1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/pagevec.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
51 #include <linux/psi.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
75 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
78 struct mem_cgroup *target_mem_cgroup;
81 * Scan pressure balancing between anon and file LRUs
83 unsigned long anon_cost;
84 unsigned long file_cost;
86 /* Can active pages be deactivated as part of reclaim? */
87 #define DEACTIVATE_ANON 1
88 #define DEACTIVATE_FILE 2
89 unsigned int may_deactivate:2;
90 unsigned int force_deactivate:1;
91 unsigned int skipped_deactivate:1;
93 /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 unsigned int may_writepage:1;
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap:1;
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap:1;
103 * Cgroups are not reclaimed below their configured memory.low,
104 * unless we threaten to OOM. If any cgroups are skipped due to
105 * memory.low and nothing was reclaimed, go back for memory.low.
107 unsigned int memcg_low_reclaim:1;
108 unsigned int memcg_low_skipped:1;
110 unsigned int hibernation_mode:1;
112 /* One of the zones is ready for compaction */
113 unsigned int compaction_ready:1;
115 /* There is easily reclaimable cold cache in the current node */
116 unsigned int cache_trim_mode:1;
118 /* The file pages on the current node are dangerously low */
119 unsigned int file_is_tiny:1;
121 /* Allocation order */
124 /* Scan (total_size >> priority) pages at once */
127 /* The highest zone to isolate pages for reclaim from */
130 /* This context's GFP mask */
133 /* Incremented by the number of inactive pages that were scanned */
134 unsigned long nr_scanned;
136 /* Number of pages freed so far during a call to shrink_zones() */
137 unsigned long nr_reclaimed;
141 unsigned int unqueued_dirty;
142 unsigned int congested;
143 unsigned int writeback;
144 unsigned int immediate;
145 unsigned int file_taken;
149 /* for recording the reclaimed slab by now */
150 struct reclaim_state reclaim_state;
153 #ifdef ARCH_HAS_PREFETCHW
154 #define prefetchw_prev_lru_page(_page, _base, _field) \
156 if ((_page)->lru.prev != _base) { \
159 prev = lru_to_page(&(_page->lru)); \
160 prefetchw(&prev->_field); \
164 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168 * From 0 .. 200. Higher means more swappy.
170 int vm_swappiness = 60;
172 static void set_task_reclaim_state(struct task_struct *task,
173 struct reclaim_state *rs)
175 /* Check for an overwrite */
176 WARN_ON_ONCE(rs && task->reclaim_state);
178 /* Check for the nulling of an already-nulled member */
179 WARN_ON_ONCE(!rs && !task->reclaim_state);
181 task->reclaim_state = rs;
184 static LIST_HEAD(shrinker_list);
185 static DECLARE_RWSEM(shrinker_rwsem);
189 static int memcg_shrinker_map_size;
190 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
192 static void free_shrinker_map_rcu(struct rcu_head *head)
194 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
197 static int expand_one_shrinker_map(struct mem_cgroup *memcg,
198 int size, int old_size)
200 struct memcg_shrinker_map *new, *old;
201 struct mem_cgroup_per_node *pn;
204 lockdep_assert_held(&memcg_shrinker_map_mutex);
207 pn = memcg->nodeinfo[nid];
208 old = rcu_dereference_protected(pn->shrinker_map, true);
209 /* Not yet online memcg */
213 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
217 /* Set all old bits, clear all new bits */
218 memset(new->map, (int)0xff, old_size);
219 memset((void *)new->map + old_size, 0, size - old_size);
221 rcu_assign_pointer(pn->shrinker_map, new);
222 call_rcu(&old->rcu, free_shrinker_map_rcu);
228 void free_shrinker_maps(struct mem_cgroup *memcg)
230 struct mem_cgroup_per_node *pn;
231 struct memcg_shrinker_map *map;
234 if (mem_cgroup_is_root(memcg))
238 pn = memcg->nodeinfo[nid];
239 map = rcu_dereference_protected(pn->shrinker_map, true);
241 rcu_assign_pointer(pn->shrinker_map, NULL);
245 int alloc_shrinker_maps(struct mem_cgroup *memcg)
247 struct memcg_shrinker_map *map;
248 int nid, size, ret = 0;
250 if (mem_cgroup_is_root(memcg))
253 mutex_lock(&memcg_shrinker_map_mutex);
254 size = memcg_shrinker_map_size;
256 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
258 free_shrinker_maps(memcg);
262 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
264 mutex_unlock(&memcg_shrinker_map_mutex);
269 static int expand_shrinker_maps(int new_id)
271 int size, old_size, ret = 0;
272 struct mem_cgroup *memcg;
274 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
275 old_size = memcg_shrinker_map_size;
276 if (size <= old_size)
279 mutex_lock(&memcg_shrinker_map_mutex);
280 if (!root_mem_cgroup)
283 memcg = mem_cgroup_iter(NULL, NULL, NULL);
285 if (mem_cgroup_is_root(memcg))
287 ret = expand_one_shrinker_map(memcg, size, old_size);
289 mem_cgroup_iter_break(NULL, memcg);
292 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
295 memcg_shrinker_map_size = size;
296 mutex_unlock(&memcg_shrinker_map_mutex);
300 void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
302 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
303 struct memcg_shrinker_map *map;
306 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
307 /* Pairs with smp mb in shrink_slab() */
308 smp_mb__before_atomic();
309 set_bit(shrinker_id, map->map);
315 * We allow subsystems to populate their shrinker-related
316 * LRU lists before register_shrinker_prepared() is called
317 * for the shrinker, since we don't want to impose
318 * restrictions on their internal registration order.
319 * In this case shrink_slab_memcg() may find corresponding
320 * bit is set in the shrinkers map.
322 * This value is used by the function to detect registering
323 * shrinkers and to skip do_shrink_slab() calls for them.
325 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
327 static DEFINE_IDR(shrinker_idr);
328 static int shrinker_nr_max;
330 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
332 int id, ret = -ENOMEM;
334 down_write(&shrinker_rwsem);
335 /* This may call shrinker, so it must use down_read_trylock() */
336 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
340 if (id >= shrinker_nr_max) {
341 if (expand_shrinker_maps(id)) {
342 idr_remove(&shrinker_idr, id);
346 shrinker_nr_max = id + 1;
351 up_write(&shrinker_rwsem);
355 static void unregister_memcg_shrinker(struct shrinker *shrinker)
357 int id = shrinker->id;
361 down_write(&shrinker_rwsem);
362 idr_remove(&shrinker_idr, id);
363 up_write(&shrinker_rwsem);
366 static bool cgroup_reclaim(struct scan_control *sc)
368 return sc->target_mem_cgroup;
372 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
373 * @sc: scan_control in question
375 * The normal page dirty throttling mechanism in balance_dirty_pages() is
376 * completely broken with the legacy memcg and direct stalling in
377 * shrink_page_list() is used for throttling instead, which lacks all the
378 * niceties such as fairness, adaptive pausing, bandwidth proportional
379 * allocation and configurability.
381 * This function tests whether the vmscan currently in progress can assume
382 * that the normal dirty throttling mechanism is operational.
384 static bool writeback_throttling_sane(struct scan_control *sc)
386 if (!cgroup_reclaim(sc))
388 #ifdef CONFIG_CGROUP_WRITEBACK
389 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
395 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
400 static void unregister_memcg_shrinker(struct shrinker *shrinker)
404 static bool cgroup_reclaim(struct scan_control *sc)
409 static bool writeback_throttling_sane(struct scan_control *sc)
416 * This misses isolated pages which are not accounted for to save counters.
417 * As the data only determines if reclaim or compaction continues, it is
418 * not expected that isolated pages will be a dominating factor.
420 unsigned long zone_reclaimable_pages(struct zone *zone)
424 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
425 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
426 if (get_nr_swap_pages() > 0)
427 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
428 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
434 * lruvec_lru_size - Returns the number of pages on the given LRU list.
435 * @lruvec: lru vector
437 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
439 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
442 unsigned long size = 0;
445 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
446 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
448 if (!managed_zone(zone))
451 if (!mem_cgroup_disabled())
452 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
454 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
460 * Add a shrinker callback to be called from the vm.
462 int prealloc_shrinker(struct shrinker *shrinker)
464 unsigned int size = sizeof(*shrinker->nr_deferred);
466 if (shrinker->flags & SHRINKER_NUMA_AWARE)
469 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
470 if (!shrinker->nr_deferred)
473 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
474 if (prealloc_memcg_shrinker(shrinker))
481 kfree(shrinker->nr_deferred);
482 shrinker->nr_deferred = NULL;
486 void free_prealloced_shrinker(struct shrinker *shrinker)
488 if (!shrinker->nr_deferred)
491 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
492 unregister_memcg_shrinker(shrinker);
494 kfree(shrinker->nr_deferred);
495 shrinker->nr_deferred = NULL;
498 void register_shrinker_prepared(struct shrinker *shrinker)
500 down_write(&shrinker_rwsem);
501 list_add_tail(&shrinker->list, &shrinker_list);
503 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
504 idr_replace(&shrinker_idr, shrinker, shrinker->id);
506 up_write(&shrinker_rwsem);
509 int register_shrinker(struct shrinker *shrinker)
511 int err = prealloc_shrinker(shrinker);
515 register_shrinker_prepared(shrinker);
518 EXPORT_SYMBOL(register_shrinker);
523 void unregister_shrinker(struct shrinker *shrinker)
525 if (!shrinker->nr_deferred)
527 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
528 unregister_memcg_shrinker(shrinker);
529 down_write(&shrinker_rwsem);
530 list_del(&shrinker->list);
531 up_write(&shrinker_rwsem);
532 kfree(shrinker->nr_deferred);
533 shrinker->nr_deferred = NULL;
535 EXPORT_SYMBOL(unregister_shrinker);
537 #define SHRINK_BATCH 128
539 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
540 struct shrinker *shrinker, int priority)
542 unsigned long freed = 0;
543 unsigned long long delta;
548 int nid = shrinkctl->nid;
549 long batch_size = shrinker->batch ? shrinker->batch
551 long scanned = 0, next_deferred;
553 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
556 freeable = shrinker->count_objects(shrinker, shrinkctl);
557 if (freeable == 0 || freeable == SHRINK_EMPTY)
561 * copy the current shrinker scan count into a local variable
562 * and zero it so that other concurrent shrinker invocations
563 * don't also do this scanning work.
565 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
568 if (shrinker->seeks) {
569 delta = freeable >> priority;
571 do_div(delta, shrinker->seeks);
574 * These objects don't require any IO to create. Trim
575 * them aggressively under memory pressure to keep
576 * them from causing refetches in the IO caches.
578 delta = freeable / 2;
582 if (total_scan < 0) {
583 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
584 shrinker->scan_objects, total_scan);
585 total_scan = freeable;
588 next_deferred = total_scan;
591 * We need to avoid excessive windup on filesystem shrinkers
592 * due to large numbers of GFP_NOFS allocations causing the
593 * shrinkers to return -1 all the time. This results in a large
594 * nr being built up so when a shrink that can do some work
595 * comes along it empties the entire cache due to nr >>>
596 * freeable. This is bad for sustaining a working set in
599 * Hence only allow the shrinker to scan the entire cache when
600 * a large delta change is calculated directly.
602 if (delta < freeable / 4)
603 total_scan = min(total_scan, freeable / 2);
606 * Avoid risking looping forever due to too large nr value:
607 * never try to free more than twice the estimate number of
610 if (total_scan > freeable * 2)
611 total_scan = freeable * 2;
613 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
614 freeable, delta, total_scan, priority);
617 * Normally, we should not scan less than batch_size objects in one
618 * pass to avoid too frequent shrinker calls, but if the slab has less
619 * than batch_size objects in total and we are really tight on memory,
620 * we will try to reclaim all available objects, otherwise we can end
621 * up failing allocations although there are plenty of reclaimable
622 * objects spread over several slabs with usage less than the
625 * We detect the "tight on memory" situations by looking at the total
626 * number of objects we want to scan (total_scan). If it is greater
627 * than the total number of objects on slab (freeable), we must be
628 * scanning at high prio and therefore should try to reclaim as much as
631 while (total_scan >= batch_size ||
632 total_scan >= freeable) {
634 unsigned long nr_to_scan = min(batch_size, total_scan);
636 shrinkctl->nr_to_scan = nr_to_scan;
637 shrinkctl->nr_scanned = nr_to_scan;
638 ret = shrinker->scan_objects(shrinker, shrinkctl);
639 if (ret == SHRINK_STOP)
643 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
644 total_scan -= shrinkctl->nr_scanned;
645 scanned += shrinkctl->nr_scanned;
650 if (next_deferred >= scanned)
651 next_deferred -= scanned;
655 * move the unused scan count back into the shrinker in a
656 * manner that handles concurrent updates. If we exhausted the
657 * scan, there is no need to do an update.
659 if (next_deferred > 0)
660 new_nr = atomic_long_add_return(next_deferred,
661 &shrinker->nr_deferred[nid]);
663 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
665 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
670 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
671 struct mem_cgroup *memcg, int priority)
673 struct memcg_shrinker_map *map;
674 unsigned long ret, freed = 0;
677 if (!mem_cgroup_online(memcg))
680 if (!down_read_trylock(&shrinker_rwsem))
683 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
688 for_each_set_bit(i, map->map, shrinker_nr_max) {
689 struct shrink_control sc = {
690 .gfp_mask = gfp_mask,
694 struct shrinker *shrinker;
696 shrinker = idr_find(&shrinker_idr, i);
697 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
699 clear_bit(i, map->map);
703 /* Call non-slab shrinkers even though kmem is disabled */
704 if (!memcg_kmem_enabled() &&
705 !(shrinker->flags & SHRINKER_NONSLAB))
708 ret = do_shrink_slab(&sc, shrinker, priority);
709 if (ret == SHRINK_EMPTY) {
710 clear_bit(i, map->map);
712 * After the shrinker reported that it had no objects to
713 * free, but before we cleared the corresponding bit in
714 * the memcg shrinker map, a new object might have been
715 * added. To make sure, we have the bit set in this
716 * case, we invoke the shrinker one more time and reset
717 * the bit if it reports that it is not empty anymore.
718 * The memory barrier here pairs with the barrier in
719 * set_shrinker_bit():
721 * list_lru_add() shrink_slab_memcg()
722 * list_add_tail() clear_bit()
724 * set_bit() do_shrink_slab()
726 smp_mb__after_atomic();
727 ret = do_shrink_slab(&sc, shrinker, priority);
728 if (ret == SHRINK_EMPTY)
731 set_shrinker_bit(memcg, nid, i);
735 if (rwsem_is_contended(&shrinker_rwsem)) {
741 up_read(&shrinker_rwsem);
744 #else /* CONFIG_MEMCG */
745 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
746 struct mem_cgroup *memcg, int priority)
750 #endif /* CONFIG_MEMCG */
753 * shrink_slab - shrink slab caches
754 * @gfp_mask: allocation context
755 * @nid: node whose slab caches to target
756 * @memcg: memory cgroup whose slab caches to target
757 * @priority: the reclaim priority
759 * Call the shrink functions to age shrinkable caches.
761 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
762 * unaware shrinkers will receive a node id of 0 instead.
764 * @memcg specifies the memory cgroup to target. Unaware shrinkers
765 * are called only if it is the root cgroup.
767 * @priority is sc->priority, we take the number of objects and >> by priority
768 * in order to get the scan target.
770 * Returns the number of reclaimed slab objects.
772 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
773 struct mem_cgroup *memcg,
776 unsigned long ret, freed = 0;
777 struct shrinker *shrinker;
780 * The root memcg might be allocated even though memcg is disabled
781 * via "cgroup_disable=memory" boot parameter. This could make
782 * mem_cgroup_is_root() return false, then just run memcg slab
783 * shrink, but skip global shrink. This may result in premature
786 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
787 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
789 if (!down_read_trylock(&shrinker_rwsem))
792 list_for_each_entry(shrinker, &shrinker_list, list) {
793 struct shrink_control sc = {
794 .gfp_mask = gfp_mask,
799 ret = do_shrink_slab(&sc, shrinker, priority);
800 if (ret == SHRINK_EMPTY)
804 * Bail out if someone want to register a new shrinker to
805 * prevent the registration from being stalled for long periods
806 * by parallel ongoing shrinking.
808 if (rwsem_is_contended(&shrinker_rwsem)) {
814 up_read(&shrinker_rwsem);
820 void drop_slab_node(int nid)
825 struct mem_cgroup *memcg = NULL;
827 if (fatal_signal_pending(current))
831 memcg = mem_cgroup_iter(NULL, NULL, NULL);
833 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
834 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
835 } while (freed > 10);
842 for_each_online_node(nid)
846 static inline int is_page_cache_freeable(struct page *page)
849 * A freeable page cache page is referenced only by the caller
850 * that isolated the page, the page cache and optional buffer
851 * heads at page->private.
853 int page_cache_pins = thp_nr_pages(page);
854 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
857 static int may_write_to_inode(struct inode *inode)
859 if (current->flags & PF_SWAPWRITE)
861 if (!inode_write_congested(inode))
863 if (inode_to_bdi(inode) == current->backing_dev_info)
869 * We detected a synchronous write error writing a page out. Probably
870 * -ENOSPC. We need to propagate that into the address_space for a subsequent
871 * fsync(), msync() or close().
873 * The tricky part is that after writepage we cannot touch the mapping: nothing
874 * prevents it from being freed up. But we have a ref on the page and once
875 * that page is locked, the mapping is pinned.
877 * We're allowed to run sleeping lock_page() here because we know the caller has
880 static void handle_write_error(struct address_space *mapping,
881 struct page *page, int error)
884 if (page_mapping(page) == mapping)
885 mapping_set_error(mapping, error);
889 /* possible outcome of pageout() */
891 /* failed to write page out, page is locked */
893 /* move page to the active list, page is locked */
895 /* page has been sent to the disk successfully, page is unlocked */
897 /* page is clean and locked */
902 * pageout is called by shrink_page_list() for each dirty page.
903 * Calls ->writepage().
905 static pageout_t pageout(struct page *page, struct address_space *mapping)
908 * If the page is dirty, only perform writeback if that write
909 * will be non-blocking. To prevent this allocation from being
910 * stalled by pagecache activity. But note that there may be
911 * stalls if we need to run get_block(). We could test
912 * PagePrivate for that.
914 * If this process is currently in __generic_file_write_iter() against
915 * this page's queue, we can perform writeback even if that
918 * If the page is swapcache, write it back even if that would
919 * block, for some throttling. This happens by accident, because
920 * swap_backing_dev_info is bust: it doesn't reflect the
921 * congestion state of the swapdevs. Easy to fix, if needed.
923 if (!is_page_cache_freeable(page))
927 * Some data journaling orphaned pages can have
928 * page->mapping == NULL while being dirty with clean buffers.
930 if (page_has_private(page)) {
931 if (try_to_free_buffers(page)) {
932 ClearPageDirty(page);
933 pr_info("%s: orphaned page\n", __func__);
939 if (mapping->a_ops->writepage == NULL)
940 return PAGE_ACTIVATE;
941 if (!may_write_to_inode(mapping->host))
944 if (clear_page_dirty_for_io(page)) {
946 struct writeback_control wbc = {
947 .sync_mode = WB_SYNC_NONE,
948 .nr_to_write = SWAP_CLUSTER_MAX,
950 .range_end = LLONG_MAX,
954 SetPageReclaim(page);
955 res = mapping->a_ops->writepage(page, &wbc);
957 handle_write_error(mapping, page, res);
958 if (res == AOP_WRITEPAGE_ACTIVATE) {
959 ClearPageReclaim(page);
960 return PAGE_ACTIVATE;
963 if (!PageWriteback(page)) {
964 /* synchronous write or broken a_ops? */
965 ClearPageReclaim(page);
967 trace_mm_vmscan_writepage(page);
968 inc_node_page_state(page, NR_VMSCAN_WRITE);
976 * Same as remove_mapping, but if the page is removed from the mapping, it
977 * gets returned with a refcount of 0.
979 static int __remove_mapping(struct address_space *mapping, struct page *page,
980 bool reclaimed, struct mem_cgroup *target_memcg)
986 BUG_ON(!PageLocked(page));
987 BUG_ON(mapping != page_mapping(page));
989 xa_lock_irqsave(&mapping->i_pages, flags);
991 * The non racy check for a busy page.
993 * Must be careful with the order of the tests. When someone has
994 * a ref to the page, it may be possible that they dirty it then
995 * drop the reference. So if PageDirty is tested before page_count
996 * here, then the following race may occur:
998 * get_user_pages(&page);
999 * [user mapping goes away]
1001 * !PageDirty(page) [good]
1002 * SetPageDirty(page);
1004 * !page_count(page) [good, discard it]
1006 * [oops, our write_to data is lost]
1008 * Reversing the order of the tests ensures such a situation cannot
1009 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1010 * load is not satisfied before that of page->_refcount.
1012 * Note that if SetPageDirty is always performed via set_page_dirty,
1013 * and thus under the i_pages lock, then this ordering is not required.
1015 refcount = 1 + compound_nr(page);
1016 if (!page_ref_freeze(page, refcount))
1018 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1019 if (unlikely(PageDirty(page))) {
1020 page_ref_unfreeze(page, refcount);
1024 if (PageSwapCache(page)) {
1025 swp_entry_t swap = { .val = page_private(page) };
1026 mem_cgroup_swapout(page, swap);
1027 if (reclaimed && !mapping_exiting(mapping))
1028 shadow = workingset_eviction(page, target_memcg);
1029 __delete_from_swap_cache(page, swap, shadow);
1030 xa_unlock_irqrestore(&mapping->i_pages, flags);
1031 put_swap_page(page, swap);
1033 void (*freepage)(struct page *);
1035 freepage = mapping->a_ops->freepage;
1037 * Remember a shadow entry for reclaimed file cache in
1038 * order to detect refaults, thus thrashing, later on.
1040 * But don't store shadows in an address space that is
1041 * already exiting. This is not just an optimization,
1042 * inode reclaim needs to empty out the radix tree or
1043 * the nodes are lost. Don't plant shadows behind its
1046 * We also don't store shadows for DAX mappings because the
1047 * only page cache pages found in these are zero pages
1048 * covering holes, and because we don't want to mix DAX
1049 * exceptional entries and shadow exceptional entries in the
1050 * same address_space.
1052 if (reclaimed && page_is_file_lru(page) &&
1053 !mapping_exiting(mapping) && !dax_mapping(mapping))
1054 shadow = workingset_eviction(page, target_memcg);
1055 __delete_from_page_cache(page, shadow);
1056 xa_unlock_irqrestore(&mapping->i_pages, flags);
1058 if (freepage != NULL)
1065 xa_unlock_irqrestore(&mapping->i_pages, flags);
1070 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1071 * someone else has a ref on the page, abort and return 0. If it was
1072 * successfully detached, return 1. Assumes the caller has a single ref on
1075 int remove_mapping(struct address_space *mapping, struct page *page)
1077 if (__remove_mapping(mapping, page, false, NULL)) {
1079 * Unfreezing the refcount with 1 rather than 2 effectively
1080 * drops the pagecache ref for us without requiring another
1083 page_ref_unfreeze(page, 1);
1090 * putback_lru_page - put previously isolated page onto appropriate LRU list
1091 * @page: page to be put back to appropriate lru list
1093 * Add previously isolated @page to appropriate LRU list.
1094 * Page may still be unevictable for other reasons.
1096 * lru_lock must not be held, interrupts must be enabled.
1098 void putback_lru_page(struct page *page)
1100 lru_cache_add(page);
1101 put_page(page); /* drop ref from isolate */
1104 enum page_references {
1106 PAGEREF_RECLAIM_CLEAN,
1111 static enum page_references page_check_references(struct page *page,
1112 struct scan_control *sc)
1114 int referenced_ptes, referenced_page;
1115 unsigned long vm_flags;
1117 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1119 referenced_page = TestClearPageReferenced(page);
1122 * Mlock lost the isolation race with us. Let try_to_unmap()
1123 * move the page to the unevictable list.
1125 if (vm_flags & VM_LOCKED)
1126 return PAGEREF_RECLAIM;
1128 if (referenced_ptes) {
1130 * All mapped pages start out with page table
1131 * references from the instantiating fault, so we need
1132 * to look twice if a mapped file page is used more
1135 * Mark it and spare it for another trip around the
1136 * inactive list. Another page table reference will
1137 * lead to its activation.
1139 * Note: the mark is set for activated pages as well
1140 * so that recently deactivated but used pages are
1141 * quickly recovered.
1143 SetPageReferenced(page);
1145 if (referenced_page || referenced_ptes > 1)
1146 return PAGEREF_ACTIVATE;
1149 * Activate file-backed executable pages after first usage.
1151 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1152 return PAGEREF_ACTIVATE;
1154 return PAGEREF_KEEP;
1157 /* Reclaim if clean, defer dirty pages to writeback */
1158 if (referenced_page && !PageSwapBacked(page))
1159 return PAGEREF_RECLAIM_CLEAN;
1161 return PAGEREF_RECLAIM;
1164 /* Check if a page is dirty or under writeback */
1165 static void page_check_dirty_writeback(struct page *page,
1166 bool *dirty, bool *writeback)
1168 struct address_space *mapping;
1171 * Anonymous pages are not handled by flushers and must be written
1172 * from reclaim context. Do not stall reclaim based on them
1174 if (!page_is_file_lru(page) ||
1175 (PageAnon(page) && !PageSwapBacked(page))) {
1181 /* By default assume that the page flags are accurate */
1182 *dirty = PageDirty(page);
1183 *writeback = PageWriteback(page);
1185 /* Verify dirty/writeback state if the filesystem supports it */
1186 if (!page_has_private(page))
1189 mapping = page_mapping(page);
1190 if (mapping && mapping->a_ops->is_dirty_writeback)
1191 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1195 * shrink_page_list() returns the number of reclaimed pages
1197 static unsigned int shrink_page_list(struct list_head *page_list,
1198 struct pglist_data *pgdat,
1199 struct scan_control *sc,
1200 struct reclaim_stat *stat,
1201 bool ignore_references)
1203 LIST_HEAD(ret_pages);
1204 LIST_HEAD(free_pages);
1205 unsigned int nr_reclaimed = 0;
1206 unsigned int pgactivate = 0;
1208 memset(stat, 0, sizeof(*stat));
1211 while (!list_empty(page_list)) {
1212 struct address_space *mapping;
1214 enum page_references references = PAGEREF_RECLAIM;
1215 bool dirty, writeback, may_enter_fs;
1216 unsigned int nr_pages;
1220 page = lru_to_page(page_list);
1221 list_del(&page->lru);
1223 if (!trylock_page(page))
1226 VM_BUG_ON_PAGE(PageActive(page), page);
1228 nr_pages = compound_nr(page);
1230 /* Account the number of base pages even though THP */
1231 sc->nr_scanned += nr_pages;
1233 if (unlikely(!page_evictable(page)))
1234 goto activate_locked;
1236 if (!sc->may_unmap && page_mapped(page))
1239 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1240 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1243 * The number of dirty pages determines if a node is marked
1244 * reclaim_congested which affects wait_iff_congested. kswapd
1245 * will stall and start writing pages if the tail of the LRU
1246 * is all dirty unqueued pages.
1248 page_check_dirty_writeback(page, &dirty, &writeback);
1249 if (dirty || writeback)
1252 if (dirty && !writeback)
1253 stat->nr_unqueued_dirty++;
1256 * Treat this page as congested if the underlying BDI is or if
1257 * pages are cycling through the LRU so quickly that the
1258 * pages marked for immediate reclaim are making it to the
1259 * end of the LRU a second time.
1261 mapping = page_mapping(page);
1262 if (((dirty || writeback) && mapping &&
1263 inode_write_congested(mapping->host)) ||
1264 (writeback && PageReclaim(page)))
1265 stat->nr_congested++;
1268 * If a page at the tail of the LRU is under writeback, there
1269 * are three cases to consider.
1271 * 1) If reclaim is encountering an excessive number of pages
1272 * under writeback and this page is both under writeback and
1273 * PageReclaim then it indicates that pages are being queued
1274 * for IO but are being recycled through the LRU before the
1275 * IO can complete. Waiting on the page itself risks an
1276 * indefinite stall if it is impossible to writeback the
1277 * page due to IO error or disconnected storage so instead
1278 * note that the LRU is being scanned too quickly and the
1279 * caller can stall after page list has been processed.
1281 * 2) Global or new memcg reclaim encounters a page that is
1282 * not marked for immediate reclaim, or the caller does not
1283 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1284 * not to fs). In this case mark the page for immediate
1285 * reclaim and continue scanning.
1287 * Require may_enter_fs because we would wait on fs, which
1288 * may not have submitted IO yet. And the loop driver might
1289 * enter reclaim, and deadlock if it waits on a page for
1290 * which it is needed to do the write (loop masks off
1291 * __GFP_IO|__GFP_FS for this reason); but more thought
1292 * would probably show more reasons.
1294 * 3) Legacy memcg encounters a page that is already marked
1295 * PageReclaim. memcg does not have any dirty pages
1296 * throttling so we could easily OOM just because too many
1297 * pages are in writeback and there is nothing else to
1298 * reclaim. Wait for the writeback to complete.
1300 * In cases 1) and 2) we activate the pages to get them out of
1301 * the way while we continue scanning for clean pages on the
1302 * inactive list and refilling from the active list. The
1303 * observation here is that waiting for disk writes is more
1304 * expensive than potentially causing reloads down the line.
1305 * Since they're marked for immediate reclaim, they won't put
1306 * memory pressure on the cache working set any longer than it
1307 * takes to write them to disk.
1309 if (PageWriteback(page)) {
1311 if (current_is_kswapd() &&
1312 PageReclaim(page) &&
1313 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1314 stat->nr_immediate++;
1315 goto activate_locked;
1318 } else if (writeback_throttling_sane(sc) ||
1319 !PageReclaim(page) || !may_enter_fs) {
1321 * This is slightly racy - end_page_writeback()
1322 * might have just cleared PageReclaim, then
1323 * setting PageReclaim here end up interpreted
1324 * as PageReadahead - but that does not matter
1325 * enough to care. What we do want is for this
1326 * page to have PageReclaim set next time memcg
1327 * reclaim reaches the tests above, so it will
1328 * then wait_on_page_writeback() to avoid OOM;
1329 * and it's also appropriate in global reclaim.
1331 SetPageReclaim(page);
1332 stat->nr_writeback++;
1333 goto activate_locked;
1338 wait_on_page_writeback(page);
1339 /* then go back and try same page again */
1340 list_add_tail(&page->lru, page_list);
1345 if (!ignore_references)
1346 references = page_check_references(page, sc);
1348 switch (references) {
1349 case PAGEREF_ACTIVATE:
1350 goto activate_locked;
1352 stat->nr_ref_keep += nr_pages;
1354 case PAGEREF_RECLAIM:
1355 case PAGEREF_RECLAIM_CLEAN:
1356 ; /* try to reclaim the page below */
1360 * Anonymous process memory has backing store?
1361 * Try to allocate it some swap space here.
1362 * Lazyfree page could be freed directly
1364 if (PageAnon(page) && PageSwapBacked(page)) {
1365 if (!PageSwapCache(page)) {
1366 if (!(sc->gfp_mask & __GFP_IO))
1368 if (page_maybe_dma_pinned(page))
1370 if (PageTransHuge(page)) {
1371 /* cannot split THP, skip it */
1372 if (!can_split_huge_page(page, NULL))
1373 goto activate_locked;
1375 * Split pages without a PMD map right
1376 * away. Chances are some or all of the
1377 * tail pages can be freed without IO.
1379 if (!compound_mapcount(page) &&
1380 split_huge_page_to_list(page,
1382 goto activate_locked;
1384 if (!add_to_swap(page)) {
1385 if (!PageTransHuge(page))
1386 goto activate_locked_split;
1387 /* Fallback to swap normal pages */
1388 if (split_huge_page_to_list(page,
1390 goto activate_locked;
1391 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1392 count_vm_event(THP_SWPOUT_FALLBACK);
1394 if (!add_to_swap(page))
1395 goto activate_locked_split;
1398 may_enter_fs = true;
1400 /* Adding to swap updated mapping */
1401 mapping = page_mapping(page);
1403 } else if (unlikely(PageTransHuge(page))) {
1404 /* Split file THP */
1405 if (split_huge_page_to_list(page, page_list))
1410 * THP may get split above, need minus tail pages and update
1411 * nr_pages to avoid accounting tail pages twice.
1413 * The tail pages that are added into swap cache successfully
1416 if ((nr_pages > 1) && !PageTransHuge(page)) {
1417 sc->nr_scanned -= (nr_pages - 1);
1422 * The page is mapped into the page tables of one or more
1423 * processes. Try to unmap it here.
1425 if (page_mapped(page)) {
1426 enum ttu_flags flags = TTU_BATCH_FLUSH;
1427 bool was_swapbacked = PageSwapBacked(page);
1429 if (unlikely(PageTransHuge(page)))
1430 flags |= TTU_SPLIT_HUGE_PMD;
1432 if (!try_to_unmap(page, flags)) {
1433 stat->nr_unmap_fail += nr_pages;
1434 if (!was_swapbacked && PageSwapBacked(page))
1435 stat->nr_lazyfree_fail += nr_pages;
1436 goto activate_locked;
1440 if (PageDirty(page)) {
1442 * Only kswapd can writeback filesystem pages
1443 * to avoid risk of stack overflow. But avoid
1444 * injecting inefficient single-page IO into
1445 * flusher writeback as much as possible: only
1446 * write pages when we've encountered many
1447 * dirty pages, and when we've already scanned
1448 * the rest of the LRU for clean pages and see
1449 * the same dirty pages again (PageReclaim).
1451 if (page_is_file_lru(page) &&
1452 (!current_is_kswapd() || !PageReclaim(page) ||
1453 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1455 * Immediately reclaim when written back.
1456 * Similar in principal to deactivate_page()
1457 * except we already have the page isolated
1458 * and know it's dirty
1460 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1461 SetPageReclaim(page);
1463 goto activate_locked;
1466 if (references == PAGEREF_RECLAIM_CLEAN)
1470 if (!sc->may_writepage)
1474 * Page is dirty. Flush the TLB if a writable entry
1475 * potentially exists to avoid CPU writes after IO
1476 * starts and then write it out here.
1478 try_to_unmap_flush_dirty();
1479 switch (pageout(page, mapping)) {
1483 goto activate_locked;
1485 stat->nr_pageout += thp_nr_pages(page);
1487 if (PageWriteback(page))
1489 if (PageDirty(page))
1493 * A synchronous write - probably a ramdisk. Go
1494 * ahead and try to reclaim the page.
1496 if (!trylock_page(page))
1498 if (PageDirty(page) || PageWriteback(page))
1500 mapping = page_mapping(page);
1503 ; /* try to free the page below */
1508 * If the page has buffers, try to free the buffer mappings
1509 * associated with this page. If we succeed we try to free
1512 * We do this even if the page is PageDirty().
1513 * try_to_release_page() does not perform I/O, but it is
1514 * possible for a page to have PageDirty set, but it is actually
1515 * clean (all its buffers are clean). This happens if the
1516 * buffers were written out directly, with submit_bh(). ext3
1517 * will do this, as well as the blockdev mapping.
1518 * try_to_release_page() will discover that cleanness and will
1519 * drop the buffers and mark the page clean - it can be freed.
1521 * Rarely, pages can have buffers and no ->mapping. These are
1522 * the pages which were not successfully invalidated in
1523 * truncate_cleanup_page(). We try to drop those buffers here
1524 * and if that worked, and the page is no longer mapped into
1525 * process address space (page_count == 1) it can be freed.
1526 * Otherwise, leave the page on the LRU so it is swappable.
1528 if (page_has_private(page)) {
1529 if (!try_to_release_page(page, sc->gfp_mask))
1530 goto activate_locked;
1531 if (!mapping && page_count(page) == 1) {
1533 if (put_page_testzero(page))
1537 * rare race with speculative reference.
1538 * the speculative reference will free
1539 * this page shortly, so we may
1540 * increment nr_reclaimed here (and
1541 * leave it off the LRU).
1549 if (PageAnon(page) && !PageSwapBacked(page)) {
1550 /* follow __remove_mapping for reference */
1551 if (!page_ref_freeze(page, 1))
1553 if (PageDirty(page)) {
1554 page_ref_unfreeze(page, 1);
1558 count_vm_event(PGLAZYFREED);
1559 count_memcg_page_event(page, PGLAZYFREED);
1560 } else if (!mapping || !__remove_mapping(mapping, page, true,
1561 sc->target_mem_cgroup))
1567 * THP may get swapped out in a whole, need account
1570 nr_reclaimed += nr_pages;
1573 * Is there need to periodically free_page_list? It would
1574 * appear not as the counts should be low
1576 if (unlikely(PageTransHuge(page)))
1577 destroy_compound_page(page);
1579 list_add(&page->lru, &free_pages);
1582 activate_locked_split:
1584 * The tail pages that are failed to add into swap cache
1585 * reach here. Fixup nr_scanned and nr_pages.
1588 sc->nr_scanned -= (nr_pages - 1);
1592 /* Not a candidate for swapping, so reclaim swap space. */
1593 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1595 try_to_free_swap(page);
1596 VM_BUG_ON_PAGE(PageActive(page), page);
1597 if (!PageMlocked(page)) {
1598 int type = page_is_file_lru(page);
1599 SetPageActive(page);
1600 stat->nr_activate[type] += nr_pages;
1601 count_memcg_page_event(page, PGACTIVATE);
1606 list_add(&page->lru, &ret_pages);
1607 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1610 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1612 mem_cgroup_uncharge_list(&free_pages);
1613 try_to_unmap_flush();
1614 free_unref_page_list(&free_pages);
1616 list_splice(&ret_pages, page_list);
1617 count_vm_events(PGACTIVATE, pgactivate);
1619 return nr_reclaimed;
1622 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1623 struct list_head *page_list)
1625 struct scan_control sc = {
1626 .gfp_mask = GFP_KERNEL,
1627 .priority = DEF_PRIORITY,
1630 struct reclaim_stat stat;
1631 unsigned int nr_reclaimed;
1632 struct page *page, *next;
1633 LIST_HEAD(clean_pages);
1635 list_for_each_entry_safe(page, next, page_list, lru) {
1636 if (!PageHuge(page) && page_is_file_lru(page) &&
1637 !PageDirty(page) && !__PageMovable(page) &&
1638 !PageUnevictable(page)) {
1639 ClearPageActive(page);
1640 list_move(&page->lru, &clean_pages);
1644 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1646 list_splice(&clean_pages, page_list);
1647 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1648 -(long)nr_reclaimed);
1650 * Since lazyfree pages are isolated from file LRU from the beginning,
1651 * they will rotate back to anonymous LRU in the end if it failed to
1652 * discard so isolated count will be mismatched.
1653 * Compensate the isolated count for both LRU lists.
1655 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1656 stat.nr_lazyfree_fail);
1657 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1658 -(long)stat.nr_lazyfree_fail);
1659 return nr_reclaimed;
1663 * Attempt to remove the specified page from its LRU. Only take this page
1664 * if it is of the appropriate PageActive status. Pages which are being
1665 * freed elsewhere are also ignored.
1667 * page: page to consider
1668 * mode: one of the LRU isolation modes defined above
1670 * returns true on success, false on failure.
1672 bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
1674 /* Only take pages on the LRU. */
1678 /* Compaction should not handle unevictable pages but CMA can do so */
1679 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1683 * To minimise LRU disruption, the caller can indicate that it only
1684 * wants to isolate pages it will be able to operate on without
1685 * blocking - clean pages for the most part.
1687 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1688 * that it is possible to migrate without blocking
1690 if (mode & ISOLATE_ASYNC_MIGRATE) {
1691 /* All the caller can do on PageWriteback is block */
1692 if (PageWriteback(page))
1695 if (PageDirty(page)) {
1696 struct address_space *mapping;
1700 * Only pages without mappings or that have a
1701 * ->migratepage callback are possible to migrate
1702 * without blocking. However, we can be racing with
1703 * truncation so it's necessary to lock the page
1704 * to stabilise the mapping as truncation holds
1705 * the page lock until after the page is removed
1706 * from the page cache.
1708 if (!trylock_page(page))
1711 mapping = page_mapping(page);
1712 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1719 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1726 * Update LRU sizes after isolating pages. The LRU size updates must
1727 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1729 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1730 enum lru_list lru, unsigned long *nr_zone_taken)
1734 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1735 if (!nr_zone_taken[zid])
1738 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1744 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1746 * lruvec->lru_lock is heavily contended. Some of the functions that
1747 * shrink the lists perform better by taking out a batch of pages
1748 * and working on them outside the LRU lock.
1750 * For pagecache intensive workloads, this function is the hottest
1751 * spot in the kernel (apart from copy_*_user functions).
1753 * Lru_lock must be held before calling this function.
1755 * @nr_to_scan: The number of eligible pages to look through on the list.
1756 * @lruvec: The LRU vector to pull pages from.
1757 * @dst: The temp list to put pages on to.
1758 * @nr_scanned: The number of pages that were scanned.
1759 * @sc: The scan_control struct for this reclaim session
1760 * @lru: LRU list id for isolating
1762 * returns how many pages were moved onto *@dst.
1764 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1765 struct lruvec *lruvec, struct list_head *dst,
1766 unsigned long *nr_scanned, struct scan_control *sc,
1769 struct list_head *src = &lruvec->lists[lru];
1770 unsigned long nr_taken = 0;
1771 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1772 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1773 unsigned long skipped = 0;
1774 unsigned long scan, total_scan, nr_pages;
1775 LIST_HEAD(pages_skipped);
1776 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1780 while (scan < nr_to_scan && !list_empty(src)) {
1783 page = lru_to_page(src);
1784 prefetchw_prev_lru_page(page, src, flags);
1786 nr_pages = compound_nr(page);
1787 total_scan += nr_pages;
1789 if (page_zonenum(page) > sc->reclaim_idx) {
1790 list_move(&page->lru, &pages_skipped);
1791 nr_skipped[page_zonenum(page)] += nr_pages;
1796 * Do not count skipped pages because that makes the function
1797 * return with no isolated pages if the LRU mostly contains
1798 * ineligible pages. This causes the VM to not reclaim any
1799 * pages, triggering a premature OOM.
1801 * Account all tail pages of THP. This would not cause
1802 * premature OOM since __isolate_lru_page() returns -EBUSY
1803 * only when the page is being freed somewhere else.
1806 if (!__isolate_lru_page_prepare(page, mode)) {
1807 /* It is being freed elsewhere */
1808 list_move(&page->lru, src);
1812 * Be careful not to clear PageLRU until after we're
1813 * sure the page is not being freed elsewhere -- the
1814 * page release code relies on it.
1816 if (unlikely(!get_page_unless_zero(page))) {
1817 list_move(&page->lru, src);
1821 if (!TestClearPageLRU(page)) {
1822 /* Another thread is already isolating this page */
1824 list_move(&page->lru, src);
1828 nr_taken += nr_pages;
1829 nr_zone_taken[page_zonenum(page)] += nr_pages;
1830 list_move(&page->lru, dst);
1834 * Splice any skipped pages to the start of the LRU list. Note that
1835 * this disrupts the LRU order when reclaiming for lower zones but
1836 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1837 * scanning would soon rescan the same pages to skip and put the
1838 * system at risk of premature OOM.
1840 if (!list_empty(&pages_skipped)) {
1843 list_splice(&pages_skipped, src);
1844 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1845 if (!nr_skipped[zid])
1848 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1849 skipped += nr_skipped[zid];
1852 *nr_scanned = total_scan;
1853 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1854 total_scan, skipped, nr_taken, mode, lru);
1855 update_lru_sizes(lruvec, lru, nr_zone_taken);
1860 * isolate_lru_page - tries to isolate a page from its LRU list
1861 * @page: page to isolate from its LRU list
1863 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1864 * vmstat statistic corresponding to whatever LRU list the page was on.
1866 * Returns 0 if the page was removed from an LRU list.
1867 * Returns -EBUSY if the page was not on an LRU list.
1869 * The returned page will have PageLRU() cleared. If it was found on
1870 * the active list, it will have PageActive set. If it was found on
1871 * the unevictable list, it will have the PageUnevictable bit set. That flag
1872 * may need to be cleared by the caller before letting the page go.
1874 * The vmstat statistic corresponding to the list on which the page was
1875 * found will be decremented.
1879 * (1) Must be called with an elevated refcount on the page. This is a
1880 * fundamental difference from isolate_lru_pages (which is called
1881 * without a stable reference).
1882 * (2) the lru_lock must not be held.
1883 * (3) interrupts must be enabled.
1885 int isolate_lru_page(struct page *page)
1889 VM_BUG_ON_PAGE(!page_count(page), page);
1890 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1892 if (TestClearPageLRU(page)) {
1893 struct lruvec *lruvec;
1896 lruvec = lock_page_lruvec_irq(page);
1897 del_page_from_lru_list(page, lruvec);
1898 unlock_page_lruvec_irq(lruvec);
1906 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1907 * then get rescheduled. When there are massive number of tasks doing page
1908 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1909 * the LRU list will go small and be scanned faster than necessary, leading to
1910 * unnecessary swapping, thrashing and OOM.
1912 static int too_many_isolated(struct pglist_data *pgdat, int file,
1913 struct scan_control *sc)
1915 unsigned long inactive, isolated;
1917 if (current_is_kswapd())
1920 if (!writeback_throttling_sane(sc))
1924 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1925 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1927 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1928 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1932 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1933 * won't get blocked by normal direct-reclaimers, forming a circular
1936 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1939 return isolated > inactive;
1943 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
1944 * On return, @list is reused as a list of pages to be freed by the caller.
1946 * Returns the number of pages moved to the given lruvec.
1948 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1949 struct list_head *list)
1951 int nr_pages, nr_moved = 0;
1952 LIST_HEAD(pages_to_free);
1955 while (!list_empty(list)) {
1956 page = lru_to_page(list);
1957 VM_BUG_ON_PAGE(PageLRU(page), page);
1958 list_del(&page->lru);
1959 if (unlikely(!page_evictable(page))) {
1960 spin_unlock_irq(&lruvec->lru_lock);
1961 putback_lru_page(page);
1962 spin_lock_irq(&lruvec->lru_lock);
1967 * The SetPageLRU needs to be kept here for list integrity.
1969 * #0 move_pages_to_lru #1 release_pages
1970 * if !put_page_testzero
1971 * if (put_page_testzero())
1972 * !PageLRU //skip lru_lock
1974 * list_add(&page->lru,)
1975 * list_add(&page->lru,)
1979 if (unlikely(put_page_testzero(page))) {
1980 __clear_page_lru_flags(page);
1982 if (unlikely(PageCompound(page))) {
1983 spin_unlock_irq(&lruvec->lru_lock);
1984 destroy_compound_page(page);
1985 spin_lock_irq(&lruvec->lru_lock);
1987 list_add(&page->lru, &pages_to_free);
1993 * All pages were isolated from the same lruvec (and isolation
1994 * inhibits memcg migration).
1996 VM_BUG_ON_PAGE(!lruvec_holds_page_lru_lock(page, lruvec), page);
1997 add_page_to_lru_list(page, lruvec);
1998 nr_pages = thp_nr_pages(page);
1999 nr_moved += nr_pages;
2000 if (PageActive(page))
2001 workingset_age_nonresident(lruvec, nr_pages);
2005 * To save our caller's stack, now use input list for pages to free.
2007 list_splice(&pages_to_free, list);
2013 * If a kernel thread (such as nfsd for loop-back mounts) services
2014 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2015 * In that case we should only throttle if the backing device it is
2016 * writing to is congested. In other cases it is safe to throttle.
2018 static int current_may_throttle(void)
2020 return !(current->flags & PF_LOCAL_THROTTLE) ||
2021 current->backing_dev_info == NULL ||
2022 bdi_write_congested(current->backing_dev_info);
2026 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2027 * of reclaimed pages
2029 static noinline_for_stack unsigned long
2030 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2031 struct scan_control *sc, enum lru_list lru)
2033 LIST_HEAD(page_list);
2034 unsigned long nr_scanned;
2035 unsigned int nr_reclaimed = 0;
2036 unsigned long nr_taken;
2037 struct reclaim_stat stat;
2038 bool file = is_file_lru(lru);
2039 enum vm_event_item item;
2040 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2041 bool stalled = false;
2043 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2047 /* wait a bit for the reclaimer. */
2051 /* We are about to die and free our memory. Return now. */
2052 if (fatal_signal_pending(current))
2053 return SWAP_CLUSTER_MAX;
2058 spin_lock_irq(&lruvec->lru_lock);
2060 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2061 &nr_scanned, sc, lru);
2063 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2064 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2065 if (!cgroup_reclaim(sc))
2066 __count_vm_events(item, nr_scanned);
2067 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2068 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2070 spin_unlock_irq(&lruvec->lru_lock);
2075 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2077 spin_lock_irq(&lruvec->lru_lock);
2078 move_pages_to_lru(lruvec, &page_list);
2080 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2081 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2082 if (!cgroup_reclaim(sc))
2083 __count_vm_events(item, nr_reclaimed);
2084 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2085 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2086 spin_unlock_irq(&lruvec->lru_lock);
2088 lru_note_cost(lruvec, file, stat.nr_pageout);
2089 mem_cgroup_uncharge_list(&page_list);
2090 free_unref_page_list(&page_list);
2093 * If dirty pages are scanned that are not queued for IO, it
2094 * implies that flushers are not doing their job. This can
2095 * happen when memory pressure pushes dirty pages to the end of
2096 * the LRU before the dirty limits are breached and the dirty
2097 * data has expired. It can also happen when the proportion of
2098 * dirty pages grows not through writes but through memory
2099 * pressure reclaiming all the clean cache. And in some cases,
2100 * the flushers simply cannot keep up with the allocation
2101 * rate. Nudge the flusher threads in case they are asleep.
2103 if (stat.nr_unqueued_dirty == nr_taken)
2104 wakeup_flusher_threads(WB_REASON_VMSCAN);
2106 sc->nr.dirty += stat.nr_dirty;
2107 sc->nr.congested += stat.nr_congested;
2108 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2109 sc->nr.writeback += stat.nr_writeback;
2110 sc->nr.immediate += stat.nr_immediate;
2111 sc->nr.taken += nr_taken;
2113 sc->nr.file_taken += nr_taken;
2115 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2116 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2117 return nr_reclaimed;
2121 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2123 * We move them the other way if the page is referenced by one or more
2126 * If the pages are mostly unmapped, the processing is fast and it is
2127 * appropriate to hold lru_lock across the whole operation. But if
2128 * the pages are mapped, the processing is slow (page_referenced()), so
2129 * we should drop lru_lock around each page. It's impossible to balance
2130 * this, so instead we remove the pages from the LRU while processing them.
2131 * It is safe to rely on PG_active against the non-LRU pages in here because
2132 * nobody will play with that bit on a non-LRU page.
2134 * The downside is that we have to touch page->_refcount against each page.
2135 * But we had to alter page->flags anyway.
2137 static void shrink_active_list(unsigned long nr_to_scan,
2138 struct lruvec *lruvec,
2139 struct scan_control *sc,
2142 unsigned long nr_taken;
2143 unsigned long nr_scanned;
2144 unsigned long vm_flags;
2145 LIST_HEAD(l_hold); /* The pages which were snipped off */
2146 LIST_HEAD(l_active);
2147 LIST_HEAD(l_inactive);
2149 unsigned nr_deactivate, nr_activate;
2150 unsigned nr_rotated = 0;
2151 int file = is_file_lru(lru);
2152 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2156 spin_lock_irq(&lruvec->lru_lock);
2158 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2159 &nr_scanned, sc, lru);
2161 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2163 if (!cgroup_reclaim(sc))
2164 __count_vm_events(PGREFILL, nr_scanned);
2165 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2167 spin_unlock_irq(&lruvec->lru_lock);
2169 while (!list_empty(&l_hold)) {
2171 page = lru_to_page(&l_hold);
2172 list_del(&page->lru);
2174 if (unlikely(!page_evictable(page))) {
2175 putback_lru_page(page);
2179 if (unlikely(buffer_heads_over_limit)) {
2180 if (page_has_private(page) && trylock_page(page)) {
2181 if (page_has_private(page))
2182 try_to_release_page(page, 0);
2187 if (page_referenced(page, 0, sc->target_mem_cgroup,
2190 * Identify referenced, file-backed active pages and
2191 * give them one more trip around the active list. So
2192 * that executable code get better chances to stay in
2193 * memory under moderate memory pressure. Anon pages
2194 * are not likely to be evicted by use-once streaming
2195 * IO, plus JVM can create lots of anon VM_EXEC pages,
2196 * so we ignore them here.
2198 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2199 nr_rotated += thp_nr_pages(page);
2200 list_add(&page->lru, &l_active);
2205 ClearPageActive(page); /* we are de-activating */
2206 SetPageWorkingset(page);
2207 list_add(&page->lru, &l_inactive);
2211 * Move pages back to the lru list.
2213 spin_lock_irq(&lruvec->lru_lock);
2215 nr_activate = move_pages_to_lru(lruvec, &l_active);
2216 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2217 /* Keep all free pages in l_active list */
2218 list_splice(&l_inactive, &l_active);
2220 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2221 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2223 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2224 spin_unlock_irq(&lruvec->lru_lock);
2226 mem_cgroup_uncharge_list(&l_active);
2227 free_unref_page_list(&l_active);
2228 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2229 nr_deactivate, nr_rotated, sc->priority, file);
2232 unsigned long reclaim_pages(struct list_head *page_list)
2234 int nid = NUMA_NO_NODE;
2235 unsigned int nr_reclaimed = 0;
2236 LIST_HEAD(node_page_list);
2237 struct reclaim_stat dummy_stat;
2239 struct scan_control sc = {
2240 .gfp_mask = GFP_KERNEL,
2241 .priority = DEF_PRIORITY,
2247 while (!list_empty(page_list)) {
2248 page = lru_to_page(page_list);
2249 if (nid == NUMA_NO_NODE) {
2250 nid = page_to_nid(page);
2251 INIT_LIST_HEAD(&node_page_list);
2254 if (nid == page_to_nid(page)) {
2255 ClearPageActive(page);
2256 list_move(&page->lru, &node_page_list);
2260 nr_reclaimed += shrink_page_list(&node_page_list,
2262 &sc, &dummy_stat, false);
2263 while (!list_empty(&node_page_list)) {
2264 page = lru_to_page(&node_page_list);
2265 list_del(&page->lru);
2266 putback_lru_page(page);
2272 if (!list_empty(&node_page_list)) {
2273 nr_reclaimed += shrink_page_list(&node_page_list,
2275 &sc, &dummy_stat, false);
2276 while (!list_empty(&node_page_list)) {
2277 page = lru_to_page(&node_page_list);
2278 list_del(&page->lru);
2279 putback_lru_page(page);
2283 return nr_reclaimed;
2286 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2287 struct lruvec *lruvec, struct scan_control *sc)
2289 if (is_active_lru(lru)) {
2290 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2291 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2293 sc->skipped_deactivate = 1;
2297 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2301 * The inactive anon list should be small enough that the VM never has
2302 * to do too much work.
2304 * The inactive file list should be small enough to leave most memory
2305 * to the established workingset on the scan-resistant active list,
2306 * but large enough to avoid thrashing the aggregate readahead window.
2308 * Both inactive lists should also be large enough that each inactive
2309 * page has a chance to be referenced again before it is reclaimed.
2311 * If that fails and refaulting is observed, the inactive list grows.
2313 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2314 * on this LRU, maintained by the pageout code. An inactive_ratio
2315 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2318 * memory ratio inactive
2319 * -------------------------------------
2328 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2330 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2331 unsigned long inactive, active;
2332 unsigned long inactive_ratio;
2335 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2336 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2338 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2340 inactive_ratio = int_sqrt(10 * gb);
2344 return inactive * inactive_ratio < active;
2355 * Determine how aggressively the anon and file LRU lists should be
2356 * scanned. The relative value of each set of LRU lists is determined
2357 * by looking at the fraction of the pages scanned we did rotate back
2358 * onto the active list instead of evict.
2360 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2361 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2363 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2366 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2367 unsigned long anon_cost, file_cost, total_cost;
2368 int swappiness = mem_cgroup_swappiness(memcg);
2369 u64 fraction[ANON_AND_FILE];
2370 u64 denominator = 0; /* gcc */
2371 enum scan_balance scan_balance;
2372 unsigned long ap, fp;
2375 /* If we have no swap space, do not bother scanning anon pages. */
2376 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2377 scan_balance = SCAN_FILE;
2382 * Global reclaim will swap to prevent OOM even with no
2383 * swappiness, but memcg users want to use this knob to
2384 * disable swapping for individual groups completely when
2385 * using the memory controller's swap limit feature would be
2388 if (cgroup_reclaim(sc) && !swappiness) {
2389 scan_balance = SCAN_FILE;
2394 * Do not apply any pressure balancing cleverness when the
2395 * system is close to OOM, scan both anon and file equally
2396 * (unless the swappiness setting disagrees with swapping).
2398 if (!sc->priority && swappiness) {
2399 scan_balance = SCAN_EQUAL;
2404 * If the system is almost out of file pages, force-scan anon.
2406 if (sc->file_is_tiny) {
2407 scan_balance = SCAN_ANON;
2412 * If there is enough inactive page cache, we do not reclaim
2413 * anything from the anonymous working right now.
2415 if (sc->cache_trim_mode) {
2416 scan_balance = SCAN_FILE;
2420 scan_balance = SCAN_FRACT;
2422 * Calculate the pressure balance between anon and file pages.
2424 * The amount of pressure we put on each LRU is inversely
2425 * proportional to the cost of reclaiming each list, as
2426 * determined by the share of pages that are refaulting, times
2427 * the relative IO cost of bringing back a swapped out
2428 * anonymous page vs reloading a filesystem page (swappiness).
2430 * Although we limit that influence to ensure no list gets
2431 * left behind completely: at least a third of the pressure is
2432 * applied, before swappiness.
2434 * With swappiness at 100, anon and file have equal IO cost.
2436 total_cost = sc->anon_cost + sc->file_cost;
2437 anon_cost = total_cost + sc->anon_cost;
2438 file_cost = total_cost + sc->file_cost;
2439 total_cost = anon_cost + file_cost;
2441 ap = swappiness * (total_cost + 1);
2442 ap /= anon_cost + 1;
2444 fp = (200 - swappiness) * (total_cost + 1);
2445 fp /= file_cost + 1;
2449 denominator = ap + fp;
2451 for_each_evictable_lru(lru) {
2452 int file = is_file_lru(lru);
2453 unsigned long lruvec_size;
2455 unsigned long protection;
2457 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2458 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2460 sc->memcg_low_reclaim);
2464 * Scale a cgroup's reclaim pressure by proportioning
2465 * its current usage to its memory.low or memory.min
2468 * This is important, as otherwise scanning aggression
2469 * becomes extremely binary -- from nothing as we
2470 * approach the memory protection threshold, to totally
2471 * nominal as we exceed it. This results in requiring
2472 * setting extremely liberal protection thresholds. It
2473 * also means we simply get no protection at all if we
2474 * set it too low, which is not ideal.
2476 * If there is any protection in place, we reduce scan
2477 * pressure by how much of the total memory used is
2478 * within protection thresholds.
2480 * There is one special case: in the first reclaim pass,
2481 * we skip over all groups that are within their low
2482 * protection. If that fails to reclaim enough pages to
2483 * satisfy the reclaim goal, we come back and override
2484 * the best-effort low protection. However, we still
2485 * ideally want to honor how well-behaved groups are in
2486 * that case instead of simply punishing them all
2487 * equally. As such, we reclaim them based on how much
2488 * memory they are using, reducing the scan pressure
2489 * again by how much of the total memory used is under
2492 unsigned long cgroup_size = mem_cgroup_size(memcg);
2494 /* Avoid TOCTOU with earlier protection check */
2495 cgroup_size = max(cgroup_size, protection);
2497 scan = lruvec_size - lruvec_size * protection /
2501 * Minimally target SWAP_CLUSTER_MAX pages to keep
2502 * reclaim moving forwards, avoiding decrementing
2503 * sc->priority further than desirable.
2505 scan = max(scan, SWAP_CLUSTER_MAX);
2510 scan >>= sc->priority;
2513 * If the cgroup's already been deleted, make sure to
2514 * scrape out the remaining cache.
2516 if (!scan && !mem_cgroup_online(memcg))
2517 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2519 switch (scan_balance) {
2521 /* Scan lists relative to size */
2525 * Scan types proportional to swappiness and
2526 * their relative recent reclaim efficiency.
2527 * Make sure we don't miss the last page on
2528 * the offlined memory cgroups because of a
2531 scan = mem_cgroup_online(memcg) ?
2532 div64_u64(scan * fraction[file], denominator) :
2533 DIV64_U64_ROUND_UP(scan * fraction[file],
2538 /* Scan one type exclusively */
2539 if ((scan_balance == SCAN_FILE) != file)
2543 /* Look ma, no brain */
2551 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2553 unsigned long nr[NR_LRU_LISTS];
2554 unsigned long targets[NR_LRU_LISTS];
2555 unsigned long nr_to_scan;
2557 unsigned long nr_reclaimed = 0;
2558 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2559 struct blk_plug plug;
2562 get_scan_count(lruvec, sc, nr);
2564 /* Record the original scan target for proportional adjustments later */
2565 memcpy(targets, nr, sizeof(nr));
2568 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2569 * event that can occur when there is little memory pressure e.g.
2570 * multiple streaming readers/writers. Hence, we do not abort scanning
2571 * when the requested number of pages are reclaimed when scanning at
2572 * DEF_PRIORITY on the assumption that the fact we are direct
2573 * reclaiming implies that kswapd is not keeping up and it is best to
2574 * do a batch of work at once. For memcg reclaim one check is made to
2575 * abort proportional reclaim if either the file or anon lru has already
2576 * dropped to zero at the first pass.
2578 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2579 sc->priority == DEF_PRIORITY);
2581 blk_start_plug(&plug);
2582 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2583 nr[LRU_INACTIVE_FILE]) {
2584 unsigned long nr_anon, nr_file, percentage;
2585 unsigned long nr_scanned;
2587 for_each_evictable_lru(lru) {
2589 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2590 nr[lru] -= nr_to_scan;
2592 nr_reclaimed += shrink_list(lru, nr_to_scan,
2599 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2603 * For kswapd and memcg, reclaim at least the number of pages
2604 * requested. Ensure that the anon and file LRUs are scanned
2605 * proportionally what was requested by get_scan_count(). We
2606 * stop reclaiming one LRU and reduce the amount scanning
2607 * proportional to the original scan target.
2609 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2610 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2613 * It's just vindictive to attack the larger once the smaller
2614 * has gone to zero. And given the way we stop scanning the
2615 * smaller below, this makes sure that we only make one nudge
2616 * towards proportionality once we've got nr_to_reclaim.
2618 if (!nr_file || !nr_anon)
2621 if (nr_file > nr_anon) {
2622 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2623 targets[LRU_ACTIVE_ANON] + 1;
2625 percentage = nr_anon * 100 / scan_target;
2627 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2628 targets[LRU_ACTIVE_FILE] + 1;
2630 percentage = nr_file * 100 / scan_target;
2633 /* Stop scanning the smaller of the LRU */
2635 nr[lru + LRU_ACTIVE] = 0;
2638 * Recalculate the other LRU scan count based on its original
2639 * scan target and the percentage scanning already complete
2641 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2642 nr_scanned = targets[lru] - nr[lru];
2643 nr[lru] = targets[lru] * (100 - percentage) / 100;
2644 nr[lru] -= min(nr[lru], nr_scanned);
2647 nr_scanned = targets[lru] - nr[lru];
2648 nr[lru] = targets[lru] * (100 - percentage) / 100;
2649 nr[lru] -= min(nr[lru], nr_scanned);
2651 scan_adjusted = true;
2653 blk_finish_plug(&plug);
2654 sc->nr_reclaimed += nr_reclaimed;
2657 * Even if we did not try to evict anon pages at all, we want to
2658 * rebalance the anon lru active/inactive ratio.
2660 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2661 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2662 sc, LRU_ACTIVE_ANON);
2665 /* Use reclaim/compaction for costly allocs or under memory pressure */
2666 static bool in_reclaim_compaction(struct scan_control *sc)
2668 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2669 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2670 sc->priority < DEF_PRIORITY - 2))
2677 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2678 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2679 * true if more pages should be reclaimed such that when the page allocator
2680 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2681 * It will give up earlier than that if there is difficulty reclaiming pages.
2683 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2684 unsigned long nr_reclaimed,
2685 struct scan_control *sc)
2687 unsigned long pages_for_compaction;
2688 unsigned long inactive_lru_pages;
2691 /* If not in reclaim/compaction mode, stop */
2692 if (!in_reclaim_compaction(sc))
2696 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2697 * number of pages that were scanned. This will return to the caller
2698 * with the risk reclaim/compaction and the resulting allocation attempt
2699 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2700 * allocations through requiring that the full LRU list has been scanned
2701 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2702 * scan, but that approximation was wrong, and there were corner cases
2703 * where always a non-zero amount of pages were scanned.
2708 /* If compaction would go ahead or the allocation would succeed, stop */
2709 for (z = 0; z <= sc->reclaim_idx; z++) {
2710 struct zone *zone = &pgdat->node_zones[z];
2711 if (!managed_zone(zone))
2714 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2715 case COMPACT_SUCCESS:
2716 case COMPACT_CONTINUE:
2719 /* check next zone */
2725 * If we have not reclaimed enough pages for compaction and the
2726 * inactive lists are large enough, continue reclaiming
2728 pages_for_compaction = compact_gap(sc->order);
2729 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2730 if (get_nr_swap_pages() > 0)
2731 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2733 return inactive_lru_pages > pages_for_compaction;
2736 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2738 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2739 struct mem_cgroup *memcg;
2741 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2743 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2744 unsigned long reclaimed;
2745 unsigned long scanned;
2748 * This loop can become CPU-bound when target memcgs
2749 * aren't eligible for reclaim - either because they
2750 * don't have any reclaimable pages, or because their
2751 * memory is explicitly protected. Avoid soft lockups.
2755 mem_cgroup_calculate_protection(target_memcg, memcg);
2757 if (mem_cgroup_below_min(memcg)) {
2760 * If there is no reclaimable memory, OOM.
2763 } else if (mem_cgroup_below_low(memcg)) {
2766 * Respect the protection only as long as
2767 * there is an unprotected supply
2768 * of reclaimable memory from other cgroups.
2770 if (!sc->memcg_low_reclaim) {
2771 sc->memcg_low_skipped = 1;
2774 memcg_memory_event(memcg, MEMCG_LOW);
2777 reclaimed = sc->nr_reclaimed;
2778 scanned = sc->nr_scanned;
2780 shrink_lruvec(lruvec, 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(target_memcg, memcg, NULL)));
2793 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2795 struct reclaim_state *reclaim_state = current->reclaim_state;
2796 unsigned long nr_reclaimed, nr_scanned;
2797 struct lruvec *target_lruvec;
2798 bool reclaimable = false;
2801 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2804 memset(&sc->nr, 0, sizeof(sc->nr));
2806 nr_reclaimed = sc->nr_reclaimed;
2807 nr_scanned = sc->nr_scanned;
2810 * Determine the scan balance between anon and file LRUs.
2812 spin_lock_irq(&target_lruvec->lru_lock);
2813 sc->anon_cost = target_lruvec->anon_cost;
2814 sc->file_cost = target_lruvec->file_cost;
2815 spin_unlock_irq(&target_lruvec->lru_lock);
2818 * Target desirable inactive:active list ratios for the anon
2819 * and file LRU lists.
2821 if (!sc->force_deactivate) {
2822 unsigned long refaults;
2824 refaults = lruvec_page_state(target_lruvec,
2825 WORKINGSET_ACTIVATE_ANON);
2826 if (refaults != target_lruvec->refaults[0] ||
2827 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2828 sc->may_deactivate |= DEACTIVATE_ANON;
2830 sc->may_deactivate &= ~DEACTIVATE_ANON;
2833 * When refaults are being observed, it means a new
2834 * workingset is being established. Deactivate to get
2835 * rid of any stale active pages quickly.
2837 refaults = lruvec_page_state(target_lruvec,
2838 WORKINGSET_ACTIVATE_FILE);
2839 if (refaults != target_lruvec->refaults[1] ||
2840 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2841 sc->may_deactivate |= DEACTIVATE_FILE;
2843 sc->may_deactivate &= ~DEACTIVATE_FILE;
2845 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2848 * If we have plenty of inactive file pages that aren't
2849 * thrashing, try to reclaim those first before touching
2852 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2853 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2854 sc->cache_trim_mode = 1;
2856 sc->cache_trim_mode = 0;
2859 * Prevent the reclaimer from falling into the cache trap: as
2860 * cache pages start out inactive, every cache fault will tip
2861 * the scan balance towards the file LRU. And as the file LRU
2862 * shrinks, so does the window for rotation from references.
2863 * This means we have a runaway feedback loop where a tiny
2864 * thrashing file LRU becomes infinitely more attractive than
2865 * anon pages. Try to detect this based on file LRU size.
2867 if (!cgroup_reclaim(sc)) {
2868 unsigned long total_high_wmark = 0;
2869 unsigned long free, anon;
2872 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2873 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2874 node_page_state(pgdat, NR_INACTIVE_FILE);
2876 for (z = 0; z < MAX_NR_ZONES; z++) {
2877 struct zone *zone = &pgdat->node_zones[z];
2878 if (!managed_zone(zone))
2881 total_high_wmark += high_wmark_pages(zone);
2885 * Consider anon: if that's low too, this isn't a
2886 * runaway file reclaim problem, but rather just
2887 * extreme pressure. Reclaim as per usual then.
2889 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2892 file + free <= total_high_wmark &&
2893 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2894 anon >> sc->priority;
2897 shrink_node_memcgs(pgdat, sc);
2899 if (reclaim_state) {
2900 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2901 reclaim_state->reclaimed_slab = 0;
2904 /* Record the subtree's reclaim efficiency */
2905 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2906 sc->nr_scanned - nr_scanned,
2907 sc->nr_reclaimed - nr_reclaimed);
2909 if (sc->nr_reclaimed - nr_reclaimed)
2912 if (current_is_kswapd()) {
2914 * If reclaim is isolating dirty pages under writeback,
2915 * it implies that the long-lived page allocation rate
2916 * is exceeding the page laundering rate. Either the
2917 * global limits are not being effective at throttling
2918 * processes due to the page distribution throughout
2919 * zones or there is heavy usage of a slow backing
2920 * device. The only option is to throttle from reclaim
2921 * context which is not ideal as there is no guarantee
2922 * the dirtying process is throttled in the same way
2923 * balance_dirty_pages() manages.
2925 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2926 * count the number of pages under pages flagged for
2927 * immediate reclaim and stall if any are encountered
2928 * in the nr_immediate check below.
2930 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2931 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2933 /* Allow kswapd to start writing pages during reclaim.*/
2934 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2935 set_bit(PGDAT_DIRTY, &pgdat->flags);
2938 * If kswapd scans pages marked for immediate
2939 * reclaim and under writeback (nr_immediate), it
2940 * implies that pages are cycling through the LRU
2941 * faster than they are written so also forcibly stall.
2943 if (sc->nr.immediate)
2944 congestion_wait(BLK_RW_ASYNC, HZ/10);
2948 * Tag a node/memcg as congested if all the dirty pages
2949 * scanned were backed by a congested BDI and
2950 * wait_iff_congested will stall.
2952 * Legacy memcg will stall in page writeback so avoid forcibly
2953 * stalling in wait_iff_congested().
2955 if ((current_is_kswapd() ||
2956 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2957 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2958 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2961 * Stall direct reclaim for IO completions if underlying BDIs
2962 * and node is congested. Allow kswapd to continue until it
2963 * starts encountering unqueued dirty pages or cycling through
2964 * the LRU too quickly.
2966 if (!current_is_kswapd() && current_may_throttle() &&
2967 !sc->hibernation_mode &&
2968 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2969 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2971 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2976 * Kswapd gives up on balancing particular nodes after too
2977 * many failures to reclaim anything from them and goes to
2978 * sleep. On reclaim progress, reset the failure counter. A
2979 * successful direct reclaim run will revive a dormant kswapd.
2982 pgdat->kswapd_failures = 0;
2986 * Returns true if compaction should go ahead for a costly-order request, or
2987 * the allocation would already succeed without compaction. Return false if we
2988 * should reclaim first.
2990 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2992 unsigned long watermark;
2993 enum compact_result suitable;
2995 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2996 if (suitable == COMPACT_SUCCESS)
2997 /* Allocation should succeed already. Don't reclaim. */
2999 if (suitable == COMPACT_SKIPPED)
3000 /* Compaction cannot yet proceed. Do reclaim. */
3004 * Compaction is already possible, but it takes time to run and there
3005 * are potentially other callers using the pages just freed. So proceed
3006 * with reclaim to make a buffer of free pages available to give
3007 * compaction a reasonable chance of completing and allocating the page.
3008 * Note that we won't actually reclaim the whole buffer in one attempt
3009 * as the target watermark in should_continue_reclaim() is lower. But if
3010 * we are already above the high+gap watermark, don't reclaim at all.
3012 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3014 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3018 * This is the direct reclaim path, for page-allocating processes. We only
3019 * try to reclaim pages from zones which will satisfy the caller's allocation
3022 * If a zone is deemed to be full of pinned pages then just give it a light
3023 * scan then give up on it.
3025 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3029 unsigned long nr_soft_reclaimed;
3030 unsigned long nr_soft_scanned;
3032 pg_data_t *last_pgdat = NULL;
3035 * If the number of buffer_heads in the machine exceeds the maximum
3036 * allowed level, force direct reclaim to scan the highmem zone as
3037 * highmem pages could be pinning lowmem pages storing buffer_heads
3039 orig_mask = sc->gfp_mask;
3040 if (buffer_heads_over_limit) {
3041 sc->gfp_mask |= __GFP_HIGHMEM;
3042 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3045 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3046 sc->reclaim_idx, sc->nodemask) {
3048 * Take care memory controller reclaiming has small influence
3051 if (!cgroup_reclaim(sc)) {
3052 if (!cpuset_zone_allowed(zone,
3053 GFP_KERNEL | __GFP_HARDWALL))
3057 * If we already have plenty of memory free for
3058 * compaction in this zone, don't free any more.
3059 * Even though compaction is invoked for any
3060 * non-zero order, only frequent costly order
3061 * reclamation is disruptive enough to become a
3062 * noticeable problem, like transparent huge
3065 if (IS_ENABLED(CONFIG_COMPACTION) &&
3066 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3067 compaction_ready(zone, sc)) {
3068 sc->compaction_ready = true;
3073 * Shrink each node in the zonelist once. If the
3074 * zonelist is ordered by zone (not the default) then a
3075 * node may be shrunk multiple times but in that case
3076 * the user prefers lower zones being preserved.
3078 if (zone->zone_pgdat == last_pgdat)
3082 * This steals pages from memory cgroups over softlimit
3083 * and returns the number of reclaimed pages and
3084 * scanned pages. This works for global memory pressure
3085 * and balancing, not for a memcg's limit.
3087 nr_soft_scanned = 0;
3088 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3089 sc->order, sc->gfp_mask,
3091 sc->nr_reclaimed += nr_soft_reclaimed;
3092 sc->nr_scanned += nr_soft_scanned;
3093 /* need some check for avoid more shrink_zone() */
3096 /* See comment about same check for global reclaim above */
3097 if (zone->zone_pgdat == last_pgdat)
3099 last_pgdat = zone->zone_pgdat;
3100 shrink_node(zone->zone_pgdat, sc);
3104 * Restore to original mask to avoid the impact on the caller if we
3105 * promoted it to __GFP_HIGHMEM.
3107 sc->gfp_mask = orig_mask;
3110 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3112 struct lruvec *target_lruvec;
3113 unsigned long refaults;
3115 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3116 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3117 target_lruvec->refaults[0] = refaults;
3118 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3119 target_lruvec->refaults[1] = refaults;
3123 * This is the main entry point to direct page reclaim.
3125 * If a full scan of the inactive list fails to free enough memory then we
3126 * are "out of memory" and something needs to be killed.
3128 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3129 * high - the zone may be full of dirty or under-writeback pages, which this
3130 * caller can't do much about. We kick the writeback threads and take explicit
3131 * naps in the hope that some of these pages can be written. But if the
3132 * allocating task holds filesystem locks which prevent writeout this might not
3133 * work, and the allocation attempt will fail.
3135 * returns: 0, if no pages reclaimed
3136 * else, the number of pages reclaimed
3138 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3139 struct scan_control *sc)
3141 int initial_priority = sc->priority;
3142 pg_data_t *last_pgdat;
3146 delayacct_freepages_start();
3148 if (!cgroup_reclaim(sc))
3149 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3152 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3155 shrink_zones(zonelist, sc);
3157 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3160 if (sc->compaction_ready)
3164 * If we're getting trouble reclaiming, start doing
3165 * writepage even in laptop mode.
3167 if (sc->priority < DEF_PRIORITY - 2)
3168 sc->may_writepage = 1;
3169 } while (--sc->priority >= 0);
3172 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3174 if (zone->zone_pgdat == last_pgdat)
3176 last_pgdat = zone->zone_pgdat;
3178 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3180 if (cgroup_reclaim(sc)) {
3181 struct lruvec *lruvec;
3183 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3185 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3189 delayacct_freepages_end();
3191 if (sc->nr_reclaimed)
3192 return sc->nr_reclaimed;
3194 /* Aborted reclaim to try compaction? don't OOM, then */
3195 if (sc->compaction_ready)
3199 * We make inactive:active ratio decisions based on the node's
3200 * composition of memory, but a restrictive reclaim_idx or a
3201 * memory.low cgroup setting can exempt large amounts of
3202 * memory from reclaim. Neither of which are very common, so
3203 * instead of doing costly eligibility calculations of the
3204 * entire cgroup subtree up front, we assume the estimates are
3205 * good, and retry with forcible deactivation if that fails.
3207 if (sc->skipped_deactivate) {
3208 sc->priority = initial_priority;
3209 sc->force_deactivate = 1;
3210 sc->skipped_deactivate = 0;
3214 /* Untapped cgroup reserves? Don't OOM, retry. */
3215 if (sc->memcg_low_skipped) {
3216 sc->priority = initial_priority;
3217 sc->force_deactivate = 0;
3218 sc->memcg_low_reclaim = 1;
3219 sc->memcg_low_skipped = 0;
3226 static bool allow_direct_reclaim(pg_data_t *pgdat)
3229 unsigned long pfmemalloc_reserve = 0;
3230 unsigned long free_pages = 0;
3234 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3237 for (i = 0; i <= ZONE_NORMAL; i++) {
3238 zone = &pgdat->node_zones[i];
3239 if (!managed_zone(zone))
3242 if (!zone_reclaimable_pages(zone))
3245 pfmemalloc_reserve += min_wmark_pages(zone);
3246 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3249 /* If there are no reserves (unexpected config) then do not throttle */
3250 if (!pfmemalloc_reserve)
3253 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3255 /* kswapd must be awake if processes are being throttled */
3256 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3257 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3258 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3260 wake_up_interruptible(&pgdat->kswapd_wait);
3267 * Throttle direct reclaimers if backing storage is backed by the network
3268 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3269 * depleted. kswapd will continue to make progress and wake the processes
3270 * when the low watermark is reached.
3272 * Returns true if a fatal signal was delivered during throttling. If this
3273 * happens, the page allocator should not consider triggering the OOM killer.
3275 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3276 nodemask_t *nodemask)
3280 pg_data_t *pgdat = NULL;
3283 * Kernel threads should not be throttled as they may be indirectly
3284 * responsible for cleaning pages necessary for reclaim to make forward
3285 * progress. kjournald for example may enter direct reclaim while
3286 * committing a transaction where throttling it could forcing other
3287 * processes to block on log_wait_commit().
3289 if (current->flags & PF_KTHREAD)
3293 * If a fatal signal is pending, this process should not throttle.
3294 * It should return quickly so it can exit and free its memory
3296 if (fatal_signal_pending(current))
3300 * Check if the pfmemalloc reserves are ok by finding the first node
3301 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3302 * GFP_KERNEL will be required for allocating network buffers when
3303 * swapping over the network so ZONE_HIGHMEM is unusable.
3305 * Throttling is based on the first usable node and throttled processes
3306 * wait on a queue until kswapd makes progress and wakes them. There
3307 * is an affinity then between processes waking up and where reclaim
3308 * progress has been made assuming the process wakes on the same node.
3309 * More importantly, processes running on remote nodes will not compete
3310 * for remote pfmemalloc reserves and processes on different nodes
3311 * should make reasonable progress.
3313 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3314 gfp_zone(gfp_mask), nodemask) {
3315 if (zone_idx(zone) > ZONE_NORMAL)
3318 /* Throttle based on the first usable node */
3319 pgdat = zone->zone_pgdat;
3320 if (allow_direct_reclaim(pgdat))
3325 /* If no zone was usable by the allocation flags then do not throttle */
3329 /* Account for the throttling */
3330 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3333 * If the caller cannot enter the filesystem, it's possible that it
3334 * is due to the caller holding an FS lock or performing a journal
3335 * transaction in the case of a filesystem like ext[3|4]. In this case,
3336 * it is not safe to block on pfmemalloc_wait as kswapd could be
3337 * blocked waiting on the same lock. Instead, throttle for up to a
3338 * second before continuing.
3340 if (!(gfp_mask & __GFP_FS)) {
3341 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3342 allow_direct_reclaim(pgdat), HZ);
3347 /* Throttle until kswapd wakes the process */
3348 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3349 allow_direct_reclaim(pgdat));
3352 if (fatal_signal_pending(current))
3359 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3360 gfp_t gfp_mask, nodemask_t *nodemask)
3362 unsigned long nr_reclaimed;
3363 struct scan_control sc = {
3364 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3365 .gfp_mask = current_gfp_context(gfp_mask),
3366 .reclaim_idx = gfp_zone(gfp_mask),
3368 .nodemask = nodemask,
3369 .priority = DEF_PRIORITY,
3370 .may_writepage = !laptop_mode,
3376 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3377 * Confirm they are large enough for max values.
3379 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3380 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3381 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3384 * Do not enter reclaim if fatal signal was delivered while throttled.
3385 * 1 is returned so that the page allocator does not OOM kill at this
3388 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3391 set_task_reclaim_state(current, &sc.reclaim_state);
3392 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3394 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3396 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3397 set_task_reclaim_state(current, NULL);
3399 return nr_reclaimed;
3404 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3405 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3406 gfp_t gfp_mask, bool noswap,
3408 unsigned long *nr_scanned)
3410 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3411 struct scan_control sc = {
3412 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3413 .target_mem_cgroup = memcg,
3414 .may_writepage = !laptop_mode,
3416 .reclaim_idx = MAX_NR_ZONES - 1,
3417 .may_swap = !noswap,
3420 WARN_ON_ONCE(!current->reclaim_state);
3422 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3423 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3425 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3429 * NOTE: Although we can get the priority field, using it
3430 * here is not a good idea, since it limits the pages we can scan.
3431 * if we don't reclaim here, the shrink_node from balance_pgdat
3432 * will pick up pages from other mem cgroup's as well. We hack
3433 * the priority and make it zero.
3435 shrink_lruvec(lruvec, &sc);
3437 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3439 *nr_scanned = sc.nr_scanned;
3441 return sc.nr_reclaimed;
3444 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3445 unsigned long nr_pages,
3449 unsigned long nr_reclaimed;
3450 unsigned int noreclaim_flag;
3451 struct scan_control sc = {
3452 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3453 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3454 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3455 .reclaim_idx = MAX_NR_ZONES - 1,
3456 .target_mem_cgroup = memcg,
3457 .priority = DEF_PRIORITY,
3458 .may_writepage = !laptop_mode,
3460 .may_swap = may_swap,
3463 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3464 * equal pressure on all the nodes. This is based on the assumption that
3465 * the reclaim does not bail out early.
3467 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3469 set_task_reclaim_state(current, &sc.reclaim_state);
3470 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3471 noreclaim_flag = memalloc_noreclaim_save();
3473 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3475 memalloc_noreclaim_restore(noreclaim_flag);
3476 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3477 set_task_reclaim_state(current, NULL);
3479 return nr_reclaimed;
3483 static void age_active_anon(struct pglist_data *pgdat,
3484 struct scan_control *sc)
3486 struct mem_cgroup *memcg;
3487 struct lruvec *lruvec;
3489 if (!total_swap_pages)
3492 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3493 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3496 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3498 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3499 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3500 sc, LRU_ACTIVE_ANON);
3501 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3505 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3511 * Check for watermark boosts top-down as the higher zones
3512 * are more likely to be boosted. Both watermarks and boosts
3513 * should not be checked at the same time as reclaim would
3514 * start prematurely when there is no boosting and a lower
3517 for (i = highest_zoneidx; i >= 0; i--) {
3518 zone = pgdat->node_zones + i;
3519 if (!managed_zone(zone))
3522 if (zone->watermark_boost)
3530 * Returns true if there is an eligible zone balanced for the request order
3531 * and highest_zoneidx
3533 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3536 unsigned long mark = -1;
3540 * Check watermarks bottom-up as lower zones are more likely to
3543 for (i = 0; i <= highest_zoneidx; i++) {
3544 zone = pgdat->node_zones + i;
3546 if (!managed_zone(zone))
3549 mark = high_wmark_pages(zone);
3550 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3555 * If a node has no populated zone within highest_zoneidx, it does not
3556 * need balancing by definition. This can happen if a zone-restricted
3557 * allocation tries to wake a remote kswapd.
3565 /* Clear pgdat state for congested, dirty or under writeback. */
3566 static void clear_pgdat_congested(pg_data_t *pgdat)
3568 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3570 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3571 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3572 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3576 * Prepare kswapd for sleeping. This verifies that there are no processes
3577 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3579 * Returns true if kswapd is ready to sleep
3581 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3582 int highest_zoneidx)
3585 * The throttled processes are normally woken up in balance_pgdat() as
3586 * soon as allow_direct_reclaim() is true. But there is a potential
3587 * race between when kswapd checks the watermarks and a process gets
3588 * throttled. There is also a potential race if processes get
3589 * throttled, kswapd wakes, a large process exits thereby balancing the
3590 * zones, which causes kswapd to exit balance_pgdat() before reaching
3591 * the wake up checks. If kswapd is going to sleep, no process should
3592 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3593 * the wake up is premature, processes will wake kswapd and get
3594 * throttled again. The difference from wake ups in balance_pgdat() is
3595 * that here we are under prepare_to_wait().
3597 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3598 wake_up_all(&pgdat->pfmemalloc_wait);
3600 /* Hopeless node, leave it to direct reclaim */
3601 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3604 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3605 clear_pgdat_congested(pgdat);
3613 * kswapd shrinks a node of pages that are at or below the highest usable
3614 * zone that is currently unbalanced.
3616 * Returns true if kswapd scanned at least the requested number of pages to
3617 * reclaim or if the lack of progress was due to pages under writeback.
3618 * This is used to determine if the scanning priority needs to be raised.
3620 static bool kswapd_shrink_node(pg_data_t *pgdat,
3621 struct scan_control *sc)
3626 /* Reclaim a number of pages proportional to the number of zones */
3627 sc->nr_to_reclaim = 0;
3628 for (z = 0; z <= sc->reclaim_idx; z++) {
3629 zone = pgdat->node_zones + z;
3630 if (!managed_zone(zone))
3633 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3637 * Historically care was taken to put equal pressure on all zones but
3638 * now pressure is applied based on node LRU order.
3640 shrink_node(pgdat, sc);
3643 * Fragmentation may mean that the system cannot be rebalanced for
3644 * high-order allocations. If twice the allocation size has been
3645 * reclaimed then recheck watermarks only at order-0 to prevent
3646 * excessive reclaim. Assume that a process requested a high-order
3647 * can direct reclaim/compact.
3649 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3652 return sc->nr_scanned >= sc->nr_to_reclaim;
3656 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3657 * that are eligible for use by the caller until at least one zone is
3660 * Returns the order kswapd finished reclaiming at.
3662 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3663 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3664 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3665 * or lower is eligible for reclaim until at least one usable zone is
3668 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3671 unsigned long nr_soft_reclaimed;
3672 unsigned long nr_soft_scanned;
3673 unsigned long pflags;
3674 unsigned long nr_boost_reclaim;
3675 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3678 struct scan_control sc = {
3679 .gfp_mask = GFP_KERNEL,
3684 set_task_reclaim_state(current, &sc.reclaim_state);
3685 psi_memstall_enter(&pflags);
3686 __fs_reclaim_acquire();
3688 count_vm_event(PAGEOUTRUN);
3691 * Account for the reclaim boost. Note that the zone boost is left in
3692 * place so that parallel allocations that are near the watermark will
3693 * stall or direct reclaim until kswapd is finished.
3695 nr_boost_reclaim = 0;
3696 for (i = 0; i <= highest_zoneidx; i++) {
3697 zone = pgdat->node_zones + i;
3698 if (!managed_zone(zone))
3701 nr_boost_reclaim += zone->watermark_boost;
3702 zone_boosts[i] = zone->watermark_boost;
3704 boosted = nr_boost_reclaim;
3707 sc.priority = DEF_PRIORITY;
3709 unsigned long nr_reclaimed = sc.nr_reclaimed;
3710 bool raise_priority = true;
3714 sc.reclaim_idx = highest_zoneidx;
3717 * If the number of buffer_heads exceeds the maximum allowed
3718 * then consider reclaiming from all zones. This has a dual
3719 * purpose -- on 64-bit systems it is expected that
3720 * buffer_heads are stripped during active rotation. On 32-bit
3721 * systems, highmem pages can pin lowmem memory and shrinking
3722 * buffers can relieve lowmem pressure. Reclaim may still not
3723 * go ahead if all eligible zones for the original allocation
3724 * request are balanced to avoid excessive reclaim from kswapd.
3726 if (buffer_heads_over_limit) {
3727 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3728 zone = pgdat->node_zones + i;
3729 if (!managed_zone(zone))
3738 * If the pgdat is imbalanced then ignore boosting and preserve
3739 * the watermarks for a later time and restart. Note that the
3740 * zone watermarks will be still reset at the end of balancing
3741 * on the grounds that the normal reclaim should be enough to
3742 * re-evaluate if boosting is required when kswapd next wakes.
3744 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3745 if (!balanced && nr_boost_reclaim) {
3746 nr_boost_reclaim = 0;
3751 * If boosting is not active then only reclaim if there are no
3752 * eligible zones. Note that sc.reclaim_idx is not used as
3753 * buffer_heads_over_limit may have adjusted it.
3755 if (!nr_boost_reclaim && balanced)
3758 /* Limit the priority of boosting to avoid reclaim writeback */
3759 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3760 raise_priority = false;
3763 * Do not writeback or swap pages for boosted reclaim. The
3764 * intent is to relieve pressure not issue sub-optimal IO
3765 * from reclaim context. If no pages are reclaimed, the
3766 * reclaim will be aborted.
3768 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3769 sc.may_swap = !nr_boost_reclaim;
3772 * Do some background aging of the anon list, to give
3773 * pages a chance to be referenced before reclaiming. All
3774 * pages are rotated regardless of classzone as this is
3775 * about consistent aging.
3777 age_active_anon(pgdat, &sc);
3780 * If we're getting trouble reclaiming, start doing writepage
3781 * even in laptop mode.
3783 if (sc.priority < DEF_PRIORITY - 2)
3784 sc.may_writepage = 1;
3786 /* Call soft limit reclaim before calling shrink_node. */
3788 nr_soft_scanned = 0;
3789 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3790 sc.gfp_mask, &nr_soft_scanned);
3791 sc.nr_reclaimed += nr_soft_reclaimed;
3794 * There should be no need to raise the scanning priority if
3795 * enough pages are already being scanned that that high
3796 * watermark would be met at 100% efficiency.
3798 if (kswapd_shrink_node(pgdat, &sc))
3799 raise_priority = false;
3802 * If the low watermark is met there is no need for processes
3803 * to be throttled on pfmemalloc_wait as they should not be
3804 * able to safely make forward progress. Wake them
3806 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3807 allow_direct_reclaim(pgdat))
3808 wake_up_all(&pgdat->pfmemalloc_wait);
3810 /* Check if kswapd should be suspending */
3811 __fs_reclaim_release();
3812 ret = try_to_freeze();
3813 __fs_reclaim_acquire();
3814 if (ret || kthread_should_stop())
3818 * Raise priority if scanning rate is too low or there was no
3819 * progress in reclaiming pages
3821 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3822 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3825 * If reclaim made no progress for a boost, stop reclaim as
3826 * IO cannot be queued and it could be an infinite loop in
3827 * extreme circumstances.
3829 if (nr_boost_reclaim && !nr_reclaimed)
3832 if (raise_priority || !nr_reclaimed)
3834 } while (sc.priority >= 1);
3836 if (!sc.nr_reclaimed)
3837 pgdat->kswapd_failures++;
3840 /* If reclaim was boosted, account for the reclaim done in this pass */
3842 unsigned long flags;
3844 for (i = 0; i <= highest_zoneidx; i++) {
3845 if (!zone_boosts[i])
3848 /* Increments are under the zone lock */
3849 zone = pgdat->node_zones + i;
3850 spin_lock_irqsave(&zone->lock, flags);
3851 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3852 spin_unlock_irqrestore(&zone->lock, flags);
3856 * As there is now likely space, wakeup kcompact to defragment
3859 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3862 snapshot_refaults(NULL, pgdat);
3863 __fs_reclaim_release();
3864 psi_memstall_leave(&pflags);
3865 set_task_reclaim_state(current, NULL);
3868 * Return the order kswapd stopped reclaiming at as
3869 * prepare_kswapd_sleep() takes it into account. If another caller
3870 * entered the allocator slow path while kswapd was awake, order will
3871 * remain at the higher level.
3877 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3878 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3879 * not a valid index then either kswapd runs for first time or kswapd couldn't
3880 * sleep after previous reclaim attempt (node is still unbalanced). In that
3881 * case return the zone index of the previous kswapd reclaim cycle.
3883 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3884 enum zone_type prev_highest_zoneidx)
3886 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3888 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3891 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3892 unsigned int highest_zoneidx)
3897 if (freezing(current) || kthread_should_stop())
3900 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3903 * Try to sleep for a short interval. Note that kcompactd will only be
3904 * woken if it is possible to sleep for a short interval. This is
3905 * deliberate on the assumption that if reclaim cannot keep an
3906 * eligible zone balanced that it's also unlikely that compaction will
3909 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3911 * Compaction records what page blocks it recently failed to
3912 * isolate pages from and skips them in the future scanning.
3913 * When kswapd is going to sleep, it is reasonable to assume
3914 * that pages and compaction may succeed so reset the cache.
3916 reset_isolation_suitable(pgdat);
3919 * We have freed the memory, now we should compact it to make
3920 * allocation of the requested order possible.
3922 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3924 remaining = schedule_timeout(HZ/10);
3927 * If woken prematurely then reset kswapd_highest_zoneidx and
3928 * order. The values will either be from a wakeup request or
3929 * the previous request that slept prematurely.
3932 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3933 kswapd_highest_zoneidx(pgdat,
3936 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3937 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3940 finish_wait(&pgdat->kswapd_wait, &wait);
3941 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3945 * After a short sleep, check if it was a premature sleep. If not, then
3946 * go fully to sleep until explicitly woken up.
3949 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3950 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3953 * vmstat counters are not perfectly accurate and the estimated
3954 * value for counters such as NR_FREE_PAGES can deviate from the
3955 * true value by nr_online_cpus * threshold. To avoid the zone
3956 * watermarks being breached while under pressure, we reduce the
3957 * per-cpu vmstat threshold while kswapd is awake and restore
3958 * them before going back to sleep.
3960 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3962 if (!kthread_should_stop())
3965 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3968 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3970 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3972 finish_wait(&pgdat->kswapd_wait, &wait);
3976 * The background pageout daemon, started as a kernel thread
3977 * from the init process.
3979 * This basically trickles out pages so that we have _some_
3980 * free memory available even if there is no other activity
3981 * that frees anything up. This is needed for things like routing
3982 * etc, where we otherwise might have all activity going on in
3983 * asynchronous contexts that cannot page things out.
3985 * If there are applications that are active memory-allocators
3986 * (most normal use), this basically shouldn't matter.
3988 static int kswapd(void *p)
3990 unsigned int alloc_order, reclaim_order;
3991 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3992 pg_data_t *pgdat = (pg_data_t*)p;
3993 struct task_struct *tsk = current;
3994 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3996 if (!cpumask_empty(cpumask))
3997 set_cpus_allowed_ptr(tsk, cpumask);
4000 * Tell the memory management that we're a "memory allocator",
4001 * and that if we need more memory we should get access to it
4002 * regardless (see "__alloc_pages()"). "kswapd" should
4003 * never get caught in the normal page freeing logic.
4005 * (Kswapd normally doesn't need memory anyway, but sometimes
4006 * you need a small amount of memory in order to be able to
4007 * page out something else, and this flag essentially protects
4008 * us from recursively trying to free more memory as we're
4009 * trying to free the first piece of memory in the first place).
4011 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
4014 WRITE_ONCE(pgdat->kswapd_order, 0);
4015 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4019 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4020 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4024 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4027 /* Read the new order and highest_zoneidx */
4028 alloc_order = READ_ONCE(pgdat->kswapd_order);
4029 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4031 WRITE_ONCE(pgdat->kswapd_order, 0);
4032 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4034 ret = try_to_freeze();
4035 if (kthread_should_stop())
4039 * We can speed up thawing tasks if we don't call balance_pgdat
4040 * after returning from the refrigerator
4046 * Reclaim begins at the requested order but if a high-order
4047 * reclaim fails then kswapd falls back to reclaiming for
4048 * order-0. If that happens, kswapd will consider sleeping
4049 * for the order it finished reclaiming at (reclaim_order)
4050 * but kcompactd is woken to compact for the original
4051 * request (alloc_order).
4053 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4055 reclaim_order = balance_pgdat(pgdat, alloc_order,
4057 if (reclaim_order < alloc_order)
4058 goto kswapd_try_sleep;
4061 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
4067 * A zone is low on free memory or too fragmented for high-order memory. If
4068 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4069 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4070 * has failed or is not needed, still wake up kcompactd if only compaction is
4073 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4074 enum zone_type highest_zoneidx)
4077 enum zone_type curr_idx;
4079 if (!managed_zone(zone))
4082 if (!cpuset_zone_allowed(zone, gfp_flags))
4085 pgdat = zone->zone_pgdat;
4086 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4088 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4089 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4091 if (READ_ONCE(pgdat->kswapd_order) < order)
4092 WRITE_ONCE(pgdat->kswapd_order, order);
4094 if (!waitqueue_active(&pgdat->kswapd_wait))
4097 /* Hopeless node, leave it to direct reclaim if possible */
4098 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4099 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4100 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4102 * There may be plenty of free memory available, but it's too
4103 * fragmented for high-order allocations. Wake up kcompactd
4104 * and rely on compaction_suitable() to determine if it's
4105 * needed. If it fails, it will defer subsequent attempts to
4106 * ratelimit its work.
4108 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4109 wakeup_kcompactd(pgdat, order, highest_zoneidx);
4113 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4115 wake_up_interruptible(&pgdat->kswapd_wait);
4118 #ifdef CONFIG_HIBERNATION
4120 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4123 * Rather than trying to age LRUs the aim is to preserve the overall
4124 * LRU order by reclaiming preferentially
4125 * inactive > active > active referenced > active mapped
4127 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4129 struct scan_control sc = {
4130 .nr_to_reclaim = nr_to_reclaim,
4131 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4132 .reclaim_idx = MAX_NR_ZONES - 1,
4133 .priority = DEF_PRIORITY,
4137 .hibernation_mode = 1,
4139 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4140 unsigned long nr_reclaimed;
4141 unsigned int noreclaim_flag;
4143 fs_reclaim_acquire(sc.gfp_mask);
4144 noreclaim_flag = memalloc_noreclaim_save();
4145 set_task_reclaim_state(current, &sc.reclaim_state);
4147 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4149 set_task_reclaim_state(current, NULL);
4150 memalloc_noreclaim_restore(noreclaim_flag);
4151 fs_reclaim_release(sc.gfp_mask);
4153 return nr_reclaimed;
4155 #endif /* CONFIG_HIBERNATION */
4158 * This kswapd start function will be called by init and node-hot-add.
4159 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4161 int kswapd_run(int nid)
4163 pg_data_t *pgdat = NODE_DATA(nid);
4169 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4170 if (IS_ERR(pgdat->kswapd)) {
4171 /* failure at boot is fatal */
4172 BUG_ON(system_state < SYSTEM_RUNNING);
4173 pr_err("Failed to start kswapd on node %d\n", nid);
4174 ret = PTR_ERR(pgdat->kswapd);
4175 pgdat->kswapd = NULL;
4181 * Called by memory hotplug when all memory in a node is offlined. Caller must
4182 * hold mem_hotplug_begin/end().
4184 void kswapd_stop(int nid)
4186 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4189 kthread_stop(kswapd);
4190 NODE_DATA(nid)->kswapd = NULL;
4194 static int __init kswapd_init(void)
4199 for_each_node_state(nid, N_MEMORY)
4204 module_init(kswapd_init)
4210 * If non-zero call node_reclaim when the number of free pages falls below
4213 int node_reclaim_mode __read_mostly;
4216 * Priority for NODE_RECLAIM. This determines the fraction of pages
4217 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4220 #define NODE_RECLAIM_PRIORITY 4
4223 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4226 int sysctl_min_unmapped_ratio = 1;
4229 * If the number of slab pages in a zone grows beyond this percentage then
4230 * slab reclaim needs to occur.
4232 int sysctl_min_slab_ratio = 5;
4234 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4236 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4237 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4238 node_page_state(pgdat, NR_ACTIVE_FILE);
4241 * It's possible for there to be more file mapped pages than
4242 * accounted for by the pages on the file LRU lists because
4243 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4245 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4248 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4249 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4251 unsigned long nr_pagecache_reclaimable;
4252 unsigned long delta = 0;
4255 * If RECLAIM_UNMAP is set, then all file pages are considered
4256 * potentially reclaimable. Otherwise, we have to worry about
4257 * pages like swapcache and node_unmapped_file_pages() provides
4260 if (node_reclaim_mode & RECLAIM_UNMAP)
4261 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4263 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4265 /* If we can't clean pages, remove dirty pages from consideration */
4266 if (!(node_reclaim_mode & RECLAIM_WRITE))
4267 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4269 /* Watch for any possible underflows due to delta */
4270 if (unlikely(delta > nr_pagecache_reclaimable))
4271 delta = nr_pagecache_reclaimable;
4273 return nr_pagecache_reclaimable - delta;
4277 * Try to free up some pages from this node through reclaim.
4279 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4281 /* Minimum pages needed in order to stay on node */
4282 const unsigned long nr_pages = 1 << order;
4283 struct task_struct *p = current;
4284 unsigned int noreclaim_flag;
4285 struct scan_control sc = {
4286 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4287 .gfp_mask = current_gfp_context(gfp_mask),
4289 .priority = NODE_RECLAIM_PRIORITY,
4290 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4291 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4293 .reclaim_idx = gfp_zone(gfp_mask),
4296 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4300 fs_reclaim_acquire(sc.gfp_mask);
4302 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4303 * and we also need to be able to write out pages for RECLAIM_WRITE
4304 * and RECLAIM_UNMAP.
4306 noreclaim_flag = memalloc_noreclaim_save();
4307 p->flags |= PF_SWAPWRITE;
4308 set_task_reclaim_state(p, &sc.reclaim_state);
4310 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4312 * Free memory by calling shrink node with increasing
4313 * priorities until we have enough memory freed.
4316 shrink_node(pgdat, &sc);
4317 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4320 set_task_reclaim_state(p, NULL);
4321 current->flags &= ~PF_SWAPWRITE;
4322 memalloc_noreclaim_restore(noreclaim_flag);
4323 fs_reclaim_release(sc.gfp_mask);
4325 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4327 return sc.nr_reclaimed >= nr_pages;
4330 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4335 * Node reclaim reclaims unmapped file backed pages and
4336 * slab pages if we are over the defined limits.
4338 * A small portion of unmapped file backed pages is needed for
4339 * file I/O otherwise pages read by file I/O will be immediately
4340 * thrown out if the node is overallocated. So we do not reclaim
4341 * if less than a specified percentage of the node is used by
4342 * unmapped file backed pages.
4344 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4345 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4346 pgdat->min_slab_pages)
4347 return NODE_RECLAIM_FULL;
4350 * Do not scan if the allocation should not be delayed.
4352 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4353 return NODE_RECLAIM_NOSCAN;
4356 * Only run node reclaim on the local node or on nodes that do not
4357 * have associated processors. This will favor the local processor
4358 * over remote processors and spread off node memory allocations
4359 * as wide as possible.
4361 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4362 return NODE_RECLAIM_NOSCAN;
4364 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4365 return NODE_RECLAIM_NOSCAN;
4367 ret = __node_reclaim(pgdat, gfp_mask, order);
4368 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4371 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4378 * check_move_unevictable_pages - check pages for evictability and move to
4379 * appropriate zone lru list
4380 * @pvec: pagevec with lru pages to check
4382 * Checks pages for evictability, if an evictable page is in the unevictable
4383 * lru list, moves it to the appropriate evictable lru list. This function
4384 * should be only used for lru pages.
4386 void check_move_unevictable_pages(struct pagevec *pvec)
4388 struct lruvec *lruvec = NULL;
4393 for (i = 0; i < pvec->nr; i++) {
4394 struct page *page = pvec->pages[i];
4397 if (PageTransTail(page))
4400 nr_pages = thp_nr_pages(page);
4401 pgscanned += nr_pages;
4403 /* block memcg migration during page moving between lru */
4404 if (!TestClearPageLRU(page))
4407 lruvec = relock_page_lruvec_irq(page, lruvec);
4408 if (page_evictable(page) && PageUnevictable(page)) {
4409 del_page_from_lru_list(page, lruvec);
4410 ClearPageUnevictable(page);
4411 add_page_to_lru_list(page, lruvec);
4412 pgrescued += nr_pages;
4418 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4419 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4420 unlock_page_lruvec_irq(lruvec);
4421 } else if (pgscanned) {
4422 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4425 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);