1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup *target_mem_cgroup;
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
91 /* e.g. boosted watermark reclaim leaves slabs alone */
92 unsigned int may_shrinkslab:1;
95 * Cgroups are not reclaimed below their configured memory.low,
96 * unless we threaten to OOM. If any cgroups are skipped due to
97 * memory.low and nothing was reclaimed, go back for memory.low.
99 unsigned int memcg_low_reclaim:1;
100 unsigned int memcg_low_skipped:1;
102 unsigned int hibernation_mode:1;
104 /* One of the zones is ready for compaction */
105 unsigned int compaction_ready:1;
107 /* Allocation order */
110 /* Scan (total_size >> priority) pages at once */
113 /* The highest zone to isolate pages for reclaim from */
116 /* This context's GFP mask */
119 /* Incremented by the number of inactive pages that were scanned */
120 unsigned long nr_scanned;
122 /* Number of pages freed so far during a call to shrink_zones() */
123 unsigned long nr_reclaimed;
127 unsigned int unqueued_dirty;
128 unsigned int congested;
129 unsigned int writeback;
130 unsigned int immediate;
131 unsigned int file_taken;
135 /* for recording the reclaimed slab by now */
136 struct reclaim_state reclaim_state;
139 #ifdef ARCH_HAS_PREFETCH
140 #define prefetch_prev_lru_page(_page, _base, _field) \
142 if ((_page)->lru.prev != _base) { \
145 prev = lru_to_page(&(_page->lru)); \
146 prefetch(&prev->_field); \
150 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
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 .. 100. Higher means more swappy.
170 int vm_swappiness = 60;
172 * The total number of pages which are beyond the high watermark within all
175 unsigned long vm_total_pages;
177 static LIST_HEAD(shrinker_list);
178 static DECLARE_RWSEM(shrinker_rwsem);
180 #ifdef CONFIG_MEMCG_KMEM
183 * We allow subsystems to populate their shrinker-related
184 * LRU lists before register_shrinker_prepared() is called
185 * for the shrinker, since we don't want to impose
186 * restrictions on their internal registration order.
187 * In this case shrink_slab_memcg() may find corresponding
188 * bit is set in the shrinkers map.
190 * This value is used by the function to detect registering
191 * shrinkers and to skip do_shrink_slab() calls for them.
193 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
195 static DEFINE_IDR(shrinker_idr);
196 static int shrinker_nr_max;
198 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
200 int id, ret = -ENOMEM;
202 down_write(&shrinker_rwsem);
203 /* This may call shrinker, so it must use down_read_trylock() */
204 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
208 if (id >= shrinker_nr_max) {
209 if (memcg_expand_shrinker_maps(id)) {
210 idr_remove(&shrinker_idr, id);
214 shrinker_nr_max = id + 1;
219 up_write(&shrinker_rwsem);
223 static void unregister_memcg_shrinker(struct shrinker *shrinker)
225 int id = shrinker->id;
229 down_write(&shrinker_rwsem);
230 idr_remove(&shrinker_idr, id);
231 up_write(&shrinker_rwsem);
233 #else /* CONFIG_MEMCG_KMEM */
234 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
239 static void unregister_memcg_shrinker(struct shrinker *shrinker)
242 #endif /* CONFIG_MEMCG_KMEM */
244 static void set_task_reclaim_state(struct task_struct *task,
245 struct reclaim_state *rs)
247 /* Check for an overwrite */
248 WARN_ON_ONCE(rs && task->reclaim_state);
250 /* Check for the nulling of an already-nulled member */
251 WARN_ON_ONCE(!rs && !task->reclaim_state);
253 task->reclaim_state = rs;
257 static bool global_reclaim(struct scan_control *sc)
259 return !sc->target_mem_cgroup;
263 * sane_reclaim - is the usual dirty throttling mechanism operational?
264 * @sc: scan_control in question
266 * The normal page dirty throttling mechanism in balance_dirty_pages() is
267 * completely broken with the legacy memcg and direct stalling in
268 * shrink_page_list() is used for throttling instead, which lacks all the
269 * niceties such as fairness, adaptive pausing, bandwidth proportional
270 * allocation and configurability.
272 * This function tests whether the vmscan currently in progress can assume
273 * that the normal dirty throttling mechanism is operational.
275 static bool sane_reclaim(struct scan_control *sc)
277 struct mem_cgroup *memcg = sc->target_mem_cgroup;
281 #ifdef CONFIG_CGROUP_WRITEBACK
282 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
288 static void set_memcg_congestion(pg_data_t *pgdat,
289 struct mem_cgroup *memcg,
292 struct mem_cgroup_per_node *mn;
297 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
298 WRITE_ONCE(mn->congested, congested);
301 static bool memcg_congested(pg_data_t *pgdat,
302 struct mem_cgroup *memcg)
304 struct mem_cgroup_per_node *mn;
306 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
307 return READ_ONCE(mn->congested);
311 static bool global_reclaim(struct scan_control *sc)
316 static bool sane_reclaim(struct scan_control *sc)
321 static inline void set_memcg_congestion(struct pglist_data *pgdat,
322 struct mem_cgroup *memcg, bool congested)
326 static inline bool memcg_congested(struct pglist_data *pgdat,
327 struct mem_cgroup *memcg)
335 * This misses isolated pages which are not accounted for to save counters.
336 * As the data only determines if reclaim or compaction continues, it is
337 * not expected that isolated pages will be a dominating factor.
339 unsigned long zone_reclaimable_pages(struct zone *zone)
343 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
344 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
345 if (get_nr_swap_pages() > 0)
346 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
347 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
353 * lruvec_lru_size - Returns the number of pages on the given LRU list.
354 * @lruvec: lru vector
356 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
358 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
360 unsigned long lru_size;
363 if (!mem_cgroup_disabled())
364 lru_size = lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
366 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
368 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
369 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
372 if (!managed_zone(zone))
375 if (!mem_cgroup_disabled())
376 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
378 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
379 NR_ZONE_LRU_BASE + lru);
380 lru_size -= min(size, lru_size);
388 * Add a shrinker callback to be called from the vm.
390 int prealloc_shrinker(struct shrinker *shrinker)
392 unsigned int size = sizeof(*shrinker->nr_deferred);
394 if (shrinker->flags & SHRINKER_NUMA_AWARE)
397 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
398 if (!shrinker->nr_deferred)
401 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
402 if (prealloc_memcg_shrinker(shrinker))
409 kfree(shrinker->nr_deferred);
410 shrinker->nr_deferred = NULL;
414 void free_prealloced_shrinker(struct shrinker *shrinker)
416 if (!shrinker->nr_deferred)
419 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
420 unregister_memcg_shrinker(shrinker);
422 kfree(shrinker->nr_deferred);
423 shrinker->nr_deferred = NULL;
426 void register_shrinker_prepared(struct shrinker *shrinker)
428 down_write(&shrinker_rwsem);
429 list_add_tail(&shrinker->list, &shrinker_list);
430 #ifdef CONFIG_MEMCG_KMEM
431 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
432 idr_replace(&shrinker_idr, shrinker, shrinker->id);
434 up_write(&shrinker_rwsem);
437 int register_shrinker(struct shrinker *shrinker)
439 int err = prealloc_shrinker(shrinker);
443 register_shrinker_prepared(shrinker);
446 EXPORT_SYMBOL(register_shrinker);
451 void unregister_shrinker(struct shrinker *shrinker)
453 if (!shrinker->nr_deferred)
455 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
456 unregister_memcg_shrinker(shrinker);
457 down_write(&shrinker_rwsem);
458 list_del(&shrinker->list);
459 up_write(&shrinker_rwsem);
460 kfree(shrinker->nr_deferred);
461 shrinker->nr_deferred = NULL;
463 EXPORT_SYMBOL(unregister_shrinker);
465 #define SHRINK_BATCH 128
467 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
468 struct shrinker *shrinker, int priority)
470 unsigned long freed = 0;
471 unsigned long long delta;
476 int nid = shrinkctl->nid;
477 long batch_size = shrinker->batch ? shrinker->batch
479 long scanned = 0, next_deferred;
481 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
484 freeable = shrinker->count_objects(shrinker, shrinkctl);
485 if (freeable == 0 || freeable == SHRINK_EMPTY)
489 * copy the current shrinker scan count into a local variable
490 * and zero it so that other concurrent shrinker invocations
491 * don't also do this scanning work.
493 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
496 if (shrinker->seeks) {
497 delta = freeable >> priority;
499 do_div(delta, shrinker->seeks);
502 * These objects don't require any IO to create. Trim
503 * them aggressively under memory pressure to keep
504 * them from causing refetches in the IO caches.
506 delta = freeable / 2;
510 if (total_scan < 0) {
511 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
512 shrinker->scan_objects, total_scan);
513 total_scan = freeable;
516 next_deferred = total_scan;
519 * We need to avoid excessive windup on filesystem shrinkers
520 * due to large numbers of GFP_NOFS allocations causing the
521 * shrinkers to return -1 all the time. This results in a large
522 * nr being built up so when a shrink that can do some work
523 * comes along it empties the entire cache due to nr >>>
524 * freeable. This is bad for sustaining a working set in
527 * Hence only allow the shrinker to scan the entire cache when
528 * a large delta change is calculated directly.
530 if (delta < freeable / 4)
531 total_scan = min(total_scan, freeable / 2);
534 * Avoid risking looping forever due to too large nr value:
535 * never try to free more than twice the estimate number of
538 if (total_scan > freeable * 2)
539 total_scan = freeable * 2;
541 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
542 freeable, delta, total_scan, priority);
545 * Normally, we should not scan less than batch_size objects in one
546 * pass to avoid too frequent shrinker calls, but if the slab has less
547 * than batch_size objects in total and we are really tight on memory,
548 * we will try to reclaim all available objects, otherwise we can end
549 * up failing allocations although there are plenty of reclaimable
550 * objects spread over several slabs with usage less than the
553 * We detect the "tight on memory" situations by looking at the total
554 * number of objects we want to scan (total_scan). If it is greater
555 * than the total number of objects on slab (freeable), we must be
556 * scanning at high prio and therefore should try to reclaim as much as
559 while (total_scan >= batch_size ||
560 total_scan >= freeable) {
562 unsigned long nr_to_scan = min(batch_size, total_scan);
564 shrinkctl->nr_to_scan = nr_to_scan;
565 shrinkctl->nr_scanned = nr_to_scan;
566 ret = shrinker->scan_objects(shrinker, shrinkctl);
567 if (ret == SHRINK_STOP)
571 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
572 total_scan -= shrinkctl->nr_scanned;
573 scanned += shrinkctl->nr_scanned;
578 if (next_deferred >= scanned)
579 next_deferred -= scanned;
583 * move the unused scan count back into the shrinker in a
584 * manner that handles concurrent updates. If we exhausted the
585 * scan, there is no need to do an update.
587 if (next_deferred > 0)
588 new_nr = atomic_long_add_return(next_deferred,
589 &shrinker->nr_deferred[nid]);
591 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
593 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
597 #ifdef CONFIG_MEMCG_KMEM
598 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
599 struct mem_cgroup *memcg, int priority)
601 struct memcg_shrinker_map *map;
602 unsigned long ret, freed = 0;
605 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
608 if (!down_read_trylock(&shrinker_rwsem))
611 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
616 for_each_set_bit(i, map->map, shrinker_nr_max) {
617 struct shrink_control sc = {
618 .gfp_mask = gfp_mask,
622 struct shrinker *shrinker;
624 shrinker = idr_find(&shrinker_idr, i);
625 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
627 clear_bit(i, map->map);
631 ret = do_shrink_slab(&sc, shrinker, priority);
632 if (ret == SHRINK_EMPTY) {
633 clear_bit(i, map->map);
635 * After the shrinker reported that it had no objects to
636 * free, but before we cleared the corresponding bit in
637 * the memcg shrinker map, a new object might have been
638 * added. To make sure, we have the bit set in this
639 * case, we invoke the shrinker one more time and reset
640 * the bit if it reports that it is not empty anymore.
641 * The memory barrier here pairs with the barrier in
642 * memcg_set_shrinker_bit():
644 * list_lru_add() shrink_slab_memcg()
645 * list_add_tail() clear_bit()
647 * set_bit() do_shrink_slab()
649 smp_mb__after_atomic();
650 ret = do_shrink_slab(&sc, shrinker, priority);
651 if (ret == SHRINK_EMPTY)
654 memcg_set_shrinker_bit(memcg, nid, i);
658 if (rwsem_is_contended(&shrinker_rwsem)) {
664 up_read(&shrinker_rwsem);
667 #else /* CONFIG_MEMCG_KMEM */
668 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
669 struct mem_cgroup *memcg, int priority)
673 #endif /* CONFIG_MEMCG_KMEM */
676 * shrink_slab - shrink slab caches
677 * @gfp_mask: allocation context
678 * @nid: node whose slab caches to target
679 * @memcg: memory cgroup whose slab caches to target
680 * @priority: the reclaim priority
682 * Call the shrink functions to age shrinkable caches.
684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 * unaware shrinkers will receive a node id of 0 instead.
687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 * are called only if it is the root cgroup.
690 * @priority is sc->priority, we take the number of objects and >> by priority
691 * in order to get the scan target.
693 * Returns the number of reclaimed slab objects.
695 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
696 struct mem_cgroup *memcg,
699 unsigned long ret, freed = 0;
700 struct shrinker *shrinker;
703 * The root memcg might be allocated even though memcg is disabled
704 * via "cgroup_disable=memory" boot parameter. This could make
705 * mem_cgroup_is_root() return false, then just run memcg slab
706 * shrink, but skip global shrink. This may result in premature
709 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
710 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
712 if (!down_read_trylock(&shrinker_rwsem))
715 list_for_each_entry(shrinker, &shrinker_list, list) {
716 struct shrink_control sc = {
717 .gfp_mask = gfp_mask,
722 ret = do_shrink_slab(&sc, shrinker, priority);
723 if (ret == SHRINK_EMPTY)
727 * Bail out if someone want to register a new shrinker to
728 * prevent the regsitration from being stalled for long periods
729 * by parallel ongoing shrinking.
731 if (rwsem_is_contended(&shrinker_rwsem)) {
737 up_read(&shrinker_rwsem);
743 void drop_slab_node(int nid)
748 struct mem_cgroup *memcg = NULL;
751 memcg = mem_cgroup_iter(NULL, NULL, NULL);
753 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
754 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
755 } while (freed > 10);
762 for_each_online_node(nid)
766 static inline int is_page_cache_freeable(struct page *page)
769 * A freeable page cache page is referenced only by the caller
770 * that isolated the page, the page cache and optional buffer
771 * heads at page->private.
773 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
775 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
778 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
780 if (current->flags & PF_SWAPWRITE)
782 if (!inode_write_congested(inode))
784 if (inode_to_bdi(inode) == current->backing_dev_info)
790 * We detected a synchronous write error writing a page out. Probably
791 * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 * fsync(), msync() or close().
794 * The tricky part is that after writepage we cannot touch the mapping: nothing
795 * prevents it from being freed up. But we have a ref on the page and once
796 * that page is locked, the mapping is pinned.
798 * We're allowed to run sleeping lock_page() here because we know the caller has
801 static void handle_write_error(struct address_space *mapping,
802 struct page *page, int error)
805 if (page_mapping(page) == mapping)
806 mapping_set_error(mapping, error);
810 /* possible outcome of pageout() */
812 /* failed to write page out, page is locked */
814 /* move page to the active list, page is locked */
816 /* page has been sent to the disk successfully, page is unlocked */
818 /* page is clean and locked */
823 * pageout is called by shrink_page_list() for each dirty page.
824 * Calls ->writepage().
826 static pageout_t pageout(struct page *page, struct address_space *mapping,
827 struct scan_control *sc)
830 * If the page is dirty, only perform writeback if that write
831 * will be non-blocking. To prevent this allocation from being
832 * stalled by pagecache activity. But note that there may be
833 * stalls if we need to run get_block(). We could test
834 * PagePrivate for that.
836 * If this process is currently in __generic_file_write_iter() against
837 * this page's queue, we can perform writeback even if that
840 * If the page is swapcache, write it back even if that would
841 * block, for some throttling. This happens by accident, because
842 * swap_backing_dev_info is bust: it doesn't reflect the
843 * congestion state of the swapdevs. Easy to fix, if needed.
845 if (!is_page_cache_freeable(page))
849 * Some data journaling orphaned pages can have
850 * page->mapping == NULL while being dirty with clean buffers.
852 if (page_has_private(page)) {
853 if (try_to_free_buffers(page)) {
854 ClearPageDirty(page);
855 pr_info("%s: orphaned page\n", __func__);
861 if (mapping->a_ops->writepage == NULL)
862 return PAGE_ACTIVATE;
863 if (!may_write_to_inode(mapping->host, sc))
866 if (clear_page_dirty_for_io(page)) {
868 struct writeback_control wbc = {
869 .sync_mode = WB_SYNC_NONE,
870 .nr_to_write = SWAP_CLUSTER_MAX,
872 .range_end = LLONG_MAX,
876 SetPageReclaim(page);
877 res = mapping->a_ops->writepage(page, &wbc);
879 handle_write_error(mapping, page, res);
880 if (res == AOP_WRITEPAGE_ACTIVATE) {
881 ClearPageReclaim(page);
882 return PAGE_ACTIVATE;
885 if (!PageWriteback(page)) {
886 /* synchronous write or broken a_ops? */
887 ClearPageReclaim(page);
889 trace_mm_vmscan_writepage(page);
890 inc_node_page_state(page, NR_VMSCAN_WRITE);
898 * Same as remove_mapping, but if the page is removed from the mapping, it
899 * gets returned with a refcount of 0.
901 static int __remove_mapping(struct address_space *mapping, struct page *page,
907 BUG_ON(!PageLocked(page));
908 BUG_ON(mapping != page_mapping(page));
910 xa_lock_irqsave(&mapping->i_pages, flags);
912 * The non racy check for a busy page.
914 * Must be careful with the order of the tests. When someone has
915 * a ref to the page, it may be possible that they dirty it then
916 * drop the reference. So if PageDirty is tested before page_count
917 * here, then the following race may occur:
919 * get_user_pages(&page);
920 * [user mapping goes away]
922 * !PageDirty(page) [good]
923 * SetPageDirty(page);
925 * !page_count(page) [good, discard it]
927 * [oops, our write_to data is lost]
929 * Reversing the order of the tests ensures such a situation cannot
930 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
931 * load is not satisfied before that of page->_refcount.
933 * Note that if SetPageDirty is always performed via set_page_dirty,
934 * and thus under the i_pages lock, then this ordering is not required.
936 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
937 refcount = 1 + HPAGE_PMD_NR;
940 if (!page_ref_freeze(page, refcount))
942 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
943 if (unlikely(PageDirty(page))) {
944 page_ref_unfreeze(page, refcount);
948 if (PageSwapCache(page)) {
949 swp_entry_t swap = { .val = page_private(page) };
950 mem_cgroup_swapout(page, swap);
951 __delete_from_swap_cache(page, swap);
952 xa_unlock_irqrestore(&mapping->i_pages, flags);
953 put_swap_page(page, swap);
955 void (*freepage)(struct page *);
958 freepage = mapping->a_ops->freepage;
960 * Remember a shadow entry for reclaimed file cache in
961 * order to detect refaults, thus thrashing, later on.
963 * But don't store shadows in an address space that is
964 * already exiting. This is not just an optizimation,
965 * inode reclaim needs to empty out the radix tree or
966 * the nodes are lost. Don't plant shadows behind its
969 * We also don't store shadows for DAX mappings because the
970 * only page cache pages found in these are zero pages
971 * covering holes, and because we don't want to mix DAX
972 * exceptional entries and shadow exceptional entries in the
973 * same address_space.
975 if (reclaimed && page_is_file_cache(page) &&
976 !mapping_exiting(mapping) && !dax_mapping(mapping))
977 shadow = workingset_eviction(page);
978 __delete_from_page_cache(page, shadow);
979 xa_unlock_irqrestore(&mapping->i_pages, flags);
981 if (freepage != NULL)
988 xa_unlock_irqrestore(&mapping->i_pages, flags);
993 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
994 * someone else has a ref on the page, abort and return 0. If it was
995 * successfully detached, return 1. Assumes the caller has a single ref on
998 int remove_mapping(struct address_space *mapping, struct page *page)
1000 if (__remove_mapping(mapping, page, false)) {
1002 * Unfreezing the refcount with 1 rather than 2 effectively
1003 * drops the pagecache ref for us without requiring another
1006 page_ref_unfreeze(page, 1);
1013 * putback_lru_page - put previously isolated page onto appropriate LRU list
1014 * @page: page to be put back to appropriate lru list
1016 * Add previously isolated @page to appropriate LRU list.
1017 * Page may still be unevictable for other reasons.
1019 * lru_lock must not be held, interrupts must be enabled.
1021 void putback_lru_page(struct page *page)
1023 lru_cache_add(page);
1024 put_page(page); /* drop ref from isolate */
1027 enum page_references {
1029 PAGEREF_RECLAIM_CLEAN,
1034 static enum page_references page_check_references(struct page *page,
1035 struct scan_control *sc)
1037 int referenced_ptes, referenced_page;
1038 unsigned long vm_flags;
1040 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1042 referenced_page = TestClearPageReferenced(page);
1045 * Mlock lost the isolation race with us. Let try_to_unmap()
1046 * move the page to the unevictable list.
1048 if (vm_flags & VM_LOCKED)
1049 return PAGEREF_RECLAIM;
1051 if (referenced_ptes) {
1052 if (PageSwapBacked(page))
1053 return PAGEREF_ACTIVATE;
1055 * All mapped pages start out with page table
1056 * references from the instantiating fault, so we need
1057 * to look twice if a mapped file page is used more
1060 * Mark it and spare it for another trip around the
1061 * inactive list. Another page table reference will
1062 * lead to its activation.
1064 * Note: the mark is set for activated pages as well
1065 * so that recently deactivated but used pages are
1066 * quickly recovered.
1068 SetPageReferenced(page);
1070 if (referenced_page || referenced_ptes > 1)
1071 return PAGEREF_ACTIVATE;
1074 * Activate file-backed executable pages after first usage.
1076 if (vm_flags & VM_EXEC)
1077 return PAGEREF_ACTIVATE;
1079 return PAGEREF_KEEP;
1082 /* Reclaim if clean, defer dirty pages to writeback */
1083 if (referenced_page && !PageSwapBacked(page))
1084 return PAGEREF_RECLAIM_CLEAN;
1086 return PAGEREF_RECLAIM;
1089 /* Check if a page is dirty or under writeback */
1090 static void page_check_dirty_writeback(struct page *page,
1091 bool *dirty, bool *writeback)
1093 struct address_space *mapping;
1096 * Anonymous pages are not handled by flushers and must be written
1097 * from reclaim context. Do not stall reclaim based on them
1099 if (!page_is_file_cache(page) ||
1100 (PageAnon(page) && !PageSwapBacked(page))) {
1106 /* By default assume that the page flags are accurate */
1107 *dirty = PageDirty(page);
1108 *writeback = PageWriteback(page);
1110 /* Verify dirty/writeback state if the filesystem supports it */
1111 if (!page_has_private(page))
1114 mapping = page_mapping(page);
1115 if (mapping && mapping->a_ops->is_dirty_writeback)
1116 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1120 * shrink_page_list() returns the number of reclaimed pages
1122 static unsigned long shrink_page_list(struct list_head *page_list,
1123 struct pglist_data *pgdat,
1124 struct scan_control *sc,
1125 enum ttu_flags ttu_flags,
1126 struct reclaim_stat *stat,
1129 LIST_HEAD(ret_pages);
1130 LIST_HEAD(free_pages);
1131 unsigned nr_reclaimed = 0;
1132 unsigned pgactivate = 0;
1134 memset(stat, 0, sizeof(*stat));
1137 while (!list_empty(page_list)) {
1138 struct address_space *mapping;
1141 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1142 bool dirty, writeback;
1143 unsigned int nr_pages;
1147 page = lru_to_page(page_list);
1148 list_del(&page->lru);
1150 if (!trylock_page(page))
1153 VM_BUG_ON_PAGE(PageActive(page), page);
1155 nr_pages = 1 << compound_order(page);
1157 /* Account the number of base pages even though THP */
1158 sc->nr_scanned += nr_pages;
1160 if (unlikely(!page_evictable(page)))
1161 goto activate_locked;
1163 if (!sc->may_unmap && page_mapped(page))
1166 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1167 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1170 * The number of dirty pages determines if a node is marked
1171 * reclaim_congested which affects wait_iff_congested. kswapd
1172 * will stall and start writing pages if the tail of the LRU
1173 * is all dirty unqueued pages.
1175 page_check_dirty_writeback(page, &dirty, &writeback);
1176 if (dirty || writeback)
1179 if (dirty && !writeback)
1180 stat->nr_unqueued_dirty++;
1183 * Treat this page as congested if the underlying BDI is or if
1184 * pages are cycling through the LRU so quickly that the
1185 * pages marked for immediate reclaim are making it to the
1186 * end of the LRU a second time.
1188 mapping = page_mapping(page);
1189 if (((dirty || writeback) && mapping &&
1190 inode_write_congested(mapping->host)) ||
1191 (writeback && PageReclaim(page)))
1192 stat->nr_congested++;
1195 * If a page at the tail of the LRU is under writeback, there
1196 * are three cases to consider.
1198 * 1) If reclaim is encountering an excessive number of pages
1199 * under writeback and this page is both under writeback and
1200 * PageReclaim then it indicates that pages are being queued
1201 * for IO but are being recycled through the LRU before the
1202 * IO can complete. Waiting on the page itself risks an
1203 * indefinite stall if it is impossible to writeback the
1204 * page due to IO error or disconnected storage so instead
1205 * note that the LRU is being scanned too quickly and the
1206 * caller can stall after page list has been processed.
1208 * 2) Global or new memcg reclaim encounters a page that is
1209 * not marked for immediate reclaim, or the caller does not
1210 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1211 * not to fs). In this case mark the page for immediate
1212 * reclaim and continue scanning.
1214 * Require may_enter_fs because we would wait on fs, which
1215 * may not have submitted IO yet. And the loop driver might
1216 * enter reclaim, and deadlock if it waits on a page for
1217 * which it is needed to do the write (loop masks off
1218 * __GFP_IO|__GFP_FS for this reason); but more thought
1219 * would probably show more reasons.
1221 * 3) Legacy memcg encounters a page that is already marked
1222 * PageReclaim. memcg does not have any dirty pages
1223 * throttling so we could easily OOM just because too many
1224 * pages are in writeback and there is nothing else to
1225 * reclaim. Wait for the writeback to complete.
1227 * In cases 1) and 2) we activate the pages to get them out of
1228 * the way while we continue scanning for clean pages on the
1229 * inactive list and refilling from the active list. The
1230 * observation here is that waiting for disk writes is more
1231 * expensive than potentially causing reloads down the line.
1232 * Since they're marked for immediate reclaim, they won't put
1233 * memory pressure on the cache working set any longer than it
1234 * takes to write them to disk.
1236 if (PageWriteback(page)) {
1238 if (current_is_kswapd() &&
1239 PageReclaim(page) &&
1240 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1241 stat->nr_immediate++;
1242 goto activate_locked;
1245 } else if (sane_reclaim(sc) ||
1246 !PageReclaim(page) || !may_enter_fs) {
1248 * This is slightly racy - end_page_writeback()
1249 * might have just cleared PageReclaim, then
1250 * setting PageReclaim here end up interpreted
1251 * as PageReadahead - but that does not matter
1252 * enough to care. What we do want is for this
1253 * page to have PageReclaim set next time memcg
1254 * reclaim reaches the tests above, so it will
1255 * then wait_on_page_writeback() to avoid OOM;
1256 * and it's also appropriate in global reclaim.
1258 SetPageReclaim(page);
1259 stat->nr_writeback++;
1260 goto activate_locked;
1265 wait_on_page_writeback(page);
1266 /* then go back and try same page again */
1267 list_add_tail(&page->lru, page_list);
1273 references = page_check_references(page, sc);
1275 switch (references) {
1276 case PAGEREF_ACTIVATE:
1277 goto activate_locked;
1279 stat->nr_ref_keep += nr_pages;
1281 case PAGEREF_RECLAIM:
1282 case PAGEREF_RECLAIM_CLEAN:
1283 ; /* try to reclaim the page below */
1287 * Anonymous process memory has backing store?
1288 * Try to allocate it some swap space here.
1289 * Lazyfree page could be freed directly
1291 if (PageAnon(page) && PageSwapBacked(page)) {
1292 if (!PageSwapCache(page)) {
1293 if (!(sc->gfp_mask & __GFP_IO))
1295 if (PageTransHuge(page)) {
1296 /* cannot split THP, skip it */
1297 if (!can_split_huge_page(page, NULL))
1298 goto activate_locked;
1300 * Split pages without a PMD map right
1301 * away. Chances are some or all of the
1302 * tail pages can be freed without IO.
1304 if (!compound_mapcount(page) &&
1305 split_huge_page_to_list(page,
1307 goto activate_locked;
1309 if (!add_to_swap(page)) {
1310 if (!PageTransHuge(page))
1311 goto activate_locked_split;
1312 /* Fallback to swap normal pages */
1313 if (split_huge_page_to_list(page,
1315 goto activate_locked;
1316 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1317 count_vm_event(THP_SWPOUT_FALLBACK);
1319 if (!add_to_swap(page))
1320 goto activate_locked_split;
1325 /* Adding to swap updated mapping */
1326 mapping = page_mapping(page);
1328 } else if (unlikely(PageTransHuge(page))) {
1329 /* Split file THP */
1330 if (split_huge_page_to_list(page, page_list))
1335 * THP may get split above, need minus tail pages and update
1336 * nr_pages to avoid accounting tail pages twice.
1338 * The tail pages that are added into swap cache successfully
1341 if ((nr_pages > 1) && !PageTransHuge(page)) {
1342 sc->nr_scanned -= (nr_pages - 1);
1347 * The page is mapped into the page tables of one or more
1348 * processes. Try to unmap it here.
1350 if (page_mapped(page)) {
1351 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1353 if (unlikely(PageTransHuge(page)))
1354 flags |= TTU_SPLIT_HUGE_PMD;
1355 if (!try_to_unmap(page, flags)) {
1356 stat->nr_unmap_fail += nr_pages;
1357 goto activate_locked;
1361 if (PageDirty(page)) {
1363 * Only kswapd can writeback filesystem pages
1364 * to avoid risk of stack overflow. But avoid
1365 * injecting inefficient single-page IO into
1366 * flusher writeback as much as possible: only
1367 * write pages when we've encountered many
1368 * dirty pages, and when we've already scanned
1369 * the rest of the LRU for clean pages and see
1370 * the same dirty pages again (PageReclaim).
1372 if (page_is_file_cache(page) &&
1373 (!current_is_kswapd() || !PageReclaim(page) ||
1374 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1376 * Immediately reclaim when written back.
1377 * Similar in principal to deactivate_page()
1378 * except we already have the page isolated
1379 * and know it's dirty
1381 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1382 SetPageReclaim(page);
1384 goto activate_locked;
1387 if (references == PAGEREF_RECLAIM_CLEAN)
1391 if (!sc->may_writepage)
1395 * Page is dirty. Flush the TLB if a writable entry
1396 * potentially exists to avoid CPU writes after IO
1397 * starts and then write it out here.
1399 try_to_unmap_flush_dirty();
1400 switch (pageout(page, mapping, sc)) {
1404 goto activate_locked;
1406 if (PageWriteback(page))
1408 if (PageDirty(page))
1412 * A synchronous write - probably a ramdisk. Go
1413 * ahead and try to reclaim the page.
1415 if (!trylock_page(page))
1417 if (PageDirty(page) || PageWriteback(page))
1419 mapping = page_mapping(page);
1421 ; /* try to free the page below */
1426 * If the page has buffers, try to free the buffer mappings
1427 * associated with this page. If we succeed we try to free
1430 * We do this even if the page is PageDirty().
1431 * try_to_release_page() does not perform I/O, but it is
1432 * possible for a page to have PageDirty set, but it is actually
1433 * clean (all its buffers are clean). This happens if the
1434 * buffers were written out directly, with submit_bh(). ext3
1435 * will do this, as well as the blockdev mapping.
1436 * try_to_release_page() will discover that cleanness and will
1437 * drop the buffers and mark the page clean - it can be freed.
1439 * Rarely, pages can have buffers and no ->mapping. These are
1440 * the pages which were not successfully invalidated in
1441 * truncate_complete_page(). We try to drop those buffers here
1442 * and if that worked, and the page is no longer mapped into
1443 * process address space (page_count == 1) it can be freed.
1444 * Otherwise, leave the page on the LRU so it is swappable.
1446 if (page_has_private(page)) {
1447 if (!try_to_release_page(page, sc->gfp_mask))
1448 goto activate_locked;
1449 if (!mapping && page_count(page) == 1) {
1451 if (put_page_testzero(page))
1455 * rare race with speculative reference.
1456 * the speculative reference will free
1457 * this page shortly, so we may
1458 * increment nr_reclaimed here (and
1459 * leave it off the LRU).
1467 if (PageAnon(page) && !PageSwapBacked(page)) {
1468 /* follow __remove_mapping for reference */
1469 if (!page_ref_freeze(page, 1))
1471 if (PageDirty(page)) {
1472 page_ref_unfreeze(page, 1);
1476 count_vm_event(PGLAZYFREED);
1477 count_memcg_page_event(page, PGLAZYFREED);
1478 } else if (!mapping || !__remove_mapping(mapping, page, true))
1484 * THP may get swapped out in a whole, need account
1487 nr_reclaimed += nr_pages;
1490 * Is there need to periodically free_page_list? It would
1491 * appear not as the counts should be low
1493 if (unlikely(PageTransHuge(page))) {
1494 mem_cgroup_uncharge(page);
1495 (*get_compound_page_dtor(page))(page);
1497 list_add(&page->lru, &free_pages);
1500 activate_locked_split:
1502 * The tail pages that are failed to add into swap cache
1503 * reach here. Fixup nr_scanned and nr_pages.
1506 sc->nr_scanned -= (nr_pages - 1);
1510 /* Not a candidate for swapping, so reclaim swap space. */
1511 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1513 try_to_free_swap(page);
1514 VM_BUG_ON_PAGE(PageActive(page), page);
1515 if (!PageMlocked(page)) {
1516 int type = page_is_file_cache(page);
1517 SetPageActive(page);
1518 stat->nr_activate[type] += nr_pages;
1519 count_memcg_page_event(page, PGACTIVATE);
1524 list_add(&page->lru, &ret_pages);
1525 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1528 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1530 mem_cgroup_uncharge_list(&free_pages);
1531 try_to_unmap_flush();
1532 free_unref_page_list(&free_pages);
1534 list_splice(&ret_pages, page_list);
1535 count_vm_events(PGACTIVATE, pgactivate);
1537 return nr_reclaimed;
1540 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1541 struct list_head *page_list)
1543 struct scan_control sc = {
1544 .gfp_mask = GFP_KERNEL,
1545 .priority = DEF_PRIORITY,
1548 struct reclaim_stat dummy_stat;
1550 struct page *page, *next;
1551 LIST_HEAD(clean_pages);
1553 list_for_each_entry_safe(page, next, page_list, lru) {
1554 if (page_is_file_cache(page) && !PageDirty(page) &&
1555 !__PageMovable(page) && !PageUnevictable(page)) {
1556 ClearPageActive(page);
1557 list_move(&page->lru, &clean_pages);
1561 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1562 TTU_IGNORE_ACCESS, &dummy_stat, true);
1563 list_splice(&clean_pages, page_list);
1564 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1569 * Attempt to remove the specified page from its LRU. Only take this page
1570 * if it is of the appropriate PageActive status. Pages which are being
1571 * freed elsewhere are also ignored.
1573 * page: page to consider
1574 * mode: one of the LRU isolation modes defined above
1576 * returns 0 on success, -ve errno on failure.
1578 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1582 /* Only take pages on the LRU. */
1586 /* Compaction should not handle unevictable pages but CMA can do so */
1587 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1593 * To minimise LRU disruption, the caller can indicate that it only
1594 * wants to isolate pages it will be able to operate on without
1595 * blocking - clean pages for the most part.
1597 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1598 * that it is possible to migrate without blocking
1600 if (mode & ISOLATE_ASYNC_MIGRATE) {
1601 /* All the caller can do on PageWriteback is block */
1602 if (PageWriteback(page))
1605 if (PageDirty(page)) {
1606 struct address_space *mapping;
1610 * Only pages without mappings or that have a
1611 * ->migratepage callback are possible to migrate
1612 * without blocking. However, we can be racing with
1613 * truncation so it's necessary to lock the page
1614 * to stabilise the mapping as truncation holds
1615 * the page lock until after the page is removed
1616 * from the page cache.
1618 if (!trylock_page(page))
1621 mapping = page_mapping(page);
1622 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1629 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1632 if (likely(get_page_unless_zero(page))) {
1634 * Be careful not to clear PageLRU until after we're
1635 * sure the page is not being freed elsewhere -- the
1636 * page release code relies on it.
1647 * Update LRU sizes after isolating pages. The LRU size updates must
1648 * be complete before mem_cgroup_update_lru_size due to a santity check.
1650 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1651 enum lru_list lru, unsigned long *nr_zone_taken)
1655 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1656 if (!nr_zone_taken[zid])
1659 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1661 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1668 * pgdat->lru_lock is heavily contended. Some of the functions that
1669 * shrink the lists perform better by taking out a batch of pages
1670 * and working on them outside the LRU lock.
1672 * For pagecache intensive workloads, this function is the hottest
1673 * spot in the kernel (apart from copy_*_user functions).
1675 * Appropriate locks must be held before calling this function.
1677 * @nr_to_scan: The number of eligible pages to look through on the list.
1678 * @lruvec: The LRU vector to pull pages from.
1679 * @dst: The temp list to put pages on to.
1680 * @nr_scanned: The number of pages that were scanned.
1681 * @sc: The scan_control struct for this reclaim session
1682 * @mode: One of the LRU isolation modes
1683 * @lru: LRU list id for isolating
1685 * returns how many pages were moved onto *@dst.
1687 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1688 struct lruvec *lruvec, struct list_head *dst,
1689 unsigned long *nr_scanned, struct scan_control *sc,
1692 struct list_head *src = &lruvec->lists[lru];
1693 unsigned long nr_taken = 0;
1694 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1695 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1696 unsigned long skipped = 0;
1697 unsigned long scan, total_scan, nr_pages;
1698 LIST_HEAD(pages_skipped);
1699 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1703 while (scan < nr_to_scan && !list_empty(src)) {
1706 page = lru_to_page(src);
1707 prefetchw_prev_lru_page(page, src, flags);
1709 VM_BUG_ON_PAGE(!PageLRU(page), page);
1711 nr_pages = 1 << compound_order(page);
1712 total_scan += nr_pages;
1714 if (page_zonenum(page) > sc->reclaim_idx) {
1715 list_move(&page->lru, &pages_skipped);
1716 nr_skipped[page_zonenum(page)] += nr_pages;
1721 * Do not count skipped pages because that makes the function
1722 * return with no isolated pages if the LRU mostly contains
1723 * ineligible pages. This causes the VM to not reclaim any
1724 * pages, triggering a premature OOM.
1726 * Account all tail pages of THP. This would not cause
1727 * premature OOM since __isolate_lru_page() returns -EBUSY
1728 * only when the page is being freed somewhere else.
1731 switch (__isolate_lru_page(page, mode)) {
1733 nr_taken += nr_pages;
1734 nr_zone_taken[page_zonenum(page)] += nr_pages;
1735 list_move(&page->lru, dst);
1739 /* else it is being freed elsewhere */
1740 list_move(&page->lru, src);
1749 * Splice any skipped pages to the start of the LRU list. Note that
1750 * this disrupts the LRU order when reclaiming for lower zones but
1751 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1752 * scanning would soon rescan the same pages to skip and put the
1753 * system at risk of premature OOM.
1755 if (!list_empty(&pages_skipped)) {
1758 list_splice(&pages_skipped, src);
1759 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1760 if (!nr_skipped[zid])
1763 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1764 skipped += nr_skipped[zid];
1767 *nr_scanned = total_scan;
1768 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1769 total_scan, skipped, nr_taken, mode, lru);
1770 update_lru_sizes(lruvec, lru, nr_zone_taken);
1775 * isolate_lru_page - tries to isolate a page from its LRU list
1776 * @page: page to isolate from its LRU list
1778 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1779 * vmstat statistic corresponding to whatever LRU list the page was on.
1781 * Returns 0 if the page was removed from an LRU list.
1782 * Returns -EBUSY if the page was not on an LRU list.
1784 * The returned page will have PageLRU() cleared. If it was found on
1785 * the active list, it will have PageActive set. If it was found on
1786 * the unevictable list, it will have the PageUnevictable bit set. That flag
1787 * may need to be cleared by the caller before letting the page go.
1789 * The vmstat statistic corresponding to the list on which the page was
1790 * found will be decremented.
1794 * (1) Must be called with an elevated refcount on the page. This is a
1795 * fundamentnal difference from isolate_lru_pages (which is called
1796 * without a stable reference).
1797 * (2) the lru_lock must not be held.
1798 * (3) interrupts must be enabled.
1800 int isolate_lru_page(struct page *page)
1804 VM_BUG_ON_PAGE(!page_count(page), page);
1805 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1807 if (PageLRU(page)) {
1808 pg_data_t *pgdat = page_pgdat(page);
1809 struct lruvec *lruvec;
1811 spin_lock_irq(&pgdat->lru_lock);
1812 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1813 if (PageLRU(page)) {
1814 int lru = page_lru(page);
1817 del_page_from_lru_list(page, lruvec, lru);
1820 spin_unlock_irq(&pgdat->lru_lock);
1826 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1827 * then get resheduled. When there are massive number of tasks doing page
1828 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1829 * the LRU list will go small and be scanned faster than necessary, leading to
1830 * unnecessary swapping, thrashing and OOM.
1832 static int too_many_isolated(struct pglist_data *pgdat, int file,
1833 struct scan_control *sc)
1835 unsigned long inactive, isolated;
1837 if (current_is_kswapd())
1840 if (!sane_reclaim(sc))
1844 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1845 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1847 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1848 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1852 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1853 * won't get blocked by normal direct-reclaimers, forming a circular
1856 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1859 return isolated > inactive;
1863 * This moves pages from @list to corresponding LRU list.
1865 * We move them the other way if the page is referenced by one or more
1866 * processes, from rmap.
1868 * If the pages are mostly unmapped, the processing is fast and it is
1869 * appropriate to hold zone_lru_lock across the whole operation. But if
1870 * the pages are mapped, the processing is slow (page_referenced()) so we
1871 * should drop zone_lru_lock around each page. It's impossible to balance
1872 * this, so instead we remove the pages from the LRU while processing them.
1873 * It is safe to rely on PG_active against the non-LRU pages in here because
1874 * nobody will play with that bit on a non-LRU page.
1876 * The downside is that we have to touch page->_refcount against each page.
1877 * But we had to alter page->flags anyway.
1879 * Returns the number of pages moved to the given lruvec.
1882 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1883 struct list_head *list)
1885 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1886 int nr_pages, nr_moved = 0;
1887 LIST_HEAD(pages_to_free);
1891 while (!list_empty(list)) {
1892 page = lru_to_page(list);
1893 VM_BUG_ON_PAGE(PageLRU(page), page);
1894 if (unlikely(!page_evictable(page))) {
1895 list_del(&page->lru);
1896 spin_unlock_irq(&pgdat->lru_lock);
1897 putback_lru_page(page);
1898 spin_lock_irq(&pgdat->lru_lock);
1901 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1904 lru = page_lru(page);
1906 nr_pages = hpage_nr_pages(page);
1907 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1908 list_move(&page->lru, &lruvec->lists[lru]);
1910 if (put_page_testzero(page)) {
1911 __ClearPageLRU(page);
1912 __ClearPageActive(page);
1913 del_page_from_lru_list(page, lruvec, lru);
1915 if (unlikely(PageCompound(page))) {
1916 spin_unlock_irq(&pgdat->lru_lock);
1917 mem_cgroup_uncharge(page);
1918 (*get_compound_page_dtor(page))(page);
1919 spin_lock_irq(&pgdat->lru_lock);
1921 list_add(&page->lru, &pages_to_free);
1923 nr_moved += nr_pages;
1928 * To save our caller's stack, now use input list for pages to free.
1930 list_splice(&pages_to_free, list);
1936 * If a kernel thread (such as nfsd for loop-back mounts) services
1937 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1938 * In that case we should only throttle if the backing device it is
1939 * writing to is congested. In other cases it is safe to throttle.
1941 static int current_may_throttle(void)
1943 return !(current->flags & PF_LESS_THROTTLE) ||
1944 current->backing_dev_info == NULL ||
1945 bdi_write_congested(current->backing_dev_info);
1949 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1950 * of reclaimed pages
1952 static noinline_for_stack unsigned long
1953 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1954 struct scan_control *sc, enum lru_list lru)
1956 LIST_HEAD(page_list);
1957 unsigned long nr_scanned;
1958 unsigned long nr_reclaimed = 0;
1959 unsigned long nr_taken;
1960 struct reclaim_stat stat;
1961 int file = is_file_lru(lru);
1962 enum vm_event_item item;
1963 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1964 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1965 bool stalled = false;
1967 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1971 /* wait a bit for the reclaimer. */
1975 /* We are about to die and free our memory. Return now. */
1976 if (fatal_signal_pending(current))
1977 return SWAP_CLUSTER_MAX;
1982 spin_lock_irq(&pgdat->lru_lock);
1984 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1985 &nr_scanned, sc, lru);
1987 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1988 reclaim_stat->recent_scanned[file] += nr_taken;
1990 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1991 if (global_reclaim(sc))
1992 __count_vm_events(item, nr_scanned);
1993 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1994 spin_unlock_irq(&pgdat->lru_lock);
1999 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
2002 spin_lock_irq(&pgdat->lru_lock);
2004 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2005 if (global_reclaim(sc))
2006 __count_vm_events(item, nr_reclaimed);
2007 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2008 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2009 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2011 move_pages_to_lru(lruvec, &page_list);
2013 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2015 spin_unlock_irq(&pgdat->lru_lock);
2017 mem_cgroup_uncharge_list(&page_list);
2018 free_unref_page_list(&page_list);
2021 * If dirty pages are scanned that are not queued for IO, it
2022 * implies that flushers are not doing their job. This can
2023 * happen when memory pressure pushes dirty pages to the end of
2024 * the LRU before the dirty limits are breached and the dirty
2025 * data has expired. It can also happen when the proportion of
2026 * dirty pages grows not through writes but through memory
2027 * pressure reclaiming all the clean cache. And in some cases,
2028 * the flushers simply cannot keep up with the allocation
2029 * rate. Nudge the flusher threads in case they are asleep.
2031 if (stat.nr_unqueued_dirty == nr_taken)
2032 wakeup_flusher_threads(WB_REASON_VMSCAN);
2034 sc->nr.dirty += stat.nr_dirty;
2035 sc->nr.congested += stat.nr_congested;
2036 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2037 sc->nr.writeback += stat.nr_writeback;
2038 sc->nr.immediate += stat.nr_immediate;
2039 sc->nr.taken += nr_taken;
2041 sc->nr.file_taken += nr_taken;
2043 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2044 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2045 return nr_reclaimed;
2048 static void shrink_active_list(unsigned long nr_to_scan,
2049 struct lruvec *lruvec,
2050 struct scan_control *sc,
2053 unsigned long nr_taken;
2054 unsigned long nr_scanned;
2055 unsigned long vm_flags;
2056 LIST_HEAD(l_hold); /* The pages which were snipped off */
2057 LIST_HEAD(l_active);
2058 LIST_HEAD(l_inactive);
2060 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2061 unsigned nr_deactivate, nr_activate;
2062 unsigned nr_rotated = 0;
2063 int file = is_file_lru(lru);
2064 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2068 spin_lock_irq(&pgdat->lru_lock);
2070 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2071 &nr_scanned, sc, lru);
2073 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2074 reclaim_stat->recent_scanned[file] += nr_taken;
2076 __count_vm_events(PGREFILL, nr_scanned);
2077 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2079 spin_unlock_irq(&pgdat->lru_lock);
2081 while (!list_empty(&l_hold)) {
2083 page = lru_to_page(&l_hold);
2084 list_del(&page->lru);
2086 if (unlikely(!page_evictable(page))) {
2087 putback_lru_page(page);
2091 if (unlikely(buffer_heads_over_limit)) {
2092 if (page_has_private(page) && trylock_page(page)) {
2093 if (page_has_private(page))
2094 try_to_release_page(page, 0);
2099 if (page_referenced(page, 0, sc->target_mem_cgroup,
2101 nr_rotated += hpage_nr_pages(page);
2103 * Identify referenced, file-backed active pages and
2104 * give them one more trip around the active list. So
2105 * that executable code get better chances to stay in
2106 * memory under moderate memory pressure. Anon pages
2107 * are not likely to be evicted by use-once streaming
2108 * IO, plus JVM can create lots of anon VM_EXEC pages,
2109 * so we ignore them here.
2111 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2112 list_add(&page->lru, &l_active);
2117 ClearPageActive(page); /* we are de-activating */
2118 SetPageWorkingset(page);
2119 list_add(&page->lru, &l_inactive);
2123 * Move pages back to the lru list.
2125 spin_lock_irq(&pgdat->lru_lock);
2127 * Count referenced pages from currently used mappings as rotated,
2128 * even though only some of them are actually re-activated. This
2129 * helps balance scan pressure between file and anonymous pages in
2132 reclaim_stat->recent_rotated[file] += nr_rotated;
2134 nr_activate = move_pages_to_lru(lruvec, &l_active);
2135 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2136 /* Keep all free pages in l_active list */
2137 list_splice(&l_inactive, &l_active);
2139 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2140 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2142 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2143 spin_unlock_irq(&pgdat->lru_lock);
2145 mem_cgroup_uncharge_list(&l_active);
2146 free_unref_page_list(&l_active);
2147 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2148 nr_deactivate, nr_rotated, sc->priority, file);
2152 * The inactive anon list should be small enough that the VM never has
2153 * to do too much work.
2155 * The inactive file list should be small enough to leave most memory
2156 * to the established workingset on the scan-resistant active list,
2157 * but large enough to avoid thrashing the aggregate readahead window.
2159 * Both inactive lists should also be large enough that each inactive
2160 * page has a chance to be referenced again before it is reclaimed.
2162 * If that fails and refaulting is observed, the inactive list grows.
2164 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2165 * on this LRU, maintained by the pageout code. An inactive_ratio
2166 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2169 * memory ratio inactive
2170 * -------------------------------------
2179 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2180 struct scan_control *sc, bool trace)
2182 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2183 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2184 enum lru_list inactive_lru = file * LRU_FILE;
2185 unsigned long inactive, active;
2186 unsigned long inactive_ratio;
2187 unsigned long refaults;
2191 * If we don't have swap space, anonymous page deactivation
2194 if (!file && !total_swap_pages)
2197 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2198 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2201 * When refaults are being observed, it means a new workingset
2202 * is being established. Disable active list protection to get
2203 * rid of the stale workingset quickly.
2205 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2206 if (file && lruvec->refaults != refaults) {
2209 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2211 inactive_ratio = int_sqrt(10 * gb);
2217 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2218 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2219 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2220 inactive_ratio, file);
2222 return inactive * inactive_ratio < active;
2225 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2226 struct lruvec *lruvec, struct scan_control *sc)
2228 if (is_active_lru(lru)) {
2229 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2230 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2234 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2245 * Determine how aggressively the anon and file LRU lists should be
2246 * scanned. The relative value of each set of LRU lists is determined
2247 * by looking at the fraction of the pages scanned we did rotate back
2248 * onto the active list instead of evict.
2250 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2251 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2253 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2254 struct scan_control *sc, unsigned long *nr,
2255 unsigned long *lru_pages)
2257 int swappiness = mem_cgroup_swappiness(memcg);
2258 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2260 u64 denominator = 0; /* gcc */
2261 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2262 unsigned long anon_prio, file_prio;
2263 enum scan_balance scan_balance;
2264 unsigned long anon, file;
2265 unsigned long ap, fp;
2268 /* If we have no swap space, do not bother scanning anon pages. */
2269 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2270 scan_balance = SCAN_FILE;
2275 * Global reclaim will swap to prevent OOM even with no
2276 * swappiness, but memcg users want to use this knob to
2277 * disable swapping for individual groups completely when
2278 * using the memory controller's swap limit feature would be
2281 if (!global_reclaim(sc) && !swappiness) {
2282 scan_balance = SCAN_FILE;
2287 * Do not apply any pressure balancing cleverness when the
2288 * system is close to OOM, scan both anon and file equally
2289 * (unless the swappiness setting disagrees with swapping).
2291 if (!sc->priority && swappiness) {
2292 scan_balance = SCAN_EQUAL;
2297 * Prevent the reclaimer from falling into the cache trap: as
2298 * cache pages start out inactive, every cache fault will tip
2299 * the scan balance towards the file LRU. And as the file LRU
2300 * shrinks, so does the window for rotation from references.
2301 * This means we have a runaway feedback loop where a tiny
2302 * thrashing file LRU becomes infinitely more attractive than
2303 * anon pages. Try to detect this based on file LRU size.
2305 if (global_reclaim(sc)) {
2306 unsigned long pgdatfile;
2307 unsigned long pgdatfree;
2309 unsigned long total_high_wmark = 0;
2311 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2312 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2313 node_page_state(pgdat, NR_INACTIVE_FILE);
2315 for (z = 0; z < MAX_NR_ZONES; z++) {
2316 struct zone *zone = &pgdat->node_zones[z];
2317 if (!managed_zone(zone))
2320 total_high_wmark += high_wmark_pages(zone);
2323 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2325 * Force SCAN_ANON if there are enough inactive
2326 * anonymous pages on the LRU in eligible zones.
2327 * Otherwise, the small LRU gets thrashed.
2329 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2330 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2332 scan_balance = SCAN_ANON;
2339 * If there is enough inactive page cache, i.e. if the size of the
2340 * inactive list is greater than that of the active list *and* the
2341 * inactive list actually has some pages to scan on this priority, we
2342 * do not reclaim anything from the anonymous working set right now.
2343 * Without the second condition we could end up never scanning an
2344 * lruvec even if it has plenty of old anonymous pages unless the
2345 * system is under heavy pressure.
2347 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2348 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2349 scan_balance = SCAN_FILE;
2353 scan_balance = SCAN_FRACT;
2356 * With swappiness at 100, anonymous and file have the same priority.
2357 * This scanning priority is essentially the inverse of IO cost.
2359 anon_prio = swappiness;
2360 file_prio = 200 - anon_prio;
2363 * OK, so we have swap space and a fair amount of page cache
2364 * pages. We use the recently rotated / recently scanned
2365 * ratios to determine how valuable each cache is.
2367 * Because workloads change over time (and to avoid overflow)
2368 * we keep these statistics as a floating average, which ends
2369 * up weighing recent references more than old ones.
2371 * anon in [0], file in [1]
2374 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2375 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2376 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2377 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2379 spin_lock_irq(&pgdat->lru_lock);
2380 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2381 reclaim_stat->recent_scanned[0] /= 2;
2382 reclaim_stat->recent_rotated[0] /= 2;
2385 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2386 reclaim_stat->recent_scanned[1] /= 2;
2387 reclaim_stat->recent_rotated[1] /= 2;
2391 * The amount of pressure on anon vs file pages is inversely
2392 * proportional to the fraction of recently scanned pages on
2393 * each list that were recently referenced and in active use.
2395 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2396 ap /= reclaim_stat->recent_rotated[0] + 1;
2398 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2399 fp /= reclaim_stat->recent_rotated[1] + 1;
2400 spin_unlock_irq(&pgdat->lru_lock);
2404 denominator = ap + fp + 1;
2407 for_each_evictable_lru(lru) {
2408 int file = is_file_lru(lru);
2412 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2413 scan = size >> sc->priority;
2415 * If the cgroup's already been deleted, make sure to
2416 * scrape out the remaining cache.
2418 if (!scan && !mem_cgroup_online(memcg))
2419 scan = min(size, SWAP_CLUSTER_MAX);
2421 switch (scan_balance) {
2423 /* Scan lists relative to size */
2427 * Scan types proportional to swappiness and
2428 * their relative recent reclaim efficiency.
2429 * Make sure we don't miss the last page
2430 * because of a round-off error.
2432 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2437 /* Scan one type exclusively */
2438 if ((scan_balance == SCAN_FILE) != file) {
2444 /* Look ma, no brain */
2454 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2456 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2457 struct scan_control *sc, unsigned long *lru_pages)
2459 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2460 unsigned long nr[NR_LRU_LISTS];
2461 unsigned long targets[NR_LRU_LISTS];
2462 unsigned long nr_to_scan;
2464 unsigned long nr_reclaimed = 0;
2465 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2466 struct blk_plug plug;
2469 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2471 /* Record the original scan target for proportional adjustments later */
2472 memcpy(targets, nr, sizeof(nr));
2475 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2476 * event that can occur when there is little memory pressure e.g.
2477 * multiple streaming readers/writers. Hence, we do not abort scanning
2478 * when the requested number of pages are reclaimed when scanning at
2479 * DEF_PRIORITY on the assumption that the fact we are direct
2480 * reclaiming implies that kswapd is not keeping up and it is best to
2481 * do a batch of work at once. For memcg reclaim one check is made to
2482 * abort proportional reclaim if either the file or anon lru has already
2483 * dropped to zero at the first pass.
2485 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2486 sc->priority == DEF_PRIORITY);
2488 blk_start_plug(&plug);
2489 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2490 nr[LRU_INACTIVE_FILE]) {
2491 unsigned long nr_anon, nr_file, percentage;
2492 unsigned long nr_scanned;
2494 for_each_evictable_lru(lru) {
2496 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2497 nr[lru] -= nr_to_scan;
2499 nr_reclaimed += shrink_list(lru, nr_to_scan,
2506 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2510 * For kswapd and memcg, reclaim at least the number of pages
2511 * requested. Ensure that the anon and file LRUs are scanned
2512 * proportionally what was requested by get_scan_count(). We
2513 * stop reclaiming one LRU and reduce the amount scanning
2514 * proportional to the original scan target.
2516 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2517 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2520 * It's just vindictive to attack the larger once the smaller
2521 * has gone to zero. And given the way we stop scanning the
2522 * smaller below, this makes sure that we only make one nudge
2523 * towards proportionality once we've got nr_to_reclaim.
2525 if (!nr_file || !nr_anon)
2528 if (nr_file > nr_anon) {
2529 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2530 targets[LRU_ACTIVE_ANON] + 1;
2532 percentage = nr_anon * 100 / scan_target;
2534 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2535 targets[LRU_ACTIVE_FILE] + 1;
2537 percentage = nr_file * 100 / scan_target;
2540 /* Stop scanning the smaller of the LRU */
2542 nr[lru + LRU_ACTIVE] = 0;
2545 * Recalculate the other LRU scan count based on its original
2546 * scan target and the percentage scanning already complete
2548 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2549 nr_scanned = targets[lru] - nr[lru];
2550 nr[lru] = targets[lru] * (100 - percentage) / 100;
2551 nr[lru] -= min(nr[lru], nr_scanned);
2554 nr_scanned = targets[lru] - nr[lru];
2555 nr[lru] = targets[lru] * (100 - percentage) / 100;
2556 nr[lru] -= min(nr[lru], nr_scanned);
2558 scan_adjusted = true;
2560 blk_finish_plug(&plug);
2561 sc->nr_reclaimed += nr_reclaimed;
2564 * Even if we did not try to evict anon pages at all, we want to
2565 * rebalance the anon lru active/inactive ratio.
2567 if (inactive_list_is_low(lruvec, false, sc, true))
2568 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2569 sc, LRU_ACTIVE_ANON);
2572 /* Use reclaim/compaction for costly allocs or under memory pressure */
2573 static bool in_reclaim_compaction(struct scan_control *sc)
2575 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2576 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2577 sc->priority < DEF_PRIORITY - 2))
2584 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2585 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2586 * true if more pages should be reclaimed such that when the page allocator
2587 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2588 * It will give up earlier than that if there is difficulty reclaiming pages.
2590 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2591 unsigned long nr_reclaimed,
2592 unsigned long nr_scanned,
2593 struct scan_control *sc)
2595 unsigned long pages_for_compaction;
2596 unsigned long inactive_lru_pages;
2599 /* If not in reclaim/compaction mode, stop */
2600 if (!in_reclaim_compaction(sc))
2603 /* Consider stopping depending on scan and reclaim activity */
2604 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2606 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2607 * full LRU list has been scanned and we are still failing
2608 * to reclaim pages. This full LRU scan is potentially
2609 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2611 if (!nr_reclaimed && !nr_scanned)
2615 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2616 * fail without consequence, stop if we failed to reclaim
2617 * any pages from the last SWAP_CLUSTER_MAX number of
2618 * pages that were scanned. This will return to the
2619 * caller faster at the risk reclaim/compaction and
2620 * the resulting allocation attempt fails
2627 * If we have not reclaimed enough pages for compaction and the
2628 * inactive lists are large enough, continue reclaiming
2630 pages_for_compaction = compact_gap(sc->order);
2631 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2632 if (get_nr_swap_pages() > 0)
2633 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2634 if (sc->nr_reclaimed < pages_for_compaction &&
2635 inactive_lru_pages > pages_for_compaction)
2638 /* If compaction would go ahead or the allocation would succeed, stop */
2639 for (z = 0; z <= sc->reclaim_idx; z++) {
2640 struct zone *zone = &pgdat->node_zones[z];
2641 if (!managed_zone(zone))
2644 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2645 case COMPACT_SUCCESS:
2646 case COMPACT_CONTINUE:
2649 /* check next zone */
2656 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2658 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2659 (memcg && memcg_congested(pgdat, memcg));
2662 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2664 struct reclaim_state *reclaim_state = current->reclaim_state;
2665 unsigned long nr_reclaimed, nr_scanned;
2666 bool reclaimable = false;
2669 struct mem_cgroup *root = sc->target_mem_cgroup;
2670 struct mem_cgroup_reclaim_cookie reclaim = {
2672 .priority = sc->priority,
2674 unsigned long node_lru_pages = 0;
2675 struct mem_cgroup *memcg;
2677 memset(&sc->nr, 0, sizeof(sc->nr));
2679 nr_reclaimed = sc->nr_reclaimed;
2680 nr_scanned = sc->nr_scanned;
2682 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2684 unsigned long lru_pages;
2685 unsigned long reclaimed;
2686 unsigned long scanned;
2688 switch (mem_cgroup_protected(root, memcg)) {
2689 case MEMCG_PROT_MIN:
2692 * If there is no reclaimable memory, OOM.
2695 case MEMCG_PROT_LOW:
2698 * Respect the protection only as long as
2699 * there is an unprotected supply
2700 * of reclaimable memory from other cgroups.
2702 if (!sc->memcg_low_reclaim) {
2703 sc->memcg_low_skipped = 1;
2706 memcg_memory_event(memcg, MEMCG_LOW);
2708 case MEMCG_PROT_NONE:
2712 reclaimed = sc->nr_reclaimed;
2713 scanned = sc->nr_scanned;
2714 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2715 node_lru_pages += lru_pages;
2717 if (sc->may_shrinkslab) {
2718 shrink_slab(sc->gfp_mask, pgdat->node_id,
2719 memcg, sc->priority);
2722 /* Record the group's reclaim efficiency */
2723 vmpressure(sc->gfp_mask, memcg, false,
2724 sc->nr_scanned - scanned,
2725 sc->nr_reclaimed - reclaimed);
2728 * Kswapd have to scan all memory cgroups to fulfill
2729 * the overall scan target for the node.
2731 * Limit reclaim, on the other hand, only cares about
2732 * nr_to_reclaim pages to be reclaimed and it will
2733 * retry with decreasing priority if one round over the
2734 * whole hierarchy is not sufficient.
2736 if (!current_is_kswapd() &&
2737 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2738 mem_cgroup_iter_break(root, memcg);
2741 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2743 if (reclaim_state) {
2744 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2745 reclaim_state->reclaimed_slab = 0;
2748 /* Record the subtree's reclaim efficiency */
2749 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2750 sc->nr_scanned - nr_scanned,
2751 sc->nr_reclaimed - nr_reclaimed);
2753 if (sc->nr_reclaimed - nr_reclaimed)
2756 if (current_is_kswapd()) {
2758 * If reclaim is isolating dirty pages under writeback,
2759 * it implies that the long-lived page allocation rate
2760 * is exceeding the page laundering rate. Either the
2761 * global limits are not being effective at throttling
2762 * processes due to the page distribution throughout
2763 * zones or there is heavy usage of a slow backing
2764 * device. The only option is to throttle from reclaim
2765 * context which is not ideal as there is no guarantee
2766 * the dirtying process is throttled in the same way
2767 * balance_dirty_pages() manages.
2769 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2770 * count the number of pages under pages flagged for
2771 * immediate reclaim and stall if any are encountered
2772 * in the nr_immediate check below.
2774 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2775 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2778 * Tag a node as congested if all the dirty pages
2779 * scanned were backed by a congested BDI and
2780 * wait_iff_congested will stall.
2782 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2783 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2785 /* Allow kswapd to start writing pages during reclaim.*/
2786 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2787 set_bit(PGDAT_DIRTY, &pgdat->flags);
2790 * If kswapd scans pages marked marked for immediate
2791 * reclaim and under writeback (nr_immediate), it
2792 * implies that pages are cycling through the LRU
2793 * faster than they are written so also forcibly stall.
2795 if (sc->nr.immediate)
2796 congestion_wait(BLK_RW_ASYNC, HZ/10);
2800 * Legacy memcg will stall in page writeback so avoid forcibly
2801 * stalling in wait_iff_congested().
2803 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2804 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2805 set_memcg_congestion(pgdat, root, true);
2808 * Stall direct reclaim for IO completions if underlying BDIs
2809 * and node is congested. Allow kswapd to continue until it
2810 * starts encountering unqueued dirty pages or cycling through
2811 * the LRU too quickly.
2813 if (!sc->hibernation_mode && !current_is_kswapd() &&
2814 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2815 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2817 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2818 sc->nr_scanned - nr_scanned, sc));
2821 * Kswapd gives up on balancing particular nodes after too
2822 * many failures to reclaim anything from them and goes to
2823 * sleep. On reclaim progress, reset the failure counter. A
2824 * successful direct reclaim run will revive a dormant kswapd.
2827 pgdat->kswapd_failures = 0;
2833 * Returns true if compaction should go ahead for a costly-order request, or
2834 * the allocation would already succeed without compaction. Return false if we
2835 * should reclaim first.
2837 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2839 unsigned long watermark;
2840 enum compact_result suitable;
2842 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2843 if (suitable == COMPACT_SUCCESS)
2844 /* Allocation should succeed already. Don't reclaim. */
2846 if (suitable == COMPACT_SKIPPED)
2847 /* Compaction cannot yet proceed. Do reclaim. */
2851 * Compaction is already possible, but it takes time to run and there
2852 * are potentially other callers using the pages just freed. So proceed
2853 * with reclaim to make a buffer of free pages available to give
2854 * compaction a reasonable chance of completing and allocating the page.
2855 * Note that we won't actually reclaim the whole buffer in one attempt
2856 * as the target watermark in should_continue_reclaim() is lower. But if
2857 * we are already above the high+gap watermark, don't reclaim at all.
2859 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2861 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2865 * This is the direct reclaim path, for page-allocating processes. We only
2866 * try to reclaim pages from zones which will satisfy the caller's allocation
2869 * If a zone is deemed to be full of pinned pages then just give it a light
2870 * scan then give up on it.
2872 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2876 unsigned long nr_soft_reclaimed;
2877 unsigned long nr_soft_scanned;
2879 pg_data_t *last_pgdat = NULL;
2882 * If the number of buffer_heads in the machine exceeds the maximum
2883 * allowed level, force direct reclaim to scan the highmem zone as
2884 * highmem pages could be pinning lowmem pages storing buffer_heads
2886 orig_mask = sc->gfp_mask;
2887 if (buffer_heads_over_limit) {
2888 sc->gfp_mask |= __GFP_HIGHMEM;
2889 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2892 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2893 sc->reclaim_idx, sc->nodemask) {
2895 * Take care memory controller reclaiming has small influence
2898 if (global_reclaim(sc)) {
2899 if (!cpuset_zone_allowed(zone,
2900 GFP_KERNEL | __GFP_HARDWALL))
2904 * If we already have plenty of memory free for
2905 * compaction in this zone, don't free any more.
2906 * Even though compaction is invoked for any
2907 * non-zero order, only frequent costly order
2908 * reclamation is disruptive enough to become a
2909 * noticeable problem, like transparent huge
2912 if (IS_ENABLED(CONFIG_COMPACTION) &&
2913 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2914 compaction_ready(zone, sc)) {
2915 sc->compaction_ready = true;
2920 * Shrink each node in the zonelist once. If the
2921 * zonelist is ordered by zone (not the default) then a
2922 * node may be shrunk multiple times but in that case
2923 * the user prefers lower zones being preserved.
2925 if (zone->zone_pgdat == last_pgdat)
2929 * This steals pages from memory cgroups over softlimit
2930 * and returns the number of reclaimed pages and
2931 * scanned pages. This works for global memory pressure
2932 * and balancing, not for a memcg's limit.
2934 nr_soft_scanned = 0;
2935 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2936 sc->order, sc->gfp_mask,
2938 sc->nr_reclaimed += nr_soft_reclaimed;
2939 sc->nr_scanned += nr_soft_scanned;
2940 /* need some check for avoid more shrink_zone() */
2943 /* See comment about same check for global reclaim above */
2944 if (zone->zone_pgdat == last_pgdat)
2946 last_pgdat = zone->zone_pgdat;
2947 shrink_node(zone->zone_pgdat, sc);
2951 * Restore to original mask to avoid the impact on the caller if we
2952 * promoted it to __GFP_HIGHMEM.
2954 sc->gfp_mask = orig_mask;
2957 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2959 struct mem_cgroup *memcg;
2961 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2963 unsigned long refaults;
2964 struct lruvec *lruvec;
2966 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2967 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2968 lruvec->refaults = refaults;
2969 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2973 * This is the main entry point to direct page reclaim.
2975 * If a full scan of the inactive list fails to free enough memory then we
2976 * are "out of memory" and something needs to be killed.
2978 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2979 * high - the zone may be full of dirty or under-writeback pages, which this
2980 * caller can't do much about. We kick the writeback threads and take explicit
2981 * naps in the hope that some of these pages can be written. But if the
2982 * allocating task holds filesystem locks which prevent writeout this might not
2983 * work, and the allocation attempt will fail.
2985 * returns: 0, if no pages reclaimed
2986 * else, the number of pages reclaimed
2988 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2989 struct scan_control *sc)
2991 int initial_priority = sc->priority;
2992 pg_data_t *last_pgdat;
2996 delayacct_freepages_start();
2998 if (global_reclaim(sc))
2999 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3002 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3005 shrink_zones(zonelist, sc);
3007 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3010 if (sc->compaction_ready)
3014 * If we're getting trouble reclaiming, start doing
3015 * writepage even in laptop mode.
3017 if (sc->priority < DEF_PRIORITY - 2)
3018 sc->may_writepage = 1;
3019 } while (--sc->priority >= 0);
3022 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3024 if (zone->zone_pgdat == last_pgdat)
3026 last_pgdat = zone->zone_pgdat;
3027 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3028 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3031 delayacct_freepages_end();
3033 if (sc->nr_reclaimed)
3034 return sc->nr_reclaimed;
3036 /* Aborted reclaim to try compaction? don't OOM, then */
3037 if (sc->compaction_ready)
3040 /* Untapped cgroup reserves? Don't OOM, retry. */
3041 if (sc->memcg_low_skipped) {
3042 sc->priority = initial_priority;
3043 sc->memcg_low_reclaim = 1;
3044 sc->memcg_low_skipped = 0;
3051 static bool allow_direct_reclaim(pg_data_t *pgdat)
3054 unsigned long pfmemalloc_reserve = 0;
3055 unsigned long free_pages = 0;
3059 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3062 for (i = 0; i <= ZONE_NORMAL; i++) {
3063 zone = &pgdat->node_zones[i];
3064 if (!managed_zone(zone))
3067 if (!zone_reclaimable_pages(zone))
3070 pfmemalloc_reserve += min_wmark_pages(zone);
3071 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3074 /* If there are no reserves (unexpected config) then do not throttle */
3075 if (!pfmemalloc_reserve)
3078 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3080 /* kswapd must be awake if processes are being throttled */
3081 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3082 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3083 (enum zone_type)ZONE_NORMAL);
3084 wake_up_interruptible(&pgdat->kswapd_wait);
3091 * Throttle direct reclaimers if backing storage is backed by the network
3092 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3093 * depleted. kswapd will continue to make progress and wake the processes
3094 * when the low watermark is reached.
3096 * Returns true if a fatal signal was delivered during throttling. If this
3097 * happens, the page allocator should not consider triggering the OOM killer.
3099 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3100 nodemask_t *nodemask)
3104 pg_data_t *pgdat = NULL;
3107 * Kernel threads should not be throttled as they may be indirectly
3108 * responsible for cleaning pages necessary for reclaim to make forward
3109 * progress. kjournald for example may enter direct reclaim while
3110 * committing a transaction where throttling it could forcing other
3111 * processes to block on log_wait_commit().
3113 if (current->flags & PF_KTHREAD)
3117 * If a fatal signal is pending, this process should not throttle.
3118 * It should return quickly so it can exit and free its memory
3120 if (fatal_signal_pending(current))
3124 * Check if the pfmemalloc reserves are ok by finding the first node
3125 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3126 * GFP_KERNEL will be required for allocating network buffers when
3127 * swapping over the network so ZONE_HIGHMEM is unusable.
3129 * Throttling is based on the first usable node and throttled processes
3130 * wait on a queue until kswapd makes progress and wakes them. There
3131 * is an affinity then between processes waking up and where reclaim
3132 * progress has been made assuming the process wakes on the same node.
3133 * More importantly, processes running on remote nodes will not compete
3134 * for remote pfmemalloc reserves and processes on different nodes
3135 * should make reasonable progress.
3137 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3138 gfp_zone(gfp_mask), nodemask) {
3139 if (zone_idx(zone) > ZONE_NORMAL)
3142 /* Throttle based on the first usable node */
3143 pgdat = zone->zone_pgdat;
3144 if (allow_direct_reclaim(pgdat))
3149 /* If no zone was usable by the allocation flags then do not throttle */
3153 /* Account for the throttling */
3154 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3157 * If the caller cannot enter the filesystem, it's possible that it
3158 * is due to the caller holding an FS lock or performing a journal
3159 * transaction in the case of a filesystem like ext[3|4]. In this case,
3160 * it is not safe to block on pfmemalloc_wait as kswapd could be
3161 * blocked waiting on the same lock. Instead, throttle for up to a
3162 * second before continuing.
3164 if (!(gfp_mask & __GFP_FS)) {
3165 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3166 allow_direct_reclaim(pgdat), HZ);
3171 /* Throttle until kswapd wakes the process */
3172 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3173 allow_direct_reclaim(pgdat));
3176 if (fatal_signal_pending(current))
3183 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3184 gfp_t gfp_mask, nodemask_t *nodemask)
3186 unsigned long nr_reclaimed;
3187 struct scan_control sc = {
3188 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3189 .gfp_mask = current_gfp_context(gfp_mask),
3190 .reclaim_idx = gfp_zone(gfp_mask),
3192 .nodemask = nodemask,
3193 .priority = DEF_PRIORITY,
3194 .may_writepage = !laptop_mode,
3197 .may_shrinkslab = 1,
3201 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3202 * Confirm they are large enough for max values.
3204 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3205 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3206 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3209 * Do not enter reclaim if fatal signal was delivered while throttled.
3210 * 1 is returned so that the page allocator does not OOM kill at this
3213 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3216 set_task_reclaim_state(current, &sc.reclaim_state);
3217 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3219 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3221 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3222 set_task_reclaim_state(current, NULL);
3224 return nr_reclaimed;
3229 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3230 gfp_t gfp_mask, bool noswap,
3232 unsigned long *nr_scanned)
3234 struct scan_control sc = {
3235 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3236 .target_mem_cgroup = memcg,
3237 .may_writepage = !laptop_mode,
3239 .reclaim_idx = MAX_NR_ZONES - 1,
3240 .may_swap = !noswap,
3241 .may_shrinkslab = 1,
3243 unsigned long lru_pages;
3245 set_task_reclaim_state(current, &sc.reclaim_state);
3246 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3247 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3249 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3253 * NOTE: Although we can get the priority field, using it
3254 * here is not a good idea, since it limits the pages we can scan.
3255 * if we don't reclaim here, the shrink_node from balance_pgdat
3256 * will pick up pages from other mem cgroup's as well. We hack
3257 * the priority and make it zero.
3259 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3261 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3263 set_task_reclaim_state(current, NULL);
3264 *nr_scanned = sc.nr_scanned;
3266 return sc.nr_reclaimed;
3269 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3270 unsigned long nr_pages,
3274 struct zonelist *zonelist;
3275 unsigned long nr_reclaimed;
3276 unsigned long pflags;
3278 unsigned int noreclaim_flag;
3279 struct scan_control sc = {
3280 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3281 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3282 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3283 .reclaim_idx = MAX_NR_ZONES - 1,
3284 .target_mem_cgroup = memcg,
3285 .priority = DEF_PRIORITY,
3286 .may_writepage = !laptop_mode,
3288 .may_swap = may_swap,
3289 .may_shrinkslab = 1,
3292 set_task_reclaim_state(current, &sc.reclaim_state);
3294 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3295 * take care of from where we get pages. So the node where we start the
3296 * scan does not need to be the current node.
3298 nid = mem_cgroup_select_victim_node(memcg);
3300 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3302 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3304 psi_memstall_enter(&pflags);
3305 noreclaim_flag = memalloc_noreclaim_save();
3307 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3309 memalloc_noreclaim_restore(noreclaim_flag);
3310 psi_memstall_leave(&pflags);
3312 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3313 set_task_reclaim_state(current, NULL);
3315 return nr_reclaimed;
3319 static void age_active_anon(struct pglist_data *pgdat,
3320 struct scan_control *sc)
3322 struct mem_cgroup *memcg;
3324 if (!total_swap_pages)
3327 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3329 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3331 if (inactive_list_is_low(lruvec, false, sc, true))
3332 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3333 sc, LRU_ACTIVE_ANON);
3335 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3339 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3345 * Check for watermark boosts top-down as the higher zones
3346 * are more likely to be boosted. Both watermarks and boosts
3347 * should not be checked at the time time as reclaim would
3348 * start prematurely when there is no boosting and a lower
3351 for (i = classzone_idx; i >= 0; i--) {
3352 zone = pgdat->node_zones + i;
3353 if (!managed_zone(zone))
3356 if (zone->watermark_boost)
3364 * Returns true if there is an eligible zone balanced for the request order
3367 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3370 unsigned long mark = -1;
3374 * Check watermarks bottom-up as lower zones are more likely to
3377 for (i = 0; i <= classzone_idx; i++) {
3378 zone = pgdat->node_zones + i;
3380 if (!managed_zone(zone))
3383 mark = high_wmark_pages(zone);
3384 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3389 * If a node has no populated zone within classzone_idx, it does not
3390 * need balancing by definition. This can happen if a zone-restricted
3391 * allocation tries to wake a remote kswapd.
3399 /* Clear pgdat state for congested, dirty or under writeback. */
3400 static void clear_pgdat_congested(pg_data_t *pgdat)
3402 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3403 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3404 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3408 * Prepare kswapd for sleeping. This verifies that there are no processes
3409 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3411 * Returns true if kswapd is ready to sleep
3413 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3416 * The throttled processes are normally woken up in balance_pgdat() as
3417 * soon as allow_direct_reclaim() is true. But there is a potential
3418 * race between when kswapd checks the watermarks and a process gets
3419 * throttled. There is also a potential race if processes get
3420 * throttled, kswapd wakes, a large process exits thereby balancing the
3421 * zones, which causes kswapd to exit balance_pgdat() before reaching
3422 * the wake up checks. If kswapd is going to sleep, no process should
3423 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3424 * the wake up is premature, processes will wake kswapd and get
3425 * throttled again. The difference from wake ups in balance_pgdat() is
3426 * that here we are under prepare_to_wait().
3428 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3429 wake_up_all(&pgdat->pfmemalloc_wait);
3431 /* Hopeless node, leave it to direct reclaim */
3432 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3435 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3436 clear_pgdat_congested(pgdat);
3444 * kswapd shrinks a node of pages that are at or below the highest usable
3445 * zone that is currently unbalanced.
3447 * Returns true if kswapd scanned at least the requested number of pages to
3448 * reclaim or if the lack of progress was due to pages under writeback.
3449 * This is used to determine if the scanning priority needs to be raised.
3451 static bool kswapd_shrink_node(pg_data_t *pgdat,
3452 struct scan_control *sc)
3457 /* Reclaim a number of pages proportional to the number of zones */
3458 sc->nr_to_reclaim = 0;
3459 for (z = 0; z <= sc->reclaim_idx; z++) {
3460 zone = pgdat->node_zones + z;
3461 if (!managed_zone(zone))
3464 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3468 * Historically care was taken to put equal pressure on all zones but
3469 * now pressure is applied based on node LRU order.
3471 shrink_node(pgdat, sc);
3474 * Fragmentation may mean that the system cannot be rebalanced for
3475 * high-order allocations. If twice the allocation size has been
3476 * reclaimed then recheck watermarks only at order-0 to prevent
3477 * excessive reclaim. Assume that a process requested a high-order
3478 * can direct reclaim/compact.
3480 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3483 return sc->nr_scanned >= sc->nr_to_reclaim;
3487 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3488 * that are eligible for use by the caller until at least one zone is
3491 * Returns the order kswapd finished reclaiming at.
3493 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3494 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3495 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3496 * or lower is eligible for reclaim until at least one usable zone is
3499 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3502 unsigned long nr_soft_reclaimed;
3503 unsigned long nr_soft_scanned;
3504 unsigned long pflags;
3505 unsigned long nr_boost_reclaim;
3506 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3509 struct scan_control sc = {
3510 .gfp_mask = GFP_KERNEL,
3515 set_task_reclaim_state(current, &sc.reclaim_state);
3516 psi_memstall_enter(&pflags);
3517 __fs_reclaim_acquire();
3519 count_vm_event(PAGEOUTRUN);
3522 * Account for the reclaim boost. Note that the zone boost is left in
3523 * place so that parallel allocations that are near the watermark will
3524 * stall or direct reclaim until kswapd is finished.
3526 nr_boost_reclaim = 0;
3527 for (i = 0; i <= classzone_idx; i++) {
3528 zone = pgdat->node_zones + i;
3529 if (!managed_zone(zone))
3532 nr_boost_reclaim += zone->watermark_boost;
3533 zone_boosts[i] = zone->watermark_boost;
3535 boosted = nr_boost_reclaim;
3538 sc.priority = DEF_PRIORITY;
3540 unsigned long nr_reclaimed = sc.nr_reclaimed;
3541 bool raise_priority = true;
3545 sc.reclaim_idx = classzone_idx;
3548 * If the number of buffer_heads exceeds the maximum allowed
3549 * then consider reclaiming from all zones. This has a dual
3550 * purpose -- on 64-bit systems it is expected that
3551 * buffer_heads are stripped during active rotation. On 32-bit
3552 * systems, highmem pages can pin lowmem memory and shrinking
3553 * buffers can relieve lowmem pressure. Reclaim may still not
3554 * go ahead if all eligible zones for the original allocation
3555 * request are balanced to avoid excessive reclaim from kswapd.
3557 if (buffer_heads_over_limit) {
3558 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3559 zone = pgdat->node_zones + i;
3560 if (!managed_zone(zone))
3569 * If the pgdat is imbalanced then ignore boosting and preserve
3570 * the watermarks for a later time and restart. Note that the
3571 * zone watermarks will be still reset at the end of balancing
3572 * on the grounds that the normal reclaim should be enough to
3573 * re-evaluate if boosting is required when kswapd next wakes.
3575 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3576 if (!balanced && nr_boost_reclaim) {
3577 nr_boost_reclaim = 0;
3582 * If boosting is not active then only reclaim if there are no
3583 * eligible zones. Note that sc.reclaim_idx is not used as
3584 * buffer_heads_over_limit may have adjusted it.
3586 if (!nr_boost_reclaim && balanced)
3589 /* Limit the priority of boosting to avoid reclaim writeback */
3590 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3591 raise_priority = false;
3594 * Do not writeback or swap pages for boosted reclaim. The
3595 * intent is to relieve pressure not issue sub-optimal IO
3596 * from reclaim context. If no pages are reclaimed, the
3597 * reclaim will be aborted.
3599 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3600 sc.may_swap = !nr_boost_reclaim;
3601 sc.may_shrinkslab = !nr_boost_reclaim;
3604 * Do some background aging of the anon list, to give
3605 * pages a chance to be referenced before reclaiming. All
3606 * pages are rotated regardless of classzone as this is
3607 * about consistent aging.
3609 age_active_anon(pgdat, &sc);
3612 * If we're getting trouble reclaiming, start doing writepage
3613 * even in laptop mode.
3615 if (sc.priority < DEF_PRIORITY - 2)
3616 sc.may_writepage = 1;
3618 /* Call soft limit reclaim before calling shrink_node. */
3620 nr_soft_scanned = 0;
3621 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3622 sc.gfp_mask, &nr_soft_scanned);
3623 sc.nr_reclaimed += nr_soft_reclaimed;
3626 * There should be no need to raise the scanning priority if
3627 * enough pages are already being scanned that that high
3628 * watermark would be met at 100% efficiency.
3630 if (kswapd_shrink_node(pgdat, &sc))
3631 raise_priority = false;
3634 * If the low watermark is met there is no need for processes
3635 * to be throttled on pfmemalloc_wait as they should not be
3636 * able to safely make forward progress. Wake them
3638 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3639 allow_direct_reclaim(pgdat))
3640 wake_up_all(&pgdat->pfmemalloc_wait);
3642 /* Check if kswapd should be suspending */
3643 __fs_reclaim_release();
3644 ret = try_to_freeze();
3645 __fs_reclaim_acquire();
3646 if (ret || kthread_should_stop())
3650 * Raise priority if scanning rate is too low or there was no
3651 * progress in reclaiming pages
3653 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3654 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3657 * If reclaim made no progress for a boost, stop reclaim as
3658 * IO cannot be queued and it could be an infinite loop in
3659 * extreme circumstances.
3661 if (nr_boost_reclaim && !nr_reclaimed)
3664 if (raise_priority || !nr_reclaimed)
3666 } while (sc.priority >= 1);
3668 if (!sc.nr_reclaimed)
3669 pgdat->kswapd_failures++;
3672 /* If reclaim was boosted, account for the reclaim done in this pass */
3674 unsigned long flags;
3676 for (i = 0; i <= classzone_idx; i++) {
3677 if (!zone_boosts[i])
3680 /* Increments are under the zone lock */
3681 zone = pgdat->node_zones + i;
3682 spin_lock_irqsave(&zone->lock, flags);
3683 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3684 spin_unlock_irqrestore(&zone->lock, flags);
3688 * As there is now likely space, wakeup kcompact to defragment
3691 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3694 snapshot_refaults(NULL, pgdat);
3695 __fs_reclaim_release();
3696 psi_memstall_leave(&pflags);
3697 set_task_reclaim_state(current, NULL);
3700 * Return the order kswapd stopped reclaiming at as
3701 * prepare_kswapd_sleep() takes it into account. If another caller
3702 * entered the allocator slow path while kswapd was awake, order will
3703 * remain at the higher level.
3709 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3710 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3711 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3712 * after previous reclaim attempt (node is still unbalanced). In that case
3713 * return the zone index of the previous kswapd reclaim cycle.
3715 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3716 enum zone_type prev_classzone_idx)
3718 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3719 return prev_classzone_idx;
3720 return pgdat->kswapd_classzone_idx;
3723 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3724 unsigned int classzone_idx)
3729 if (freezing(current) || kthread_should_stop())
3732 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3735 * Try to sleep for a short interval. Note that kcompactd will only be
3736 * woken if it is possible to sleep for a short interval. This is
3737 * deliberate on the assumption that if reclaim cannot keep an
3738 * eligible zone balanced that it's also unlikely that compaction will
3741 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3743 * Compaction records what page blocks it recently failed to
3744 * isolate pages from and skips them in the future scanning.
3745 * When kswapd is going to sleep, it is reasonable to assume
3746 * that pages and compaction may succeed so reset the cache.
3748 reset_isolation_suitable(pgdat);
3751 * We have freed the memory, now we should compact it to make
3752 * allocation of the requested order possible.
3754 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3756 remaining = schedule_timeout(HZ/10);
3759 * If woken prematurely then reset kswapd_classzone_idx and
3760 * order. The values will either be from a wakeup request or
3761 * the previous request that slept prematurely.
3764 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3765 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3768 finish_wait(&pgdat->kswapd_wait, &wait);
3769 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3773 * After a short sleep, check if it was a premature sleep. If not, then
3774 * go fully to sleep until explicitly woken up.
3777 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3778 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3781 * vmstat counters are not perfectly accurate and the estimated
3782 * value for counters such as NR_FREE_PAGES can deviate from the
3783 * true value by nr_online_cpus * threshold. To avoid the zone
3784 * watermarks being breached while under pressure, we reduce the
3785 * per-cpu vmstat threshold while kswapd is awake and restore
3786 * them before going back to sleep.
3788 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3790 if (!kthread_should_stop())
3793 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3796 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3798 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3800 finish_wait(&pgdat->kswapd_wait, &wait);
3804 * The background pageout daemon, started as a kernel thread
3805 * from the init process.
3807 * This basically trickles out pages so that we have _some_
3808 * free memory available even if there is no other activity
3809 * that frees anything up. This is needed for things like routing
3810 * etc, where we otherwise might have all activity going on in
3811 * asynchronous contexts that cannot page things out.
3813 * If there are applications that are active memory-allocators
3814 * (most normal use), this basically shouldn't matter.
3816 static int kswapd(void *p)
3818 unsigned int alloc_order, reclaim_order;
3819 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3820 pg_data_t *pgdat = (pg_data_t*)p;
3821 struct task_struct *tsk = current;
3822 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3824 if (!cpumask_empty(cpumask))
3825 set_cpus_allowed_ptr(tsk, cpumask);
3828 * Tell the memory management that we're a "memory allocator",
3829 * and that if we need more memory we should get access to it
3830 * regardless (see "__alloc_pages()"). "kswapd" should
3831 * never get caught in the normal page freeing logic.
3833 * (Kswapd normally doesn't need memory anyway, but sometimes
3834 * you need a small amount of memory in order to be able to
3835 * page out something else, and this flag essentially protects
3836 * us from recursively trying to free more memory as we're
3837 * trying to free the first piece of memory in the first place).
3839 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3842 pgdat->kswapd_order = 0;
3843 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3847 alloc_order = reclaim_order = pgdat->kswapd_order;
3848 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3851 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3854 /* Read the new order and classzone_idx */
3855 alloc_order = reclaim_order = pgdat->kswapd_order;
3856 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3857 pgdat->kswapd_order = 0;
3858 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3860 ret = try_to_freeze();
3861 if (kthread_should_stop())
3865 * We can speed up thawing tasks if we don't call balance_pgdat
3866 * after returning from the refrigerator
3872 * Reclaim begins at the requested order but if a high-order
3873 * reclaim fails then kswapd falls back to reclaiming for
3874 * order-0. If that happens, kswapd will consider sleeping
3875 * for the order it finished reclaiming at (reclaim_order)
3876 * but kcompactd is woken to compact for the original
3877 * request (alloc_order).
3879 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3881 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3882 if (reclaim_order < alloc_order)
3883 goto kswapd_try_sleep;
3886 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3892 * A zone is low on free memory or too fragmented for high-order memory. If
3893 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3894 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3895 * has failed or is not needed, still wake up kcompactd if only compaction is
3898 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3899 enum zone_type classzone_idx)
3903 if (!managed_zone(zone))
3906 if (!cpuset_zone_allowed(zone, gfp_flags))
3908 pgdat = zone->zone_pgdat;
3910 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3911 pgdat->kswapd_classzone_idx = classzone_idx;
3913 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3915 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3916 if (!waitqueue_active(&pgdat->kswapd_wait))
3919 /* Hopeless node, leave it to direct reclaim if possible */
3920 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3921 (pgdat_balanced(pgdat, order, classzone_idx) &&
3922 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3924 * There may be plenty of free memory available, but it's too
3925 * fragmented for high-order allocations. Wake up kcompactd
3926 * and rely on compaction_suitable() to determine if it's
3927 * needed. If it fails, it will defer subsequent attempts to
3928 * ratelimit its work.
3930 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3931 wakeup_kcompactd(pgdat, order, classzone_idx);
3935 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3937 wake_up_interruptible(&pgdat->kswapd_wait);
3940 #ifdef CONFIG_HIBERNATION
3942 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3945 * Rather than trying to age LRUs the aim is to preserve the overall
3946 * LRU order by reclaiming preferentially
3947 * inactive > active > active referenced > active mapped
3949 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3951 struct scan_control sc = {
3952 .nr_to_reclaim = nr_to_reclaim,
3953 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3954 .reclaim_idx = MAX_NR_ZONES - 1,
3955 .priority = DEF_PRIORITY,
3959 .hibernation_mode = 1,
3961 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3962 unsigned long nr_reclaimed;
3963 unsigned int noreclaim_flag;
3965 fs_reclaim_acquire(sc.gfp_mask);
3966 noreclaim_flag = memalloc_noreclaim_save();
3967 set_task_reclaim_state(current, &sc.reclaim_state);
3969 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3971 set_task_reclaim_state(current, NULL);
3972 memalloc_noreclaim_restore(noreclaim_flag);
3973 fs_reclaim_release(sc.gfp_mask);
3975 return nr_reclaimed;
3977 #endif /* CONFIG_HIBERNATION */
3979 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3980 not required for correctness. So if the last cpu in a node goes
3981 away, we get changed to run anywhere: as the first one comes back,
3982 restore their cpu bindings. */
3983 static int kswapd_cpu_online(unsigned int cpu)
3987 for_each_node_state(nid, N_MEMORY) {
3988 pg_data_t *pgdat = NODE_DATA(nid);
3989 const struct cpumask *mask;
3991 mask = cpumask_of_node(pgdat->node_id);
3993 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3994 /* One of our CPUs online: restore mask */
3995 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4001 * This kswapd start function will be called by init and node-hot-add.
4002 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4004 int kswapd_run(int nid)
4006 pg_data_t *pgdat = NODE_DATA(nid);
4012 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4013 if (IS_ERR(pgdat->kswapd)) {
4014 /* failure at boot is fatal */
4015 BUG_ON(system_state < SYSTEM_RUNNING);
4016 pr_err("Failed to start kswapd on node %d\n", nid);
4017 ret = PTR_ERR(pgdat->kswapd);
4018 pgdat->kswapd = NULL;
4024 * Called by memory hotplug when all memory in a node is offlined. Caller must
4025 * hold mem_hotplug_begin/end().
4027 void kswapd_stop(int nid)
4029 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4032 kthread_stop(kswapd);
4033 NODE_DATA(nid)->kswapd = NULL;
4037 static int __init kswapd_init(void)
4042 for_each_node_state(nid, N_MEMORY)
4044 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4045 "mm/vmscan:online", kswapd_cpu_online,
4051 module_init(kswapd_init)
4057 * If non-zero call node_reclaim when the number of free pages falls below
4060 int node_reclaim_mode __read_mostly;
4062 #define RECLAIM_OFF 0
4063 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4064 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4065 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4068 * Priority for NODE_RECLAIM. This determines the fraction of pages
4069 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4072 #define NODE_RECLAIM_PRIORITY 4
4075 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4078 int sysctl_min_unmapped_ratio = 1;
4081 * If the number of slab pages in a zone grows beyond this percentage then
4082 * slab reclaim needs to occur.
4084 int sysctl_min_slab_ratio = 5;
4086 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4088 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4089 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4090 node_page_state(pgdat, NR_ACTIVE_FILE);
4093 * It's possible for there to be more file mapped pages than
4094 * accounted for by the pages on the file LRU lists because
4095 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4097 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4100 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4101 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4103 unsigned long nr_pagecache_reclaimable;
4104 unsigned long delta = 0;
4107 * If RECLAIM_UNMAP is set, then all file pages are considered
4108 * potentially reclaimable. Otherwise, we have to worry about
4109 * pages like swapcache and node_unmapped_file_pages() provides
4112 if (node_reclaim_mode & RECLAIM_UNMAP)
4113 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4115 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4117 /* If we can't clean pages, remove dirty pages from consideration */
4118 if (!(node_reclaim_mode & RECLAIM_WRITE))
4119 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4121 /* Watch for any possible underflows due to delta */
4122 if (unlikely(delta > nr_pagecache_reclaimable))
4123 delta = nr_pagecache_reclaimable;
4125 return nr_pagecache_reclaimable - delta;
4129 * Try to free up some pages from this node through reclaim.
4131 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4133 /* Minimum pages needed in order to stay on node */
4134 const unsigned long nr_pages = 1 << order;
4135 struct task_struct *p = current;
4136 unsigned int noreclaim_flag;
4137 struct scan_control sc = {
4138 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4139 .gfp_mask = current_gfp_context(gfp_mask),
4141 .priority = NODE_RECLAIM_PRIORITY,
4142 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4143 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4145 .reclaim_idx = gfp_zone(gfp_mask),
4148 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4152 fs_reclaim_acquire(sc.gfp_mask);
4154 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4155 * and we also need to be able to write out pages for RECLAIM_WRITE
4156 * and RECLAIM_UNMAP.
4158 noreclaim_flag = memalloc_noreclaim_save();
4159 p->flags |= PF_SWAPWRITE;
4160 set_task_reclaim_state(p, &sc.reclaim_state);
4162 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4164 * Free memory by calling shrink node with increasing
4165 * priorities until we have enough memory freed.
4168 shrink_node(pgdat, &sc);
4169 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4172 set_task_reclaim_state(p, NULL);
4173 current->flags &= ~PF_SWAPWRITE;
4174 memalloc_noreclaim_restore(noreclaim_flag);
4175 fs_reclaim_release(sc.gfp_mask);
4177 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4179 return sc.nr_reclaimed >= nr_pages;
4182 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4187 * Node reclaim reclaims unmapped file backed pages and
4188 * slab pages if we are over the defined limits.
4190 * A small portion of unmapped file backed pages is needed for
4191 * file I/O otherwise pages read by file I/O will be immediately
4192 * thrown out if the node is overallocated. So we do not reclaim
4193 * if less than a specified percentage of the node is used by
4194 * unmapped file backed pages.
4196 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4197 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4198 return NODE_RECLAIM_FULL;
4201 * Do not scan if the allocation should not be delayed.
4203 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4204 return NODE_RECLAIM_NOSCAN;
4207 * Only run node reclaim on the local node or on nodes that do not
4208 * have associated processors. This will favor the local processor
4209 * over remote processors and spread off node memory allocations
4210 * as wide as possible.
4212 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4213 return NODE_RECLAIM_NOSCAN;
4215 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4216 return NODE_RECLAIM_NOSCAN;
4218 ret = __node_reclaim(pgdat, gfp_mask, order);
4219 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4222 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4229 * page_evictable - test whether a page is evictable
4230 * @page: the page to test
4232 * Test whether page is evictable--i.e., should be placed on active/inactive
4233 * lists vs unevictable list.
4235 * Reasons page might not be evictable:
4236 * (1) page's mapping marked unevictable
4237 * (2) page is part of an mlocked VMA
4240 int page_evictable(struct page *page)
4244 /* Prevent address_space of inode and swap cache from being freed */
4246 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4252 * check_move_unevictable_pages - check pages for evictability and move to
4253 * appropriate zone lru list
4254 * @pvec: pagevec with lru pages to check
4256 * Checks pages for evictability, if an evictable page is in the unevictable
4257 * lru list, moves it to the appropriate evictable lru list. This function
4258 * should be only used for lru pages.
4260 void check_move_unevictable_pages(struct pagevec *pvec)
4262 struct lruvec *lruvec;
4263 struct pglist_data *pgdat = NULL;
4268 for (i = 0; i < pvec->nr; i++) {
4269 struct page *page = pvec->pages[i];
4270 struct pglist_data *pagepgdat = page_pgdat(page);
4273 if (pagepgdat != pgdat) {
4275 spin_unlock_irq(&pgdat->lru_lock);
4277 spin_lock_irq(&pgdat->lru_lock);
4279 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4281 if (!PageLRU(page) || !PageUnevictable(page))
4284 if (page_evictable(page)) {
4285 enum lru_list lru = page_lru_base_type(page);
4287 VM_BUG_ON_PAGE(PageActive(page), page);
4288 ClearPageUnevictable(page);
4289 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4290 add_page_to_lru_list(page, lruvec, lru);
4296 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4297 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4298 spin_unlock_irq(&pgdat->lru_lock);
4301 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);