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 cgroup_reclaim(struct scan_control *sc)
244 return sc->target_mem_cgroup;
248 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
249 * @sc: scan_control in question
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
260 static bool writeback_throttling_sane(struct scan_control *sc)
262 if (!cgroup_reclaim(sc))
264 #ifdef CONFIG_CGROUP_WRITEBACK
265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
271 static void set_memcg_congestion(pg_data_t *pgdat,
272 struct mem_cgroup *memcg,
275 struct mem_cgroup_per_node *mn;
280 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
281 WRITE_ONCE(mn->congested, congested);
284 static bool memcg_congested(pg_data_t *pgdat,
285 struct mem_cgroup *memcg)
287 struct mem_cgroup_per_node *mn;
289 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
290 return READ_ONCE(mn->congested);
294 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
299 static void unregister_memcg_shrinker(struct shrinker *shrinker)
303 static bool cgroup_reclaim(struct scan_control *sc)
308 static bool writeback_throttling_sane(struct scan_control *sc)
313 static inline void set_memcg_congestion(struct pglist_data *pgdat,
314 struct mem_cgroup *memcg, bool congested)
318 static inline bool memcg_congested(struct pglist_data *pgdat,
319 struct mem_cgroup *memcg)
327 * This misses isolated pages which are not accounted for to save counters.
328 * As the data only determines if reclaim or compaction continues, it is
329 * not expected that isolated pages will be a dominating factor.
331 unsigned long zone_reclaimable_pages(struct zone *zone)
335 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
336 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
337 if (get_nr_swap_pages() > 0)
338 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
339 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
345 * lruvec_lru_size - Returns the number of pages on the given LRU list.
346 * @lruvec: lru vector
348 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
350 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
352 unsigned long size = 0;
355 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
356 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
358 if (!managed_zone(zone))
361 if (!mem_cgroup_disabled())
362 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
364 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
370 * Add a shrinker callback to be called from the vm.
372 int prealloc_shrinker(struct shrinker *shrinker)
374 unsigned int size = sizeof(*shrinker->nr_deferred);
376 if (shrinker->flags & SHRINKER_NUMA_AWARE)
379 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
380 if (!shrinker->nr_deferred)
383 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
384 if (prealloc_memcg_shrinker(shrinker))
391 kfree(shrinker->nr_deferred);
392 shrinker->nr_deferred = NULL;
396 void free_prealloced_shrinker(struct shrinker *shrinker)
398 if (!shrinker->nr_deferred)
401 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
402 unregister_memcg_shrinker(shrinker);
404 kfree(shrinker->nr_deferred);
405 shrinker->nr_deferred = NULL;
408 void register_shrinker_prepared(struct shrinker *shrinker)
410 down_write(&shrinker_rwsem);
411 list_add_tail(&shrinker->list, &shrinker_list);
412 #ifdef CONFIG_MEMCG_KMEM
413 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
414 idr_replace(&shrinker_idr, shrinker, shrinker->id);
416 up_write(&shrinker_rwsem);
419 int register_shrinker(struct shrinker *shrinker)
421 int err = prealloc_shrinker(shrinker);
425 register_shrinker_prepared(shrinker);
428 EXPORT_SYMBOL(register_shrinker);
433 void unregister_shrinker(struct shrinker *shrinker)
435 if (!shrinker->nr_deferred)
437 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
438 unregister_memcg_shrinker(shrinker);
439 down_write(&shrinker_rwsem);
440 list_del(&shrinker->list);
441 up_write(&shrinker_rwsem);
442 kfree(shrinker->nr_deferred);
443 shrinker->nr_deferred = NULL;
445 EXPORT_SYMBOL(unregister_shrinker);
447 #define SHRINK_BATCH 128
449 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
450 struct shrinker *shrinker, int priority)
452 unsigned long freed = 0;
453 unsigned long long delta;
458 int nid = shrinkctl->nid;
459 long batch_size = shrinker->batch ? shrinker->batch
461 long scanned = 0, next_deferred;
463 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
466 freeable = shrinker->count_objects(shrinker, shrinkctl);
467 if (freeable == 0 || freeable == SHRINK_EMPTY)
471 * copy the current shrinker scan count into a local variable
472 * and zero it so that other concurrent shrinker invocations
473 * don't also do this scanning work.
475 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
478 if (shrinker->seeks) {
479 delta = freeable >> priority;
481 do_div(delta, shrinker->seeks);
484 * These objects don't require any IO to create. Trim
485 * them aggressively under memory pressure to keep
486 * them from causing refetches in the IO caches.
488 delta = freeable / 2;
492 if (total_scan < 0) {
493 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
494 shrinker->scan_objects, total_scan);
495 total_scan = freeable;
498 next_deferred = total_scan;
501 * We need to avoid excessive windup on filesystem shrinkers
502 * due to large numbers of GFP_NOFS allocations causing the
503 * shrinkers to return -1 all the time. This results in a large
504 * nr being built up so when a shrink that can do some work
505 * comes along it empties the entire cache due to nr >>>
506 * freeable. This is bad for sustaining a working set in
509 * Hence only allow the shrinker to scan the entire cache when
510 * a large delta change is calculated directly.
512 if (delta < freeable / 4)
513 total_scan = min(total_scan, freeable / 2);
516 * Avoid risking looping forever due to too large nr value:
517 * never try to free more than twice the estimate number of
520 if (total_scan > freeable * 2)
521 total_scan = freeable * 2;
523 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
524 freeable, delta, total_scan, priority);
527 * Normally, we should not scan less than batch_size objects in one
528 * pass to avoid too frequent shrinker calls, but if the slab has less
529 * than batch_size objects in total and we are really tight on memory,
530 * we will try to reclaim all available objects, otherwise we can end
531 * up failing allocations although there are plenty of reclaimable
532 * objects spread over several slabs with usage less than the
535 * We detect the "tight on memory" situations by looking at the total
536 * number of objects we want to scan (total_scan). If it is greater
537 * than the total number of objects on slab (freeable), we must be
538 * scanning at high prio and therefore should try to reclaim as much as
541 while (total_scan >= batch_size ||
542 total_scan >= freeable) {
544 unsigned long nr_to_scan = min(batch_size, total_scan);
546 shrinkctl->nr_to_scan = nr_to_scan;
547 shrinkctl->nr_scanned = nr_to_scan;
548 ret = shrinker->scan_objects(shrinker, shrinkctl);
549 if (ret == SHRINK_STOP)
553 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
554 total_scan -= shrinkctl->nr_scanned;
555 scanned += shrinkctl->nr_scanned;
560 if (next_deferred >= scanned)
561 next_deferred -= scanned;
565 * move the unused scan count back into the shrinker in a
566 * manner that handles concurrent updates. If we exhausted the
567 * scan, there is no need to do an update.
569 if (next_deferred > 0)
570 new_nr = atomic_long_add_return(next_deferred,
571 &shrinker->nr_deferred[nid]);
573 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
575 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
580 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
581 struct mem_cgroup *memcg, int priority)
583 struct memcg_shrinker_map *map;
584 unsigned long ret, freed = 0;
587 if (!mem_cgroup_online(memcg))
590 if (!down_read_trylock(&shrinker_rwsem))
593 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
598 for_each_set_bit(i, map->map, shrinker_nr_max) {
599 struct shrink_control sc = {
600 .gfp_mask = gfp_mask,
604 struct shrinker *shrinker;
606 shrinker = idr_find(&shrinker_idr, i);
607 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
609 clear_bit(i, map->map);
613 /* Call non-slab shrinkers even though kmem is disabled */
614 if (!memcg_kmem_enabled() &&
615 !(shrinker->flags & SHRINKER_NONSLAB))
618 ret = do_shrink_slab(&sc, shrinker, priority);
619 if (ret == SHRINK_EMPTY) {
620 clear_bit(i, map->map);
622 * After the shrinker reported that it had no objects to
623 * free, but before we cleared the corresponding bit in
624 * the memcg shrinker map, a new object might have been
625 * added. To make sure, we have the bit set in this
626 * case, we invoke the shrinker one more time and reset
627 * the bit if it reports that it is not empty anymore.
628 * The memory barrier here pairs with the barrier in
629 * memcg_set_shrinker_bit():
631 * list_lru_add() shrink_slab_memcg()
632 * list_add_tail() clear_bit()
634 * set_bit() do_shrink_slab()
636 smp_mb__after_atomic();
637 ret = do_shrink_slab(&sc, shrinker, priority);
638 if (ret == SHRINK_EMPTY)
641 memcg_set_shrinker_bit(memcg, nid, i);
645 if (rwsem_is_contended(&shrinker_rwsem)) {
651 up_read(&shrinker_rwsem);
654 #else /* CONFIG_MEMCG */
655 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
656 struct mem_cgroup *memcg, int priority)
660 #endif /* CONFIG_MEMCG */
663 * shrink_slab - shrink slab caches
664 * @gfp_mask: allocation context
665 * @nid: node whose slab caches to target
666 * @memcg: memory cgroup whose slab caches to target
667 * @priority: the reclaim priority
669 * Call the shrink functions to age shrinkable caches.
671 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
672 * unaware shrinkers will receive a node id of 0 instead.
674 * @memcg specifies the memory cgroup to target. Unaware shrinkers
675 * are called only if it is the root cgroup.
677 * @priority is sc->priority, we take the number of objects and >> by priority
678 * in order to get the scan target.
680 * Returns the number of reclaimed slab objects.
682 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
683 struct mem_cgroup *memcg,
686 unsigned long ret, freed = 0;
687 struct shrinker *shrinker;
690 * The root memcg might be allocated even though memcg is disabled
691 * via "cgroup_disable=memory" boot parameter. This could make
692 * mem_cgroup_is_root() return false, then just run memcg slab
693 * shrink, but skip global shrink. This may result in premature
696 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
697 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
699 if (!down_read_trylock(&shrinker_rwsem))
702 list_for_each_entry(shrinker, &shrinker_list, list) {
703 struct shrink_control sc = {
704 .gfp_mask = gfp_mask,
709 ret = do_shrink_slab(&sc, shrinker, priority);
710 if (ret == SHRINK_EMPTY)
714 * Bail out if someone want to register a new shrinker to
715 * prevent the regsitration from being stalled for long periods
716 * by parallel ongoing shrinking.
718 if (rwsem_is_contended(&shrinker_rwsem)) {
724 up_read(&shrinker_rwsem);
730 void drop_slab_node(int nid)
735 struct mem_cgroup *memcg = NULL;
738 memcg = mem_cgroup_iter(NULL, NULL, NULL);
740 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
741 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
742 } while (freed > 10);
749 for_each_online_node(nid)
753 static inline int is_page_cache_freeable(struct page *page)
756 * A freeable page cache page is referenced only by the caller
757 * that isolated the page, the page cache and optional buffer
758 * heads at page->private.
760 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
762 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
765 static int may_write_to_inode(struct inode *inode)
767 if (current->flags & PF_SWAPWRITE)
769 if (!inode_write_congested(inode))
771 if (inode_to_bdi(inode) == current->backing_dev_info)
777 * We detected a synchronous write error writing a page out. Probably
778 * -ENOSPC. We need to propagate that into the address_space for a subsequent
779 * fsync(), msync() or close().
781 * The tricky part is that after writepage we cannot touch the mapping: nothing
782 * prevents it from being freed up. But we have a ref on the page and once
783 * that page is locked, the mapping is pinned.
785 * We're allowed to run sleeping lock_page() here because we know the caller has
788 static void handle_write_error(struct address_space *mapping,
789 struct page *page, int error)
792 if (page_mapping(page) == mapping)
793 mapping_set_error(mapping, error);
797 /* possible outcome of pageout() */
799 /* failed to write page out, page is locked */
801 /* move page to the active list, page is locked */
803 /* page has been sent to the disk successfully, page is unlocked */
805 /* page is clean and locked */
810 * pageout is called by shrink_page_list() for each dirty page.
811 * Calls ->writepage().
813 static pageout_t pageout(struct page *page, struct address_space *mapping)
816 * If the page is dirty, only perform writeback if that write
817 * will be non-blocking. To prevent this allocation from being
818 * stalled by pagecache activity. But note that there may be
819 * stalls if we need to run get_block(). We could test
820 * PagePrivate for that.
822 * If this process is currently in __generic_file_write_iter() against
823 * this page's queue, we can perform writeback even if that
826 * If the page is swapcache, write it back even if that would
827 * block, for some throttling. This happens by accident, because
828 * swap_backing_dev_info is bust: it doesn't reflect the
829 * congestion state of the swapdevs. Easy to fix, if needed.
831 if (!is_page_cache_freeable(page))
835 * Some data journaling orphaned pages can have
836 * page->mapping == NULL while being dirty with clean buffers.
838 if (page_has_private(page)) {
839 if (try_to_free_buffers(page)) {
840 ClearPageDirty(page);
841 pr_info("%s: orphaned page\n", __func__);
847 if (mapping->a_ops->writepage == NULL)
848 return PAGE_ACTIVATE;
849 if (!may_write_to_inode(mapping->host))
852 if (clear_page_dirty_for_io(page)) {
854 struct writeback_control wbc = {
855 .sync_mode = WB_SYNC_NONE,
856 .nr_to_write = SWAP_CLUSTER_MAX,
858 .range_end = LLONG_MAX,
862 SetPageReclaim(page);
863 res = mapping->a_ops->writepage(page, &wbc);
865 handle_write_error(mapping, page, res);
866 if (res == AOP_WRITEPAGE_ACTIVATE) {
867 ClearPageReclaim(page);
868 return PAGE_ACTIVATE;
871 if (!PageWriteback(page)) {
872 /* synchronous write or broken a_ops? */
873 ClearPageReclaim(page);
875 trace_mm_vmscan_writepage(page);
876 inc_node_page_state(page, NR_VMSCAN_WRITE);
884 * Same as remove_mapping, but if the page is removed from the mapping, it
885 * gets returned with a refcount of 0.
887 static int __remove_mapping(struct address_space *mapping, struct page *page,
893 BUG_ON(!PageLocked(page));
894 BUG_ON(mapping != page_mapping(page));
896 xa_lock_irqsave(&mapping->i_pages, flags);
898 * The non racy check for a busy page.
900 * Must be careful with the order of the tests. When someone has
901 * a ref to the page, it may be possible that they dirty it then
902 * drop the reference. So if PageDirty is tested before page_count
903 * here, then the following race may occur:
905 * get_user_pages(&page);
906 * [user mapping goes away]
908 * !PageDirty(page) [good]
909 * SetPageDirty(page);
911 * !page_count(page) [good, discard it]
913 * [oops, our write_to data is lost]
915 * Reversing the order of the tests ensures such a situation cannot
916 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
917 * load is not satisfied before that of page->_refcount.
919 * Note that if SetPageDirty is always performed via set_page_dirty,
920 * and thus under the i_pages lock, then this ordering is not required.
922 refcount = 1 + compound_nr(page);
923 if (!page_ref_freeze(page, refcount))
925 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
926 if (unlikely(PageDirty(page))) {
927 page_ref_unfreeze(page, refcount);
931 if (PageSwapCache(page)) {
932 swp_entry_t swap = { .val = page_private(page) };
933 mem_cgroup_swapout(page, swap);
934 __delete_from_swap_cache(page, swap);
935 xa_unlock_irqrestore(&mapping->i_pages, flags);
936 put_swap_page(page, swap);
938 void (*freepage)(struct page *);
941 freepage = mapping->a_ops->freepage;
943 * Remember a shadow entry for reclaimed file cache in
944 * order to detect refaults, thus thrashing, later on.
946 * But don't store shadows in an address space that is
947 * already exiting. This is not just an optizimation,
948 * inode reclaim needs to empty out the radix tree or
949 * the nodes are lost. Don't plant shadows behind its
952 * We also don't store shadows for DAX mappings because the
953 * only page cache pages found in these are zero pages
954 * covering holes, and because we don't want to mix DAX
955 * exceptional entries and shadow exceptional entries in the
956 * same address_space.
958 if (reclaimed && page_is_file_cache(page) &&
959 !mapping_exiting(mapping) && !dax_mapping(mapping))
960 shadow = workingset_eviction(page);
961 __delete_from_page_cache(page, shadow);
962 xa_unlock_irqrestore(&mapping->i_pages, flags);
964 if (freepage != NULL)
971 xa_unlock_irqrestore(&mapping->i_pages, flags);
976 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
977 * someone else has a ref on the page, abort and return 0. If it was
978 * successfully detached, return 1. Assumes the caller has a single ref on
981 int remove_mapping(struct address_space *mapping, struct page *page)
983 if (__remove_mapping(mapping, page, false)) {
985 * Unfreezing the refcount with 1 rather than 2 effectively
986 * drops the pagecache ref for us without requiring another
989 page_ref_unfreeze(page, 1);
996 * putback_lru_page - put previously isolated page onto appropriate LRU list
997 * @page: page to be put back to appropriate lru list
999 * Add previously isolated @page to appropriate LRU list.
1000 * Page may still be unevictable for other reasons.
1002 * lru_lock must not be held, interrupts must be enabled.
1004 void putback_lru_page(struct page *page)
1006 lru_cache_add(page);
1007 put_page(page); /* drop ref from isolate */
1010 enum page_references {
1012 PAGEREF_RECLAIM_CLEAN,
1017 static enum page_references page_check_references(struct page *page,
1018 struct scan_control *sc)
1020 int referenced_ptes, referenced_page;
1021 unsigned long vm_flags;
1023 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1025 referenced_page = TestClearPageReferenced(page);
1028 * Mlock lost the isolation race with us. Let try_to_unmap()
1029 * move the page to the unevictable list.
1031 if (vm_flags & VM_LOCKED)
1032 return PAGEREF_RECLAIM;
1034 if (referenced_ptes) {
1035 if (PageSwapBacked(page))
1036 return PAGEREF_ACTIVATE;
1038 * All mapped pages start out with page table
1039 * references from the instantiating fault, so we need
1040 * to look twice if a mapped file page is used more
1043 * Mark it and spare it for another trip around the
1044 * inactive list. Another page table reference will
1045 * lead to its activation.
1047 * Note: the mark is set for activated pages as well
1048 * so that recently deactivated but used pages are
1049 * quickly recovered.
1051 SetPageReferenced(page);
1053 if (referenced_page || referenced_ptes > 1)
1054 return PAGEREF_ACTIVATE;
1057 * Activate file-backed executable pages after first usage.
1059 if (vm_flags & VM_EXEC)
1060 return PAGEREF_ACTIVATE;
1062 return PAGEREF_KEEP;
1065 /* Reclaim if clean, defer dirty pages to writeback */
1066 if (referenced_page && !PageSwapBacked(page))
1067 return PAGEREF_RECLAIM_CLEAN;
1069 return PAGEREF_RECLAIM;
1072 /* Check if a page is dirty or under writeback */
1073 static void page_check_dirty_writeback(struct page *page,
1074 bool *dirty, bool *writeback)
1076 struct address_space *mapping;
1079 * Anonymous pages are not handled by flushers and must be written
1080 * from reclaim context. Do not stall reclaim based on them
1082 if (!page_is_file_cache(page) ||
1083 (PageAnon(page) && !PageSwapBacked(page))) {
1089 /* By default assume that the page flags are accurate */
1090 *dirty = PageDirty(page);
1091 *writeback = PageWriteback(page);
1093 /* Verify dirty/writeback state if the filesystem supports it */
1094 if (!page_has_private(page))
1097 mapping = page_mapping(page);
1098 if (mapping && mapping->a_ops->is_dirty_writeback)
1099 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1103 * shrink_page_list() returns the number of reclaimed pages
1105 static unsigned long shrink_page_list(struct list_head *page_list,
1106 struct pglist_data *pgdat,
1107 struct scan_control *sc,
1108 enum ttu_flags ttu_flags,
1109 struct reclaim_stat *stat,
1110 bool ignore_references)
1112 LIST_HEAD(ret_pages);
1113 LIST_HEAD(free_pages);
1114 unsigned nr_reclaimed = 0;
1115 unsigned pgactivate = 0;
1117 memset(stat, 0, sizeof(*stat));
1120 while (!list_empty(page_list)) {
1121 struct address_space *mapping;
1124 enum page_references references = PAGEREF_RECLAIM;
1125 bool dirty, writeback;
1126 unsigned int nr_pages;
1130 page = lru_to_page(page_list);
1131 list_del(&page->lru);
1133 if (!trylock_page(page))
1136 VM_BUG_ON_PAGE(PageActive(page), page);
1138 nr_pages = compound_nr(page);
1140 /* Account the number of base pages even though THP */
1141 sc->nr_scanned += nr_pages;
1143 if (unlikely(!page_evictable(page)))
1144 goto activate_locked;
1146 if (!sc->may_unmap && page_mapped(page))
1149 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1150 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1153 * The number of dirty pages determines if a node is marked
1154 * reclaim_congested which affects wait_iff_congested. kswapd
1155 * will stall and start writing pages if the tail of the LRU
1156 * is all dirty unqueued pages.
1158 page_check_dirty_writeback(page, &dirty, &writeback);
1159 if (dirty || writeback)
1162 if (dirty && !writeback)
1163 stat->nr_unqueued_dirty++;
1166 * Treat this page as congested if the underlying BDI is or if
1167 * pages are cycling through the LRU so quickly that the
1168 * pages marked for immediate reclaim are making it to the
1169 * end of the LRU a second time.
1171 mapping = page_mapping(page);
1172 if (((dirty || writeback) && mapping &&
1173 inode_write_congested(mapping->host)) ||
1174 (writeback && PageReclaim(page)))
1175 stat->nr_congested++;
1178 * If a page at the tail of the LRU is under writeback, there
1179 * are three cases to consider.
1181 * 1) If reclaim is encountering an excessive number of pages
1182 * under writeback and this page is both under writeback and
1183 * PageReclaim then it indicates that pages are being queued
1184 * for IO but are being recycled through the LRU before the
1185 * IO can complete. Waiting on the page itself risks an
1186 * indefinite stall if it is impossible to writeback the
1187 * page due to IO error or disconnected storage so instead
1188 * note that the LRU is being scanned too quickly and the
1189 * caller can stall after page list has been processed.
1191 * 2) Global or new memcg reclaim encounters a page that is
1192 * not marked for immediate reclaim, or the caller does not
1193 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1194 * not to fs). In this case mark the page for immediate
1195 * reclaim and continue scanning.
1197 * Require may_enter_fs because we would wait on fs, which
1198 * may not have submitted IO yet. And the loop driver might
1199 * enter reclaim, and deadlock if it waits on a page for
1200 * which it is needed to do the write (loop masks off
1201 * __GFP_IO|__GFP_FS for this reason); but more thought
1202 * would probably show more reasons.
1204 * 3) Legacy memcg encounters a page that is already marked
1205 * PageReclaim. memcg does not have any dirty pages
1206 * throttling so we could easily OOM just because too many
1207 * pages are in writeback and there is nothing else to
1208 * reclaim. Wait for the writeback to complete.
1210 * In cases 1) and 2) we activate the pages to get them out of
1211 * the way while we continue scanning for clean pages on the
1212 * inactive list and refilling from the active list. The
1213 * observation here is that waiting for disk writes is more
1214 * expensive than potentially causing reloads down the line.
1215 * Since they're marked for immediate reclaim, they won't put
1216 * memory pressure on the cache working set any longer than it
1217 * takes to write them to disk.
1219 if (PageWriteback(page)) {
1221 if (current_is_kswapd() &&
1222 PageReclaim(page) &&
1223 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1224 stat->nr_immediate++;
1225 goto activate_locked;
1228 } else if (writeback_throttling_sane(sc) ||
1229 !PageReclaim(page) || !may_enter_fs) {
1231 * This is slightly racy - end_page_writeback()
1232 * might have just cleared PageReclaim, then
1233 * setting PageReclaim here end up interpreted
1234 * as PageReadahead - but that does not matter
1235 * enough to care. What we do want is for this
1236 * page to have PageReclaim set next time memcg
1237 * reclaim reaches the tests above, so it will
1238 * then wait_on_page_writeback() to avoid OOM;
1239 * and it's also appropriate in global reclaim.
1241 SetPageReclaim(page);
1242 stat->nr_writeback++;
1243 goto activate_locked;
1248 wait_on_page_writeback(page);
1249 /* then go back and try same page again */
1250 list_add_tail(&page->lru, page_list);
1255 if (!ignore_references)
1256 references = page_check_references(page, sc);
1258 switch (references) {
1259 case PAGEREF_ACTIVATE:
1260 goto activate_locked;
1262 stat->nr_ref_keep += nr_pages;
1264 case PAGEREF_RECLAIM:
1265 case PAGEREF_RECLAIM_CLEAN:
1266 ; /* try to reclaim the page below */
1270 * Anonymous process memory has backing store?
1271 * Try to allocate it some swap space here.
1272 * Lazyfree page could be freed directly
1274 if (PageAnon(page) && PageSwapBacked(page)) {
1275 if (!PageSwapCache(page)) {
1276 if (!(sc->gfp_mask & __GFP_IO))
1278 if (PageTransHuge(page)) {
1279 /* cannot split THP, skip it */
1280 if (!can_split_huge_page(page, NULL))
1281 goto activate_locked;
1283 * Split pages without a PMD map right
1284 * away. Chances are some or all of the
1285 * tail pages can be freed without IO.
1287 if (!compound_mapcount(page) &&
1288 split_huge_page_to_list(page,
1290 goto activate_locked;
1292 if (!add_to_swap(page)) {
1293 if (!PageTransHuge(page))
1294 goto activate_locked_split;
1295 /* Fallback to swap normal pages */
1296 if (split_huge_page_to_list(page,
1298 goto activate_locked;
1299 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1300 count_vm_event(THP_SWPOUT_FALLBACK);
1302 if (!add_to_swap(page))
1303 goto activate_locked_split;
1308 /* Adding to swap updated mapping */
1309 mapping = page_mapping(page);
1311 } else if (unlikely(PageTransHuge(page))) {
1312 /* Split file THP */
1313 if (split_huge_page_to_list(page, page_list))
1318 * THP may get split above, need minus tail pages and update
1319 * nr_pages to avoid accounting tail pages twice.
1321 * The tail pages that are added into swap cache successfully
1324 if ((nr_pages > 1) && !PageTransHuge(page)) {
1325 sc->nr_scanned -= (nr_pages - 1);
1330 * The page is mapped into the page tables of one or more
1331 * processes. Try to unmap it here.
1333 if (page_mapped(page)) {
1334 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1336 if (unlikely(PageTransHuge(page)))
1337 flags |= TTU_SPLIT_HUGE_PMD;
1338 if (!try_to_unmap(page, flags)) {
1339 stat->nr_unmap_fail += nr_pages;
1340 goto activate_locked;
1344 if (PageDirty(page)) {
1346 * Only kswapd can writeback filesystem pages
1347 * to avoid risk of stack overflow. But avoid
1348 * injecting inefficient single-page IO into
1349 * flusher writeback as much as possible: only
1350 * write pages when we've encountered many
1351 * dirty pages, and when we've already scanned
1352 * the rest of the LRU for clean pages and see
1353 * the same dirty pages again (PageReclaim).
1355 if (page_is_file_cache(page) &&
1356 (!current_is_kswapd() || !PageReclaim(page) ||
1357 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1359 * Immediately reclaim when written back.
1360 * Similar in principal to deactivate_page()
1361 * except we already have the page isolated
1362 * and know it's dirty
1364 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1365 SetPageReclaim(page);
1367 goto activate_locked;
1370 if (references == PAGEREF_RECLAIM_CLEAN)
1374 if (!sc->may_writepage)
1378 * Page is dirty. Flush the TLB if a writable entry
1379 * potentially exists to avoid CPU writes after IO
1380 * starts and then write it out here.
1382 try_to_unmap_flush_dirty();
1383 switch (pageout(page, mapping)) {
1387 goto activate_locked;
1389 if (PageWriteback(page))
1391 if (PageDirty(page))
1395 * A synchronous write - probably a ramdisk. Go
1396 * ahead and try to reclaim the page.
1398 if (!trylock_page(page))
1400 if (PageDirty(page) || PageWriteback(page))
1402 mapping = page_mapping(page);
1404 ; /* try to free the page below */
1409 * If the page has buffers, try to free the buffer mappings
1410 * associated with this page. If we succeed we try to free
1413 * We do this even if the page is PageDirty().
1414 * try_to_release_page() does not perform I/O, but it is
1415 * possible for a page to have PageDirty set, but it is actually
1416 * clean (all its buffers are clean). This happens if the
1417 * buffers were written out directly, with submit_bh(). ext3
1418 * will do this, as well as the blockdev mapping.
1419 * try_to_release_page() will discover that cleanness and will
1420 * drop the buffers and mark the page clean - it can be freed.
1422 * Rarely, pages can have buffers and no ->mapping. These are
1423 * the pages which were not successfully invalidated in
1424 * truncate_complete_page(). We try to drop those buffers here
1425 * and if that worked, and the page is no longer mapped into
1426 * process address space (page_count == 1) it can be freed.
1427 * Otherwise, leave the page on the LRU so it is swappable.
1429 if (page_has_private(page)) {
1430 if (!try_to_release_page(page, sc->gfp_mask))
1431 goto activate_locked;
1432 if (!mapping && page_count(page) == 1) {
1434 if (put_page_testzero(page))
1438 * rare race with speculative reference.
1439 * the speculative reference will free
1440 * this page shortly, so we may
1441 * increment nr_reclaimed here (and
1442 * leave it off the LRU).
1450 if (PageAnon(page) && !PageSwapBacked(page)) {
1451 /* follow __remove_mapping for reference */
1452 if (!page_ref_freeze(page, 1))
1454 if (PageDirty(page)) {
1455 page_ref_unfreeze(page, 1);
1459 count_vm_event(PGLAZYFREED);
1460 count_memcg_page_event(page, PGLAZYFREED);
1461 } else if (!mapping || !__remove_mapping(mapping, page, true))
1467 * THP may get swapped out in a whole, need account
1470 nr_reclaimed += nr_pages;
1473 * Is there need to periodically free_page_list? It would
1474 * appear not as the counts should be low
1476 if (unlikely(PageTransHuge(page)))
1477 (*get_compound_page_dtor(page))(page);
1479 list_add(&page->lru, &free_pages);
1482 activate_locked_split:
1484 * The tail pages that are failed to add into swap cache
1485 * reach here. Fixup nr_scanned and nr_pages.
1488 sc->nr_scanned -= (nr_pages - 1);
1492 /* Not a candidate for swapping, so reclaim swap space. */
1493 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1495 try_to_free_swap(page);
1496 VM_BUG_ON_PAGE(PageActive(page), page);
1497 if (!PageMlocked(page)) {
1498 int type = page_is_file_cache(page);
1499 SetPageActive(page);
1500 stat->nr_activate[type] += nr_pages;
1501 count_memcg_page_event(page, PGACTIVATE);
1506 list_add(&page->lru, &ret_pages);
1507 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1510 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1512 mem_cgroup_uncharge_list(&free_pages);
1513 try_to_unmap_flush();
1514 free_unref_page_list(&free_pages);
1516 list_splice(&ret_pages, page_list);
1517 count_vm_events(PGACTIVATE, pgactivate);
1519 return nr_reclaimed;
1522 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1523 struct list_head *page_list)
1525 struct scan_control sc = {
1526 .gfp_mask = GFP_KERNEL,
1527 .priority = DEF_PRIORITY,
1530 struct reclaim_stat dummy_stat;
1532 struct page *page, *next;
1533 LIST_HEAD(clean_pages);
1535 list_for_each_entry_safe(page, next, page_list, lru) {
1536 if (page_is_file_cache(page) && !PageDirty(page) &&
1537 !__PageMovable(page) && !PageUnevictable(page)) {
1538 ClearPageActive(page);
1539 list_move(&page->lru, &clean_pages);
1543 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1544 TTU_IGNORE_ACCESS, &dummy_stat, true);
1545 list_splice(&clean_pages, page_list);
1546 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1551 * Attempt to remove the specified page from its LRU. Only take this page
1552 * if it is of the appropriate PageActive status. Pages which are being
1553 * freed elsewhere are also ignored.
1555 * page: page to consider
1556 * mode: one of the LRU isolation modes defined above
1558 * returns 0 on success, -ve errno on failure.
1560 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1564 /* Only take pages on the LRU. */
1568 /* Compaction should not handle unevictable pages but CMA can do so */
1569 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1575 * To minimise LRU disruption, the caller can indicate that it only
1576 * wants to isolate pages it will be able to operate on without
1577 * blocking - clean pages for the most part.
1579 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1580 * that it is possible to migrate without blocking
1582 if (mode & ISOLATE_ASYNC_MIGRATE) {
1583 /* All the caller can do on PageWriteback is block */
1584 if (PageWriteback(page))
1587 if (PageDirty(page)) {
1588 struct address_space *mapping;
1592 * Only pages without mappings or that have a
1593 * ->migratepage callback are possible to migrate
1594 * without blocking. However, we can be racing with
1595 * truncation so it's necessary to lock the page
1596 * to stabilise the mapping as truncation holds
1597 * the page lock until after the page is removed
1598 * from the page cache.
1600 if (!trylock_page(page))
1603 mapping = page_mapping(page);
1604 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1611 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1614 if (likely(get_page_unless_zero(page))) {
1616 * Be careful not to clear PageLRU until after we're
1617 * sure the page is not being freed elsewhere -- the
1618 * page release code relies on it.
1629 * Update LRU sizes after isolating pages. The LRU size updates must
1630 * be complete before mem_cgroup_update_lru_size due to a santity check.
1632 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1633 enum lru_list lru, unsigned long *nr_zone_taken)
1637 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1638 if (!nr_zone_taken[zid])
1641 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1643 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1650 * pgdat->lru_lock is heavily contended. Some of the functions that
1651 * shrink the lists perform better by taking out a batch of pages
1652 * and working on them outside the LRU lock.
1654 * For pagecache intensive workloads, this function is the hottest
1655 * spot in the kernel (apart from copy_*_user functions).
1657 * Appropriate locks must be held before calling this function.
1659 * @nr_to_scan: The number of eligible pages to look through on the list.
1660 * @lruvec: The LRU vector to pull pages from.
1661 * @dst: The temp list to put pages on to.
1662 * @nr_scanned: The number of pages that were scanned.
1663 * @sc: The scan_control struct for this reclaim session
1664 * @mode: One of the LRU isolation modes
1665 * @lru: LRU list id for isolating
1667 * returns how many pages were moved onto *@dst.
1669 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1670 struct lruvec *lruvec, struct list_head *dst,
1671 unsigned long *nr_scanned, struct scan_control *sc,
1674 struct list_head *src = &lruvec->lists[lru];
1675 unsigned long nr_taken = 0;
1676 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1677 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1678 unsigned long skipped = 0;
1679 unsigned long scan, total_scan, nr_pages;
1680 LIST_HEAD(pages_skipped);
1681 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1685 while (scan < nr_to_scan && !list_empty(src)) {
1688 page = lru_to_page(src);
1689 prefetchw_prev_lru_page(page, src, flags);
1691 VM_BUG_ON_PAGE(!PageLRU(page), page);
1693 nr_pages = compound_nr(page);
1694 total_scan += nr_pages;
1696 if (page_zonenum(page) > sc->reclaim_idx) {
1697 list_move(&page->lru, &pages_skipped);
1698 nr_skipped[page_zonenum(page)] += nr_pages;
1703 * Do not count skipped pages because that makes the function
1704 * return with no isolated pages if the LRU mostly contains
1705 * ineligible pages. This causes the VM to not reclaim any
1706 * pages, triggering a premature OOM.
1708 * Account all tail pages of THP. This would not cause
1709 * premature OOM since __isolate_lru_page() returns -EBUSY
1710 * only when the page is being freed somewhere else.
1713 switch (__isolate_lru_page(page, mode)) {
1715 nr_taken += nr_pages;
1716 nr_zone_taken[page_zonenum(page)] += nr_pages;
1717 list_move(&page->lru, dst);
1721 /* else it is being freed elsewhere */
1722 list_move(&page->lru, src);
1731 * Splice any skipped pages to the start of the LRU list. Note that
1732 * this disrupts the LRU order when reclaiming for lower zones but
1733 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1734 * scanning would soon rescan the same pages to skip and put the
1735 * system at risk of premature OOM.
1737 if (!list_empty(&pages_skipped)) {
1740 list_splice(&pages_skipped, src);
1741 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1742 if (!nr_skipped[zid])
1745 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1746 skipped += nr_skipped[zid];
1749 *nr_scanned = total_scan;
1750 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1751 total_scan, skipped, nr_taken, mode, lru);
1752 update_lru_sizes(lruvec, lru, nr_zone_taken);
1757 * isolate_lru_page - tries to isolate a page from its LRU list
1758 * @page: page to isolate from its LRU list
1760 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1761 * vmstat statistic corresponding to whatever LRU list the page was on.
1763 * Returns 0 if the page was removed from an LRU list.
1764 * Returns -EBUSY if the page was not on an LRU list.
1766 * The returned page will have PageLRU() cleared. If it was found on
1767 * the active list, it will have PageActive set. If it was found on
1768 * the unevictable list, it will have the PageUnevictable bit set. That flag
1769 * may need to be cleared by the caller before letting the page go.
1771 * The vmstat statistic corresponding to the list on which the page was
1772 * found will be decremented.
1776 * (1) Must be called with an elevated refcount on the page. This is a
1777 * fundamentnal difference from isolate_lru_pages (which is called
1778 * without a stable reference).
1779 * (2) the lru_lock must not be held.
1780 * (3) interrupts must be enabled.
1782 int isolate_lru_page(struct page *page)
1786 VM_BUG_ON_PAGE(!page_count(page), page);
1787 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1789 if (PageLRU(page)) {
1790 pg_data_t *pgdat = page_pgdat(page);
1791 struct lruvec *lruvec;
1793 spin_lock_irq(&pgdat->lru_lock);
1794 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1795 if (PageLRU(page)) {
1796 int lru = page_lru(page);
1799 del_page_from_lru_list(page, lruvec, lru);
1802 spin_unlock_irq(&pgdat->lru_lock);
1808 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1809 * then get resheduled. When there are massive number of tasks doing page
1810 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1811 * the LRU list will go small and be scanned faster than necessary, leading to
1812 * unnecessary swapping, thrashing and OOM.
1814 static int too_many_isolated(struct pglist_data *pgdat, int file,
1815 struct scan_control *sc)
1817 unsigned long inactive, isolated;
1819 if (current_is_kswapd())
1822 if (!writeback_throttling_sane(sc))
1826 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1827 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1829 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1830 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1834 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1835 * won't get blocked by normal direct-reclaimers, forming a circular
1838 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1841 return isolated > inactive;
1845 * This moves pages from @list to corresponding LRU list.
1847 * We move them the other way if the page is referenced by one or more
1848 * processes, from rmap.
1850 * If the pages are mostly unmapped, the processing is fast and it is
1851 * appropriate to hold zone_lru_lock across the whole operation. But if
1852 * the pages are mapped, the processing is slow (page_referenced()) so we
1853 * should drop zone_lru_lock around each page. It's impossible to balance
1854 * this, so instead we remove the pages from the LRU while processing them.
1855 * It is safe to rely on PG_active against the non-LRU pages in here because
1856 * nobody will play with that bit on a non-LRU page.
1858 * The downside is that we have to touch page->_refcount against each page.
1859 * But we had to alter page->flags anyway.
1861 * Returns the number of pages moved to the given lruvec.
1864 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1865 struct list_head *list)
1867 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1868 int nr_pages, nr_moved = 0;
1869 LIST_HEAD(pages_to_free);
1873 while (!list_empty(list)) {
1874 page = lru_to_page(list);
1875 VM_BUG_ON_PAGE(PageLRU(page), page);
1876 if (unlikely(!page_evictable(page))) {
1877 list_del(&page->lru);
1878 spin_unlock_irq(&pgdat->lru_lock);
1879 putback_lru_page(page);
1880 spin_lock_irq(&pgdat->lru_lock);
1883 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1886 lru = page_lru(page);
1888 nr_pages = hpage_nr_pages(page);
1889 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1890 list_move(&page->lru, &lruvec->lists[lru]);
1892 if (put_page_testzero(page)) {
1893 __ClearPageLRU(page);
1894 __ClearPageActive(page);
1895 del_page_from_lru_list(page, lruvec, lru);
1897 if (unlikely(PageCompound(page))) {
1898 spin_unlock_irq(&pgdat->lru_lock);
1899 (*get_compound_page_dtor(page))(page);
1900 spin_lock_irq(&pgdat->lru_lock);
1902 list_add(&page->lru, &pages_to_free);
1904 nr_moved += nr_pages;
1909 * To save our caller's stack, now use input list for pages to free.
1911 list_splice(&pages_to_free, list);
1917 * If a kernel thread (such as nfsd for loop-back mounts) services
1918 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1919 * In that case we should only throttle if the backing device it is
1920 * writing to is congested. In other cases it is safe to throttle.
1922 static int current_may_throttle(void)
1924 return !(current->flags & PF_LESS_THROTTLE) ||
1925 current->backing_dev_info == NULL ||
1926 bdi_write_congested(current->backing_dev_info);
1930 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1931 * of reclaimed pages
1933 static noinline_for_stack unsigned long
1934 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1935 struct scan_control *sc, enum lru_list lru)
1937 LIST_HEAD(page_list);
1938 unsigned long nr_scanned;
1939 unsigned long nr_reclaimed = 0;
1940 unsigned long nr_taken;
1941 struct reclaim_stat stat;
1942 int file = is_file_lru(lru);
1943 enum vm_event_item item;
1944 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1945 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1946 bool stalled = false;
1948 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1952 /* wait a bit for the reclaimer. */
1956 /* We are about to die and free our memory. Return now. */
1957 if (fatal_signal_pending(current))
1958 return SWAP_CLUSTER_MAX;
1963 spin_lock_irq(&pgdat->lru_lock);
1965 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1966 &nr_scanned, sc, lru);
1968 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1969 reclaim_stat->recent_scanned[file] += nr_taken;
1971 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1972 if (!cgroup_reclaim(sc))
1973 __count_vm_events(item, nr_scanned);
1974 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1975 spin_unlock_irq(&pgdat->lru_lock);
1980 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1983 spin_lock_irq(&pgdat->lru_lock);
1985 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1986 if (!cgroup_reclaim(sc))
1987 __count_vm_events(item, nr_reclaimed);
1988 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1989 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1990 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1992 move_pages_to_lru(lruvec, &page_list);
1994 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1996 spin_unlock_irq(&pgdat->lru_lock);
1998 mem_cgroup_uncharge_list(&page_list);
1999 free_unref_page_list(&page_list);
2002 * If dirty pages are scanned that are not queued for IO, it
2003 * implies that flushers are not doing their job. This can
2004 * happen when memory pressure pushes dirty pages to the end of
2005 * the LRU before the dirty limits are breached and the dirty
2006 * data has expired. It can also happen when the proportion of
2007 * dirty pages grows not through writes but through memory
2008 * pressure reclaiming all the clean cache. And in some cases,
2009 * the flushers simply cannot keep up with the allocation
2010 * rate. Nudge the flusher threads in case they are asleep.
2012 if (stat.nr_unqueued_dirty == nr_taken)
2013 wakeup_flusher_threads(WB_REASON_VMSCAN);
2015 sc->nr.dirty += stat.nr_dirty;
2016 sc->nr.congested += stat.nr_congested;
2017 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2018 sc->nr.writeback += stat.nr_writeback;
2019 sc->nr.immediate += stat.nr_immediate;
2020 sc->nr.taken += nr_taken;
2022 sc->nr.file_taken += nr_taken;
2024 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2025 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2026 return nr_reclaimed;
2029 static void shrink_active_list(unsigned long nr_to_scan,
2030 struct lruvec *lruvec,
2031 struct scan_control *sc,
2034 unsigned long nr_taken;
2035 unsigned long nr_scanned;
2036 unsigned long vm_flags;
2037 LIST_HEAD(l_hold); /* The pages which were snipped off */
2038 LIST_HEAD(l_active);
2039 LIST_HEAD(l_inactive);
2041 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2042 unsigned nr_deactivate, nr_activate;
2043 unsigned nr_rotated = 0;
2044 int file = is_file_lru(lru);
2045 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2049 spin_lock_irq(&pgdat->lru_lock);
2051 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2052 &nr_scanned, sc, lru);
2054 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2055 reclaim_stat->recent_scanned[file] += nr_taken;
2057 __count_vm_events(PGREFILL, nr_scanned);
2058 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2060 spin_unlock_irq(&pgdat->lru_lock);
2062 while (!list_empty(&l_hold)) {
2064 page = lru_to_page(&l_hold);
2065 list_del(&page->lru);
2067 if (unlikely(!page_evictable(page))) {
2068 putback_lru_page(page);
2072 if (unlikely(buffer_heads_over_limit)) {
2073 if (page_has_private(page) && trylock_page(page)) {
2074 if (page_has_private(page))
2075 try_to_release_page(page, 0);
2080 if (page_referenced(page, 0, sc->target_mem_cgroup,
2082 nr_rotated += hpage_nr_pages(page);
2084 * Identify referenced, file-backed active pages and
2085 * give them one more trip around the active list. So
2086 * that executable code get better chances to stay in
2087 * memory under moderate memory pressure. Anon pages
2088 * are not likely to be evicted by use-once streaming
2089 * IO, plus JVM can create lots of anon VM_EXEC pages,
2090 * so we ignore them here.
2092 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2093 list_add(&page->lru, &l_active);
2098 ClearPageActive(page); /* we are de-activating */
2099 SetPageWorkingset(page);
2100 list_add(&page->lru, &l_inactive);
2104 * Move pages back to the lru list.
2106 spin_lock_irq(&pgdat->lru_lock);
2108 * Count referenced pages from currently used mappings as rotated,
2109 * even though only some of them are actually re-activated. This
2110 * helps balance scan pressure between file and anonymous pages in
2113 reclaim_stat->recent_rotated[file] += nr_rotated;
2115 nr_activate = move_pages_to_lru(lruvec, &l_active);
2116 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2117 /* Keep all free pages in l_active list */
2118 list_splice(&l_inactive, &l_active);
2120 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2121 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2123 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2124 spin_unlock_irq(&pgdat->lru_lock);
2126 mem_cgroup_uncharge_list(&l_active);
2127 free_unref_page_list(&l_active);
2128 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2129 nr_deactivate, nr_rotated, sc->priority, file);
2132 unsigned long reclaim_pages(struct list_head *page_list)
2135 unsigned long nr_reclaimed = 0;
2136 LIST_HEAD(node_page_list);
2137 struct reclaim_stat dummy_stat;
2139 struct scan_control sc = {
2140 .gfp_mask = GFP_KERNEL,
2141 .priority = DEF_PRIORITY,
2147 while (!list_empty(page_list)) {
2148 page = lru_to_page(page_list);
2150 nid = page_to_nid(page);
2151 INIT_LIST_HEAD(&node_page_list);
2154 if (nid == page_to_nid(page)) {
2155 ClearPageActive(page);
2156 list_move(&page->lru, &node_page_list);
2160 nr_reclaimed += shrink_page_list(&node_page_list,
2163 &dummy_stat, false);
2164 while (!list_empty(&node_page_list)) {
2165 page = lru_to_page(&node_page_list);
2166 list_del(&page->lru);
2167 putback_lru_page(page);
2173 if (!list_empty(&node_page_list)) {
2174 nr_reclaimed += shrink_page_list(&node_page_list,
2177 &dummy_stat, false);
2178 while (!list_empty(&node_page_list)) {
2179 page = lru_to_page(&node_page_list);
2180 list_del(&page->lru);
2181 putback_lru_page(page);
2185 return nr_reclaimed;
2189 * The inactive anon list should be small enough that the VM never has
2190 * to do too much work.
2192 * The inactive file list should be small enough to leave most memory
2193 * to the established workingset on the scan-resistant active list,
2194 * but large enough to avoid thrashing the aggregate readahead window.
2196 * Both inactive lists should also be large enough that each inactive
2197 * page has a chance to be referenced again before it is reclaimed.
2199 * If that fails and refaulting is observed, the inactive list grows.
2201 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2202 * on this LRU, maintained by the pageout code. An inactive_ratio
2203 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2206 * memory ratio inactive
2207 * -------------------------------------
2216 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2217 struct scan_control *sc, bool trace)
2219 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2220 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2221 enum lru_list inactive_lru = file * LRU_FILE;
2222 unsigned long inactive, active;
2223 unsigned long inactive_ratio;
2224 unsigned long refaults;
2227 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2228 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2231 * When refaults are being observed, it means a new workingset
2232 * is being established. Disable active list protection to get
2233 * rid of the stale workingset quickly.
2235 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2236 if (file && lruvec->refaults != refaults) {
2239 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2241 inactive_ratio = int_sqrt(10 * gb);
2247 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2248 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2249 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2250 inactive_ratio, file);
2252 return inactive * inactive_ratio < active;
2255 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2256 struct lruvec *lruvec, struct scan_control *sc)
2258 if (is_active_lru(lru)) {
2259 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2260 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2264 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2275 * Determine how aggressively the anon and file LRU lists should be
2276 * scanned. The relative value of each set of LRU lists is determined
2277 * by looking at the fraction of the pages scanned we did rotate back
2278 * onto the active list instead of evict.
2280 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2281 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2283 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2284 struct scan_control *sc, unsigned long *nr)
2286 int swappiness = mem_cgroup_swappiness(memcg);
2287 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2289 u64 denominator = 0; /* gcc */
2290 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2291 unsigned long anon_prio, file_prio;
2292 enum scan_balance scan_balance;
2293 unsigned long anon, file;
2294 unsigned long ap, fp;
2297 /* If we have no swap space, do not bother scanning anon pages. */
2298 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2299 scan_balance = SCAN_FILE;
2304 * Global reclaim will swap to prevent OOM even with no
2305 * swappiness, but memcg users want to use this knob to
2306 * disable swapping for individual groups completely when
2307 * using the memory controller's swap limit feature would be
2310 if (cgroup_reclaim(sc) && !swappiness) {
2311 scan_balance = SCAN_FILE;
2316 * Do not apply any pressure balancing cleverness when the
2317 * system is close to OOM, scan both anon and file equally
2318 * (unless the swappiness setting disagrees with swapping).
2320 if (!sc->priority && swappiness) {
2321 scan_balance = SCAN_EQUAL;
2326 * Prevent the reclaimer from falling into the cache trap: as
2327 * cache pages start out inactive, every cache fault will tip
2328 * the scan balance towards the file LRU. And as the file LRU
2329 * shrinks, so does the window for rotation from references.
2330 * This means we have a runaway feedback loop where a tiny
2331 * thrashing file LRU becomes infinitely more attractive than
2332 * anon pages. Try to detect this based on file LRU size.
2334 if (!cgroup_reclaim(sc)) {
2335 unsigned long pgdatfile;
2336 unsigned long pgdatfree;
2338 unsigned long total_high_wmark = 0;
2340 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2341 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2342 node_page_state(pgdat, NR_INACTIVE_FILE);
2344 for (z = 0; z < MAX_NR_ZONES; z++) {
2345 struct zone *zone = &pgdat->node_zones[z];
2346 if (!managed_zone(zone))
2349 total_high_wmark += high_wmark_pages(zone);
2352 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2354 * Force SCAN_ANON if there are enough inactive
2355 * anonymous pages on the LRU in eligible zones.
2356 * Otherwise, the small LRU gets thrashed.
2358 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2359 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2361 scan_balance = SCAN_ANON;
2368 * If there is enough inactive page cache, i.e. if the size of the
2369 * inactive list is greater than that of the active list *and* the
2370 * inactive list actually has some pages to scan on this priority, we
2371 * do not reclaim anything from the anonymous working set right now.
2372 * Without the second condition we could end up never scanning an
2373 * lruvec even if it has plenty of old anonymous pages unless the
2374 * system is under heavy pressure.
2376 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2377 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2378 scan_balance = SCAN_FILE;
2382 scan_balance = SCAN_FRACT;
2385 * With swappiness at 100, anonymous and file have the same priority.
2386 * This scanning priority is essentially the inverse of IO cost.
2388 anon_prio = swappiness;
2389 file_prio = 200 - anon_prio;
2392 * OK, so we have swap space and a fair amount of page cache
2393 * pages. We use the recently rotated / recently scanned
2394 * ratios to determine how valuable each cache is.
2396 * Because workloads change over time (and to avoid overflow)
2397 * we keep these statistics as a floating average, which ends
2398 * up weighing recent references more than old ones.
2400 * anon in [0], file in [1]
2403 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2404 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2405 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2406 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2408 spin_lock_irq(&pgdat->lru_lock);
2409 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2410 reclaim_stat->recent_scanned[0] /= 2;
2411 reclaim_stat->recent_rotated[0] /= 2;
2414 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2415 reclaim_stat->recent_scanned[1] /= 2;
2416 reclaim_stat->recent_rotated[1] /= 2;
2420 * The amount of pressure on anon vs file pages is inversely
2421 * proportional to the fraction of recently scanned pages on
2422 * each list that were recently referenced and in active use.
2424 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2425 ap /= reclaim_stat->recent_rotated[0] + 1;
2427 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2428 fp /= reclaim_stat->recent_rotated[1] + 1;
2429 spin_unlock_irq(&pgdat->lru_lock);
2433 denominator = ap + fp + 1;
2435 for_each_evictable_lru(lru) {
2436 int file = is_file_lru(lru);
2437 unsigned long lruvec_size;
2439 unsigned long protection;
2441 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2442 protection = mem_cgroup_protection(memcg,
2443 sc->memcg_low_reclaim);
2447 * Scale a cgroup's reclaim pressure by proportioning
2448 * its current usage to its memory.low or memory.min
2451 * This is important, as otherwise scanning aggression
2452 * becomes extremely binary -- from nothing as we
2453 * approach the memory protection threshold, to totally
2454 * nominal as we exceed it. This results in requiring
2455 * setting extremely liberal protection thresholds. It
2456 * also means we simply get no protection at all if we
2457 * set it too low, which is not ideal.
2459 * If there is any protection in place, we reduce scan
2460 * pressure by how much of the total memory used is
2461 * within protection thresholds.
2463 * There is one special case: in the first reclaim pass,
2464 * we skip over all groups that are within their low
2465 * protection. If that fails to reclaim enough pages to
2466 * satisfy the reclaim goal, we come back and override
2467 * the best-effort low protection. However, we still
2468 * ideally want to honor how well-behaved groups are in
2469 * that case instead of simply punishing them all
2470 * equally. As such, we reclaim them based on how much
2471 * memory they are using, reducing the scan pressure
2472 * again by how much of the total memory used is under
2475 unsigned long cgroup_size = mem_cgroup_size(memcg);
2477 /* Avoid TOCTOU with earlier protection check */
2478 cgroup_size = max(cgroup_size, protection);
2480 scan = lruvec_size - lruvec_size * protection /
2484 * Minimally target SWAP_CLUSTER_MAX pages to keep
2485 * reclaim moving forwards, avoiding decremeting
2486 * sc->priority further than desirable.
2488 scan = max(scan, SWAP_CLUSTER_MAX);
2493 scan >>= sc->priority;
2496 * If the cgroup's already been deleted, make sure to
2497 * scrape out the remaining cache.
2499 if (!scan && !mem_cgroup_online(memcg))
2500 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2502 switch (scan_balance) {
2504 /* Scan lists relative to size */
2508 * Scan types proportional to swappiness and
2509 * their relative recent reclaim efficiency.
2510 * Make sure we don't miss the last page
2511 * because of a round-off error.
2513 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2518 /* Scan one type exclusively */
2519 if ((scan_balance == SCAN_FILE) != file) {
2525 /* Look ma, no brain */
2534 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2536 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2537 struct scan_control *sc)
2539 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2540 unsigned long nr[NR_LRU_LISTS];
2541 unsigned long targets[NR_LRU_LISTS];
2542 unsigned long nr_to_scan;
2544 unsigned long nr_reclaimed = 0;
2545 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2546 struct blk_plug plug;
2549 get_scan_count(lruvec, memcg, sc, nr);
2551 /* Record the original scan target for proportional adjustments later */
2552 memcpy(targets, nr, sizeof(nr));
2555 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2556 * event that can occur when there is little memory pressure e.g.
2557 * multiple streaming readers/writers. Hence, we do not abort scanning
2558 * when the requested number of pages are reclaimed when scanning at
2559 * DEF_PRIORITY on the assumption that the fact we are direct
2560 * reclaiming implies that kswapd is not keeping up and it is best to
2561 * do a batch of work at once. For memcg reclaim one check is made to
2562 * abort proportional reclaim if either the file or anon lru has already
2563 * dropped to zero at the first pass.
2565 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2566 sc->priority == DEF_PRIORITY);
2568 blk_start_plug(&plug);
2569 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2570 nr[LRU_INACTIVE_FILE]) {
2571 unsigned long nr_anon, nr_file, percentage;
2572 unsigned long nr_scanned;
2574 for_each_evictable_lru(lru) {
2576 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2577 nr[lru] -= nr_to_scan;
2579 nr_reclaimed += shrink_list(lru, nr_to_scan,
2586 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2590 * For kswapd and memcg, reclaim at least the number of pages
2591 * requested. Ensure that the anon and file LRUs are scanned
2592 * proportionally what was requested by get_scan_count(). We
2593 * stop reclaiming one LRU and reduce the amount scanning
2594 * proportional to the original scan target.
2596 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2597 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2600 * It's just vindictive to attack the larger once the smaller
2601 * has gone to zero. And given the way we stop scanning the
2602 * smaller below, this makes sure that we only make one nudge
2603 * towards proportionality once we've got nr_to_reclaim.
2605 if (!nr_file || !nr_anon)
2608 if (nr_file > nr_anon) {
2609 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2610 targets[LRU_ACTIVE_ANON] + 1;
2612 percentage = nr_anon * 100 / scan_target;
2614 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2615 targets[LRU_ACTIVE_FILE] + 1;
2617 percentage = nr_file * 100 / scan_target;
2620 /* Stop scanning the smaller of the LRU */
2622 nr[lru + LRU_ACTIVE] = 0;
2625 * Recalculate the other LRU scan count based on its original
2626 * scan target and the percentage scanning already complete
2628 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2629 nr_scanned = targets[lru] - nr[lru];
2630 nr[lru] = targets[lru] * (100 - percentage) / 100;
2631 nr[lru] -= min(nr[lru], nr_scanned);
2634 nr_scanned = targets[lru] - nr[lru];
2635 nr[lru] = targets[lru] * (100 - percentage) / 100;
2636 nr[lru] -= min(nr[lru], nr_scanned);
2638 scan_adjusted = true;
2640 blk_finish_plug(&plug);
2641 sc->nr_reclaimed += nr_reclaimed;
2644 * Even if we did not try to evict anon pages at all, we want to
2645 * rebalance the anon lru active/inactive ratio.
2647 if (total_swap_pages && inactive_list_is_low(lruvec, false, sc, true))
2648 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2649 sc, LRU_ACTIVE_ANON);
2652 /* Use reclaim/compaction for costly allocs or under memory pressure */
2653 static bool in_reclaim_compaction(struct scan_control *sc)
2655 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2656 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2657 sc->priority < DEF_PRIORITY - 2))
2664 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2665 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2666 * true if more pages should be reclaimed such that when the page allocator
2667 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2668 * It will give up earlier than that if there is difficulty reclaiming pages.
2670 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2671 unsigned long nr_reclaimed,
2672 struct scan_control *sc)
2674 unsigned long pages_for_compaction;
2675 unsigned long inactive_lru_pages;
2678 /* If not in reclaim/compaction mode, stop */
2679 if (!in_reclaim_compaction(sc))
2683 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2684 * number of pages that were scanned. This will return to the caller
2685 * with the risk reclaim/compaction and the resulting allocation attempt
2686 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2687 * allocations through requiring that the full LRU list has been scanned
2688 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2689 * scan, but that approximation was wrong, and there were corner cases
2690 * where always a non-zero amount of pages were scanned.
2695 /* If compaction would go ahead or the allocation would succeed, stop */
2696 for (z = 0; z <= sc->reclaim_idx; z++) {
2697 struct zone *zone = &pgdat->node_zones[z];
2698 if (!managed_zone(zone))
2701 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2702 case COMPACT_SUCCESS:
2703 case COMPACT_CONTINUE:
2706 /* check next zone */
2712 * If we have not reclaimed enough pages for compaction and the
2713 * inactive lists are large enough, continue reclaiming
2715 pages_for_compaction = compact_gap(sc->order);
2716 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2717 if (get_nr_swap_pages() > 0)
2718 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2720 return inactive_lru_pages > pages_for_compaction;
2723 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2725 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2726 (memcg && memcg_congested(pgdat, memcg));
2729 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2731 struct reclaim_state *reclaim_state = current->reclaim_state;
2732 struct mem_cgroup *root = sc->target_mem_cgroup;
2733 unsigned long nr_reclaimed, nr_scanned;
2734 bool reclaimable = false;
2735 struct mem_cgroup *memcg;
2737 memset(&sc->nr, 0, sizeof(sc->nr));
2739 nr_reclaimed = sc->nr_reclaimed;
2740 nr_scanned = sc->nr_scanned;
2742 memcg = mem_cgroup_iter(root, NULL, NULL);
2744 unsigned long reclaimed;
2745 unsigned long scanned;
2747 switch (mem_cgroup_protected(root, memcg)) {
2748 case MEMCG_PROT_MIN:
2751 * If there is no reclaimable memory, OOM.
2754 case MEMCG_PROT_LOW:
2757 * Respect the protection only as long as
2758 * there is an unprotected supply
2759 * of reclaimable memory from other cgroups.
2761 if (!sc->memcg_low_reclaim) {
2762 sc->memcg_low_skipped = 1;
2765 memcg_memory_event(memcg, MEMCG_LOW);
2767 case MEMCG_PROT_NONE:
2769 * All protection thresholds breached. We may
2770 * still choose to vary the scan pressure
2771 * applied based on by how much the cgroup in
2772 * question has exceeded its protection
2773 * thresholds (see get_scan_count).
2778 reclaimed = sc->nr_reclaimed;
2779 scanned = sc->nr_scanned;
2780 shrink_node_memcg(pgdat, memcg, sc);
2782 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2785 /* Record the group's reclaim efficiency */
2786 vmpressure(sc->gfp_mask, memcg, false,
2787 sc->nr_scanned - scanned,
2788 sc->nr_reclaimed - reclaimed);
2790 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2792 if (reclaim_state) {
2793 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2794 reclaim_state->reclaimed_slab = 0;
2797 /* Record the subtree's reclaim efficiency */
2798 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2799 sc->nr_scanned - nr_scanned,
2800 sc->nr_reclaimed - nr_reclaimed);
2802 if (sc->nr_reclaimed - nr_reclaimed)
2805 if (current_is_kswapd()) {
2807 * If reclaim is isolating dirty pages under writeback,
2808 * it implies that the long-lived page allocation rate
2809 * is exceeding the page laundering rate. Either the
2810 * global limits are not being effective at throttling
2811 * processes due to the page distribution throughout
2812 * zones or there is heavy usage of a slow backing
2813 * device. The only option is to throttle from reclaim
2814 * context which is not ideal as there is no guarantee
2815 * the dirtying process is throttled in the same way
2816 * balance_dirty_pages() manages.
2818 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2819 * count the number of pages under pages flagged for
2820 * immediate reclaim and stall if any are encountered
2821 * in the nr_immediate check below.
2823 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2824 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2827 * Tag a node as congested if all the dirty pages
2828 * scanned were backed by a congested BDI and
2829 * wait_iff_congested will stall.
2831 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2832 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2834 /* Allow kswapd to start writing pages during reclaim.*/
2835 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2836 set_bit(PGDAT_DIRTY, &pgdat->flags);
2839 * If kswapd scans pages marked marked for immediate
2840 * reclaim and under writeback (nr_immediate), it
2841 * implies that pages are cycling through the LRU
2842 * faster than they are written so also forcibly stall.
2844 if (sc->nr.immediate)
2845 congestion_wait(BLK_RW_ASYNC, HZ/10);
2849 * Legacy memcg will stall in page writeback so avoid forcibly
2850 * stalling in wait_iff_congested().
2852 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc) &&
2853 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2854 set_memcg_congestion(pgdat, root, true);
2857 * Stall direct reclaim for IO completions if underlying BDIs
2858 * and node is congested. Allow kswapd to continue until it
2859 * starts encountering unqueued dirty pages or cycling through
2860 * the LRU too quickly.
2862 if (!sc->hibernation_mode && !current_is_kswapd() &&
2863 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2864 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2866 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2871 * Kswapd gives up on balancing particular nodes after too
2872 * many failures to reclaim anything from them and goes to
2873 * sleep. On reclaim progress, reset the failure counter. A
2874 * successful direct reclaim run will revive a dormant kswapd.
2877 pgdat->kswapd_failures = 0;
2883 * Returns true if compaction should go ahead for a costly-order request, or
2884 * the allocation would already succeed without compaction. Return false if we
2885 * should reclaim first.
2887 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2889 unsigned long watermark;
2890 enum compact_result suitable;
2892 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2893 if (suitable == COMPACT_SUCCESS)
2894 /* Allocation should succeed already. Don't reclaim. */
2896 if (suitable == COMPACT_SKIPPED)
2897 /* Compaction cannot yet proceed. Do reclaim. */
2901 * Compaction is already possible, but it takes time to run and there
2902 * are potentially other callers using the pages just freed. So proceed
2903 * with reclaim to make a buffer of free pages available to give
2904 * compaction a reasonable chance of completing and allocating the page.
2905 * Note that we won't actually reclaim the whole buffer in one attempt
2906 * as the target watermark in should_continue_reclaim() is lower. But if
2907 * we are already above the high+gap watermark, don't reclaim at all.
2909 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2911 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2915 * This is the direct reclaim path, for page-allocating processes. We only
2916 * try to reclaim pages from zones which will satisfy the caller's allocation
2919 * If a zone is deemed to be full of pinned pages then just give it a light
2920 * scan then give up on it.
2922 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2926 unsigned long nr_soft_reclaimed;
2927 unsigned long nr_soft_scanned;
2929 pg_data_t *last_pgdat = NULL;
2932 * If the number of buffer_heads in the machine exceeds the maximum
2933 * allowed level, force direct reclaim to scan the highmem zone as
2934 * highmem pages could be pinning lowmem pages storing buffer_heads
2936 orig_mask = sc->gfp_mask;
2937 if (buffer_heads_over_limit) {
2938 sc->gfp_mask |= __GFP_HIGHMEM;
2939 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2942 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2943 sc->reclaim_idx, sc->nodemask) {
2945 * Take care memory controller reclaiming has small influence
2948 if (!cgroup_reclaim(sc)) {
2949 if (!cpuset_zone_allowed(zone,
2950 GFP_KERNEL | __GFP_HARDWALL))
2954 * If we already have plenty of memory free for
2955 * compaction in this zone, don't free any more.
2956 * Even though compaction is invoked for any
2957 * non-zero order, only frequent costly order
2958 * reclamation is disruptive enough to become a
2959 * noticeable problem, like transparent huge
2962 if (IS_ENABLED(CONFIG_COMPACTION) &&
2963 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2964 compaction_ready(zone, sc)) {
2965 sc->compaction_ready = true;
2970 * Shrink each node in the zonelist once. If the
2971 * zonelist is ordered by zone (not the default) then a
2972 * node may be shrunk multiple times but in that case
2973 * the user prefers lower zones being preserved.
2975 if (zone->zone_pgdat == last_pgdat)
2979 * This steals pages from memory cgroups over softlimit
2980 * and returns the number of reclaimed pages and
2981 * scanned pages. This works for global memory pressure
2982 * and balancing, not for a memcg's limit.
2984 nr_soft_scanned = 0;
2985 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2986 sc->order, sc->gfp_mask,
2988 sc->nr_reclaimed += nr_soft_reclaimed;
2989 sc->nr_scanned += nr_soft_scanned;
2990 /* need some check for avoid more shrink_zone() */
2993 /* See comment about same check for global reclaim above */
2994 if (zone->zone_pgdat == last_pgdat)
2996 last_pgdat = zone->zone_pgdat;
2997 shrink_node(zone->zone_pgdat, sc);
3001 * Restore to original mask to avoid the impact on the caller if we
3002 * promoted it to __GFP_HIGHMEM.
3004 sc->gfp_mask = orig_mask;
3007 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3009 struct mem_cgroup *memcg;
3011 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3013 unsigned long refaults;
3014 struct lruvec *lruvec;
3016 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3017 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3018 lruvec->refaults = refaults;
3019 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3023 * This is the main entry point to direct page reclaim.
3025 * If a full scan of the inactive list fails to free enough memory then we
3026 * are "out of memory" and something needs to be killed.
3028 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3029 * high - the zone may be full of dirty or under-writeback pages, which this
3030 * caller can't do much about. We kick the writeback threads and take explicit
3031 * naps in the hope that some of these pages can be written. But if the
3032 * allocating task holds filesystem locks which prevent writeout this might not
3033 * work, and the allocation attempt will fail.
3035 * returns: 0, if no pages reclaimed
3036 * else, the number of pages reclaimed
3038 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3039 struct scan_control *sc)
3041 int initial_priority = sc->priority;
3042 pg_data_t *last_pgdat;
3046 delayacct_freepages_start();
3048 if (!cgroup_reclaim(sc))
3049 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3052 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3055 shrink_zones(zonelist, sc);
3057 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3060 if (sc->compaction_ready)
3064 * If we're getting trouble reclaiming, start doing
3065 * writepage even in laptop mode.
3067 if (sc->priority < DEF_PRIORITY - 2)
3068 sc->may_writepage = 1;
3069 } while (--sc->priority >= 0);
3072 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3074 if (zone->zone_pgdat == last_pgdat)
3076 last_pgdat = zone->zone_pgdat;
3077 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3078 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3081 delayacct_freepages_end();
3083 if (sc->nr_reclaimed)
3084 return sc->nr_reclaimed;
3086 /* Aborted reclaim to try compaction? don't OOM, then */
3087 if (sc->compaction_ready)
3090 /* Untapped cgroup reserves? Don't OOM, retry. */
3091 if (sc->memcg_low_skipped) {
3092 sc->priority = initial_priority;
3093 sc->memcg_low_reclaim = 1;
3094 sc->memcg_low_skipped = 0;
3101 static bool allow_direct_reclaim(pg_data_t *pgdat)
3104 unsigned long pfmemalloc_reserve = 0;
3105 unsigned long free_pages = 0;
3109 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3112 for (i = 0; i <= ZONE_NORMAL; i++) {
3113 zone = &pgdat->node_zones[i];
3114 if (!managed_zone(zone))
3117 if (!zone_reclaimable_pages(zone))
3120 pfmemalloc_reserve += min_wmark_pages(zone);
3121 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3124 /* If there are no reserves (unexpected config) then do not throttle */
3125 if (!pfmemalloc_reserve)
3128 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3130 /* kswapd must be awake if processes are being throttled */
3131 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3132 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3133 (enum zone_type)ZONE_NORMAL);
3134 wake_up_interruptible(&pgdat->kswapd_wait);
3141 * Throttle direct reclaimers if backing storage is backed by the network
3142 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3143 * depleted. kswapd will continue to make progress and wake the processes
3144 * when the low watermark is reached.
3146 * Returns true if a fatal signal was delivered during throttling. If this
3147 * happens, the page allocator should not consider triggering the OOM killer.
3149 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3150 nodemask_t *nodemask)
3154 pg_data_t *pgdat = NULL;
3157 * Kernel threads should not be throttled as they may be indirectly
3158 * responsible for cleaning pages necessary for reclaim to make forward
3159 * progress. kjournald for example may enter direct reclaim while
3160 * committing a transaction where throttling it could forcing other
3161 * processes to block on log_wait_commit().
3163 if (current->flags & PF_KTHREAD)
3167 * If a fatal signal is pending, this process should not throttle.
3168 * It should return quickly so it can exit and free its memory
3170 if (fatal_signal_pending(current))
3174 * Check if the pfmemalloc reserves are ok by finding the first node
3175 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3176 * GFP_KERNEL will be required for allocating network buffers when
3177 * swapping over the network so ZONE_HIGHMEM is unusable.
3179 * Throttling is based on the first usable node and throttled processes
3180 * wait on a queue until kswapd makes progress and wakes them. There
3181 * is an affinity then between processes waking up and where reclaim
3182 * progress has been made assuming the process wakes on the same node.
3183 * More importantly, processes running on remote nodes will not compete
3184 * for remote pfmemalloc reserves and processes on different nodes
3185 * should make reasonable progress.
3187 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3188 gfp_zone(gfp_mask), nodemask) {
3189 if (zone_idx(zone) > ZONE_NORMAL)
3192 /* Throttle based on the first usable node */
3193 pgdat = zone->zone_pgdat;
3194 if (allow_direct_reclaim(pgdat))
3199 /* If no zone was usable by the allocation flags then do not throttle */
3203 /* Account for the throttling */
3204 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3207 * If the caller cannot enter the filesystem, it's possible that it
3208 * is due to the caller holding an FS lock or performing a journal
3209 * transaction in the case of a filesystem like ext[3|4]. In this case,
3210 * it is not safe to block on pfmemalloc_wait as kswapd could be
3211 * blocked waiting on the same lock. Instead, throttle for up to a
3212 * second before continuing.
3214 if (!(gfp_mask & __GFP_FS)) {
3215 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3216 allow_direct_reclaim(pgdat), HZ);
3221 /* Throttle until kswapd wakes the process */
3222 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3223 allow_direct_reclaim(pgdat));
3226 if (fatal_signal_pending(current))
3233 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3234 gfp_t gfp_mask, nodemask_t *nodemask)
3236 unsigned long nr_reclaimed;
3237 struct scan_control sc = {
3238 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3239 .gfp_mask = current_gfp_context(gfp_mask),
3240 .reclaim_idx = gfp_zone(gfp_mask),
3242 .nodemask = nodemask,
3243 .priority = DEF_PRIORITY,
3244 .may_writepage = !laptop_mode,
3250 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3251 * Confirm they are large enough for max values.
3253 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3254 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3255 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3258 * Do not enter reclaim if fatal signal was delivered while throttled.
3259 * 1 is returned so that the page allocator does not OOM kill at this
3262 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3265 set_task_reclaim_state(current, &sc.reclaim_state);
3266 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3268 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3270 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3271 set_task_reclaim_state(current, NULL);
3273 return nr_reclaimed;
3278 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3279 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3280 gfp_t gfp_mask, bool noswap,
3282 unsigned long *nr_scanned)
3284 struct scan_control sc = {
3285 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3286 .target_mem_cgroup = memcg,
3287 .may_writepage = !laptop_mode,
3289 .reclaim_idx = MAX_NR_ZONES - 1,
3290 .may_swap = !noswap,
3293 WARN_ON_ONCE(!current->reclaim_state);
3295 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3296 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3298 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3302 * NOTE: Although we can get the priority field, using it
3303 * here is not a good idea, since it limits the pages we can scan.
3304 * if we don't reclaim here, the shrink_node from balance_pgdat
3305 * will pick up pages from other mem cgroup's as well. We hack
3306 * the priority and make it zero.
3308 shrink_node_memcg(pgdat, memcg, &sc);
3310 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3312 *nr_scanned = sc.nr_scanned;
3314 return sc.nr_reclaimed;
3317 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3318 unsigned long nr_pages,
3322 unsigned long nr_reclaimed;
3323 unsigned long pflags;
3324 unsigned int noreclaim_flag;
3325 struct scan_control sc = {
3326 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3327 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3328 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3329 .reclaim_idx = MAX_NR_ZONES - 1,
3330 .target_mem_cgroup = memcg,
3331 .priority = DEF_PRIORITY,
3332 .may_writepage = !laptop_mode,
3334 .may_swap = may_swap,
3337 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3338 * equal pressure on all the nodes. This is based on the assumption that
3339 * the reclaim does not bail out early.
3341 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3343 set_task_reclaim_state(current, &sc.reclaim_state);
3345 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3347 psi_memstall_enter(&pflags);
3348 noreclaim_flag = memalloc_noreclaim_save();
3350 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3352 memalloc_noreclaim_restore(noreclaim_flag);
3353 psi_memstall_leave(&pflags);
3355 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3356 set_task_reclaim_state(current, NULL);
3358 return nr_reclaimed;
3362 static void age_active_anon(struct pglist_data *pgdat,
3363 struct scan_control *sc)
3365 struct mem_cgroup *memcg;
3367 if (!total_swap_pages)
3370 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3372 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3374 if (inactive_list_is_low(lruvec, false, sc, true))
3375 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3376 sc, LRU_ACTIVE_ANON);
3378 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3382 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3388 * Check for watermark boosts top-down as the higher zones
3389 * are more likely to be boosted. Both watermarks and boosts
3390 * should not be checked at the time time as reclaim would
3391 * start prematurely when there is no boosting and a lower
3394 for (i = classzone_idx; i >= 0; i--) {
3395 zone = pgdat->node_zones + i;
3396 if (!managed_zone(zone))
3399 if (zone->watermark_boost)
3407 * Returns true if there is an eligible zone balanced for the request order
3410 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3413 unsigned long mark = -1;
3417 * Check watermarks bottom-up as lower zones are more likely to
3420 for (i = 0; i <= classzone_idx; i++) {
3421 zone = pgdat->node_zones + i;
3423 if (!managed_zone(zone))
3426 mark = high_wmark_pages(zone);
3427 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3432 * If a node has no populated zone within classzone_idx, it does not
3433 * need balancing by definition. This can happen if a zone-restricted
3434 * allocation tries to wake a remote kswapd.
3442 /* Clear pgdat state for congested, dirty or under writeback. */
3443 static void clear_pgdat_congested(pg_data_t *pgdat)
3445 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3446 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3447 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3451 * Prepare kswapd for sleeping. This verifies that there are no processes
3452 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3454 * Returns true if kswapd is ready to sleep
3456 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3459 * The throttled processes are normally woken up in balance_pgdat() as
3460 * soon as allow_direct_reclaim() is true. But there is a potential
3461 * race between when kswapd checks the watermarks and a process gets
3462 * throttled. There is also a potential race if processes get
3463 * throttled, kswapd wakes, a large process exits thereby balancing the
3464 * zones, which causes kswapd to exit balance_pgdat() before reaching
3465 * the wake up checks. If kswapd is going to sleep, no process should
3466 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3467 * the wake up is premature, processes will wake kswapd and get
3468 * throttled again. The difference from wake ups in balance_pgdat() is
3469 * that here we are under prepare_to_wait().
3471 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3472 wake_up_all(&pgdat->pfmemalloc_wait);
3474 /* Hopeless node, leave it to direct reclaim */
3475 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3478 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3479 clear_pgdat_congested(pgdat);
3487 * kswapd shrinks a node of pages that are at or below the highest usable
3488 * zone that is currently unbalanced.
3490 * Returns true if kswapd scanned at least the requested number of pages to
3491 * reclaim or if the lack of progress was due to pages under writeback.
3492 * This is used to determine if the scanning priority needs to be raised.
3494 static bool kswapd_shrink_node(pg_data_t *pgdat,
3495 struct scan_control *sc)
3500 /* Reclaim a number of pages proportional to the number of zones */
3501 sc->nr_to_reclaim = 0;
3502 for (z = 0; z <= sc->reclaim_idx; z++) {
3503 zone = pgdat->node_zones + z;
3504 if (!managed_zone(zone))
3507 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3511 * Historically care was taken to put equal pressure on all zones but
3512 * now pressure is applied based on node LRU order.
3514 shrink_node(pgdat, sc);
3517 * Fragmentation may mean that the system cannot be rebalanced for
3518 * high-order allocations. If twice the allocation size has been
3519 * reclaimed then recheck watermarks only at order-0 to prevent
3520 * excessive reclaim. Assume that a process requested a high-order
3521 * can direct reclaim/compact.
3523 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3526 return sc->nr_scanned >= sc->nr_to_reclaim;
3530 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3531 * that are eligible for use by the caller until at least one zone is
3534 * Returns the order kswapd finished reclaiming at.
3536 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3537 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3538 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3539 * or lower is eligible for reclaim until at least one usable zone is
3542 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3545 unsigned long nr_soft_reclaimed;
3546 unsigned long nr_soft_scanned;
3547 unsigned long pflags;
3548 unsigned long nr_boost_reclaim;
3549 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3552 struct scan_control sc = {
3553 .gfp_mask = GFP_KERNEL,
3558 set_task_reclaim_state(current, &sc.reclaim_state);
3559 psi_memstall_enter(&pflags);
3560 __fs_reclaim_acquire();
3562 count_vm_event(PAGEOUTRUN);
3565 * Account for the reclaim boost. Note that the zone boost is left in
3566 * place so that parallel allocations that are near the watermark will
3567 * stall or direct reclaim until kswapd is finished.
3569 nr_boost_reclaim = 0;
3570 for (i = 0; i <= classzone_idx; i++) {
3571 zone = pgdat->node_zones + i;
3572 if (!managed_zone(zone))
3575 nr_boost_reclaim += zone->watermark_boost;
3576 zone_boosts[i] = zone->watermark_boost;
3578 boosted = nr_boost_reclaim;
3581 sc.priority = DEF_PRIORITY;
3583 unsigned long nr_reclaimed = sc.nr_reclaimed;
3584 bool raise_priority = true;
3588 sc.reclaim_idx = classzone_idx;
3591 * If the number of buffer_heads exceeds the maximum allowed
3592 * then consider reclaiming from all zones. This has a dual
3593 * purpose -- on 64-bit systems it is expected that
3594 * buffer_heads are stripped during active rotation. On 32-bit
3595 * systems, highmem pages can pin lowmem memory and shrinking
3596 * buffers can relieve lowmem pressure. Reclaim may still not
3597 * go ahead if all eligible zones for the original allocation
3598 * request are balanced to avoid excessive reclaim from kswapd.
3600 if (buffer_heads_over_limit) {
3601 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3602 zone = pgdat->node_zones + i;
3603 if (!managed_zone(zone))
3612 * If the pgdat is imbalanced then ignore boosting and preserve
3613 * the watermarks for a later time and restart. Note that the
3614 * zone watermarks will be still reset at the end of balancing
3615 * on the grounds that the normal reclaim should be enough to
3616 * re-evaluate if boosting is required when kswapd next wakes.
3618 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3619 if (!balanced && nr_boost_reclaim) {
3620 nr_boost_reclaim = 0;
3625 * If boosting is not active then only reclaim if there are no
3626 * eligible zones. Note that sc.reclaim_idx is not used as
3627 * buffer_heads_over_limit may have adjusted it.
3629 if (!nr_boost_reclaim && balanced)
3632 /* Limit the priority of boosting to avoid reclaim writeback */
3633 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3634 raise_priority = false;
3637 * Do not writeback or swap pages for boosted reclaim. The
3638 * intent is to relieve pressure not issue sub-optimal IO
3639 * from reclaim context. If no pages are reclaimed, the
3640 * reclaim will be aborted.
3642 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3643 sc.may_swap = !nr_boost_reclaim;
3646 * Do some background aging of the anon list, to give
3647 * pages a chance to be referenced before reclaiming. All
3648 * pages are rotated regardless of classzone as this is
3649 * about consistent aging.
3651 age_active_anon(pgdat, &sc);
3654 * If we're getting trouble reclaiming, start doing writepage
3655 * even in laptop mode.
3657 if (sc.priority < DEF_PRIORITY - 2)
3658 sc.may_writepage = 1;
3660 /* Call soft limit reclaim before calling shrink_node. */
3662 nr_soft_scanned = 0;
3663 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3664 sc.gfp_mask, &nr_soft_scanned);
3665 sc.nr_reclaimed += nr_soft_reclaimed;
3668 * There should be no need to raise the scanning priority if
3669 * enough pages are already being scanned that that high
3670 * watermark would be met at 100% efficiency.
3672 if (kswapd_shrink_node(pgdat, &sc))
3673 raise_priority = false;
3676 * If the low watermark is met there is no need for processes
3677 * to be throttled on pfmemalloc_wait as they should not be
3678 * able to safely make forward progress. Wake them
3680 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3681 allow_direct_reclaim(pgdat))
3682 wake_up_all(&pgdat->pfmemalloc_wait);
3684 /* Check if kswapd should be suspending */
3685 __fs_reclaim_release();
3686 ret = try_to_freeze();
3687 __fs_reclaim_acquire();
3688 if (ret || kthread_should_stop())
3692 * Raise priority if scanning rate is too low or there was no
3693 * progress in reclaiming pages
3695 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3696 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3699 * If reclaim made no progress for a boost, stop reclaim as
3700 * IO cannot be queued and it could be an infinite loop in
3701 * extreme circumstances.
3703 if (nr_boost_reclaim && !nr_reclaimed)
3706 if (raise_priority || !nr_reclaimed)
3708 } while (sc.priority >= 1);
3710 if (!sc.nr_reclaimed)
3711 pgdat->kswapd_failures++;
3714 /* If reclaim was boosted, account for the reclaim done in this pass */
3716 unsigned long flags;
3718 for (i = 0; i <= classzone_idx; i++) {
3719 if (!zone_boosts[i])
3722 /* Increments are under the zone lock */
3723 zone = pgdat->node_zones + i;
3724 spin_lock_irqsave(&zone->lock, flags);
3725 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3726 spin_unlock_irqrestore(&zone->lock, flags);
3730 * As there is now likely space, wakeup kcompact to defragment
3733 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3736 snapshot_refaults(NULL, pgdat);
3737 __fs_reclaim_release();
3738 psi_memstall_leave(&pflags);
3739 set_task_reclaim_state(current, NULL);
3742 * Return the order kswapd stopped reclaiming at as
3743 * prepare_kswapd_sleep() takes it into account. If another caller
3744 * entered the allocator slow path while kswapd was awake, order will
3745 * remain at the higher level.
3751 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3752 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3753 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3754 * after previous reclaim attempt (node is still unbalanced). In that case
3755 * return the zone index of the previous kswapd reclaim cycle.
3757 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3758 enum zone_type prev_classzone_idx)
3760 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3761 return prev_classzone_idx;
3762 return pgdat->kswapd_classzone_idx;
3765 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3766 unsigned int classzone_idx)
3771 if (freezing(current) || kthread_should_stop())
3774 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3777 * Try to sleep for a short interval. Note that kcompactd will only be
3778 * woken if it is possible to sleep for a short interval. This is
3779 * deliberate on the assumption that if reclaim cannot keep an
3780 * eligible zone balanced that it's also unlikely that compaction will
3783 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3785 * Compaction records what page blocks it recently failed to
3786 * isolate pages from and skips them in the future scanning.
3787 * When kswapd is going to sleep, it is reasonable to assume
3788 * that pages and compaction may succeed so reset the cache.
3790 reset_isolation_suitable(pgdat);
3793 * We have freed the memory, now we should compact it to make
3794 * allocation of the requested order possible.
3796 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3798 remaining = schedule_timeout(HZ/10);
3801 * If woken prematurely then reset kswapd_classzone_idx and
3802 * order. The values will either be from a wakeup request or
3803 * the previous request that slept prematurely.
3806 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3807 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3810 finish_wait(&pgdat->kswapd_wait, &wait);
3811 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3815 * After a short sleep, check if it was a premature sleep. If not, then
3816 * go fully to sleep until explicitly woken up.
3819 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3820 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3823 * vmstat counters are not perfectly accurate and the estimated
3824 * value for counters such as NR_FREE_PAGES can deviate from the
3825 * true value by nr_online_cpus * threshold. To avoid the zone
3826 * watermarks being breached while under pressure, we reduce the
3827 * per-cpu vmstat threshold while kswapd is awake and restore
3828 * them before going back to sleep.
3830 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3832 if (!kthread_should_stop())
3835 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3838 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3840 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3842 finish_wait(&pgdat->kswapd_wait, &wait);
3846 * The background pageout daemon, started as a kernel thread
3847 * from the init process.
3849 * This basically trickles out pages so that we have _some_
3850 * free memory available even if there is no other activity
3851 * that frees anything up. This is needed for things like routing
3852 * etc, where we otherwise might have all activity going on in
3853 * asynchronous contexts that cannot page things out.
3855 * If there are applications that are active memory-allocators
3856 * (most normal use), this basically shouldn't matter.
3858 static int kswapd(void *p)
3860 unsigned int alloc_order, reclaim_order;
3861 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3862 pg_data_t *pgdat = (pg_data_t*)p;
3863 struct task_struct *tsk = current;
3864 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3866 if (!cpumask_empty(cpumask))
3867 set_cpus_allowed_ptr(tsk, cpumask);
3870 * Tell the memory management that we're a "memory allocator",
3871 * and that if we need more memory we should get access to it
3872 * regardless (see "__alloc_pages()"). "kswapd" should
3873 * never get caught in the normal page freeing logic.
3875 * (Kswapd normally doesn't need memory anyway, but sometimes
3876 * you need a small amount of memory in order to be able to
3877 * page out something else, and this flag essentially protects
3878 * us from recursively trying to free more memory as we're
3879 * trying to free the first piece of memory in the first place).
3881 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3884 pgdat->kswapd_order = 0;
3885 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3889 alloc_order = reclaim_order = pgdat->kswapd_order;
3890 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3893 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3896 /* Read the new order and classzone_idx */
3897 alloc_order = reclaim_order = pgdat->kswapd_order;
3898 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3899 pgdat->kswapd_order = 0;
3900 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3902 ret = try_to_freeze();
3903 if (kthread_should_stop())
3907 * We can speed up thawing tasks if we don't call balance_pgdat
3908 * after returning from the refrigerator
3914 * Reclaim begins at the requested order but if a high-order
3915 * reclaim fails then kswapd falls back to reclaiming for
3916 * order-0. If that happens, kswapd will consider sleeping
3917 * for the order it finished reclaiming at (reclaim_order)
3918 * but kcompactd is woken to compact for the original
3919 * request (alloc_order).
3921 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3923 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3924 if (reclaim_order < alloc_order)
3925 goto kswapd_try_sleep;
3928 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3934 * A zone is low on free memory or too fragmented for high-order memory. If
3935 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3936 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3937 * has failed or is not needed, still wake up kcompactd if only compaction is
3940 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3941 enum zone_type classzone_idx)
3945 if (!managed_zone(zone))
3948 if (!cpuset_zone_allowed(zone, gfp_flags))
3950 pgdat = zone->zone_pgdat;
3952 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3953 pgdat->kswapd_classzone_idx = classzone_idx;
3955 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3957 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3958 if (!waitqueue_active(&pgdat->kswapd_wait))
3961 /* Hopeless node, leave it to direct reclaim if possible */
3962 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3963 (pgdat_balanced(pgdat, order, classzone_idx) &&
3964 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3966 * There may be plenty of free memory available, but it's too
3967 * fragmented for high-order allocations. Wake up kcompactd
3968 * and rely on compaction_suitable() to determine if it's
3969 * needed. If it fails, it will defer subsequent attempts to
3970 * ratelimit its work.
3972 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3973 wakeup_kcompactd(pgdat, order, classzone_idx);
3977 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3979 wake_up_interruptible(&pgdat->kswapd_wait);
3982 #ifdef CONFIG_HIBERNATION
3984 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3987 * Rather than trying to age LRUs the aim is to preserve the overall
3988 * LRU order by reclaiming preferentially
3989 * inactive > active > active referenced > active mapped
3991 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3993 struct scan_control sc = {
3994 .nr_to_reclaim = nr_to_reclaim,
3995 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3996 .reclaim_idx = MAX_NR_ZONES - 1,
3997 .priority = DEF_PRIORITY,
4001 .hibernation_mode = 1,
4003 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4004 unsigned long nr_reclaimed;
4005 unsigned int noreclaim_flag;
4007 fs_reclaim_acquire(sc.gfp_mask);
4008 noreclaim_flag = memalloc_noreclaim_save();
4009 set_task_reclaim_state(current, &sc.reclaim_state);
4011 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4013 set_task_reclaim_state(current, NULL);
4014 memalloc_noreclaim_restore(noreclaim_flag);
4015 fs_reclaim_release(sc.gfp_mask);
4017 return nr_reclaimed;
4019 #endif /* CONFIG_HIBERNATION */
4021 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4022 not required for correctness. So if the last cpu in a node goes
4023 away, we get changed to run anywhere: as the first one comes back,
4024 restore their cpu bindings. */
4025 static int kswapd_cpu_online(unsigned int cpu)
4029 for_each_node_state(nid, N_MEMORY) {
4030 pg_data_t *pgdat = NODE_DATA(nid);
4031 const struct cpumask *mask;
4033 mask = cpumask_of_node(pgdat->node_id);
4035 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4036 /* One of our CPUs online: restore mask */
4037 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4043 * This kswapd start function will be called by init and node-hot-add.
4044 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4046 int kswapd_run(int nid)
4048 pg_data_t *pgdat = NODE_DATA(nid);
4054 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4055 if (IS_ERR(pgdat->kswapd)) {
4056 /* failure at boot is fatal */
4057 BUG_ON(system_state < SYSTEM_RUNNING);
4058 pr_err("Failed to start kswapd on node %d\n", nid);
4059 ret = PTR_ERR(pgdat->kswapd);
4060 pgdat->kswapd = NULL;
4066 * Called by memory hotplug when all memory in a node is offlined. Caller must
4067 * hold mem_hotplug_begin/end().
4069 void kswapd_stop(int nid)
4071 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4074 kthread_stop(kswapd);
4075 NODE_DATA(nid)->kswapd = NULL;
4079 static int __init kswapd_init(void)
4084 for_each_node_state(nid, N_MEMORY)
4086 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4087 "mm/vmscan:online", kswapd_cpu_online,
4093 module_init(kswapd_init)
4099 * If non-zero call node_reclaim when the number of free pages falls below
4102 int node_reclaim_mode __read_mostly;
4104 #define RECLAIM_OFF 0
4105 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4106 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4107 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4110 * Priority for NODE_RECLAIM. This determines the fraction of pages
4111 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4114 #define NODE_RECLAIM_PRIORITY 4
4117 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4120 int sysctl_min_unmapped_ratio = 1;
4123 * If the number of slab pages in a zone grows beyond this percentage then
4124 * slab reclaim needs to occur.
4126 int sysctl_min_slab_ratio = 5;
4128 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4130 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4131 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4132 node_page_state(pgdat, NR_ACTIVE_FILE);
4135 * It's possible for there to be more file mapped pages than
4136 * accounted for by the pages on the file LRU lists because
4137 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4139 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4142 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4143 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4145 unsigned long nr_pagecache_reclaimable;
4146 unsigned long delta = 0;
4149 * If RECLAIM_UNMAP is set, then all file pages are considered
4150 * potentially reclaimable. Otherwise, we have to worry about
4151 * pages like swapcache and node_unmapped_file_pages() provides
4154 if (node_reclaim_mode & RECLAIM_UNMAP)
4155 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4157 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4159 /* If we can't clean pages, remove dirty pages from consideration */
4160 if (!(node_reclaim_mode & RECLAIM_WRITE))
4161 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4163 /* Watch for any possible underflows due to delta */
4164 if (unlikely(delta > nr_pagecache_reclaimable))
4165 delta = nr_pagecache_reclaimable;
4167 return nr_pagecache_reclaimable - delta;
4171 * Try to free up some pages from this node through reclaim.
4173 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4175 /* Minimum pages needed in order to stay on node */
4176 const unsigned long nr_pages = 1 << order;
4177 struct task_struct *p = current;
4178 unsigned int noreclaim_flag;
4179 struct scan_control sc = {
4180 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4181 .gfp_mask = current_gfp_context(gfp_mask),
4183 .priority = NODE_RECLAIM_PRIORITY,
4184 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4185 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4187 .reclaim_idx = gfp_zone(gfp_mask),
4190 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4194 fs_reclaim_acquire(sc.gfp_mask);
4196 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4197 * and we also need to be able to write out pages for RECLAIM_WRITE
4198 * and RECLAIM_UNMAP.
4200 noreclaim_flag = memalloc_noreclaim_save();
4201 p->flags |= PF_SWAPWRITE;
4202 set_task_reclaim_state(p, &sc.reclaim_state);
4204 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4206 * Free memory by calling shrink node with increasing
4207 * priorities until we have enough memory freed.
4210 shrink_node(pgdat, &sc);
4211 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4214 set_task_reclaim_state(p, NULL);
4215 current->flags &= ~PF_SWAPWRITE;
4216 memalloc_noreclaim_restore(noreclaim_flag);
4217 fs_reclaim_release(sc.gfp_mask);
4219 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4221 return sc.nr_reclaimed >= nr_pages;
4224 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4229 * Node reclaim reclaims unmapped file backed pages and
4230 * slab pages if we are over the defined limits.
4232 * A small portion of unmapped file backed pages is needed for
4233 * file I/O otherwise pages read by file I/O will be immediately
4234 * thrown out if the node is overallocated. So we do not reclaim
4235 * if less than a specified percentage of the node is used by
4236 * unmapped file backed pages.
4238 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4239 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4240 return NODE_RECLAIM_FULL;
4243 * Do not scan if the allocation should not be delayed.
4245 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4246 return NODE_RECLAIM_NOSCAN;
4249 * Only run node reclaim on the local node or on nodes that do not
4250 * have associated processors. This will favor the local processor
4251 * over remote processors and spread off node memory allocations
4252 * as wide as possible.
4254 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4255 return NODE_RECLAIM_NOSCAN;
4257 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4258 return NODE_RECLAIM_NOSCAN;
4260 ret = __node_reclaim(pgdat, gfp_mask, order);
4261 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4264 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4271 * page_evictable - test whether a page is evictable
4272 * @page: the page to test
4274 * Test whether page is evictable--i.e., should be placed on active/inactive
4275 * lists vs unevictable list.
4277 * Reasons page might not be evictable:
4278 * (1) page's mapping marked unevictable
4279 * (2) page is part of an mlocked VMA
4282 int page_evictable(struct page *page)
4286 /* Prevent address_space of inode and swap cache from being freed */
4288 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4294 * check_move_unevictable_pages - check pages for evictability and move to
4295 * appropriate zone lru list
4296 * @pvec: pagevec with lru pages to check
4298 * Checks pages for evictability, if an evictable page is in the unevictable
4299 * lru list, moves it to the appropriate evictable lru list. This function
4300 * should be only used for lru pages.
4302 void check_move_unevictable_pages(struct pagevec *pvec)
4304 struct lruvec *lruvec;
4305 struct pglist_data *pgdat = NULL;
4310 for (i = 0; i < pvec->nr; i++) {
4311 struct page *page = pvec->pages[i];
4312 struct pglist_data *pagepgdat = page_pgdat(page);
4315 if (pagepgdat != pgdat) {
4317 spin_unlock_irq(&pgdat->lru_lock);
4319 spin_lock_irq(&pgdat->lru_lock);
4321 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4323 if (!PageLRU(page) || !PageUnevictable(page))
4326 if (page_evictable(page)) {
4327 enum lru_list lru = page_lru_base_type(page);
4329 VM_BUG_ON_PAGE(PageActive(page), page);
4330 ClearPageUnevictable(page);
4331 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4332 add_page_to_lru_list(page, lruvec, lru);
4338 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4339 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4340 spin_unlock_irq(&pgdat->lru_lock);
4343 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);