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 lru_size = 0;
357 if (!mem_cgroup_disabled()) {
358 for (zid = 0; zid < MAX_NR_ZONES; zid++)
359 lru_size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
361 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
363 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
364 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
367 if (!managed_zone(zone))
370 if (!mem_cgroup_disabled())
371 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
373 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
374 NR_ZONE_LRU_BASE + lru);
375 lru_size -= min(size, lru_size);
383 * Add a shrinker callback to be called from the vm.
385 int prealloc_shrinker(struct shrinker *shrinker)
387 unsigned int size = sizeof(*shrinker->nr_deferred);
389 if (shrinker->flags & SHRINKER_NUMA_AWARE)
392 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
393 if (!shrinker->nr_deferred)
396 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
397 if (prealloc_memcg_shrinker(shrinker))
404 kfree(shrinker->nr_deferred);
405 shrinker->nr_deferred = NULL;
409 void free_prealloced_shrinker(struct shrinker *shrinker)
411 if (!shrinker->nr_deferred)
414 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
415 unregister_memcg_shrinker(shrinker);
417 kfree(shrinker->nr_deferred);
418 shrinker->nr_deferred = NULL;
421 void register_shrinker_prepared(struct shrinker *shrinker)
423 down_write(&shrinker_rwsem);
424 list_add_tail(&shrinker->list, &shrinker_list);
425 #ifdef CONFIG_MEMCG_KMEM
426 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
427 idr_replace(&shrinker_idr, shrinker, shrinker->id);
429 up_write(&shrinker_rwsem);
432 int register_shrinker(struct shrinker *shrinker)
434 int err = prealloc_shrinker(shrinker);
438 register_shrinker_prepared(shrinker);
441 EXPORT_SYMBOL(register_shrinker);
446 void unregister_shrinker(struct shrinker *shrinker)
448 if (!shrinker->nr_deferred)
450 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
451 unregister_memcg_shrinker(shrinker);
452 down_write(&shrinker_rwsem);
453 list_del(&shrinker->list);
454 up_write(&shrinker_rwsem);
455 kfree(shrinker->nr_deferred);
456 shrinker->nr_deferred = NULL;
458 EXPORT_SYMBOL(unregister_shrinker);
460 #define SHRINK_BATCH 128
462 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
463 struct shrinker *shrinker, int priority)
465 unsigned long freed = 0;
466 unsigned long long delta;
471 int nid = shrinkctl->nid;
472 long batch_size = shrinker->batch ? shrinker->batch
474 long scanned = 0, next_deferred;
476 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
479 freeable = shrinker->count_objects(shrinker, shrinkctl);
480 if (freeable == 0 || freeable == SHRINK_EMPTY)
484 * copy the current shrinker scan count into a local variable
485 * and zero it so that other concurrent shrinker invocations
486 * don't also do this scanning work.
488 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
491 if (shrinker->seeks) {
492 delta = freeable >> priority;
494 do_div(delta, shrinker->seeks);
497 * These objects don't require any IO to create. Trim
498 * them aggressively under memory pressure to keep
499 * them from causing refetches in the IO caches.
501 delta = freeable / 2;
505 if (total_scan < 0) {
506 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
507 shrinker->scan_objects, total_scan);
508 total_scan = freeable;
511 next_deferred = total_scan;
514 * We need to avoid excessive windup on filesystem shrinkers
515 * due to large numbers of GFP_NOFS allocations causing the
516 * shrinkers to return -1 all the time. This results in a large
517 * nr being built up so when a shrink that can do some work
518 * comes along it empties the entire cache due to nr >>>
519 * freeable. This is bad for sustaining a working set in
522 * Hence only allow the shrinker to scan the entire cache when
523 * a large delta change is calculated directly.
525 if (delta < freeable / 4)
526 total_scan = min(total_scan, freeable / 2);
529 * Avoid risking looping forever due to too large nr value:
530 * never try to free more than twice the estimate number of
533 if (total_scan > freeable * 2)
534 total_scan = freeable * 2;
536 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
537 freeable, delta, total_scan, priority);
540 * Normally, we should not scan less than batch_size objects in one
541 * pass to avoid too frequent shrinker calls, but if the slab has less
542 * than batch_size objects in total and we are really tight on memory,
543 * we will try to reclaim all available objects, otherwise we can end
544 * up failing allocations although there are plenty of reclaimable
545 * objects spread over several slabs with usage less than the
548 * We detect the "tight on memory" situations by looking at the total
549 * number of objects we want to scan (total_scan). If it is greater
550 * than the total number of objects on slab (freeable), we must be
551 * scanning at high prio and therefore should try to reclaim as much as
554 while (total_scan >= batch_size ||
555 total_scan >= freeable) {
557 unsigned long nr_to_scan = min(batch_size, total_scan);
559 shrinkctl->nr_to_scan = nr_to_scan;
560 shrinkctl->nr_scanned = nr_to_scan;
561 ret = shrinker->scan_objects(shrinker, shrinkctl);
562 if (ret == SHRINK_STOP)
566 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
567 total_scan -= shrinkctl->nr_scanned;
568 scanned += shrinkctl->nr_scanned;
573 if (next_deferred >= scanned)
574 next_deferred -= scanned;
578 * move the unused scan count back into the shrinker in a
579 * manner that handles concurrent updates. If we exhausted the
580 * scan, there is no need to do an update.
582 if (next_deferred > 0)
583 new_nr = atomic_long_add_return(next_deferred,
584 &shrinker->nr_deferred[nid]);
586 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
588 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
593 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
594 struct mem_cgroup *memcg, int priority)
596 struct memcg_shrinker_map *map;
597 unsigned long ret, freed = 0;
600 if (!mem_cgroup_online(memcg))
603 if (!down_read_trylock(&shrinker_rwsem))
606 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
611 for_each_set_bit(i, map->map, shrinker_nr_max) {
612 struct shrink_control sc = {
613 .gfp_mask = gfp_mask,
617 struct shrinker *shrinker;
619 shrinker = idr_find(&shrinker_idr, i);
620 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
622 clear_bit(i, map->map);
626 /* Call non-slab shrinkers even though kmem is disabled */
627 if (!memcg_kmem_enabled() &&
628 !(shrinker->flags & SHRINKER_NONSLAB))
631 ret = do_shrink_slab(&sc, shrinker, priority);
632 if (ret == SHRINK_EMPTY) {
633 clear_bit(i, map->map);
635 * After the shrinker reported that it had no objects to
636 * free, but before we cleared the corresponding bit in
637 * the memcg shrinker map, a new object might have been
638 * added. To make sure, we have the bit set in this
639 * case, we invoke the shrinker one more time and reset
640 * the bit if it reports that it is not empty anymore.
641 * The memory barrier here pairs with the barrier in
642 * memcg_set_shrinker_bit():
644 * list_lru_add() shrink_slab_memcg()
645 * list_add_tail() clear_bit()
647 * set_bit() do_shrink_slab()
649 smp_mb__after_atomic();
650 ret = do_shrink_slab(&sc, shrinker, priority);
651 if (ret == SHRINK_EMPTY)
654 memcg_set_shrinker_bit(memcg, nid, i);
658 if (rwsem_is_contended(&shrinker_rwsem)) {
664 up_read(&shrinker_rwsem);
667 #else /* CONFIG_MEMCG */
668 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
669 struct mem_cgroup *memcg, int priority)
673 #endif /* CONFIG_MEMCG */
676 * shrink_slab - shrink slab caches
677 * @gfp_mask: allocation context
678 * @nid: node whose slab caches to target
679 * @memcg: memory cgroup whose slab caches to target
680 * @priority: the reclaim priority
682 * Call the shrink functions to age shrinkable caches.
684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 * unaware shrinkers will receive a node id of 0 instead.
687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 * are called only if it is the root cgroup.
690 * @priority is sc->priority, we take the number of objects and >> by priority
691 * in order to get the scan target.
693 * Returns the number of reclaimed slab objects.
695 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
696 struct mem_cgroup *memcg,
699 unsigned long ret, freed = 0;
700 struct shrinker *shrinker;
703 * The root memcg might be allocated even though memcg is disabled
704 * via "cgroup_disable=memory" boot parameter. This could make
705 * mem_cgroup_is_root() return false, then just run memcg slab
706 * shrink, but skip global shrink. This may result in premature
709 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
710 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
712 if (!down_read_trylock(&shrinker_rwsem))
715 list_for_each_entry(shrinker, &shrinker_list, list) {
716 struct shrink_control sc = {
717 .gfp_mask = gfp_mask,
722 ret = do_shrink_slab(&sc, shrinker, priority);
723 if (ret == SHRINK_EMPTY)
727 * Bail out if someone want to register a new shrinker to
728 * prevent the regsitration from being stalled for long periods
729 * by parallel ongoing shrinking.
731 if (rwsem_is_contended(&shrinker_rwsem)) {
737 up_read(&shrinker_rwsem);
743 void drop_slab_node(int nid)
748 struct mem_cgroup *memcg = NULL;
751 memcg = mem_cgroup_iter(NULL, NULL, NULL);
753 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
754 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
755 } while (freed > 10);
762 for_each_online_node(nid)
766 static inline int is_page_cache_freeable(struct page *page)
769 * A freeable page cache page is referenced only by the caller
770 * that isolated the page, the page cache and optional buffer
771 * heads at page->private.
773 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
775 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
778 static int may_write_to_inode(struct inode *inode)
780 if (current->flags & PF_SWAPWRITE)
782 if (!inode_write_congested(inode))
784 if (inode_to_bdi(inode) == current->backing_dev_info)
790 * We detected a synchronous write error writing a page out. Probably
791 * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 * fsync(), msync() or close().
794 * The tricky part is that after writepage we cannot touch the mapping: nothing
795 * prevents it from being freed up. But we have a ref on the page and once
796 * that page is locked, the mapping is pinned.
798 * We're allowed to run sleeping lock_page() here because we know the caller has
801 static void handle_write_error(struct address_space *mapping,
802 struct page *page, int error)
805 if (page_mapping(page) == mapping)
806 mapping_set_error(mapping, error);
810 /* possible outcome of pageout() */
812 /* failed to write page out, page is locked */
814 /* move page to the active list, page is locked */
816 /* page has been sent to the disk successfully, page is unlocked */
818 /* page is clean and locked */
823 * pageout is called by shrink_page_list() for each dirty page.
824 * Calls ->writepage().
826 static pageout_t pageout(struct page *page, struct address_space *mapping)
829 * If the page is dirty, only perform writeback if that write
830 * will be non-blocking. To prevent this allocation from being
831 * stalled by pagecache activity. But note that there may be
832 * stalls if we need to run get_block(). We could test
833 * PagePrivate for that.
835 * If this process is currently in __generic_file_write_iter() against
836 * this page's queue, we can perform writeback even if that
839 * If the page is swapcache, write it back even if that would
840 * block, for some throttling. This happens by accident, because
841 * swap_backing_dev_info is bust: it doesn't reflect the
842 * congestion state of the swapdevs. Easy to fix, if needed.
844 if (!is_page_cache_freeable(page))
848 * Some data journaling orphaned pages can have
849 * page->mapping == NULL while being dirty with clean buffers.
851 if (page_has_private(page)) {
852 if (try_to_free_buffers(page)) {
853 ClearPageDirty(page);
854 pr_info("%s: orphaned page\n", __func__);
860 if (mapping->a_ops->writepage == NULL)
861 return PAGE_ACTIVATE;
862 if (!may_write_to_inode(mapping->host))
865 if (clear_page_dirty_for_io(page)) {
867 struct writeback_control wbc = {
868 .sync_mode = WB_SYNC_NONE,
869 .nr_to_write = SWAP_CLUSTER_MAX,
871 .range_end = LLONG_MAX,
875 SetPageReclaim(page);
876 res = mapping->a_ops->writepage(page, &wbc);
878 handle_write_error(mapping, page, res);
879 if (res == AOP_WRITEPAGE_ACTIVATE) {
880 ClearPageReclaim(page);
881 return PAGE_ACTIVATE;
884 if (!PageWriteback(page)) {
885 /* synchronous write or broken a_ops? */
886 ClearPageReclaim(page);
888 trace_mm_vmscan_writepage(page);
889 inc_node_page_state(page, NR_VMSCAN_WRITE);
897 * Same as remove_mapping, but if the page is removed from the mapping, it
898 * gets returned with a refcount of 0.
900 static int __remove_mapping(struct address_space *mapping, struct page *page,
906 BUG_ON(!PageLocked(page));
907 BUG_ON(mapping != page_mapping(page));
909 xa_lock_irqsave(&mapping->i_pages, flags);
911 * The non racy check for a busy page.
913 * Must be careful with the order of the tests. When someone has
914 * a ref to the page, it may be possible that they dirty it then
915 * drop the reference. So if PageDirty is tested before page_count
916 * here, then the following race may occur:
918 * get_user_pages(&page);
919 * [user mapping goes away]
921 * !PageDirty(page) [good]
922 * SetPageDirty(page);
924 * !page_count(page) [good, discard it]
926 * [oops, our write_to data is lost]
928 * Reversing the order of the tests ensures such a situation cannot
929 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
930 * load is not satisfied before that of page->_refcount.
932 * Note that if SetPageDirty is always performed via set_page_dirty,
933 * and thus under the i_pages lock, then this ordering is not required.
935 refcount = 1 + compound_nr(page);
936 if (!page_ref_freeze(page, refcount))
938 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
939 if (unlikely(PageDirty(page))) {
940 page_ref_unfreeze(page, refcount);
944 if (PageSwapCache(page)) {
945 swp_entry_t swap = { .val = page_private(page) };
946 mem_cgroup_swapout(page, swap);
947 __delete_from_swap_cache(page, swap);
948 xa_unlock_irqrestore(&mapping->i_pages, flags);
949 put_swap_page(page, swap);
951 void (*freepage)(struct page *);
954 freepage = mapping->a_ops->freepage;
956 * Remember a shadow entry for reclaimed file cache in
957 * order to detect refaults, thus thrashing, later on.
959 * But don't store shadows in an address space that is
960 * already exiting. This is not just an optizimation,
961 * inode reclaim needs to empty out the radix tree or
962 * the nodes are lost. Don't plant shadows behind its
965 * We also don't store shadows for DAX mappings because the
966 * only page cache pages found in these are zero pages
967 * covering holes, and because we don't want to mix DAX
968 * exceptional entries and shadow exceptional entries in the
969 * same address_space.
971 if (reclaimed && page_is_file_cache(page) &&
972 !mapping_exiting(mapping) && !dax_mapping(mapping))
973 shadow = workingset_eviction(page);
974 __delete_from_page_cache(page, shadow);
975 xa_unlock_irqrestore(&mapping->i_pages, flags);
977 if (freepage != NULL)
984 xa_unlock_irqrestore(&mapping->i_pages, flags);
989 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
990 * someone else has a ref on the page, abort and return 0. If it was
991 * successfully detached, return 1. Assumes the caller has a single ref on
994 int remove_mapping(struct address_space *mapping, struct page *page)
996 if (__remove_mapping(mapping, page, false)) {
998 * Unfreezing the refcount with 1 rather than 2 effectively
999 * drops the pagecache ref for us without requiring another
1002 page_ref_unfreeze(page, 1);
1009 * putback_lru_page - put previously isolated page onto appropriate LRU list
1010 * @page: page to be put back to appropriate lru list
1012 * Add previously isolated @page to appropriate LRU list.
1013 * Page may still be unevictable for other reasons.
1015 * lru_lock must not be held, interrupts must be enabled.
1017 void putback_lru_page(struct page *page)
1019 lru_cache_add(page);
1020 put_page(page); /* drop ref from isolate */
1023 enum page_references {
1025 PAGEREF_RECLAIM_CLEAN,
1030 static enum page_references page_check_references(struct page *page,
1031 struct scan_control *sc)
1033 int referenced_ptes, referenced_page;
1034 unsigned long vm_flags;
1036 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1038 referenced_page = TestClearPageReferenced(page);
1041 * Mlock lost the isolation race with us. Let try_to_unmap()
1042 * move the page to the unevictable list.
1044 if (vm_flags & VM_LOCKED)
1045 return PAGEREF_RECLAIM;
1047 if (referenced_ptes) {
1048 if (PageSwapBacked(page))
1049 return PAGEREF_ACTIVATE;
1051 * All mapped pages start out with page table
1052 * references from the instantiating fault, so we need
1053 * to look twice if a mapped file page is used more
1056 * Mark it and spare it for another trip around the
1057 * inactive list. Another page table reference will
1058 * lead to its activation.
1060 * Note: the mark is set for activated pages as well
1061 * so that recently deactivated but used pages are
1062 * quickly recovered.
1064 SetPageReferenced(page);
1066 if (referenced_page || referenced_ptes > 1)
1067 return PAGEREF_ACTIVATE;
1070 * Activate file-backed executable pages after first usage.
1072 if (vm_flags & VM_EXEC)
1073 return PAGEREF_ACTIVATE;
1075 return PAGEREF_KEEP;
1078 /* Reclaim if clean, defer dirty pages to writeback */
1079 if (referenced_page && !PageSwapBacked(page))
1080 return PAGEREF_RECLAIM_CLEAN;
1082 return PAGEREF_RECLAIM;
1085 /* Check if a page is dirty or under writeback */
1086 static void page_check_dirty_writeback(struct page *page,
1087 bool *dirty, bool *writeback)
1089 struct address_space *mapping;
1092 * Anonymous pages are not handled by flushers and must be written
1093 * from reclaim context. Do not stall reclaim based on them
1095 if (!page_is_file_cache(page) ||
1096 (PageAnon(page) && !PageSwapBacked(page))) {
1102 /* By default assume that the page flags are accurate */
1103 *dirty = PageDirty(page);
1104 *writeback = PageWriteback(page);
1106 /* Verify dirty/writeback state if the filesystem supports it */
1107 if (!page_has_private(page))
1110 mapping = page_mapping(page);
1111 if (mapping && mapping->a_ops->is_dirty_writeback)
1112 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1116 * shrink_page_list() returns the number of reclaimed pages
1118 static unsigned long shrink_page_list(struct list_head *page_list,
1119 struct pglist_data *pgdat,
1120 struct scan_control *sc,
1121 enum ttu_flags ttu_flags,
1122 struct reclaim_stat *stat,
1123 bool ignore_references)
1125 LIST_HEAD(ret_pages);
1126 LIST_HEAD(free_pages);
1127 unsigned nr_reclaimed = 0;
1128 unsigned pgactivate = 0;
1130 memset(stat, 0, sizeof(*stat));
1133 while (!list_empty(page_list)) {
1134 struct address_space *mapping;
1137 enum page_references references = PAGEREF_RECLAIM;
1138 bool dirty, writeback;
1139 unsigned int nr_pages;
1143 page = lru_to_page(page_list);
1144 list_del(&page->lru);
1146 if (!trylock_page(page))
1149 VM_BUG_ON_PAGE(PageActive(page), page);
1151 nr_pages = compound_nr(page);
1153 /* Account the number of base pages even though THP */
1154 sc->nr_scanned += nr_pages;
1156 if (unlikely(!page_evictable(page)))
1157 goto activate_locked;
1159 if (!sc->may_unmap && page_mapped(page))
1162 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1163 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1166 * The number of dirty pages determines if a node is marked
1167 * reclaim_congested which affects wait_iff_congested. kswapd
1168 * will stall and start writing pages if the tail of the LRU
1169 * is all dirty unqueued pages.
1171 page_check_dirty_writeback(page, &dirty, &writeback);
1172 if (dirty || writeback)
1175 if (dirty && !writeback)
1176 stat->nr_unqueued_dirty++;
1179 * Treat this page as congested if the underlying BDI is or if
1180 * pages are cycling through the LRU so quickly that the
1181 * pages marked for immediate reclaim are making it to the
1182 * end of the LRU a second time.
1184 mapping = page_mapping(page);
1185 if (((dirty || writeback) && mapping &&
1186 inode_write_congested(mapping->host)) ||
1187 (writeback && PageReclaim(page)))
1188 stat->nr_congested++;
1191 * If a page at the tail of the LRU is under writeback, there
1192 * are three cases to consider.
1194 * 1) If reclaim is encountering an excessive number of pages
1195 * under writeback and this page is both under writeback and
1196 * PageReclaim then it indicates that pages are being queued
1197 * for IO but are being recycled through the LRU before the
1198 * IO can complete. Waiting on the page itself risks an
1199 * indefinite stall if it is impossible to writeback the
1200 * page due to IO error or disconnected storage so instead
1201 * note that the LRU is being scanned too quickly and the
1202 * caller can stall after page list has been processed.
1204 * 2) Global or new memcg reclaim encounters a page that is
1205 * not marked for immediate reclaim, or the caller does not
1206 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1207 * not to fs). In this case mark the page for immediate
1208 * reclaim and continue scanning.
1210 * Require may_enter_fs because we would wait on fs, which
1211 * may not have submitted IO yet. And the loop driver might
1212 * enter reclaim, and deadlock if it waits on a page for
1213 * which it is needed to do the write (loop masks off
1214 * __GFP_IO|__GFP_FS for this reason); but more thought
1215 * would probably show more reasons.
1217 * 3) Legacy memcg encounters a page that is already marked
1218 * PageReclaim. memcg does not have any dirty pages
1219 * throttling so we could easily OOM just because too many
1220 * pages are in writeback and there is nothing else to
1221 * reclaim. Wait for the writeback to complete.
1223 * In cases 1) and 2) we activate the pages to get them out of
1224 * the way while we continue scanning for clean pages on the
1225 * inactive list and refilling from the active list. The
1226 * observation here is that waiting for disk writes is more
1227 * expensive than potentially causing reloads down the line.
1228 * Since they're marked for immediate reclaim, they won't put
1229 * memory pressure on the cache working set any longer than it
1230 * takes to write them to disk.
1232 if (PageWriteback(page)) {
1234 if (current_is_kswapd() &&
1235 PageReclaim(page) &&
1236 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1237 stat->nr_immediate++;
1238 goto activate_locked;
1241 } else if (sane_reclaim(sc) ||
1242 !PageReclaim(page) || !may_enter_fs) {
1244 * This is slightly racy - end_page_writeback()
1245 * might have just cleared PageReclaim, then
1246 * setting PageReclaim here end up interpreted
1247 * as PageReadahead - but that does not matter
1248 * enough to care. What we do want is for this
1249 * page to have PageReclaim set next time memcg
1250 * reclaim reaches the tests above, so it will
1251 * then wait_on_page_writeback() to avoid OOM;
1252 * and it's also appropriate in global reclaim.
1254 SetPageReclaim(page);
1255 stat->nr_writeback++;
1256 goto activate_locked;
1261 wait_on_page_writeback(page);
1262 /* then go back and try same page again */
1263 list_add_tail(&page->lru, page_list);
1268 if (!ignore_references)
1269 references = page_check_references(page, sc);
1271 switch (references) {
1272 case PAGEREF_ACTIVATE:
1273 goto activate_locked;
1275 stat->nr_ref_keep += nr_pages;
1277 case PAGEREF_RECLAIM:
1278 case PAGEREF_RECLAIM_CLEAN:
1279 ; /* try to reclaim the page below */
1283 * Anonymous process memory has backing store?
1284 * Try to allocate it some swap space here.
1285 * Lazyfree page could be freed directly
1287 if (PageAnon(page) && PageSwapBacked(page)) {
1288 if (!PageSwapCache(page)) {
1289 if (!(sc->gfp_mask & __GFP_IO))
1291 if (PageTransHuge(page)) {
1292 /* cannot split THP, skip it */
1293 if (!can_split_huge_page(page, NULL))
1294 goto activate_locked;
1296 * Split pages without a PMD map right
1297 * away. Chances are some or all of the
1298 * tail pages can be freed without IO.
1300 if (!compound_mapcount(page) &&
1301 split_huge_page_to_list(page,
1303 goto activate_locked;
1305 if (!add_to_swap(page)) {
1306 if (!PageTransHuge(page))
1307 goto activate_locked_split;
1308 /* Fallback to swap normal pages */
1309 if (split_huge_page_to_list(page,
1311 goto activate_locked;
1312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1313 count_vm_event(THP_SWPOUT_FALLBACK);
1315 if (!add_to_swap(page))
1316 goto activate_locked_split;
1321 /* Adding to swap updated mapping */
1322 mapping = page_mapping(page);
1324 } else if (unlikely(PageTransHuge(page))) {
1325 /* Split file THP */
1326 if (split_huge_page_to_list(page, page_list))
1331 * THP may get split above, need minus tail pages and update
1332 * nr_pages to avoid accounting tail pages twice.
1334 * The tail pages that are added into swap cache successfully
1337 if ((nr_pages > 1) && !PageTransHuge(page)) {
1338 sc->nr_scanned -= (nr_pages - 1);
1343 * The page is mapped into the page tables of one or more
1344 * processes. Try to unmap it here.
1346 if (page_mapped(page)) {
1347 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1349 if (unlikely(PageTransHuge(page)))
1350 flags |= TTU_SPLIT_HUGE_PMD;
1351 if (!try_to_unmap(page, flags)) {
1352 stat->nr_unmap_fail += nr_pages;
1353 goto activate_locked;
1357 if (PageDirty(page)) {
1359 * Only kswapd can writeback filesystem pages
1360 * to avoid risk of stack overflow. But avoid
1361 * injecting inefficient single-page IO into
1362 * flusher writeback as much as possible: only
1363 * write pages when we've encountered many
1364 * dirty pages, and when we've already scanned
1365 * the rest of the LRU for clean pages and see
1366 * the same dirty pages again (PageReclaim).
1368 if (page_is_file_cache(page) &&
1369 (!current_is_kswapd() || !PageReclaim(page) ||
1370 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1372 * Immediately reclaim when written back.
1373 * Similar in principal to deactivate_page()
1374 * except we already have the page isolated
1375 * and know it's dirty
1377 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1378 SetPageReclaim(page);
1380 goto activate_locked;
1383 if (references == PAGEREF_RECLAIM_CLEAN)
1387 if (!sc->may_writepage)
1391 * Page is dirty. Flush the TLB if a writable entry
1392 * potentially exists to avoid CPU writes after IO
1393 * starts and then write it out here.
1395 try_to_unmap_flush_dirty();
1396 switch (pageout(page, mapping)) {
1400 goto activate_locked;
1402 if (PageWriteback(page))
1404 if (PageDirty(page))
1408 * A synchronous write - probably a ramdisk. Go
1409 * ahead and try to reclaim the page.
1411 if (!trylock_page(page))
1413 if (PageDirty(page) || PageWriteback(page))
1415 mapping = page_mapping(page);
1417 ; /* try to free the page below */
1422 * If the page has buffers, try to free the buffer mappings
1423 * associated with this page. If we succeed we try to free
1426 * We do this even if the page is PageDirty().
1427 * try_to_release_page() does not perform I/O, but it is
1428 * possible for a page to have PageDirty set, but it is actually
1429 * clean (all its buffers are clean). This happens if the
1430 * buffers were written out directly, with submit_bh(). ext3
1431 * will do this, as well as the blockdev mapping.
1432 * try_to_release_page() will discover that cleanness and will
1433 * drop the buffers and mark the page clean - it can be freed.
1435 * Rarely, pages can have buffers and no ->mapping. These are
1436 * the pages which were not successfully invalidated in
1437 * truncate_complete_page(). We try to drop those buffers here
1438 * and if that worked, and the page is no longer mapped into
1439 * process address space (page_count == 1) it can be freed.
1440 * Otherwise, leave the page on the LRU so it is swappable.
1442 if (page_has_private(page)) {
1443 if (!try_to_release_page(page, sc->gfp_mask))
1444 goto activate_locked;
1445 if (!mapping && page_count(page) == 1) {
1447 if (put_page_testzero(page))
1451 * rare race with speculative reference.
1452 * the speculative reference will free
1453 * this page shortly, so we may
1454 * increment nr_reclaimed here (and
1455 * leave it off the LRU).
1463 if (PageAnon(page) && !PageSwapBacked(page)) {
1464 /* follow __remove_mapping for reference */
1465 if (!page_ref_freeze(page, 1))
1467 if (PageDirty(page)) {
1468 page_ref_unfreeze(page, 1);
1472 count_vm_event(PGLAZYFREED);
1473 count_memcg_page_event(page, PGLAZYFREED);
1474 } else if (!mapping || !__remove_mapping(mapping, page, true))
1480 * THP may get swapped out in a whole, need account
1483 nr_reclaimed += nr_pages;
1486 * Is there need to periodically free_page_list? It would
1487 * appear not as the counts should be low
1489 if (unlikely(PageTransHuge(page)))
1490 (*get_compound_page_dtor(page))(page);
1492 list_add(&page->lru, &free_pages);
1495 activate_locked_split:
1497 * The tail pages that are failed to add into swap cache
1498 * reach here. Fixup nr_scanned and nr_pages.
1501 sc->nr_scanned -= (nr_pages - 1);
1505 /* Not a candidate for swapping, so reclaim swap space. */
1506 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1508 try_to_free_swap(page);
1509 VM_BUG_ON_PAGE(PageActive(page), page);
1510 if (!PageMlocked(page)) {
1511 int type = page_is_file_cache(page);
1512 SetPageActive(page);
1513 stat->nr_activate[type] += nr_pages;
1514 count_memcg_page_event(page, PGACTIVATE);
1519 list_add(&page->lru, &ret_pages);
1520 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1523 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1525 mem_cgroup_uncharge_list(&free_pages);
1526 try_to_unmap_flush();
1527 free_unref_page_list(&free_pages);
1529 list_splice(&ret_pages, page_list);
1530 count_vm_events(PGACTIVATE, pgactivate);
1532 return nr_reclaimed;
1535 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1536 struct list_head *page_list)
1538 struct scan_control sc = {
1539 .gfp_mask = GFP_KERNEL,
1540 .priority = DEF_PRIORITY,
1543 struct reclaim_stat dummy_stat;
1545 struct page *page, *next;
1546 LIST_HEAD(clean_pages);
1548 list_for_each_entry_safe(page, next, page_list, lru) {
1549 if (page_is_file_cache(page) && !PageDirty(page) &&
1550 !__PageMovable(page) && !PageUnevictable(page)) {
1551 ClearPageActive(page);
1552 list_move(&page->lru, &clean_pages);
1556 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1557 TTU_IGNORE_ACCESS, &dummy_stat, true);
1558 list_splice(&clean_pages, page_list);
1559 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1564 * Attempt to remove the specified page from its LRU. Only take this page
1565 * if it is of the appropriate PageActive status. Pages which are being
1566 * freed elsewhere are also ignored.
1568 * page: page to consider
1569 * mode: one of the LRU isolation modes defined above
1571 * returns 0 on success, -ve errno on failure.
1573 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1577 /* Only take pages on the LRU. */
1581 /* Compaction should not handle unevictable pages but CMA can do so */
1582 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1588 * To minimise LRU disruption, the caller can indicate that it only
1589 * wants to isolate pages it will be able to operate on without
1590 * blocking - clean pages for the most part.
1592 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1593 * that it is possible to migrate without blocking
1595 if (mode & ISOLATE_ASYNC_MIGRATE) {
1596 /* All the caller can do on PageWriteback is block */
1597 if (PageWriteback(page))
1600 if (PageDirty(page)) {
1601 struct address_space *mapping;
1605 * Only pages without mappings or that have a
1606 * ->migratepage callback are possible to migrate
1607 * without blocking. However, we can be racing with
1608 * truncation so it's necessary to lock the page
1609 * to stabilise the mapping as truncation holds
1610 * the page lock until after the page is removed
1611 * from the page cache.
1613 if (!trylock_page(page))
1616 mapping = page_mapping(page);
1617 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1624 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1627 if (likely(get_page_unless_zero(page))) {
1629 * Be careful not to clear PageLRU until after we're
1630 * sure the page is not being freed elsewhere -- the
1631 * page release code relies on it.
1642 * Update LRU sizes after isolating pages. The LRU size updates must
1643 * be complete before mem_cgroup_update_lru_size due to a santity check.
1645 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1646 enum lru_list lru, unsigned long *nr_zone_taken)
1650 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1651 if (!nr_zone_taken[zid])
1654 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1656 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1663 * pgdat->lru_lock is heavily contended. Some of the functions that
1664 * shrink the lists perform better by taking out a batch of pages
1665 * and working on them outside the LRU lock.
1667 * For pagecache intensive workloads, this function is the hottest
1668 * spot in the kernel (apart from copy_*_user functions).
1670 * Appropriate locks must be held before calling this function.
1672 * @nr_to_scan: The number of eligible pages to look through on the list.
1673 * @lruvec: The LRU vector to pull pages from.
1674 * @dst: The temp list to put pages on to.
1675 * @nr_scanned: The number of pages that were scanned.
1676 * @sc: The scan_control struct for this reclaim session
1677 * @mode: One of the LRU isolation modes
1678 * @lru: LRU list id for isolating
1680 * returns how many pages were moved onto *@dst.
1682 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1683 struct lruvec *lruvec, struct list_head *dst,
1684 unsigned long *nr_scanned, struct scan_control *sc,
1687 struct list_head *src = &lruvec->lists[lru];
1688 unsigned long nr_taken = 0;
1689 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1690 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1691 unsigned long skipped = 0;
1692 unsigned long scan, total_scan, nr_pages;
1693 LIST_HEAD(pages_skipped);
1694 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1698 while (scan < nr_to_scan && !list_empty(src)) {
1701 page = lru_to_page(src);
1702 prefetchw_prev_lru_page(page, src, flags);
1704 VM_BUG_ON_PAGE(!PageLRU(page), page);
1706 nr_pages = compound_nr(page);
1707 total_scan += nr_pages;
1709 if (page_zonenum(page) > sc->reclaim_idx) {
1710 list_move(&page->lru, &pages_skipped);
1711 nr_skipped[page_zonenum(page)] += nr_pages;
1716 * Do not count skipped pages because that makes the function
1717 * return with no isolated pages if the LRU mostly contains
1718 * ineligible pages. This causes the VM to not reclaim any
1719 * pages, triggering a premature OOM.
1721 * Account all tail pages of THP. This would not cause
1722 * premature OOM since __isolate_lru_page() returns -EBUSY
1723 * only when the page is being freed somewhere else.
1726 switch (__isolate_lru_page(page, mode)) {
1728 nr_taken += nr_pages;
1729 nr_zone_taken[page_zonenum(page)] += nr_pages;
1730 list_move(&page->lru, dst);
1734 /* else it is being freed elsewhere */
1735 list_move(&page->lru, src);
1744 * Splice any skipped pages to the start of the LRU list. Note that
1745 * this disrupts the LRU order when reclaiming for lower zones but
1746 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1747 * scanning would soon rescan the same pages to skip and put the
1748 * system at risk of premature OOM.
1750 if (!list_empty(&pages_skipped)) {
1753 list_splice(&pages_skipped, src);
1754 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1755 if (!nr_skipped[zid])
1758 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1759 skipped += nr_skipped[zid];
1762 *nr_scanned = total_scan;
1763 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1764 total_scan, skipped, nr_taken, mode, lru);
1765 update_lru_sizes(lruvec, lru, nr_zone_taken);
1770 * isolate_lru_page - tries to isolate a page from its LRU list
1771 * @page: page to isolate from its LRU list
1773 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1774 * vmstat statistic corresponding to whatever LRU list the page was on.
1776 * Returns 0 if the page was removed from an LRU list.
1777 * Returns -EBUSY if the page was not on an LRU list.
1779 * The returned page will have PageLRU() cleared. If it was found on
1780 * the active list, it will have PageActive set. If it was found on
1781 * the unevictable list, it will have the PageUnevictable bit set. That flag
1782 * may need to be cleared by the caller before letting the page go.
1784 * The vmstat statistic corresponding to the list on which the page was
1785 * found will be decremented.
1789 * (1) Must be called with an elevated refcount on the page. This is a
1790 * fundamentnal difference from isolate_lru_pages (which is called
1791 * without a stable reference).
1792 * (2) the lru_lock must not be held.
1793 * (3) interrupts must be enabled.
1795 int isolate_lru_page(struct page *page)
1799 VM_BUG_ON_PAGE(!page_count(page), page);
1800 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1802 if (PageLRU(page)) {
1803 pg_data_t *pgdat = page_pgdat(page);
1804 struct lruvec *lruvec;
1806 spin_lock_irq(&pgdat->lru_lock);
1807 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1808 if (PageLRU(page)) {
1809 int lru = page_lru(page);
1812 del_page_from_lru_list(page, lruvec, lru);
1815 spin_unlock_irq(&pgdat->lru_lock);
1821 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1822 * then get resheduled. When there are massive number of tasks doing page
1823 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1824 * the LRU list will go small and be scanned faster than necessary, leading to
1825 * unnecessary swapping, thrashing and OOM.
1827 static int too_many_isolated(struct pglist_data *pgdat, int file,
1828 struct scan_control *sc)
1830 unsigned long inactive, isolated;
1832 if (current_is_kswapd())
1835 if (!sane_reclaim(sc))
1839 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1840 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1842 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1843 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1847 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1848 * won't get blocked by normal direct-reclaimers, forming a circular
1851 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1854 return isolated > inactive;
1858 * This moves pages from @list to corresponding LRU list.
1860 * We move them the other way if the page is referenced by one or more
1861 * processes, from rmap.
1863 * If the pages are mostly unmapped, the processing is fast and it is
1864 * appropriate to hold zone_lru_lock across the whole operation. But if
1865 * the pages are mapped, the processing is slow (page_referenced()) so we
1866 * should drop zone_lru_lock around each page. It's impossible to balance
1867 * this, so instead we remove the pages from the LRU while processing them.
1868 * It is safe to rely on PG_active against the non-LRU pages in here because
1869 * nobody will play with that bit on a non-LRU page.
1871 * The downside is that we have to touch page->_refcount against each page.
1872 * But we had to alter page->flags anyway.
1874 * Returns the number of pages moved to the given lruvec.
1877 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1878 struct list_head *list)
1880 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1881 int nr_pages, nr_moved = 0;
1882 LIST_HEAD(pages_to_free);
1886 while (!list_empty(list)) {
1887 page = lru_to_page(list);
1888 VM_BUG_ON_PAGE(PageLRU(page), page);
1889 if (unlikely(!page_evictable(page))) {
1890 list_del(&page->lru);
1891 spin_unlock_irq(&pgdat->lru_lock);
1892 putback_lru_page(page);
1893 spin_lock_irq(&pgdat->lru_lock);
1896 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1899 lru = page_lru(page);
1901 nr_pages = hpage_nr_pages(page);
1902 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1903 list_move(&page->lru, &lruvec->lists[lru]);
1905 if (put_page_testzero(page)) {
1906 __ClearPageLRU(page);
1907 __ClearPageActive(page);
1908 del_page_from_lru_list(page, lruvec, lru);
1910 if (unlikely(PageCompound(page))) {
1911 spin_unlock_irq(&pgdat->lru_lock);
1912 (*get_compound_page_dtor(page))(page);
1913 spin_lock_irq(&pgdat->lru_lock);
1915 list_add(&page->lru, &pages_to_free);
1917 nr_moved += nr_pages;
1922 * To save our caller's stack, now use input list for pages to free.
1924 list_splice(&pages_to_free, list);
1930 * If a kernel thread (such as nfsd for loop-back mounts) services
1931 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1932 * In that case we should only throttle if the backing device it is
1933 * writing to is congested. In other cases it is safe to throttle.
1935 static int current_may_throttle(void)
1937 return !(current->flags & PF_LESS_THROTTLE) ||
1938 current->backing_dev_info == NULL ||
1939 bdi_write_congested(current->backing_dev_info);
1943 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1944 * of reclaimed pages
1946 static noinline_for_stack unsigned long
1947 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1948 struct scan_control *sc, enum lru_list lru)
1950 LIST_HEAD(page_list);
1951 unsigned long nr_scanned;
1952 unsigned long nr_reclaimed = 0;
1953 unsigned long nr_taken;
1954 struct reclaim_stat stat;
1955 int file = is_file_lru(lru);
1956 enum vm_event_item item;
1957 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1958 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1959 bool stalled = false;
1961 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1965 /* wait a bit for the reclaimer. */
1969 /* We are about to die and free our memory. Return now. */
1970 if (fatal_signal_pending(current))
1971 return SWAP_CLUSTER_MAX;
1976 spin_lock_irq(&pgdat->lru_lock);
1978 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1979 &nr_scanned, sc, lru);
1981 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1982 reclaim_stat->recent_scanned[file] += nr_taken;
1984 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1985 if (global_reclaim(sc))
1986 __count_vm_events(item, nr_scanned);
1987 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1988 spin_unlock_irq(&pgdat->lru_lock);
1993 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1996 spin_lock_irq(&pgdat->lru_lock);
1998 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1999 if (global_reclaim(sc))
2000 __count_vm_events(item, nr_reclaimed);
2001 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2002 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2003 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2005 move_pages_to_lru(lruvec, &page_list);
2007 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2009 spin_unlock_irq(&pgdat->lru_lock);
2011 mem_cgroup_uncharge_list(&page_list);
2012 free_unref_page_list(&page_list);
2015 * If dirty pages are scanned that are not queued for IO, it
2016 * implies that flushers are not doing their job. This can
2017 * happen when memory pressure pushes dirty pages to the end of
2018 * the LRU before the dirty limits are breached and the dirty
2019 * data has expired. It can also happen when the proportion of
2020 * dirty pages grows not through writes but through memory
2021 * pressure reclaiming all the clean cache. And in some cases,
2022 * the flushers simply cannot keep up with the allocation
2023 * rate. Nudge the flusher threads in case they are asleep.
2025 if (stat.nr_unqueued_dirty == nr_taken)
2026 wakeup_flusher_threads(WB_REASON_VMSCAN);
2028 sc->nr.dirty += stat.nr_dirty;
2029 sc->nr.congested += stat.nr_congested;
2030 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2031 sc->nr.writeback += stat.nr_writeback;
2032 sc->nr.immediate += stat.nr_immediate;
2033 sc->nr.taken += nr_taken;
2035 sc->nr.file_taken += nr_taken;
2037 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2038 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2039 return nr_reclaimed;
2042 static void shrink_active_list(unsigned long nr_to_scan,
2043 struct lruvec *lruvec,
2044 struct scan_control *sc,
2047 unsigned long nr_taken;
2048 unsigned long nr_scanned;
2049 unsigned long vm_flags;
2050 LIST_HEAD(l_hold); /* The pages which were snipped off */
2051 LIST_HEAD(l_active);
2052 LIST_HEAD(l_inactive);
2054 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2055 unsigned nr_deactivate, nr_activate;
2056 unsigned nr_rotated = 0;
2057 int file = is_file_lru(lru);
2058 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2062 spin_lock_irq(&pgdat->lru_lock);
2064 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2065 &nr_scanned, sc, lru);
2067 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2068 reclaim_stat->recent_scanned[file] += nr_taken;
2070 __count_vm_events(PGREFILL, nr_scanned);
2071 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2073 spin_unlock_irq(&pgdat->lru_lock);
2075 while (!list_empty(&l_hold)) {
2077 page = lru_to_page(&l_hold);
2078 list_del(&page->lru);
2080 if (unlikely(!page_evictable(page))) {
2081 putback_lru_page(page);
2085 if (unlikely(buffer_heads_over_limit)) {
2086 if (page_has_private(page) && trylock_page(page)) {
2087 if (page_has_private(page))
2088 try_to_release_page(page, 0);
2093 if (page_referenced(page, 0, sc->target_mem_cgroup,
2095 nr_rotated += hpage_nr_pages(page);
2097 * Identify referenced, file-backed active pages and
2098 * give them one more trip around the active list. So
2099 * that executable code get better chances to stay in
2100 * memory under moderate memory pressure. Anon pages
2101 * are not likely to be evicted by use-once streaming
2102 * IO, plus JVM can create lots of anon VM_EXEC pages,
2103 * so we ignore them here.
2105 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2106 list_add(&page->lru, &l_active);
2111 ClearPageActive(page); /* we are de-activating */
2112 SetPageWorkingset(page);
2113 list_add(&page->lru, &l_inactive);
2117 * Move pages back to the lru list.
2119 spin_lock_irq(&pgdat->lru_lock);
2121 * Count referenced pages from currently used mappings as rotated,
2122 * even though only some of them are actually re-activated. This
2123 * helps balance scan pressure between file and anonymous pages in
2126 reclaim_stat->recent_rotated[file] += nr_rotated;
2128 nr_activate = move_pages_to_lru(lruvec, &l_active);
2129 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2130 /* Keep all free pages in l_active list */
2131 list_splice(&l_inactive, &l_active);
2133 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2134 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2136 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2137 spin_unlock_irq(&pgdat->lru_lock);
2139 mem_cgroup_uncharge_list(&l_active);
2140 free_unref_page_list(&l_active);
2141 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2142 nr_deactivate, nr_rotated, sc->priority, file);
2145 unsigned long reclaim_pages(struct list_head *page_list)
2148 unsigned long nr_reclaimed = 0;
2149 LIST_HEAD(node_page_list);
2150 struct reclaim_stat dummy_stat;
2152 struct scan_control sc = {
2153 .gfp_mask = GFP_KERNEL,
2154 .priority = DEF_PRIORITY,
2160 while (!list_empty(page_list)) {
2161 page = lru_to_page(page_list);
2163 nid = page_to_nid(page);
2164 INIT_LIST_HEAD(&node_page_list);
2167 if (nid == page_to_nid(page)) {
2168 ClearPageActive(page);
2169 list_move(&page->lru, &node_page_list);
2173 nr_reclaimed += shrink_page_list(&node_page_list,
2176 &dummy_stat, false);
2177 while (!list_empty(&node_page_list)) {
2178 page = lru_to_page(&node_page_list);
2179 list_del(&page->lru);
2180 putback_lru_page(page);
2186 if (!list_empty(&node_page_list)) {
2187 nr_reclaimed += shrink_page_list(&node_page_list,
2190 &dummy_stat, false);
2191 while (!list_empty(&node_page_list)) {
2192 page = lru_to_page(&node_page_list);
2193 list_del(&page->lru);
2194 putback_lru_page(page);
2198 return nr_reclaimed;
2202 * The inactive anon list should be small enough that the VM never has
2203 * to do too much work.
2205 * The inactive file list should be small enough to leave most memory
2206 * to the established workingset on the scan-resistant active list,
2207 * but large enough to avoid thrashing the aggregate readahead window.
2209 * Both inactive lists should also be large enough that each inactive
2210 * page has a chance to be referenced again before it is reclaimed.
2212 * If that fails and refaulting is observed, the inactive list grows.
2214 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2215 * on this LRU, maintained by the pageout code. An inactive_ratio
2216 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2219 * memory ratio inactive
2220 * -------------------------------------
2229 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2230 struct scan_control *sc, bool trace)
2232 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2233 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2234 enum lru_list inactive_lru = file * LRU_FILE;
2235 unsigned long inactive, active;
2236 unsigned long inactive_ratio;
2237 unsigned long refaults;
2241 * If we don't have swap space, anonymous page deactivation
2244 if (!file && !total_swap_pages)
2247 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2248 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2251 * When refaults are being observed, it means a new workingset
2252 * is being established. Disable active list protection to get
2253 * rid of the stale workingset quickly.
2255 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2256 if (file && lruvec->refaults != refaults) {
2259 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2261 inactive_ratio = int_sqrt(10 * gb);
2267 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2268 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2269 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2270 inactive_ratio, file);
2272 return inactive * inactive_ratio < active;
2275 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2276 struct lruvec *lruvec, struct scan_control *sc)
2278 if (is_active_lru(lru)) {
2279 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2280 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2284 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2295 * Determine how aggressively the anon and file LRU lists should be
2296 * scanned. The relative value of each set of LRU lists is determined
2297 * by looking at the fraction of the pages scanned we did rotate back
2298 * onto the active list instead of evict.
2300 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2301 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2303 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2304 struct scan_control *sc, unsigned long *nr)
2306 int swappiness = mem_cgroup_swappiness(memcg);
2307 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2309 u64 denominator = 0; /* gcc */
2310 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2311 unsigned long anon_prio, file_prio;
2312 enum scan_balance scan_balance;
2313 unsigned long anon, file;
2314 unsigned long ap, fp;
2317 /* If we have no swap space, do not bother scanning anon pages. */
2318 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2319 scan_balance = SCAN_FILE;
2324 * Global reclaim will swap to prevent OOM even with no
2325 * swappiness, but memcg users want to use this knob to
2326 * disable swapping for individual groups completely when
2327 * using the memory controller's swap limit feature would be
2330 if (!global_reclaim(sc) && !swappiness) {
2331 scan_balance = SCAN_FILE;
2336 * Do not apply any pressure balancing cleverness when the
2337 * system is close to OOM, scan both anon and file equally
2338 * (unless the swappiness setting disagrees with swapping).
2340 if (!sc->priority && swappiness) {
2341 scan_balance = SCAN_EQUAL;
2346 * Prevent the reclaimer from falling into the cache trap: as
2347 * cache pages start out inactive, every cache fault will tip
2348 * the scan balance towards the file LRU. And as the file LRU
2349 * shrinks, so does the window for rotation from references.
2350 * This means we have a runaway feedback loop where a tiny
2351 * thrashing file LRU becomes infinitely more attractive than
2352 * anon pages. Try to detect this based on file LRU size.
2354 if (global_reclaim(sc)) {
2355 unsigned long pgdatfile;
2356 unsigned long pgdatfree;
2358 unsigned long total_high_wmark = 0;
2360 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2361 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2362 node_page_state(pgdat, NR_INACTIVE_FILE);
2364 for (z = 0; z < MAX_NR_ZONES; z++) {
2365 struct zone *zone = &pgdat->node_zones[z];
2366 if (!managed_zone(zone))
2369 total_high_wmark += high_wmark_pages(zone);
2372 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2374 * Force SCAN_ANON if there are enough inactive
2375 * anonymous pages on the LRU in eligible zones.
2376 * Otherwise, the small LRU gets thrashed.
2378 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2379 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2381 scan_balance = SCAN_ANON;
2388 * If there is enough inactive page cache, i.e. if the size of the
2389 * inactive list is greater than that of the active list *and* the
2390 * inactive list actually has some pages to scan on this priority, we
2391 * do not reclaim anything from the anonymous working set right now.
2392 * Without the second condition we could end up never scanning an
2393 * lruvec even if it has plenty of old anonymous pages unless the
2394 * system is under heavy pressure.
2396 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2397 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2398 scan_balance = SCAN_FILE;
2402 scan_balance = SCAN_FRACT;
2405 * With swappiness at 100, anonymous and file have the same priority.
2406 * This scanning priority is essentially the inverse of IO cost.
2408 anon_prio = swappiness;
2409 file_prio = 200 - anon_prio;
2412 * OK, so we have swap space and a fair amount of page cache
2413 * pages. We use the recently rotated / recently scanned
2414 * ratios to determine how valuable each cache is.
2416 * Because workloads change over time (and to avoid overflow)
2417 * we keep these statistics as a floating average, which ends
2418 * up weighing recent references more than old ones.
2420 * anon in [0], file in [1]
2423 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2424 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2425 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2426 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2428 spin_lock_irq(&pgdat->lru_lock);
2429 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2430 reclaim_stat->recent_scanned[0] /= 2;
2431 reclaim_stat->recent_rotated[0] /= 2;
2434 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2435 reclaim_stat->recent_scanned[1] /= 2;
2436 reclaim_stat->recent_rotated[1] /= 2;
2440 * The amount of pressure on anon vs file pages is inversely
2441 * proportional to the fraction of recently scanned pages on
2442 * each list that were recently referenced and in active use.
2444 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2445 ap /= reclaim_stat->recent_rotated[0] + 1;
2447 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2448 fp /= reclaim_stat->recent_rotated[1] + 1;
2449 spin_unlock_irq(&pgdat->lru_lock);
2453 denominator = ap + fp + 1;
2455 for_each_evictable_lru(lru) {
2456 int file = is_file_lru(lru);
2457 unsigned long lruvec_size;
2459 unsigned long protection;
2461 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2462 protection = mem_cgroup_protection(memcg,
2463 sc->memcg_low_reclaim);
2467 * Scale a cgroup's reclaim pressure by proportioning
2468 * its current usage to its memory.low or memory.min
2471 * This is important, as otherwise scanning aggression
2472 * becomes extremely binary -- from nothing as we
2473 * approach the memory protection threshold, to totally
2474 * nominal as we exceed it. This results in requiring
2475 * setting extremely liberal protection thresholds. It
2476 * also means we simply get no protection at all if we
2477 * set it too low, which is not ideal.
2479 * If there is any protection in place, we reduce scan
2480 * pressure by how much of the total memory used is
2481 * within protection thresholds.
2483 * There is one special case: in the first reclaim pass,
2484 * we skip over all groups that are within their low
2485 * protection. If that fails to reclaim enough pages to
2486 * satisfy the reclaim goal, we come back and override
2487 * the best-effort low protection. However, we still
2488 * ideally want to honor how well-behaved groups are in
2489 * that case instead of simply punishing them all
2490 * equally. As such, we reclaim them based on how much
2491 * memory they are using, reducing the scan pressure
2492 * again by how much of the total memory used is under
2495 unsigned long cgroup_size = mem_cgroup_size(memcg);
2497 /* Avoid TOCTOU with earlier protection check */
2498 cgroup_size = max(cgroup_size, protection);
2500 scan = lruvec_size - lruvec_size * protection /
2504 * Minimally target SWAP_CLUSTER_MAX pages to keep
2505 * reclaim moving forwards, avoiding decremeting
2506 * sc->priority further than desirable.
2508 scan = max(scan, SWAP_CLUSTER_MAX);
2513 scan >>= sc->priority;
2516 * If the cgroup's already been deleted, make sure to
2517 * scrape out the remaining cache.
2519 if (!scan && !mem_cgroup_online(memcg))
2520 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2522 switch (scan_balance) {
2524 /* Scan lists relative to size */
2528 * Scan types proportional to swappiness and
2529 * their relative recent reclaim efficiency.
2530 * Make sure we don't miss the last page
2531 * because of a round-off error.
2533 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2538 /* Scan one type exclusively */
2539 if ((scan_balance == SCAN_FILE) != file) {
2545 /* Look ma, no brain */
2554 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2556 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2557 struct scan_control *sc)
2559 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2560 unsigned long nr[NR_LRU_LISTS];
2561 unsigned long targets[NR_LRU_LISTS];
2562 unsigned long nr_to_scan;
2564 unsigned long nr_reclaimed = 0;
2565 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2566 struct blk_plug plug;
2569 get_scan_count(lruvec, memcg, sc, nr);
2571 /* Record the original scan target for proportional adjustments later */
2572 memcpy(targets, nr, sizeof(nr));
2575 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2576 * event that can occur when there is little memory pressure e.g.
2577 * multiple streaming readers/writers. Hence, we do not abort scanning
2578 * when the requested number of pages are reclaimed when scanning at
2579 * DEF_PRIORITY on the assumption that the fact we are direct
2580 * reclaiming implies that kswapd is not keeping up and it is best to
2581 * do a batch of work at once. For memcg reclaim one check is made to
2582 * abort proportional reclaim if either the file or anon lru has already
2583 * dropped to zero at the first pass.
2585 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2586 sc->priority == DEF_PRIORITY);
2588 blk_start_plug(&plug);
2589 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2590 nr[LRU_INACTIVE_FILE]) {
2591 unsigned long nr_anon, nr_file, percentage;
2592 unsigned long nr_scanned;
2594 for_each_evictable_lru(lru) {
2596 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2597 nr[lru] -= nr_to_scan;
2599 nr_reclaimed += shrink_list(lru, nr_to_scan,
2606 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2610 * For kswapd and memcg, reclaim at least the number of pages
2611 * requested. Ensure that the anon and file LRUs are scanned
2612 * proportionally what was requested by get_scan_count(). We
2613 * stop reclaiming one LRU and reduce the amount scanning
2614 * proportional to the original scan target.
2616 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2617 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2620 * It's just vindictive to attack the larger once the smaller
2621 * has gone to zero. And given the way we stop scanning the
2622 * smaller below, this makes sure that we only make one nudge
2623 * towards proportionality once we've got nr_to_reclaim.
2625 if (!nr_file || !nr_anon)
2628 if (nr_file > nr_anon) {
2629 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2630 targets[LRU_ACTIVE_ANON] + 1;
2632 percentage = nr_anon * 100 / scan_target;
2634 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2635 targets[LRU_ACTIVE_FILE] + 1;
2637 percentage = nr_file * 100 / scan_target;
2640 /* Stop scanning the smaller of the LRU */
2642 nr[lru + LRU_ACTIVE] = 0;
2645 * Recalculate the other LRU scan count based on its original
2646 * scan target and the percentage scanning already complete
2648 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2649 nr_scanned = targets[lru] - nr[lru];
2650 nr[lru] = targets[lru] * (100 - percentage) / 100;
2651 nr[lru] -= min(nr[lru], nr_scanned);
2654 nr_scanned = targets[lru] - nr[lru];
2655 nr[lru] = targets[lru] * (100 - percentage) / 100;
2656 nr[lru] -= min(nr[lru], nr_scanned);
2658 scan_adjusted = true;
2660 blk_finish_plug(&plug);
2661 sc->nr_reclaimed += nr_reclaimed;
2664 * Even if we did not try to evict anon pages at all, we want to
2665 * rebalance the anon lru active/inactive ratio.
2667 if (inactive_list_is_low(lruvec, false, sc, true))
2668 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2669 sc, LRU_ACTIVE_ANON);
2672 /* Use reclaim/compaction for costly allocs or under memory pressure */
2673 static bool in_reclaim_compaction(struct scan_control *sc)
2675 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2676 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2677 sc->priority < DEF_PRIORITY - 2))
2684 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2685 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2686 * true if more pages should be reclaimed such that when the page allocator
2687 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2688 * It will give up earlier than that if there is difficulty reclaiming pages.
2690 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2691 unsigned long nr_reclaimed,
2692 struct scan_control *sc)
2694 unsigned long pages_for_compaction;
2695 unsigned long inactive_lru_pages;
2698 /* If not in reclaim/compaction mode, stop */
2699 if (!in_reclaim_compaction(sc))
2703 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2704 * number of pages that were scanned. This will return to the caller
2705 * with the risk reclaim/compaction and the resulting allocation attempt
2706 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2707 * allocations through requiring that the full LRU list has been scanned
2708 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2709 * scan, but that approximation was wrong, and there were corner cases
2710 * where always a non-zero amount of pages were scanned.
2715 /* If compaction would go ahead or the allocation would succeed, stop */
2716 for (z = 0; z <= sc->reclaim_idx; z++) {
2717 struct zone *zone = &pgdat->node_zones[z];
2718 if (!managed_zone(zone))
2721 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2722 case COMPACT_SUCCESS:
2723 case COMPACT_CONTINUE:
2726 /* check next zone */
2732 * If we have not reclaimed enough pages for compaction and the
2733 * inactive lists are large enough, continue reclaiming
2735 pages_for_compaction = compact_gap(sc->order);
2736 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2737 if (get_nr_swap_pages() > 0)
2738 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2740 return inactive_lru_pages > pages_for_compaction;
2743 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2745 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2746 (memcg && memcg_congested(pgdat, memcg));
2749 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2751 struct reclaim_state *reclaim_state = current->reclaim_state;
2752 unsigned long nr_reclaimed, nr_scanned;
2753 bool reclaimable = false;
2756 struct mem_cgroup *root = sc->target_mem_cgroup;
2757 struct mem_cgroup *memcg;
2759 memset(&sc->nr, 0, sizeof(sc->nr));
2761 nr_reclaimed = sc->nr_reclaimed;
2762 nr_scanned = sc->nr_scanned;
2764 memcg = mem_cgroup_iter(root, NULL, NULL);
2766 unsigned long reclaimed;
2767 unsigned long scanned;
2769 switch (mem_cgroup_protected(root, memcg)) {
2770 case MEMCG_PROT_MIN:
2773 * If there is no reclaimable memory, OOM.
2776 case MEMCG_PROT_LOW:
2779 * Respect the protection only as long as
2780 * there is an unprotected supply
2781 * of reclaimable memory from other cgroups.
2783 if (!sc->memcg_low_reclaim) {
2784 sc->memcg_low_skipped = 1;
2787 memcg_memory_event(memcg, MEMCG_LOW);
2789 case MEMCG_PROT_NONE:
2791 * All protection thresholds breached. We may
2792 * still choose to vary the scan pressure
2793 * applied based on by how much the cgroup in
2794 * question has exceeded its protection
2795 * thresholds (see get_scan_count).
2800 reclaimed = sc->nr_reclaimed;
2801 scanned = sc->nr_scanned;
2802 shrink_node_memcg(pgdat, memcg, sc);
2804 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2807 /* Record the group's reclaim efficiency */
2808 vmpressure(sc->gfp_mask, memcg, false,
2809 sc->nr_scanned - scanned,
2810 sc->nr_reclaimed - reclaimed);
2812 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2814 if (reclaim_state) {
2815 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2816 reclaim_state->reclaimed_slab = 0;
2819 /* Record the subtree's reclaim efficiency */
2820 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2821 sc->nr_scanned - nr_scanned,
2822 sc->nr_reclaimed - nr_reclaimed);
2824 if (sc->nr_reclaimed - nr_reclaimed)
2827 if (current_is_kswapd()) {
2829 * If reclaim is isolating dirty pages under writeback,
2830 * it implies that the long-lived page allocation rate
2831 * is exceeding the page laundering rate. Either the
2832 * global limits are not being effective at throttling
2833 * processes due to the page distribution throughout
2834 * zones or there is heavy usage of a slow backing
2835 * device. The only option is to throttle from reclaim
2836 * context which is not ideal as there is no guarantee
2837 * the dirtying process is throttled in the same way
2838 * balance_dirty_pages() manages.
2840 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2841 * count the number of pages under pages flagged for
2842 * immediate reclaim and stall if any are encountered
2843 * in the nr_immediate check below.
2845 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2846 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2849 * Tag a node as congested if all the dirty pages
2850 * scanned were backed by a congested BDI and
2851 * wait_iff_congested will stall.
2853 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2854 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2856 /* Allow kswapd to start writing pages during reclaim.*/
2857 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2858 set_bit(PGDAT_DIRTY, &pgdat->flags);
2861 * If kswapd scans pages marked marked for immediate
2862 * reclaim and under writeback (nr_immediate), it
2863 * implies that pages are cycling through the LRU
2864 * faster than they are written so also forcibly stall.
2866 if (sc->nr.immediate)
2867 congestion_wait(BLK_RW_ASYNC, HZ/10);
2871 * Legacy memcg will stall in page writeback so avoid forcibly
2872 * stalling in wait_iff_congested().
2874 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2875 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2876 set_memcg_congestion(pgdat, root, true);
2879 * Stall direct reclaim for IO completions if underlying BDIs
2880 * and node is congested. Allow kswapd to continue until it
2881 * starts encountering unqueued dirty pages or cycling through
2882 * the LRU too quickly.
2884 if (!sc->hibernation_mode && !current_is_kswapd() &&
2885 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2886 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2888 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2892 * Kswapd gives up on balancing particular nodes after too
2893 * many failures to reclaim anything from them and goes to
2894 * sleep. On reclaim progress, reset the failure counter. A
2895 * successful direct reclaim run will revive a dormant kswapd.
2898 pgdat->kswapd_failures = 0;
2904 * Returns true if compaction should go ahead for a costly-order request, or
2905 * the allocation would already succeed without compaction. Return false if we
2906 * should reclaim first.
2908 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2910 unsigned long watermark;
2911 enum compact_result suitable;
2913 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2914 if (suitable == COMPACT_SUCCESS)
2915 /* Allocation should succeed already. Don't reclaim. */
2917 if (suitable == COMPACT_SKIPPED)
2918 /* Compaction cannot yet proceed. Do reclaim. */
2922 * Compaction is already possible, but it takes time to run and there
2923 * are potentially other callers using the pages just freed. So proceed
2924 * with reclaim to make a buffer of free pages available to give
2925 * compaction a reasonable chance of completing and allocating the page.
2926 * Note that we won't actually reclaim the whole buffer in one attempt
2927 * as the target watermark in should_continue_reclaim() is lower. But if
2928 * we are already above the high+gap watermark, don't reclaim at all.
2930 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2932 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2936 * This is the direct reclaim path, for page-allocating processes. We only
2937 * try to reclaim pages from zones which will satisfy the caller's allocation
2940 * If a zone is deemed to be full of pinned pages then just give it a light
2941 * scan then give up on it.
2943 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2947 unsigned long nr_soft_reclaimed;
2948 unsigned long nr_soft_scanned;
2950 pg_data_t *last_pgdat = NULL;
2953 * If the number of buffer_heads in the machine exceeds the maximum
2954 * allowed level, force direct reclaim to scan the highmem zone as
2955 * highmem pages could be pinning lowmem pages storing buffer_heads
2957 orig_mask = sc->gfp_mask;
2958 if (buffer_heads_over_limit) {
2959 sc->gfp_mask |= __GFP_HIGHMEM;
2960 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2963 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2964 sc->reclaim_idx, sc->nodemask) {
2966 * Take care memory controller reclaiming has small influence
2969 if (global_reclaim(sc)) {
2970 if (!cpuset_zone_allowed(zone,
2971 GFP_KERNEL | __GFP_HARDWALL))
2975 * If we already have plenty of memory free for
2976 * compaction in this zone, don't free any more.
2977 * Even though compaction is invoked for any
2978 * non-zero order, only frequent costly order
2979 * reclamation is disruptive enough to become a
2980 * noticeable problem, like transparent huge
2983 if (IS_ENABLED(CONFIG_COMPACTION) &&
2984 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2985 compaction_ready(zone, sc)) {
2986 sc->compaction_ready = true;
2991 * Shrink each node in the zonelist once. If the
2992 * zonelist is ordered by zone (not the default) then a
2993 * node may be shrunk multiple times but in that case
2994 * the user prefers lower zones being preserved.
2996 if (zone->zone_pgdat == last_pgdat)
3000 * This steals pages from memory cgroups over softlimit
3001 * and returns the number of reclaimed pages and
3002 * scanned pages. This works for global memory pressure
3003 * and balancing, not for a memcg's limit.
3005 nr_soft_scanned = 0;
3006 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3007 sc->order, sc->gfp_mask,
3009 sc->nr_reclaimed += nr_soft_reclaimed;
3010 sc->nr_scanned += nr_soft_scanned;
3011 /* need some check for avoid more shrink_zone() */
3014 /* See comment about same check for global reclaim above */
3015 if (zone->zone_pgdat == last_pgdat)
3017 last_pgdat = zone->zone_pgdat;
3018 shrink_node(zone->zone_pgdat, sc);
3022 * Restore to original mask to avoid the impact on the caller if we
3023 * promoted it to __GFP_HIGHMEM.
3025 sc->gfp_mask = orig_mask;
3028 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3030 struct mem_cgroup *memcg;
3032 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3034 unsigned long refaults;
3035 struct lruvec *lruvec;
3037 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3038 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3039 lruvec->refaults = refaults;
3040 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3044 * This is the main entry point to direct page reclaim.
3046 * If a full scan of the inactive list fails to free enough memory then we
3047 * are "out of memory" and something needs to be killed.
3049 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3050 * high - the zone may be full of dirty or under-writeback pages, which this
3051 * caller can't do much about. We kick the writeback threads and take explicit
3052 * naps in the hope that some of these pages can be written. But if the
3053 * allocating task holds filesystem locks which prevent writeout this might not
3054 * work, and the allocation attempt will fail.
3056 * returns: 0, if no pages reclaimed
3057 * else, the number of pages reclaimed
3059 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3060 struct scan_control *sc)
3062 int initial_priority = sc->priority;
3063 pg_data_t *last_pgdat;
3067 delayacct_freepages_start();
3069 if (global_reclaim(sc))
3070 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3073 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3076 shrink_zones(zonelist, sc);
3078 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3081 if (sc->compaction_ready)
3085 * If we're getting trouble reclaiming, start doing
3086 * writepage even in laptop mode.
3088 if (sc->priority < DEF_PRIORITY - 2)
3089 sc->may_writepage = 1;
3090 } while (--sc->priority >= 0);
3093 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3095 if (zone->zone_pgdat == last_pgdat)
3097 last_pgdat = zone->zone_pgdat;
3098 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3099 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3102 delayacct_freepages_end();
3104 if (sc->nr_reclaimed)
3105 return sc->nr_reclaimed;
3107 /* Aborted reclaim to try compaction? don't OOM, then */
3108 if (sc->compaction_ready)
3111 /* Untapped cgroup reserves? Don't OOM, retry. */
3112 if (sc->memcg_low_skipped) {
3113 sc->priority = initial_priority;
3114 sc->memcg_low_reclaim = 1;
3115 sc->memcg_low_skipped = 0;
3122 static bool allow_direct_reclaim(pg_data_t *pgdat)
3125 unsigned long pfmemalloc_reserve = 0;
3126 unsigned long free_pages = 0;
3130 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3133 for (i = 0; i <= ZONE_NORMAL; i++) {
3134 zone = &pgdat->node_zones[i];
3135 if (!managed_zone(zone))
3138 if (!zone_reclaimable_pages(zone))
3141 pfmemalloc_reserve += min_wmark_pages(zone);
3142 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3145 /* If there are no reserves (unexpected config) then do not throttle */
3146 if (!pfmemalloc_reserve)
3149 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3151 /* kswapd must be awake if processes are being throttled */
3152 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3153 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3154 (enum zone_type)ZONE_NORMAL);
3155 wake_up_interruptible(&pgdat->kswapd_wait);
3162 * Throttle direct reclaimers if backing storage is backed by the network
3163 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3164 * depleted. kswapd will continue to make progress and wake the processes
3165 * when the low watermark is reached.
3167 * Returns true if a fatal signal was delivered during throttling. If this
3168 * happens, the page allocator should not consider triggering the OOM killer.
3170 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3171 nodemask_t *nodemask)
3175 pg_data_t *pgdat = NULL;
3178 * Kernel threads should not be throttled as they may be indirectly
3179 * responsible for cleaning pages necessary for reclaim to make forward
3180 * progress. kjournald for example may enter direct reclaim while
3181 * committing a transaction where throttling it could forcing other
3182 * processes to block on log_wait_commit().
3184 if (current->flags & PF_KTHREAD)
3188 * If a fatal signal is pending, this process should not throttle.
3189 * It should return quickly so it can exit and free its memory
3191 if (fatal_signal_pending(current))
3195 * Check if the pfmemalloc reserves are ok by finding the first node
3196 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3197 * GFP_KERNEL will be required for allocating network buffers when
3198 * swapping over the network so ZONE_HIGHMEM is unusable.
3200 * Throttling is based on the first usable node and throttled processes
3201 * wait on a queue until kswapd makes progress and wakes them. There
3202 * is an affinity then between processes waking up and where reclaim
3203 * progress has been made assuming the process wakes on the same node.
3204 * More importantly, processes running on remote nodes will not compete
3205 * for remote pfmemalloc reserves and processes on different nodes
3206 * should make reasonable progress.
3208 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3209 gfp_zone(gfp_mask), nodemask) {
3210 if (zone_idx(zone) > ZONE_NORMAL)
3213 /* Throttle based on the first usable node */
3214 pgdat = zone->zone_pgdat;
3215 if (allow_direct_reclaim(pgdat))
3220 /* If no zone was usable by the allocation flags then do not throttle */
3224 /* Account for the throttling */
3225 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3228 * If the caller cannot enter the filesystem, it's possible that it
3229 * is due to the caller holding an FS lock or performing a journal
3230 * transaction in the case of a filesystem like ext[3|4]. In this case,
3231 * it is not safe to block on pfmemalloc_wait as kswapd could be
3232 * blocked waiting on the same lock. Instead, throttle for up to a
3233 * second before continuing.
3235 if (!(gfp_mask & __GFP_FS)) {
3236 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3237 allow_direct_reclaim(pgdat), HZ);
3242 /* Throttle until kswapd wakes the process */
3243 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3244 allow_direct_reclaim(pgdat));
3247 if (fatal_signal_pending(current))
3254 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3255 gfp_t gfp_mask, nodemask_t *nodemask)
3257 unsigned long nr_reclaimed;
3258 struct scan_control sc = {
3259 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3260 .gfp_mask = current_gfp_context(gfp_mask),
3261 .reclaim_idx = gfp_zone(gfp_mask),
3263 .nodemask = nodemask,
3264 .priority = DEF_PRIORITY,
3265 .may_writepage = !laptop_mode,
3271 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3272 * Confirm they are large enough for max values.
3274 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3275 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3276 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3279 * Do not enter reclaim if fatal signal was delivered while throttled.
3280 * 1 is returned so that the page allocator does not OOM kill at this
3283 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3286 set_task_reclaim_state(current, &sc.reclaim_state);
3287 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3289 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3291 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3292 set_task_reclaim_state(current, NULL);
3294 return nr_reclaimed;
3299 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3300 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3301 gfp_t gfp_mask, bool noswap,
3303 unsigned long *nr_scanned)
3305 struct scan_control sc = {
3306 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3307 .target_mem_cgroup = memcg,
3308 .may_writepage = !laptop_mode,
3310 .reclaim_idx = MAX_NR_ZONES - 1,
3311 .may_swap = !noswap,
3314 WARN_ON_ONCE(!current->reclaim_state);
3316 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3317 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3319 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3323 * NOTE: Although we can get the priority field, using it
3324 * here is not a good idea, since it limits the pages we can scan.
3325 * if we don't reclaim here, the shrink_node from balance_pgdat
3326 * will pick up pages from other mem cgroup's as well. We hack
3327 * the priority and make it zero.
3329 shrink_node_memcg(pgdat, memcg, &sc);
3331 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3333 *nr_scanned = sc.nr_scanned;
3335 return sc.nr_reclaimed;
3338 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3339 unsigned long nr_pages,
3343 unsigned long nr_reclaimed;
3344 unsigned long pflags;
3345 unsigned int noreclaim_flag;
3346 struct scan_control sc = {
3347 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3348 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3349 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3350 .reclaim_idx = MAX_NR_ZONES - 1,
3351 .target_mem_cgroup = memcg,
3352 .priority = DEF_PRIORITY,
3353 .may_writepage = !laptop_mode,
3355 .may_swap = may_swap,
3358 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3359 * equal pressure on all the nodes. This is based on the assumption that
3360 * the reclaim does not bail out early.
3362 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3364 set_task_reclaim_state(current, &sc.reclaim_state);
3366 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3368 psi_memstall_enter(&pflags);
3369 noreclaim_flag = memalloc_noreclaim_save();
3371 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3373 memalloc_noreclaim_restore(noreclaim_flag);
3374 psi_memstall_leave(&pflags);
3376 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3377 set_task_reclaim_state(current, NULL);
3379 return nr_reclaimed;
3383 static void age_active_anon(struct pglist_data *pgdat,
3384 struct scan_control *sc)
3386 struct mem_cgroup *memcg;
3388 if (!total_swap_pages)
3391 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3393 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3395 if (inactive_list_is_low(lruvec, false, sc, true))
3396 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3397 sc, LRU_ACTIVE_ANON);
3399 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3403 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3409 * Check for watermark boosts top-down as the higher zones
3410 * are more likely to be boosted. Both watermarks and boosts
3411 * should not be checked at the time time as reclaim would
3412 * start prematurely when there is no boosting and a lower
3415 for (i = classzone_idx; i >= 0; i--) {
3416 zone = pgdat->node_zones + i;
3417 if (!managed_zone(zone))
3420 if (zone->watermark_boost)
3428 * Returns true if there is an eligible zone balanced for the request order
3431 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3434 unsigned long mark = -1;
3438 * Check watermarks bottom-up as lower zones are more likely to
3441 for (i = 0; i <= classzone_idx; i++) {
3442 zone = pgdat->node_zones + i;
3444 if (!managed_zone(zone))
3447 mark = high_wmark_pages(zone);
3448 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3453 * If a node has no populated zone within classzone_idx, it does not
3454 * need balancing by definition. This can happen if a zone-restricted
3455 * allocation tries to wake a remote kswapd.
3463 /* Clear pgdat state for congested, dirty or under writeback. */
3464 static void clear_pgdat_congested(pg_data_t *pgdat)
3466 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3467 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3468 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3472 * Prepare kswapd for sleeping. This verifies that there are no processes
3473 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3475 * Returns true if kswapd is ready to sleep
3477 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3480 * The throttled processes are normally woken up in balance_pgdat() as
3481 * soon as allow_direct_reclaim() is true. But there is a potential
3482 * race between when kswapd checks the watermarks and a process gets
3483 * throttled. There is also a potential race if processes get
3484 * throttled, kswapd wakes, a large process exits thereby balancing the
3485 * zones, which causes kswapd to exit balance_pgdat() before reaching
3486 * the wake up checks. If kswapd is going to sleep, no process should
3487 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3488 * the wake up is premature, processes will wake kswapd and get
3489 * throttled again. The difference from wake ups in balance_pgdat() is
3490 * that here we are under prepare_to_wait().
3492 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3493 wake_up_all(&pgdat->pfmemalloc_wait);
3495 /* Hopeless node, leave it to direct reclaim */
3496 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3499 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3500 clear_pgdat_congested(pgdat);
3508 * kswapd shrinks a node of pages that are at or below the highest usable
3509 * zone that is currently unbalanced.
3511 * Returns true if kswapd scanned at least the requested number of pages to
3512 * reclaim or if the lack of progress was due to pages under writeback.
3513 * This is used to determine if the scanning priority needs to be raised.
3515 static bool kswapd_shrink_node(pg_data_t *pgdat,
3516 struct scan_control *sc)
3521 /* Reclaim a number of pages proportional to the number of zones */
3522 sc->nr_to_reclaim = 0;
3523 for (z = 0; z <= sc->reclaim_idx; z++) {
3524 zone = pgdat->node_zones + z;
3525 if (!managed_zone(zone))
3528 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3532 * Historically care was taken to put equal pressure on all zones but
3533 * now pressure is applied based on node LRU order.
3535 shrink_node(pgdat, sc);
3538 * Fragmentation may mean that the system cannot be rebalanced for
3539 * high-order allocations. If twice the allocation size has been
3540 * reclaimed then recheck watermarks only at order-0 to prevent
3541 * excessive reclaim. Assume that a process requested a high-order
3542 * can direct reclaim/compact.
3544 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3547 return sc->nr_scanned >= sc->nr_to_reclaim;
3551 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3552 * that are eligible for use by the caller until at least one zone is
3555 * Returns the order kswapd finished reclaiming at.
3557 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3558 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3559 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3560 * or lower is eligible for reclaim until at least one usable zone is
3563 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3566 unsigned long nr_soft_reclaimed;
3567 unsigned long nr_soft_scanned;
3568 unsigned long pflags;
3569 unsigned long nr_boost_reclaim;
3570 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3573 struct scan_control sc = {
3574 .gfp_mask = GFP_KERNEL,
3579 set_task_reclaim_state(current, &sc.reclaim_state);
3580 psi_memstall_enter(&pflags);
3581 __fs_reclaim_acquire();
3583 count_vm_event(PAGEOUTRUN);
3586 * Account for the reclaim boost. Note that the zone boost is left in
3587 * place so that parallel allocations that are near the watermark will
3588 * stall or direct reclaim until kswapd is finished.
3590 nr_boost_reclaim = 0;
3591 for (i = 0; i <= classzone_idx; i++) {
3592 zone = pgdat->node_zones + i;
3593 if (!managed_zone(zone))
3596 nr_boost_reclaim += zone->watermark_boost;
3597 zone_boosts[i] = zone->watermark_boost;
3599 boosted = nr_boost_reclaim;
3602 sc.priority = DEF_PRIORITY;
3604 unsigned long nr_reclaimed = sc.nr_reclaimed;
3605 bool raise_priority = true;
3609 sc.reclaim_idx = classzone_idx;
3612 * If the number of buffer_heads exceeds the maximum allowed
3613 * then consider reclaiming from all zones. This has a dual
3614 * purpose -- on 64-bit systems it is expected that
3615 * buffer_heads are stripped during active rotation. On 32-bit
3616 * systems, highmem pages can pin lowmem memory and shrinking
3617 * buffers can relieve lowmem pressure. Reclaim may still not
3618 * go ahead if all eligible zones for the original allocation
3619 * request are balanced to avoid excessive reclaim from kswapd.
3621 if (buffer_heads_over_limit) {
3622 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3623 zone = pgdat->node_zones + i;
3624 if (!managed_zone(zone))
3633 * If the pgdat is imbalanced then ignore boosting and preserve
3634 * the watermarks for a later time and restart. Note that the
3635 * zone watermarks will be still reset at the end of balancing
3636 * on the grounds that the normal reclaim should be enough to
3637 * re-evaluate if boosting is required when kswapd next wakes.
3639 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3640 if (!balanced && nr_boost_reclaim) {
3641 nr_boost_reclaim = 0;
3646 * If boosting is not active then only reclaim if there are no
3647 * eligible zones. Note that sc.reclaim_idx is not used as
3648 * buffer_heads_over_limit may have adjusted it.
3650 if (!nr_boost_reclaim && balanced)
3653 /* Limit the priority of boosting to avoid reclaim writeback */
3654 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3655 raise_priority = false;
3658 * Do not writeback or swap pages for boosted reclaim. The
3659 * intent is to relieve pressure not issue sub-optimal IO
3660 * from reclaim context. If no pages are reclaimed, the
3661 * reclaim will be aborted.
3663 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3664 sc.may_swap = !nr_boost_reclaim;
3667 * Do some background aging of the anon list, to give
3668 * pages a chance to be referenced before reclaiming. All
3669 * pages are rotated regardless of classzone as this is
3670 * about consistent aging.
3672 age_active_anon(pgdat, &sc);
3675 * If we're getting trouble reclaiming, start doing writepage
3676 * even in laptop mode.
3678 if (sc.priority < DEF_PRIORITY - 2)
3679 sc.may_writepage = 1;
3681 /* Call soft limit reclaim before calling shrink_node. */
3683 nr_soft_scanned = 0;
3684 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3685 sc.gfp_mask, &nr_soft_scanned);
3686 sc.nr_reclaimed += nr_soft_reclaimed;
3689 * There should be no need to raise the scanning priority if
3690 * enough pages are already being scanned that that high
3691 * watermark would be met at 100% efficiency.
3693 if (kswapd_shrink_node(pgdat, &sc))
3694 raise_priority = false;
3697 * If the low watermark is met there is no need for processes
3698 * to be throttled on pfmemalloc_wait as they should not be
3699 * able to safely make forward progress. Wake them
3701 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3702 allow_direct_reclaim(pgdat))
3703 wake_up_all(&pgdat->pfmemalloc_wait);
3705 /* Check if kswapd should be suspending */
3706 __fs_reclaim_release();
3707 ret = try_to_freeze();
3708 __fs_reclaim_acquire();
3709 if (ret || kthread_should_stop())
3713 * Raise priority if scanning rate is too low or there was no
3714 * progress in reclaiming pages
3716 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3717 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3720 * If reclaim made no progress for a boost, stop reclaim as
3721 * IO cannot be queued and it could be an infinite loop in
3722 * extreme circumstances.
3724 if (nr_boost_reclaim && !nr_reclaimed)
3727 if (raise_priority || !nr_reclaimed)
3729 } while (sc.priority >= 1);
3731 if (!sc.nr_reclaimed)
3732 pgdat->kswapd_failures++;
3735 /* If reclaim was boosted, account for the reclaim done in this pass */
3737 unsigned long flags;
3739 for (i = 0; i <= classzone_idx; i++) {
3740 if (!zone_boosts[i])
3743 /* Increments are under the zone lock */
3744 zone = pgdat->node_zones + i;
3745 spin_lock_irqsave(&zone->lock, flags);
3746 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3747 spin_unlock_irqrestore(&zone->lock, flags);
3751 * As there is now likely space, wakeup kcompact to defragment
3754 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3757 snapshot_refaults(NULL, pgdat);
3758 __fs_reclaim_release();
3759 psi_memstall_leave(&pflags);
3760 set_task_reclaim_state(current, NULL);
3763 * Return the order kswapd stopped reclaiming at as
3764 * prepare_kswapd_sleep() takes it into account. If another caller
3765 * entered the allocator slow path while kswapd was awake, order will
3766 * remain at the higher level.
3772 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3773 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3774 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3775 * after previous reclaim attempt (node is still unbalanced). In that case
3776 * return the zone index of the previous kswapd reclaim cycle.
3778 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3779 enum zone_type prev_classzone_idx)
3781 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3782 return prev_classzone_idx;
3783 return pgdat->kswapd_classzone_idx;
3786 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3787 unsigned int classzone_idx)
3792 if (freezing(current) || kthread_should_stop())
3795 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3798 * Try to sleep for a short interval. Note that kcompactd will only be
3799 * woken if it is possible to sleep for a short interval. This is
3800 * deliberate on the assumption that if reclaim cannot keep an
3801 * eligible zone balanced that it's also unlikely that compaction will
3804 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3806 * Compaction records what page blocks it recently failed to
3807 * isolate pages from and skips them in the future scanning.
3808 * When kswapd is going to sleep, it is reasonable to assume
3809 * that pages and compaction may succeed so reset the cache.
3811 reset_isolation_suitable(pgdat);
3814 * We have freed the memory, now we should compact it to make
3815 * allocation of the requested order possible.
3817 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3819 remaining = schedule_timeout(HZ/10);
3822 * If woken prematurely then reset kswapd_classzone_idx and
3823 * order. The values will either be from a wakeup request or
3824 * the previous request that slept prematurely.
3827 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3828 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3831 finish_wait(&pgdat->kswapd_wait, &wait);
3832 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3836 * After a short sleep, check if it was a premature sleep. If not, then
3837 * go fully to sleep until explicitly woken up.
3840 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3841 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3844 * vmstat counters are not perfectly accurate and the estimated
3845 * value for counters such as NR_FREE_PAGES can deviate from the
3846 * true value by nr_online_cpus * threshold. To avoid the zone
3847 * watermarks being breached while under pressure, we reduce the
3848 * per-cpu vmstat threshold while kswapd is awake and restore
3849 * them before going back to sleep.
3851 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3853 if (!kthread_should_stop())
3856 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3859 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3861 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3863 finish_wait(&pgdat->kswapd_wait, &wait);
3867 * The background pageout daemon, started as a kernel thread
3868 * from the init process.
3870 * This basically trickles out pages so that we have _some_
3871 * free memory available even if there is no other activity
3872 * that frees anything up. This is needed for things like routing
3873 * etc, where we otherwise might have all activity going on in
3874 * asynchronous contexts that cannot page things out.
3876 * If there are applications that are active memory-allocators
3877 * (most normal use), this basically shouldn't matter.
3879 static int kswapd(void *p)
3881 unsigned int alloc_order, reclaim_order;
3882 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3883 pg_data_t *pgdat = (pg_data_t*)p;
3884 struct task_struct *tsk = current;
3885 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3887 if (!cpumask_empty(cpumask))
3888 set_cpus_allowed_ptr(tsk, cpumask);
3891 * Tell the memory management that we're a "memory allocator",
3892 * and that if we need more memory we should get access to it
3893 * regardless (see "__alloc_pages()"). "kswapd" should
3894 * never get caught in the normal page freeing logic.
3896 * (Kswapd normally doesn't need memory anyway, but sometimes
3897 * you need a small amount of memory in order to be able to
3898 * page out something else, and this flag essentially protects
3899 * us from recursively trying to free more memory as we're
3900 * trying to free the first piece of memory in the first place).
3902 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3905 pgdat->kswapd_order = 0;
3906 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3910 alloc_order = reclaim_order = pgdat->kswapd_order;
3911 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3914 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3917 /* Read the new order and classzone_idx */
3918 alloc_order = reclaim_order = pgdat->kswapd_order;
3919 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3920 pgdat->kswapd_order = 0;
3921 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3923 ret = try_to_freeze();
3924 if (kthread_should_stop())
3928 * We can speed up thawing tasks if we don't call balance_pgdat
3929 * after returning from the refrigerator
3935 * Reclaim begins at the requested order but if a high-order
3936 * reclaim fails then kswapd falls back to reclaiming for
3937 * order-0. If that happens, kswapd will consider sleeping
3938 * for the order it finished reclaiming at (reclaim_order)
3939 * but kcompactd is woken to compact for the original
3940 * request (alloc_order).
3942 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3944 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3945 if (reclaim_order < alloc_order)
3946 goto kswapd_try_sleep;
3949 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3955 * A zone is low on free memory or too fragmented for high-order memory. If
3956 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3957 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3958 * has failed or is not needed, still wake up kcompactd if only compaction is
3961 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3962 enum zone_type classzone_idx)
3966 if (!managed_zone(zone))
3969 if (!cpuset_zone_allowed(zone, gfp_flags))
3971 pgdat = zone->zone_pgdat;
3973 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3974 pgdat->kswapd_classzone_idx = classzone_idx;
3976 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3978 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3979 if (!waitqueue_active(&pgdat->kswapd_wait))
3982 /* Hopeless node, leave it to direct reclaim if possible */
3983 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3984 (pgdat_balanced(pgdat, order, classzone_idx) &&
3985 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3987 * There may be plenty of free memory available, but it's too
3988 * fragmented for high-order allocations. Wake up kcompactd
3989 * and rely on compaction_suitable() to determine if it's
3990 * needed. If it fails, it will defer subsequent attempts to
3991 * ratelimit its work.
3993 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3994 wakeup_kcompactd(pgdat, order, classzone_idx);
3998 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
4000 wake_up_interruptible(&pgdat->kswapd_wait);
4003 #ifdef CONFIG_HIBERNATION
4005 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4008 * Rather than trying to age LRUs the aim is to preserve the overall
4009 * LRU order by reclaiming preferentially
4010 * inactive > active > active referenced > active mapped
4012 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4014 struct scan_control sc = {
4015 .nr_to_reclaim = nr_to_reclaim,
4016 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4017 .reclaim_idx = MAX_NR_ZONES - 1,
4018 .priority = DEF_PRIORITY,
4022 .hibernation_mode = 1,
4024 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4025 unsigned long nr_reclaimed;
4026 unsigned int noreclaim_flag;
4028 fs_reclaim_acquire(sc.gfp_mask);
4029 noreclaim_flag = memalloc_noreclaim_save();
4030 set_task_reclaim_state(current, &sc.reclaim_state);
4032 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4034 set_task_reclaim_state(current, NULL);
4035 memalloc_noreclaim_restore(noreclaim_flag);
4036 fs_reclaim_release(sc.gfp_mask);
4038 return nr_reclaimed;
4040 #endif /* CONFIG_HIBERNATION */
4042 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4043 not required for correctness. So if the last cpu in a node goes
4044 away, we get changed to run anywhere: as the first one comes back,
4045 restore their cpu bindings. */
4046 static int kswapd_cpu_online(unsigned int cpu)
4050 for_each_node_state(nid, N_MEMORY) {
4051 pg_data_t *pgdat = NODE_DATA(nid);
4052 const struct cpumask *mask;
4054 mask = cpumask_of_node(pgdat->node_id);
4056 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4057 /* One of our CPUs online: restore mask */
4058 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4064 * This kswapd start function will be called by init and node-hot-add.
4065 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4067 int kswapd_run(int nid)
4069 pg_data_t *pgdat = NODE_DATA(nid);
4075 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4076 if (IS_ERR(pgdat->kswapd)) {
4077 /* failure at boot is fatal */
4078 BUG_ON(system_state < SYSTEM_RUNNING);
4079 pr_err("Failed to start kswapd on node %d\n", nid);
4080 ret = PTR_ERR(pgdat->kswapd);
4081 pgdat->kswapd = NULL;
4087 * Called by memory hotplug when all memory in a node is offlined. Caller must
4088 * hold mem_hotplug_begin/end().
4090 void kswapd_stop(int nid)
4092 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4095 kthread_stop(kswapd);
4096 NODE_DATA(nid)->kswapd = NULL;
4100 static int __init kswapd_init(void)
4105 for_each_node_state(nid, N_MEMORY)
4107 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4108 "mm/vmscan:online", kswapd_cpu_online,
4114 module_init(kswapd_init)
4120 * If non-zero call node_reclaim when the number of free pages falls below
4123 int node_reclaim_mode __read_mostly;
4125 #define RECLAIM_OFF 0
4126 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4127 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4128 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4131 * Priority for NODE_RECLAIM. This determines the fraction of pages
4132 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4135 #define NODE_RECLAIM_PRIORITY 4
4138 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4141 int sysctl_min_unmapped_ratio = 1;
4144 * If the number of slab pages in a zone grows beyond this percentage then
4145 * slab reclaim needs to occur.
4147 int sysctl_min_slab_ratio = 5;
4149 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4151 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4152 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4153 node_page_state(pgdat, NR_ACTIVE_FILE);
4156 * It's possible for there to be more file mapped pages than
4157 * accounted for by the pages on the file LRU lists because
4158 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4160 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4163 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4164 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4166 unsigned long nr_pagecache_reclaimable;
4167 unsigned long delta = 0;
4170 * If RECLAIM_UNMAP is set, then all file pages are considered
4171 * potentially reclaimable. Otherwise, we have to worry about
4172 * pages like swapcache and node_unmapped_file_pages() provides
4175 if (node_reclaim_mode & RECLAIM_UNMAP)
4176 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4178 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4180 /* If we can't clean pages, remove dirty pages from consideration */
4181 if (!(node_reclaim_mode & RECLAIM_WRITE))
4182 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4184 /* Watch for any possible underflows due to delta */
4185 if (unlikely(delta > nr_pagecache_reclaimable))
4186 delta = nr_pagecache_reclaimable;
4188 return nr_pagecache_reclaimable - delta;
4192 * Try to free up some pages from this node through reclaim.
4194 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4196 /* Minimum pages needed in order to stay on node */
4197 const unsigned long nr_pages = 1 << order;
4198 struct task_struct *p = current;
4199 unsigned int noreclaim_flag;
4200 struct scan_control sc = {
4201 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4202 .gfp_mask = current_gfp_context(gfp_mask),
4204 .priority = NODE_RECLAIM_PRIORITY,
4205 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4206 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4208 .reclaim_idx = gfp_zone(gfp_mask),
4211 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4215 fs_reclaim_acquire(sc.gfp_mask);
4217 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4218 * and we also need to be able to write out pages for RECLAIM_WRITE
4219 * and RECLAIM_UNMAP.
4221 noreclaim_flag = memalloc_noreclaim_save();
4222 p->flags |= PF_SWAPWRITE;
4223 set_task_reclaim_state(p, &sc.reclaim_state);
4225 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4227 * Free memory by calling shrink node with increasing
4228 * priorities until we have enough memory freed.
4231 shrink_node(pgdat, &sc);
4232 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4235 set_task_reclaim_state(p, NULL);
4236 current->flags &= ~PF_SWAPWRITE;
4237 memalloc_noreclaim_restore(noreclaim_flag);
4238 fs_reclaim_release(sc.gfp_mask);
4240 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4242 return sc.nr_reclaimed >= nr_pages;
4245 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4250 * Node reclaim reclaims unmapped file backed pages and
4251 * slab pages if we are over the defined limits.
4253 * A small portion of unmapped file backed pages is needed for
4254 * file I/O otherwise pages read by file I/O will be immediately
4255 * thrown out if the node is overallocated. So we do not reclaim
4256 * if less than a specified percentage of the node is used by
4257 * unmapped file backed pages.
4259 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4260 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4261 return NODE_RECLAIM_FULL;
4264 * Do not scan if the allocation should not be delayed.
4266 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4267 return NODE_RECLAIM_NOSCAN;
4270 * Only run node reclaim on the local node or on nodes that do not
4271 * have associated processors. This will favor the local processor
4272 * over remote processors and spread off node memory allocations
4273 * as wide as possible.
4275 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4276 return NODE_RECLAIM_NOSCAN;
4278 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4279 return NODE_RECLAIM_NOSCAN;
4281 ret = __node_reclaim(pgdat, gfp_mask, order);
4282 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4285 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4292 * page_evictable - test whether a page is evictable
4293 * @page: the page to test
4295 * Test whether page is evictable--i.e., should be placed on active/inactive
4296 * lists vs unevictable list.
4298 * Reasons page might not be evictable:
4299 * (1) page's mapping marked unevictable
4300 * (2) page is part of an mlocked VMA
4303 int page_evictable(struct page *page)
4307 /* Prevent address_space of inode and swap cache from being freed */
4309 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4315 * check_move_unevictable_pages - check pages for evictability and move to
4316 * appropriate zone lru list
4317 * @pvec: pagevec with lru pages to check
4319 * Checks pages for evictability, if an evictable page is in the unevictable
4320 * lru list, moves it to the appropriate evictable lru list. This function
4321 * should be only used for lru pages.
4323 void check_move_unevictable_pages(struct pagevec *pvec)
4325 struct lruvec *lruvec;
4326 struct pglist_data *pgdat = NULL;
4331 for (i = 0; i < pvec->nr; i++) {
4332 struct page *page = pvec->pages[i];
4333 struct pglist_data *pagepgdat = page_pgdat(page);
4336 if (pagepgdat != pgdat) {
4338 spin_unlock_irq(&pgdat->lru_lock);
4340 spin_lock_irq(&pgdat->lru_lock);
4342 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4344 if (!PageLRU(page) || !PageUnevictable(page))
4347 if (page_evictable(page)) {
4348 enum lru_list lru = page_lru_base_type(page);
4350 VM_BUG_ON_PAGE(PageActive(page), page);
4351 ClearPageUnevictable(page);
4352 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4353 add_page_to_lru_list(page, lruvec, lru);
4359 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4360 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4361 spin_unlock_irq(&pgdat->lru_lock);
4364 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);