold. To avoid the zone
3837 * watermarks being breached while under pressure, we reduce the
3838 * per-cpu vmstat threshold while kswapd is awake and restore
3839 * them before going back to sleep.
3841 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3843 if (!kthread_should_stop())
3846 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3849 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3851 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3853 finish_wait(&pgdat->kswapd_wait, &wait);
3857 * The background pageout daemon, started as a kernel thread
3858 * from the init process.
3860 * This basically trickles out pages so that we have _some_
3861 * free memory available even if there is no other activity
3862 * that frees anything up. This is needed for things like routing
3863 * etc, where we otherwise might have all activity going on in
3864 * asynchronous contexts that cannot page things out.
3866 * If there are applications that are active memory-allocators
3867 * (most normal use), this basically shouldn't matter.
3869 static int kswapd(void *p)
3871 unsigned int alloc_order, reclaim_order;
3872 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3873 pg_data_t *pgdat = (pg_data_t*)p;
3874 struct task_struct *tsk = current;
3875 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3877 if (!cpumask_empty(cpumask))
3878 set_cpus_allowed_ptr(tsk, cpumask);
3881 * Tell the memory management that we're a "memory allocator",
3882 * and that if we need more memory we should get access to it
3883 * regardless (see "__alloc_pages()"). "kswapd" should
3884 * never get caught in the normal page freeing logic.
3886 * (Kswapd normally doesn't need memory anyway, but sometimes
3887 * you need a small amount of memory in order to be able to
3888 * page out something else, and this flag essentially protects
3889 * us from recursively trying to free more memory as we're
3890 * trying to free the first piece of memory in the first place).
3892 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3895 pgdat->kswapd_order = 0;
3896 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3900 alloc_order = reclaim_order = pgdat->kswapd_order;
3901 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3904 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3907 /* Read the new order and classzone_idx */
3908 alloc_order = reclaim_order = pgdat->kswapd_order;
3909 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3910 pgdat->kswapd_order = 0;
3911 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3913 ret = try_to_freeze();
3914 if (kthread_should_stop())
3918 * We can speed up thawing tasks if we don't call balance_pgdat
3919 * after returning from the refrigerator
3925 * Reclaim begins at the requested order but if a high-order
3926 * reclaim fails then kswapd falls back to reclaiming for
3927 * order-0. If that happens, kswapd will consider sleeping
3928 * for the order it finished reclaiming at (reclaim_order)
3929 * but kcompactd is woken to compact for the original
3930 * request (alloc_order).
3932 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3934 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3935 if (reclaim_order < alloc_order)
3936 goto kswapd_try_sleep;
3939 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3945 * A zone is low on free memory or too fragmented for high-order memory. If
3946 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3947 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3948 * has failed or is not needed, still wake up kcompactd if only compaction is
3951 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3952 enum zone_type classzone_idx)
3956 if (!managed_zone(zone))
3959 if (!cpuset_zone_allowed(zone, gfp_flags))
3961 pgdat = zone->zone_pgdat;
3963 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3964 pgdat->kswapd_classzone_idx = classzone_idx;
3966 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3968 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3969 if (!waitqueue_active(&pgdat->kswapd_wait))
3972 /* Hopeless node, leave it to direct reclaim if possible */
3973 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3974 (pgdat_balanced(pgdat, order, classzone_idx) &&
3975 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3977 * There may be plenty of free memory available, but it's too
3978 * fragmented for high-order allocations. Wake up kcompactd
3979 * and rely on compaction_suitable() to determine if it's
3980 * needed. If it fails, it will defer subsequent attempts to
3981 * ratelimit its work.
3983 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3984 wakeup_kcompactd(pgdat, order, classzone_idx);
3988 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3990 wake_up_interruptible(&pgdat->kswapd_wait);
3993 #ifdef CONFIG_HIBERNATION
3995 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3998 * Rather than trying to age LRUs the aim is to preserve the overall
3999 * LRU order by reclaiming preferentially
4000 * inactive > active > active referenced > active mapped
4002 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4004 struct scan_control sc = {
4005 .nr_to_reclaim = nr_to_reclaim,
4006 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4007 .reclaim_idx = MAX_NR_ZONES - 1,
4008 .priority = DEF_PRIORITY,
4012 .hibernation_mode = 1,
4014 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4015 unsigned long nr_reclaimed;
4016 unsigned int noreclaim_flag;
4018 fs_reclaim_acquire(sc.gfp_mask);
4019 noreclaim_flag = memalloc_noreclaim_save();
4020 set_task_reclaim_state(current, &sc.reclaim_state);
4022 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4024 set_task_reclaim_state(current, NULL);
4025 memalloc_noreclaim_restore(noreclaim_flag);
4026 fs_reclaim_release(sc.gfp_mask);
4028 return nr_reclaimed;
4030 #endif /* CONFIG_HIBERNATION */
4032 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4033 not required for correctness. So if the last cpu in a node goes
4034 away, we get changed to run anywhere: as the first one comes back,
4035 restore their cpu bindings. */
4036 static int kswapd_cpu_online(unsigned int cpu)
4040 for_each_node_state(nid, N_MEMORY) {
4041 pg_data_t *pgdat = NODE_DATA(nid);
4042 const struct cpumask *mask;
4044 mask = cpumask_of_node(pgdat->node_id);
4046 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4047 /* One of our CPUs online: restore mask */
4048 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4054 * This kswapd start function will be called by init and node-hot-add.
4055 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4057 int kswapd_run(int nid)
4059 pg_data_t *pgdat = NODE_DATA(nid);
4065 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4066 if (IS_ERR(pgdat->kswapd)) {
4067 /* failure at boot is fatal */
4068 BUG_ON(system_state < SYSTEM_RUNNING);
4069 pr_err("Failed to start kswapd on node %d\n", nid);
4070 ret = PTR_ERR(pgdat->kswapd);
4071 pgdat->kswapd = NULL;
4077 * Called by memory hotplug when all memory in a node is offlined. Caller must
4078 * hold mem_hotplug_begin/end().
4080 void kswapd_stop(int nid)
4082 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4085 kthread_stop(kswapd);
4086 NODE_DATA(nid)->kswapd = NULL;
4090 static int __init kswapd_init(void)
4095 for_each_node_state(nid, N_MEMORY)
4097 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4098 "mm/vmscan:online", kswapd_cpu_online,
4104 module_init(kswapd_init)
4110 * If non-zero call node_reclaim when the number of free pages falls below
4113 int node_reclaim_mode __read_mostly;
4115 #define RECLAIM_OFF 0
4116 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4117 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4118 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4121 * Priority for NODE_RECLAIM. This determines the fraction of pages
4122 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4125 #define NODE_RECLAIM_PRIORITY 4
4128 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4131 int sysctl_min_unmapped_ratio = 1;
4134 * If the number of slab pages in a zone grows beyond this percentage then
4135 * slab reclaim needs to occur.
4137 int sysctl_min_slab_ratio = 5;
4139 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4141 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4142 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4143 node_page_state(pgdat, NR_ACTIVE_FILE);
4146 * It's possible for there to be more file mapped pages than
4147 * accounted for by the pages on the file LRU lists because
4148 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4150 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4153 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4154 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4156 unsigned long nr_pagecache_reclaimable;
4157 unsigned long delta = 0;
4160 * If RECLAIM_UNMAP is set, then all file pages are considered
4161 * potentially reclaimable. Otherwise, we have to worry about
4162 * pages like swapcache and node_unmapped_file_pages() provides
4165 if (node_reclaim_mode & RECLAIM_UNMAP)
4166 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4168 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4170 /* If we can't clean pages, remove dirty pages from consideration */
4171 if (!(node_reclaim_mode & RECLAIM_WRITE))
4172 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4174 /* Watch for any possible underflows due to delta */
4175 if (unlikely(delta > nr_pagecache_reclaimable))
4176 delta = nr_pagecache_reclaimable;
4178 return nr_pagecache_reclaimable - delta;
4182 * Try to free up some pages from this node through reclaim.
4184 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4186 /* Minimum pages needed in order to stay on node */
4187 const unsigned long nr_pages = 1 << order;
4188 struct task_struct *p = current;
4189 unsigned int noreclaim_flag;
4190 struct scan_control sc = {
4191 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4192 .gfp_mask = current_gfp_context(gfp_mask),
4194 .priority = NODE_RECLAIM_PRIORITY,
4195 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4196 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4198 .reclaim_idx = gfp_zone(gfp_mask),
4201 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4205 fs_reclaim_acquire(sc.gfp_mask);
4207 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4208 * and we also need to be able to write out pages for RECLAIM_WRITE
4209 * and RECLAIM_UNMAP.
4211 noreclaim_flag = memalloc_noreclaim_save();
4212 p->flags |= PF_SWAPWRITE;
4213 set_task_reclaim_state(p, &sc.reclaim_state);
4215 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4217 * Free memory by calling shrink node with increasing
4218 * priorities until we have enough memory freed.
4221 shrink_node(pgdat, &sc);
4222 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4225 set_task_reclaim_state(p, NULL);
4226 current->flags &= ~PF_SWAPWRITE;
4227 memalloc_noreclaim_restore(noreclaim_flag);
4228 fs_reclaim_release(sc.gfp_mask);
4230 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4232 return sc.nr_reclaimed >= nr_pages;
4235 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4240 * Node reclaim reclaims unmapped file backed pages and
4241 * slab pages if we are over the defined limits.
4243 * A small portion of unmapped file backed pages is needed for
4244 * file I/O otherwise pages read by file I/O will be immediately
4245 * thrown out if the node is overallocated. So we do not reclaim
4246 * if less than a specified percentage of the node is used by
4247 * unmapped file backed pages.
4249 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4250 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4251 return NODE_RECLAIM_FULL;
4254 * Do not scan if the allocation should not be delayed.
4256 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4257 return NODE_RECLAIM_NOSCAN;
4260 * Only run node reclaim on the local node or on nodes that do not
4261 * have associated processors. This will favor the local processor
4262 * over remote processors and spread off node memory allocations
4263 * as wide as possible.
4265 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4266 return NODE_RECLAIM_NOSCAN;
4268 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4269 return NODE_RECLAIM_NOSCAN;
4271 ret = __node_reclaim(pgdat, gfp_mask, order);
4272 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4275 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4282 * page_evictable - test whether a page is evictable
4283 * @page: the page to test
4285 * Test whether page is evictable--i.e., should be placed on active/inactive
4286 * lists vs unevictable list.
4288 * Reasons page might not be evictable:
4289 * (1) page's mapping marked unevictable
4290 * (2) page is part of an mlocked VMA
4293 int page_evictable(struct page *page)
4297 /* Prevent address_space of inode and swap cache from being freed */
4299 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4305 * check_move_unevictable_pages - check pages for evictability and move to
4306 * appropriate zone lru list
4307 * @pvec: pagevec with lru pages to check
4309 * Checks pages for evictability, if an evictable page is in the unevictable
4310 * lru list, moves it to the appropriate evictable lru list. This function
4311 * should be only used for lru pages.
4313 void check_move_unevictable_pages(struct pagevec *pvec)
4315 struct lruvec *lruvec;
4316 struct pglist_data *pgdat = NULL;
4321 for (i = 0; i < pvec->nr; i++) {
4322 struct page *page = pvec->pages[i];
4323 struct pglist_data *pagepgdat = page_pgdat(page);
4326 if (pagepgdat != pgdat) {
4328 spin_unlock_irq(&pgdat->lru_lock);
4330 spin_lock_irq(&pgdat->lru_lock);
4332 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4334 if (!PageLRU(page) || !PageUnevictable(page))
4337 if (page_evictable(page)) {
4338 enum lru_list lru = page_lru_base_type(page);
4340 VM_BUG_ON_PAGE(PageActive(page), page);
4341 ClearPageUnevictable(page);
4342 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4343 add_page_to_lru_list(page, lruvec, lru);
4349 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4350 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4351 spin_unlock_irq(&pgdat->lru_lock);
4354 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);