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;
136 #ifdef ARCH_HAS_PREFETCH
137 #define prefetch_prev_lru_page(_page, _base, _field) \
139 if ((_page)->lru.prev != _base) { \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
147 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
150 #ifdef ARCH_HAS_PREFETCHW
151 #define prefetchw_prev_lru_page(_page, _base, _field) \
153 if ((_page)->lru.prev != _base) { \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
161 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
165 * From 0 .. 100. Higher means more swappy.
167 int vm_swappiness = 60;
169 * The total number of pages which are beyond the high watermark within all
172 unsigned long vm_total_pages;
174 static LIST_HEAD(shrinker_list);
175 static DECLARE_RWSEM(shrinker_rwsem);
177 #ifdef CONFIG_MEMCG_KMEM
180 * We allow subsystems to populate their shrinker-related
181 * LRU lists before register_shrinker_prepared() is called
182 * for the shrinker, since we don't want to impose
183 * restrictions on their internal registration order.
184 * In this case shrink_slab_memcg() may find corresponding
185 * bit is set in the shrinkers map.
187 * This value is used by the function to detect registering
188 * shrinkers and to skip do_shrink_slab() calls for them.
190 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
192 static DEFINE_IDR(shrinker_idr);
193 static int shrinker_nr_max;
195 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
197 int id, ret = -ENOMEM;
199 down_write(&shrinker_rwsem);
200 /* This may call shrinker, so it must use down_read_trylock() */
201 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
205 if (id >= shrinker_nr_max) {
206 if (memcg_expand_shrinker_maps(id)) {
207 idr_remove(&shrinker_idr, id);
211 shrinker_nr_max = id + 1;
216 up_write(&shrinker_rwsem);
220 static void unregister_memcg_shrinker(struct shrinker *shrinker)
222 int id = shrinker->id;
226 down_write(&shrinker_rwsem);
227 idr_remove(&shrinker_idr, id);
228 up_write(&shrinker_rwsem);
230 #else /* CONFIG_MEMCG_KMEM */
231 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
236 static void unregister_memcg_shrinker(struct shrinker *shrinker)
239 #endif /* CONFIG_MEMCG_KMEM */
242 static bool global_reclaim(struct scan_control *sc)
244 return !sc->target_mem_cgroup;
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
260 static bool sane_reclaim(struct scan_control *sc)
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
266 #ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
273 static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
277 struct mem_cgroup_per_node *mn;
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
286 static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
289 struct mem_cgroup_per_node *mn;
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
296 static bool global_reclaim(struct scan_control *sc)
301 static bool sane_reclaim(struct scan_control *sc)
306 static inline void set_memcg_congestion(struct pglist_data *pgdat,
307 struct mem_cgroup *memcg, bool congested)
311 static inline bool memcg_congested(struct pglist_data *pgdat,
312 struct mem_cgroup *memcg)
320 * This misses isolated pages which are not accounted for to save counters.
321 * As the data only determines if reclaim or compaction continues, it is
322 * not expected that isolated pages will be a dominating factor.
324 unsigned long zone_reclaimable_pages(struct zone *zone)
328 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
329 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
330 if (get_nr_swap_pages() > 0)
331 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
332 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
338 * lruvec_lru_size - Returns the number of pages on the given LRU list.
339 * @lruvec: lru vector
341 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
343 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
345 unsigned long lru_size;
348 if (!mem_cgroup_disabled())
349 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
351 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
353 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
354 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
357 if (!managed_zone(zone))
360 if (!mem_cgroup_disabled())
361 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
363 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
364 NR_ZONE_LRU_BASE + lru);
365 lru_size -= min(size, lru_size);
373 * Add a shrinker callback to be called from the vm.
375 int prealloc_shrinker(struct shrinker *shrinker)
377 size_t size = sizeof(*shrinker->nr_deferred);
379 if (shrinker->flags & SHRINKER_NUMA_AWARE)
382 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
383 if (!shrinker->nr_deferred)
386 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
387 if (prealloc_memcg_shrinker(shrinker))
394 kfree(shrinker->nr_deferred);
395 shrinker->nr_deferred = NULL;
399 void free_prealloced_shrinker(struct shrinker *shrinker)
401 if (!shrinker->nr_deferred)
404 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
405 unregister_memcg_shrinker(shrinker);
407 kfree(shrinker->nr_deferred);
408 shrinker->nr_deferred = NULL;
411 void register_shrinker_prepared(struct shrinker *shrinker)
413 down_write(&shrinker_rwsem);
414 list_add_tail(&shrinker->list, &shrinker_list);
415 #ifdef CONFIG_MEMCG_KMEM
416 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
417 idr_replace(&shrinker_idr, shrinker, shrinker->id);
419 up_write(&shrinker_rwsem);
422 int register_shrinker(struct shrinker *shrinker)
424 int err = prealloc_shrinker(shrinker);
428 register_shrinker_prepared(shrinker);
431 EXPORT_SYMBOL(register_shrinker);
436 void unregister_shrinker(struct shrinker *shrinker)
438 if (!shrinker->nr_deferred)
440 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
441 unregister_memcg_shrinker(shrinker);
442 down_write(&shrinker_rwsem);
443 list_del(&shrinker->list);
444 up_write(&shrinker_rwsem);
445 kfree(shrinker->nr_deferred);
446 shrinker->nr_deferred = NULL;
448 EXPORT_SYMBOL(unregister_shrinker);
450 #define SHRINK_BATCH 128
452 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
453 struct shrinker *shrinker, int priority)
455 unsigned long freed = 0;
456 unsigned long long delta;
461 int nid = shrinkctl->nid;
462 long batch_size = shrinker->batch ? shrinker->batch
464 long scanned = 0, next_deferred;
466 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
469 freeable = shrinker->count_objects(shrinker, shrinkctl);
470 if (freeable == 0 || freeable == SHRINK_EMPTY)
474 * copy the current shrinker scan count into a local variable
475 * and zero it so that other concurrent shrinker invocations
476 * don't also do this scanning work.
478 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
481 if (shrinker->seeks) {
482 delta = freeable >> priority;
484 do_div(delta, shrinker->seeks);
487 * These objects don't require any IO to create. Trim
488 * them aggressively under memory pressure to keep
489 * them from causing refetches in the IO caches.
491 delta = freeable / 2;
495 * Make sure we apply some minimal pressure on default priority
496 * even on small cgroups. Stale objects are not only consuming memory
497 * by themselves, but can also hold a reference to a dying cgroup,
498 * preventing it from being reclaimed. A dying cgroup with all
499 * corresponding structures like per-cpu stats and kmem caches
500 * can be really big, so it may lead to a significant waste of memory.
502 delta = max_t(unsigned long long, delta, min(freeable, batch_size));
505 if (total_scan < 0) {
506 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
507 shrinker->scan_objects, total_scan);
508 total_scan = freeable;
511 next_deferred = total_scan;
514 * We need to avoid excessive windup on filesystem shrinkers
515 * due to large numbers of GFP_NOFS allocations causing the
516 * shrinkers to return -1 all the time. This results in a large
517 * nr being built up so when a shrink that can do some work
518 * comes along it empties the entire cache due to nr >>>
519 * freeable. This is bad for sustaining a working set in
522 * Hence only allow the shrinker to scan the entire cache when
523 * a large delta change is calculated directly.
525 if (delta < freeable / 4)
526 total_scan = min(total_scan, freeable / 2);
529 * Avoid risking looping forever due to too large nr value:
530 * never try to free more than twice the estimate number of
533 if (total_scan > freeable * 2)
534 total_scan = freeable * 2;
536 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
537 freeable, delta, total_scan, priority);
540 * Normally, we should not scan less than batch_size objects in one
541 * pass to avoid too frequent shrinker calls, but if the slab has less
542 * than batch_size objects in total and we are really tight on memory,
543 * we will try to reclaim all available objects, otherwise we can end
544 * up failing allocations although there are plenty of reclaimable
545 * objects spread over several slabs with usage less than the
548 * We detect the "tight on memory" situations by looking at the total
549 * number of objects we want to scan (total_scan). If it is greater
550 * than the total number of objects on slab (freeable), we must be
551 * scanning at high prio and therefore should try to reclaim as much as
554 while (total_scan >= batch_size ||
555 total_scan >= freeable) {
557 unsigned long nr_to_scan = min(batch_size, total_scan);
559 shrinkctl->nr_to_scan = nr_to_scan;
560 shrinkctl->nr_scanned = nr_to_scan;
561 ret = shrinker->scan_objects(shrinker, shrinkctl);
562 if (ret == SHRINK_STOP)
566 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
567 total_scan -= shrinkctl->nr_scanned;
568 scanned += shrinkctl->nr_scanned;
573 if (next_deferred >= scanned)
574 next_deferred -= scanned;
578 * move the unused scan count back into the shrinker in a
579 * manner that handles concurrent updates. If we exhausted the
580 * scan, there is no need to do an update.
582 if (next_deferred > 0)
583 new_nr = atomic_long_add_return(next_deferred,
584 &shrinker->nr_deferred[nid]);
586 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
588 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
592 #ifdef CONFIG_MEMCG_KMEM
593 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
594 struct mem_cgroup *memcg, int priority)
596 struct memcg_shrinker_map *map;
597 unsigned long ret, freed = 0;
600 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
603 if (!down_read_trylock(&shrinker_rwsem))
606 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
611 for_each_set_bit(i, map->map, shrinker_nr_max) {
612 struct shrink_control sc = {
613 .gfp_mask = gfp_mask,
617 struct shrinker *shrinker;
619 shrinker = idr_find(&shrinker_idr, i);
620 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
622 clear_bit(i, map->map);
626 ret = do_shrink_slab(&sc, shrinker, priority);
627 if (ret == SHRINK_EMPTY) {
628 clear_bit(i, map->map);
630 * After the shrinker reported that it had no objects to
631 * free, but before we cleared the corresponding bit in
632 * the memcg shrinker map, a new object might have been
633 * added. To make sure, we have the bit set in this
634 * case, we invoke the shrinker one more time and reset
635 * the bit if it reports that it is not empty anymore.
636 * The memory barrier here pairs with the barrier in
637 * memcg_set_shrinker_bit():
639 * list_lru_add() shrink_slab_memcg()
640 * list_add_tail() clear_bit()
642 * set_bit() do_shrink_slab()
644 smp_mb__after_atomic();
645 ret = do_shrink_slab(&sc, shrinker, priority);
646 if (ret == SHRINK_EMPTY)
649 memcg_set_shrinker_bit(memcg, nid, i);
653 if (rwsem_is_contended(&shrinker_rwsem)) {
659 up_read(&shrinker_rwsem);
662 #else /* CONFIG_MEMCG_KMEM */
663 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
664 struct mem_cgroup *memcg, int priority)
668 #endif /* CONFIG_MEMCG_KMEM */
671 * shrink_slab - shrink slab caches
672 * @gfp_mask: allocation context
673 * @nid: node whose slab caches to target
674 * @memcg: memory cgroup whose slab caches to target
675 * @priority: the reclaim priority
677 * Call the shrink functions to age shrinkable caches.
679 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
680 * unaware shrinkers will receive a node id of 0 instead.
682 * @memcg specifies the memory cgroup to target. Unaware shrinkers
683 * are called only if it is the root cgroup.
685 * @priority is sc->priority, we take the number of objects and >> by priority
686 * in order to get the scan target.
688 * Returns the number of reclaimed slab objects.
690 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
691 struct mem_cgroup *memcg,
694 unsigned long ret, freed = 0;
695 struct shrinker *shrinker;
697 if (!mem_cgroup_is_root(memcg))
698 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
700 if (!down_read_trylock(&shrinker_rwsem))
703 list_for_each_entry(shrinker, &shrinker_list, list) {
704 struct shrink_control sc = {
705 .gfp_mask = gfp_mask,
710 ret = do_shrink_slab(&sc, shrinker, priority);
711 if (ret == SHRINK_EMPTY)
715 * Bail out if someone want to register a new shrinker to
716 * prevent the regsitration from being stalled for long periods
717 * by parallel ongoing shrinking.
719 if (rwsem_is_contended(&shrinker_rwsem)) {
725 up_read(&shrinker_rwsem);
731 void drop_slab_node(int nid)
736 struct mem_cgroup *memcg = NULL;
739 memcg = mem_cgroup_iter(NULL, NULL, NULL);
741 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
742 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
743 } while (freed > 10);
750 for_each_online_node(nid)
754 static inline int is_page_cache_freeable(struct page *page)
757 * A freeable page cache page is referenced only by the caller
758 * that isolated the page, the page cache and optional buffer
759 * heads at page->private.
761 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
763 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
766 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
768 if (current->flags & PF_SWAPWRITE)
770 if (!inode_write_congested(inode))
772 if (inode_to_bdi(inode) == current->backing_dev_info)
778 * We detected a synchronous write error writing a page out. Probably
779 * -ENOSPC. We need to propagate that into the address_space for a subsequent
780 * fsync(), msync() or close().
782 * The tricky part is that after writepage we cannot touch the mapping: nothing
783 * prevents it from being freed up. But we have a ref on the page and once
784 * that page is locked, the mapping is pinned.
786 * We're allowed to run sleeping lock_page() here because we know the caller has
789 static void handle_write_error(struct address_space *mapping,
790 struct page *page, int error)
793 if (page_mapping(page) == mapping)
794 mapping_set_error(mapping, error);
798 /* possible outcome of pageout() */
800 /* failed to write page out, page is locked */
802 /* move page to the active list, page is locked */
804 /* page has been sent to the disk successfully, page is unlocked */
806 /* page is clean and locked */
811 * pageout is called by shrink_page_list() for each dirty page.
812 * Calls ->writepage().
814 static pageout_t pageout(struct page *page, struct address_space *mapping,
815 struct scan_control *sc)
818 * If the page is dirty, only perform writeback if that write
819 * will be non-blocking. To prevent this allocation from being
820 * stalled by pagecache activity. But note that there may be
821 * stalls if we need to run get_block(). We could test
822 * PagePrivate for that.
824 * If this process is currently in __generic_file_write_iter() against
825 * this page's queue, we can perform writeback even if that
828 * If the page is swapcache, write it back even if that would
829 * block, for some throttling. This happens by accident, because
830 * swap_backing_dev_info is bust: it doesn't reflect the
831 * congestion state of the swapdevs. Easy to fix, if needed.
833 if (!is_page_cache_freeable(page))
837 * Some data journaling orphaned pages can have
838 * page->mapping == NULL while being dirty with clean buffers.
840 if (page_has_private(page)) {
841 if (try_to_free_buffers(page)) {
842 ClearPageDirty(page);
843 pr_info("%s: orphaned page\n", __func__);
849 if (mapping->a_ops->writepage == NULL)
850 return PAGE_ACTIVATE;
851 if (!may_write_to_inode(mapping->host, sc))
854 if (clear_page_dirty_for_io(page)) {
856 struct writeback_control wbc = {
857 .sync_mode = WB_SYNC_NONE,
858 .nr_to_write = SWAP_CLUSTER_MAX,
860 .range_end = LLONG_MAX,
864 SetPageReclaim(page);
865 res = mapping->a_ops->writepage(page, &wbc);
867 handle_write_error(mapping, page, res);
868 if (res == AOP_WRITEPAGE_ACTIVATE) {
869 ClearPageReclaim(page);
870 return PAGE_ACTIVATE;
873 if (!PageWriteback(page)) {
874 /* synchronous write or broken a_ops? */
875 ClearPageReclaim(page);
877 trace_mm_vmscan_writepage(page);
878 inc_node_page_state(page, NR_VMSCAN_WRITE);
886 * Same as remove_mapping, but if the page is removed from the mapping, it
887 * gets returned with a refcount of 0.
889 static int __remove_mapping(struct address_space *mapping, struct page *page,
895 BUG_ON(!PageLocked(page));
896 BUG_ON(mapping != page_mapping(page));
898 xa_lock_irqsave(&mapping->i_pages, flags);
900 * The non racy check for a busy page.
902 * Must be careful with the order of the tests. When someone has
903 * a ref to the page, it may be possible that they dirty it then
904 * drop the reference. So if PageDirty is tested before page_count
905 * here, then the following race may occur:
907 * get_user_pages(&page);
908 * [user mapping goes away]
910 * !PageDirty(page) [good]
911 * SetPageDirty(page);
913 * !page_count(page) [good, discard it]
915 * [oops, our write_to data is lost]
917 * Reversing the order of the tests ensures such a situation cannot
918 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
919 * load is not satisfied before that of page->_refcount.
921 * Note that if SetPageDirty is always performed via set_page_dirty,
922 * and thus under the i_pages lock, then this ordering is not required.
924 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
925 refcount = 1 + HPAGE_PMD_NR;
928 if (!page_ref_freeze(page, refcount))
930 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
931 if (unlikely(PageDirty(page))) {
932 page_ref_unfreeze(page, refcount);
936 if (PageSwapCache(page)) {
937 swp_entry_t swap = { .val = page_private(page) };
938 mem_cgroup_swapout(page, swap);
939 __delete_from_swap_cache(page, swap);
940 xa_unlock_irqrestore(&mapping->i_pages, flags);
941 put_swap_page(page, swap);
943 void (*freepage)(struct page *);
946 freepage = mapping->a_ops->freepage;
948 * Remember a shadow entry for reclaimed file cache in
949 * order to detect refaults, thus thrashing, later on.
951 * But don't store shadows in an address space that is
952 * already exiting. This is not just an optizimation,
953 * inode reclaim needs to empty out the radix tree or
954 * the nodes are lost. Don't plant shadows behind its
957 * We also don't store shadows for DAX mappings because the
958 * only page cache pages found in these are zero pages
959 * covering holes, and because we don't want to mix DAX
960 * exceptional entries and shadow exceptional entries in the
961 * same address_space.
963 if (reclaimed && page_is_file_cache(page) &&
964 !mapping_exiting(mapping) && !dax_mapping(mapping))
965 shadow = workingset_eviction(mapping, page);
966 __delete_from_page_cache(page, shadow);
967 xa_unlock_irqrestore(&mapping->i_pages, flags);
969 if (freepage != NULL)
976 xa_unlock_irqrestore(&mapping->i_pages, flags);
981 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
982 * someone else has a ref on the page, abort and return 0. If it was
983 * successfully detached, return 1. Assumes the caller has a single ref on
986 int remove_mapping(struct address_space *mapping, struct page *page)
988 if (__remove_mapping(mapping, page, false)) {
990 * Unfreezing the refcount with 1 rather than 2 effectively
991 * drops the pagecache ref for us without requiring another
994 page_ref_unfreeze(page, 1);
1001 * putback_lru_page - put previously isolated page onto appropriate LRU list
1002 * @page: page to be put back to appropriate lru list
1004 * Add previously isolated @page to appropriate LRU list.
1005 * Page may still be unevictable for other reasons.
1007 * lru_lock must not be held, interrupts must be enabled.
1009 void putback_lru_page(struct page *page)
1011 lru_cache_add(page);
1012 put_page(page); /* drop ref from isolate */
1015 enum page_references {
1017 PAGEREF_RECLAIM_CLEAN,
1022 static enum page_references page_check_references(struct page *page,
1023 struct scan_control *sc)
1025 int referenced_ptes, referenced_page;
1026 unsigned long vm_flags;
1028 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1030 referenced_page = TestClearPageReferenced(page);
1033 * Mlock lost the isolation race with us. Let try_to_unmap()
1034 * move the page to the unevictable list.
1036 if (vm_flags & VM_LOCKED)
1037 return PAGEREF_RECLAIM;
1039 if (referenced_ptes) {
1040 if (PageSwapBacked(page))
1041 return PAGEREF_ACTIVATE;
1043 * All mapped pages start out with page table
1044 * references from the instantiating fault, so we need
1045 * to look twice if a mapped file page is used more
1048 * Mark it and spare it for another trip around the
1049 * inactive list. Another page table reference will
1050 * lead to its activation.
1052 * Note: the mark is set for activated pages as well
1053 * so that recently deactivated but used pages are
1054 * quickly recovered.
1056 SetPageReferenced(page);
1058 if (referenced_page || referenced_ptes > 1)
1059 return PAGEREF_ACTIVATE;
1062 * Activate file-backed executable pages after first usage.
1064 if (vm_flags & VM_EXEC)
1065 return PAGEREF_ACTIVATE;
1067 return PAGEREF_KEEP;
1070 /* Reclaim if clean, defer dirty pages to writeback */
1071 if (referenced_page && !PageSwapBacked(page))
1072 return PAGEREF_RECLAIM_CLEAN;
1074 return PAGEREF_RECLAIM;
1077 /* Check if a page is dirty or under writeback */
1078 static void page_check_dirty_writeback(struct page *page,
1079 bool *dirty, bool *writeback)
1081 struct address_space *mapping;
1084 * Anonymous pages are not handled by flushers and must be written
1085 * from reclaim context. Do not stall reclaim based on them
1087 if (!page_is_file_cache(page) ||
1088 (PageAnon(page) && !PageSwapBacked(page))) {
1094 /* By default assume that the page flags are accurate */
1095 *dirty = PageDirty(page);
1096 *writeback = PageWriteback(page);
1098 /* Verify dirty/writeback state if the filesystem supports it */
1099 if (!page_has_private(page))
1102 mapping = page_mapping(page);
1103 if (mapping && mapping->a_ops->is_dirty_writeback)
1104 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1108 * shrink_page_list() returns the number of reclaimed pages
1110 static unsigned long shrink_page_list(struct list_head *page_list,
1111 struct pglist_data *pgdat,
1112 struct scan_control *sc,
1113 enum ttu_flags ttu_flags,
1114 struct reclaim_stat *stat,
1117 LIST_HEAD(ret_pages);
1118 LIST_HEAD(free_pages);
1120 unsigned nr_unqueued_dirty = 0;
1121 unsigned nr_dirty = 0;
1122 unsigned nr_congested = 0;
1123 unsigned nr_reclaimed = 0;
1124 unsigned nr_writeback = 0;
1125 unsigned nr_immediate = 0;
1126 unsigned nr_ref_keep = 0;
1127 unsigned nr_unmap_fail = 0;
1131 while (!list_empty(page_list)) {
1132 struct address_space *mapping;
1135 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1136 bool dirty, writeback;
1140 page = lru_to_page(page_list);
1141 list_del(&page->lru);
1143 if (!trylock_page(page))
1146 VM_BUG_ON_PAGE(PageActive(page), page);
1150 if (unlikely(!page_evictable(page)))
1151 goto activate_locked;
1153 if (!sc->may_unmap && page_mapped(page))
1156 /* Double the slab pressure for mapped and swapcache pages */
1157 if ((page_mapped(page) || PageSwapCache(page)) &&
1158 !(PageAnon(page) && !PageSwapBacked(page)))
1161 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1162 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1165 * The number of dirty pages determines if a node is marked
1166 * reclaim_congested which affects wait_iff_congested. kswapd
1167 * will stall and start writing pages if the tail of the LRU
1168 * is all dirty unqueued pages.
1170 page_check_dirty_writeback(page, &dirty, &writeback);
1171 if (dirty || writeback)
1174 if (dirty && !writeback)
1175 nr_unqueued_dirty++;
1178 * Treat this page as congested if the underlying BDI is or if
1179 * pages are cycling through the LRU so quickly that the
1180 * pages marked for immediate reclaim are making it to the
1181 * end of the LRU a second time.
1183 mapping = page_mapping(page);
1184 if (((dirty || writeback) && mapping &&
1185 inode_write_congested(mapping->host)) ||
1186 (writeback && PageReclaim(page)))
1190 * If a page at the tail of the LRU is under writeback, there
1191 * are three cases to consider.
1193 * 1) If reclaim is encountering an excessive number of pages
1194 * under writeback and this page is both under writeback and
1195 * PageReclaim then it indicates that pages are being queued
1196 * for IO but are being recycled through the LRU before the
1197 * IO can complete. Waiting on the page itself risks an
1198 * indefinite stall if it is impossible to writeback the
1199 * page due to IO error or disconnected storage so instead
1200 * note that the LRU is being scanned too quickly and the
1201 * caller can stall after page list has been processed.
1203 * 2) Global or new memcg reclaim encounters a page that is
1204 * not marked for immediate reclaim, or the caller does not
1205 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1206 * not to fs). In this case mark the page for immediate
1207 * reclaim and continue scanning.
1209 * Require may_enter_fs because we would wait on fs, which
1210 * may not have submitted IO yet. And the loop driver might
1211 * enter reclaim, and deadlock if it waits on a page for
1212 * which it is needed to do the write (loop masks off
1213 * __GFP_IO|__GFP_FS for this reason); but more thought
1214 * would probably show more reasons.
1216 * 3) Legacy memcg encounters a page that is already marked
1217 * PageReclaim. memcg does not have any dirty pages
1218 * throttling so we could easily OOM just because too many
1219 * pages are in writeback and there is nothing else to
1220 * reclaim. Wait for the writeback to complete.
1222 * In cases 1) and 2) we activate the pages to get them out of
1223 * the way while we continue scanning for clean pages on the
1224 * inactive list and refilling from the active list. The
1225 * observation here is that waiting for disk writes is more
1226 * expensive than potentially causing reloads down the line.
1227 * Since they're marked for immediate reclaim, they won't put
1228 * memory pressure on the cache working set any longer than it
1229 * takes to write them to disk.
1231 if (PageWriteback(page)) {
1233 if (current_is_kswapd() &&
1234 PageReclaim(page) &&
1235 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1237 goto activate_locked;
1240 } else if (sane_reclaim(sc) ||
1241 !PageReclaim(page) || !may_enter_fs) {
1243 * This is slightly racy - end_page_writeback()
1244 * might have just cleared PageReclaim, then
1245 * setting PageReclaim here end up interpreted
1246 * as PageReadahead - but that does not matter
1247 * enough to care. What we do want is for this
1248 * page to have PageReclaim set next time memcg
1249 * reclaim reaches the tests above, so it will
1250 * then wait_on_page_writeback() to avoid OOM;
1251 * and it's also appropriate in global reclaim.
1253 SetPageReclaim(page);
1255 goto activate_locked;
1260 wait_on_page_writeback(page);
1261 /* then go back and try same page again */
1262 list_add_tail(&page->lru, page_list);
1268 references = page_check_references(page, sc);
1270 switch (references) {
1271 case PAGEREF_ACTIVATE:
1272 goto activate_locked;
1276 case PAGEREF_RECLAIM:
1277 case PAGEREF_RECLAIM_CLEAN:
1278 ; /* try to reclaim the page below */
1282 * Anonymous process memory has backing store?
1283 * Try to allocate it some swap space here.
1284 * Lazyfree page could be freed directly
1286 if (PageAnon(page) && PageSwapBacked(page)) {
1287 if (!PageSwapCache(page)) {
1288 if (!(sc->gfp_mask & __GFP_IO))
1290 if (PageTransHuge(page)) {
1291 /* cannot split THP, skip it */
1292 if (!can_split_huge_page(page, NULL))
1293 goto activate_locked;
1295 * Split pages without a PMD map right
1296 * away. Chances are some or all of the
1297 * tail pages can be freed without IO.
1299 if (!compound_mapcount(page) &&
1300 split_huge_page_to_list(page,
1302 goto activate_locked;
1304 if (!add_to_swap(page)) {
1305 if (!PageTransHuge(page))
1306 goto activate_locked;
1307 /* Fallback to swap normal pages */
1308 if (split_huge_page_to_list(page,
1310 goto activate_locked;
1311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1312 count_vm_event(THP_SWPOUT_FALLBACK);
1314 if (!add_to_swap(page))
1315 goto activate_locked;
1320 /* Adding to swap updated mapping */
1321 mapping = page_mapping(page);
1323 } else if (unlikely(PageTransHuge(page))) {
1324 /* Split file THP */
1325 if (split_huge_page_to_list(page, page_list))
1330 * The page is mapped into the page tables of one or more
1331 * processes. Try to unmap it here.
1333 if (page_mapped(page)) {
1334 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1336 if (unlikely(PageTransHuge(page)))
1337 flags |= TTU_SPLIT_HUGE_PMD;
1338 if (!try_to_unmap(page, flags)) {
1340 goto activate_locked;
1344 if (PageDirty(page)) {
1346 * Only kswapd can writeback filesystem pages
1347 * to avoid risk of stack overflow. But avoid
1348 * injecting inefficient single-page IO into
1349 * flusher writeback as much as possible: only
1350 * write pages when we've encountered many
1351 * dirty pages, and when we've already scanned
1352 * the rest of the LRU for clean pages and see
1353 * the same dirty pages again (PageReclaim).
1355 if (page_is_file_cache(page) &&
1356 (!current_is_kswapd() || !PageReclaim(page) ||
1357 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1359 * Immediately reclaim when written back.
1360 * Similar in principal to deactivate_page()
1361 * except we already have the page isolated
1362 * and know it's dirty
1364 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1365 SetPageReclaim(page);
1367 goto activate_locked;
1370 if (references == PAGEREF_RECLAIM_CLEAN)
1374 if (!sc->may_writepage)
1378 * Page is dirty. Flush the TLB if a writable entry
1379 * potentially exists to avoid CPU writes after IO
1380 * starts and then write it out here.
1382 try_to_unmap_flush_dirty();
1383 switch (pageout(page, mapping, sc)) {
1387 goto activate_locked;
1389 if (PageWriteback(page))
1391 if (PageDirty(page))
1395 * A synchronous write - probably a ramdisk. Go
1396 * ahead and try to reclaim the page.
1398 if (!trylock_page(page))
1400 if (PageDirty(page) || PageWriteback(page))
1402 mapping = page_mapping(page);
1404 ; /* try to free the page below */
1409 * If the page has buffers, try to free the buffer mappings
1410 * associated with this page. If we succeed we try to free
1413 * We do this even if the page is PageDirty().
1414 * try_to_release_page() does not perform I/O, but it is
1415 * possible for a page to have PageDirty set, but it is actually
1416 * clean (all its buffers are clean). This happens if the
1417 * buffers were written out directly, with submit_bh(). ext3
1418 * will do this, as well as the blockdev mapping.
1419 * try_to_release_page() will discover that cleanness and will
1420 * drop the buffers and mark the page clean - it can be freed.
1422 * Rarely, pages can have buffers and no ->mapping. These are
1423 * the pages which were not successfully invalidated in
1424 * truncate_complete_page(). We try to drop those buffers here
1425 * and if that worked, and the page is no longer mapped into
1426 * process address space (page_count == 1) it can be freed.
1427 * Otherwise, leave the page on the LRU so it is swappable.
1429 if (page_has_private(page)) {
1430 if (!try_to_release_page(page, sc->gfp_mask))
1431 goto activate_locked;
1432 if (!mapping && page_count(page) == 1) {
1434 if (put_page_testzero(page))
1438 * rare race with speculative reference.
1439 * the speculative reference will free
1440 * this page shortly, so we may
1441 * increment nr_reclaimed here (and
1442 * leave it off the LRU).
1450 if (PageAnon(page) && !PageSwapBacked(page)) {
1451 /* follow __remove_mapping for reference */
1452 if (!page_ref_freeze(page, 1))
1454 if (PageDirty(page)) {
1455 page_ref_unfreeze(page, 1);
1459 count_vm_event(PGLAZYFREED);
1460 count_memcg_page_event(page, PGLAZYFREED);
1461 } else if (!mapping || !__remove_mapping(mapping, page, true))
1469 * Is there need to periodically free_page_list? It would
1470 * appear not as the counts should be low
1472 if (unlikely(PageTransHuge(page))) {
1473 mem_cgroup_uncharge(page);
1474 (*get_compound_page_dtor(page))(page);
1476 list_add(&page->lru, &free_pages);
1480 /* Not a candidate for swapping, so reclaim swap space. */
1481 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1483 try_to_free_swap(page);
1484 VM_BUG_ON_PAGE(PageActive(page), page);
1485 if (!PageMlocked(page)) {
1486 SetPageActive(page);
1488 count_memcg_page_event(page, PGACTIVATE);
1493 list_add(&page->lru, &ret_pages);
1494 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1497 mem_cgroup_uncharge_list(&free_pages);
1498 try_to_unmap_flush();
1499 free_unref_page_list(&free_pages);
1501 list_splice(&ret_pages, page_list);
1502 count_vm_events(PGACTIVATE, pgactivate);
1505 stat->nr_dirty = nr_dirty;
1506 stat->nr_congested = nr_congested;
1507 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1508 stat->nr_writeback = nr_writeback;
1509 stat->nr_immediate = nr_immediate;
1510 stat->nr_activate = pgactivate;
1511 stat->nr_ref_keep = nr_ref_keep;
1512 stat->nr_unmap_fail = nr_unmap_fail;
1514 return nr_reclaimed;
1517 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1518 struct list_head *page_list)
1520 struct scan_control sc = {
1521 .gfp_mask = GFP_KERNEL,
1522 .priority = DEF_PRIORITY,
1526 struct page *page, *next;
1527 LIST_HEAD(clean_pages);
1529 list_for_each_entry_safe(page, next, page_list, lru) {
1530 if (page_is_file_cache(page) && !PageDirty(page) &&
1531 !__PageMovable(page)) {
1532 ClearPageActive(page);
1533 list_move(&page->lru, &clean_pages);
1537 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1538 TTU_IGNORE_ACCESS, NULL, true);
1539 list_splice(&clean_pages, page_list);
1540 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1545 * Attempt to remove the specified page from its LRU. Only take this page
1546 * if it is of the appropriate PageActive status. Pages which are being
1547 * freed elsewhere are also ignored.
1549 * page: page to consider
1550 * mode: one of the LRU isolation modes defined above
1552 * returns 0 on success, -ve errno on failure.
1554 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1558 /* Only take pages on the LRU. */
1562 /* Compaction should not handle unevictable pages but CMA can do so */
1563 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1569 * To minimise LRU disruption, the caller can indicate that it only
1570 * wants to isolate pages it will be able to operate on without
1571 * blocking - clean pages for the most part.
1573 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1574 * that it is possible to migrate without blocking
1576 if (mode & ISOLATE_ASYNC_MIGRATE) {
1577 /* All the caller can do on PageWriteback is block */
1578 if (PageWriteback(page))
1581 if (PageDirty(page)) {
1582 struct address_space *mapping;
1586 * Only pages without mappings or that have a
1587 * ->migratepage callback are possible to migrate
1588 * without blocking. However, we can be racing with
1589 * truncation so it's necessary to lock the page
1590 * to stabilise the mapping as truncation holds
1591 * the page lock until after the page is removed
1592 * from the page cache.
1594 if (!trylock_page(page))
1597 mapping = page_mapping(page);
1598 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1605 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1608 if (likely(get_page_unless_zero(page))) {
1610 * Be careful not to clear PageLRU until after we're
1611 * sure the page is not being freed elsewhere -- the
1612 * page release code relies on it.
1623 * Update LRU sizes after isolating pages. The LRU size updates must
1624 * be complete before mem_cgroup_update_lru_size due to a santity check.
1626 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1627 enum lru_list lru, unsigned long *nr_zone_taken)
1631 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1632 if (!nr_zone_taken[zid])
1635 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1637 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1644 * zone_lru_lock is heavily contended. Some of the functions that
1645 * shrink the lists perform better by taking out a batch of pages
1646 * and working on them outside the LRU lock.
1648 * For pagecache intensive workloads, this function is the hottest
1649 * spot in the kernel (apart from copy_*_user functions).
1651 * Appropriate locks must be held before calling this function.
1653 * @nr_to_scan: The number of eligible pages to look through on the list.
1654 * @lruvec: The LRU vector to pull pages from.
1655 * @dst: The temp list to put pages on to.
1656 * @nr_scanned: The number of pages that were scanned.
1657 * @sc: The scan_control struct for this reclaim session
1658 * @mode: One of the LRU isolation modes
1659 * @lru: LRU list id for isolating
1661 * returns how many pages were moved onto *@dst.
1663 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1664 struct lruvec *lruvec, struct list_head *dst,
1665 unsigned long *nr_scanned, struct scan_control *sc,
1666 isolate_mode_t mode, enum lru_list lru)
1668 struct list_head *src = &lruvec->lists[lru];
1669 unsigned long nr_taken = 0;
1670 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1671 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1672 unsigned long skipped = 0;
1673 unsigned long scan, total_scan, nr_pages;
1674 LIST_HEAD(pages_skipped);
1677 for (total_scan = 0;
1678 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1682 page = lru_to_page(src);
1683 prefetchw_prev_lru_page(page, src, flags);
1685 VM_BUG_ON_PAGE(!PageLRU(page), page);
1687 if (page_zonenum(page) > sc->reclaim_idx) {
1688 list_move(&page->lru, &pages_skipped);
1689 nr_skipped[page_zonenum(page)]++;
1694 * Do not count skipped pages because that makes the function
1695 * return with no isolated pages if the LRU mostly contains
1696 * ineligible pages. This causes the VM to not reclaim any
1697 * pages, triggering a premature OOM.
1700 switch (__isolate_lru_page(page, mode)) {
1702 nr_pages = hpage_nr_pages(page);
1703 nr_taken += nr_pages;
1704 nr_zone_taken[page_zonenum(page)] += nr_pages;
1705 list_move(&page->lru, dst);
1709 /* else it is being freed elsewhere */
1710 list_move(&page->lru, src);
1719 * Splice any skipped pages to the start of the LRU list. Note that
1720 * this disrupts the LRU order when reclaiming for lower zones but
1721 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1722 * scanning would soon rescan the same pages to skip and put the
1723 * system at risk of premature OOM.
1725 if (!list_empty(&pages_skipped)) {
1728 list_splice(&pages_skipped, src);
1729 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1730 if (!nr_skipped[zid])
1733 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1734 skipped += nr_skipped[zid];
1737 *nr_scanned = total_scan;
1738 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1739 total_scan, skipped, nr_taken, mode, lru);
1740 update_lru_sizes(lruvec, lru, nr_zone_taken);
1745 * isolate_lru_page - tries to isolate a page from its LRU list
1746 * @page: page to isolate from its LRU list
1748 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1749 * vmstat statistic corresponding to whatever LRU list the page was on.
1751 * Returns 0 if the page was removed from an LRU list.
1752 * Returns -EBUSY if the page was not on an LRU list.
1754 * The returned page will have PageLRU() cleared. If it was found on
1755 * the active list, it will have PageActive set. If it was found on
1756 * the unevictable list, it will have the PageUnevictable bit set. That flag
1757 * may need to be cleared by the caller before letting the page go.
1759 * The vmstat statistic corresponding to the list on which the page was
1760 * found will be decremented.
1764 * (1) Must be called with an elevated refcount on the page. This is a
1765 * fundamentnal difference from isolate_lru_pages (which is called
1766 * without a stable reference).
1767 * (2) the lru_lock must not be held.
1768 * (3) interrupts must be enabled.
1770 int isolate_lru_page(struct page *page)
1774 VM_BUG_ON_PAGE(!page_count(page), page);
1775 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1777 if (PageLRU(page)) {
1778 struct zone *zone = page_zone(page);
1779 struct lruvec *lruvec;
1781 spin_lock_irq(zone_lru_lock(zone));
1782 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1783 if (PageLRU(page)) {
1784 int lru = page_lru(page);
1787 del_page_from_lru_list(page, lruvec, lru);
1790 spin_unlock_irq(zone_lru_lock(zone));
1796 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1797 * then get resheduled. When there are massive number of tasks doing page
1798 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1799 * the LRU list will go small and be scanned faster than necessary, leading to
1800 * unnecessary swapping, thrashing and OOM.
1802 static int too_many_isolated(struct pglist_data *pgdat, int file,
1803 struct scan_control *sc)
1805 unsigned long inactive, isolated;
1807 if (current_is_kswapd())
1810 if (!sane_reclaim(sc))
1814 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1815 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1817 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1818 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1822 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1823 * won't get blocked by normal direct-reclaimers, forming a circular
1826 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1829 return isolated > inactive;
1832 static noinline_for_stack void
1833 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1835 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1836 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1837 LIST_HEAD(pages_to_free);
1840 * Put back any unfreeable pages.
1842 while (!list_empty(page_list)) {
1843 struct page *page = lru_to_page(page_list);
1846 VM_BUG_ON_PAGE(PageLRU(page), page);
1847 list_del(&page->lru);
1848 if (unlikely(!page_evictable(page))) {
1849 spin_unlock_irq(&pgdat->lru_lock);
1850 putback_lru_page(page);
1851 spin_lock_irq(&pgdat->lru_lock);
1855 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1858 lru = page_lru(page);
1859 add_page_to_lru_list(page, lruvec, lru);
1861 if (is_active_lru(lru)) {
1862 int file = is_file_lru(lru);
1863 int numpages = hpage_nr_pages(page);
1864 reclaim_stat->recent_rotated[file] += numpages;
1866 if (put_page_testzero(page)) {
1867 __ClearPageLRU(page);
1868 __ClearPageActive(page);
1869 del_page_from_lru_list(page, lruvec, lru);
1871 if (unlikely(PageCompound(page))) {
1872 spin_unlock_irq(&pgdat->lru_lock);
1873 mem_cgroup_uncharge(page);
1874 (*get_compound_page_dtor(page))(page);
1875 spin_lock_irq(&pgdat->lru_lock);
1877 list_add(&page->lru, &pages_to_free);
1882 * To save our caller's stack, now use input list for pages to free.
1884 list_splice(&pages_to_free, page_list);
1888 * If a kernel thread (such as nfsd for loop-back mounts) services
1889 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1890 * In that case we should only throttle if the backing device it is
1891 * writing to is congested. In other cases it is safe to throttle.
1893 static int current_may_throttle(void)
1895 return !(current->flags & PF_LESS_THROTTLE) ||
1896 current->backing_dev_info == NULL ||
1897 bdi_write_congested(current->backing_dev_info);
1901 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1902 * of reclaimed pages
1904 static noinline_for_stack unsigned long
1905 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1906 struct scan_control *sc, enum lru_list lru)
1908 LIST_HEAD(page_list);
1909 unsigned long nr_scanned;
1910 unsigned long nr_reclaimed = 0;
1911 unsigned long nr_taken;
1912 struct reclaim_stat stat = {};
1913 isolate_mode_t isolate_mode = 0;
1914 int file = is_file_lru(lru);
1915 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1916 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1917 bool stalled = false;
1919 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1923 /* wait a bit for the reclaimer. */
1927 /* We are about to die and free our memory. Return now. */
1928 if (fatal_signal_pending(current))
1929 return SWAP_CLUSTER_MAX;
1935 isolate_mode |= ISOLATE_UNMAPPED;
1937 spin_lock_irq(&pgdat->lru_lock);
1939 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1940 &nr_scanned, sc, isolate_mode, lru);
1942 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1943 reclaim_stat->recent_scanned[file] += nr_taken;
1945 if (current_is_kswapd()) {
1946 if (global_reclaim(sc))
1947 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1948 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1951 if (global_reclaim(sc))
1952 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1953 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1956 spin_unlock_irq(&pgdat->lru_lock);
1961 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1964 spin_lock_irq(&pgdat->lru_lock);
1966 if (current_is_kswapd()) {
1967 if (global_reclaim(sc))
1968 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1969 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1972 if (global_reclaim(sc))
1973 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1974 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1978 putback_inactive_pages(lruvec, &page_list);
1980 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1982 spin_unlock_irq(&pgdat->lru_lock);
1984 mem_cgroup_uncharge_list(&page_list);
1985 free_unref_page_list(&page_list);
1988 * If dirty pages are scanned that are not queued for IO, it
1989 * implies that flushers are not doing their job. This can
1990 * happen when memory pressure pushes dirty pages to the end of
1991 * the LRU before the dirty limits are breached and the dirty
1992 * data has expired. It can also happen when the proportion of
1993 * dirty pages grows not through writes but through memory
1994 * pressure reclaiming all the clean cache. And in some cases,
1995 * the flushers simply cannot keep up with the allocation
1996 * rate. Nudge the flusher threads in case they are asleep.
1998 if (stat.nr_unqueued_dirty == nr_taken)
1999 wakeup_flusher_threads(WB_REASON_VMSCAN);
2001 sc->nr.dirty += stat.nr_dirty;
2002 sc->nr.congested += stat.nr_congested;
2003 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2004 sc->nr.writeback += stat.nr_writeback;
2005 sc->nr.immediate += stat.nr_immediate;
2006 sc->nr.taken += nr_taken;
2008 sc->nr.file_taken += nr_taken;
2010 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2011 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2012 return nr_reclaimed;
2016 * This moves pages from the active list to the inactive list.
2018 * We move them the other way if the page is referenced by one or more
2019 * processes, from rmap.
2021 * If the pages are mostly unmapped, the processing is fast and it is
2022 * appropriate to hold zone_lru_lock across the whole operation. But if
2023 * the pages are mapped, the processing is slow (page_referenced()) so we
2024 * should drop zone_lru_lock around each page. It's impossible to balance
2025 * this, so instead we remove the pages from the LRU while processing them.
2026 * It is safe to rely on PG_active against the non-LRU pages in here because
2027 * nobody will play with that bit on a non-LRU page.
2029 * The downside is that we have to touch page->_refcount against each page.
2030 * But we had to alter page->flags anyway.
2032 * Returns the number of pages moved to the given lru.
2035 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
2036 struct list_head *list,
2037 struct list_head *pages_to_free,
2040 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2045 while (!list_empty(list)) {
2046 page = lru_to_page(list);
2047 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2049 VM_BUG_ON_PAGE(PageLRU(page), page);
2052 nr_pages = hpage_nr_pages(page);
2053 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
2054 list_move(&page->lru, &lruvec->lists[lru]);
2056 if (put_page_testzero(page)) {
2057 __ClearPageLRU(page);
2058 __ClearPageActive(page);
2059 del_page_from_lru_list(page, lruvec, lru);
2061 if (unlikely(PageCompound(page))) {
2062 spin_unlock_irq(&pgdat->lru_lock);
2063 mem_cgroup_uncharge(page);
2064 (*get_compound_page_dtor(page))(page);
2065 spin_lock_irq(&pgdat->lru_lock);
2067 list_add(&page->lru, pages_to_free);
2069 nr_moved += nr_pages;
2073 if (!is_active_lru(lru)) {
2074 __count_vm_events(PGDEACTIVATE, nr_moved);
2075 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
2082 static void shrink_active_list(unsigned long nr_to_scan,
2083 struct lruvec *lruvec,
2084 struct scan_control *sc,
2087 unsigned long nr_taken;
2088 unsigned long nr_scanned;
2089 unsigned long vm_flags;
2090 LIST_HEAD(l_hold); /* The pages which were snipped off */
2091 LIST_HEAD(l_active);
2092 LIST_HEAD(l_inactive);
2094 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2095 unsigned nr_deactivate, nr_activate;
2096 unsigned nr_rotated = 0;
2097 isolate_mode_t isolate_mode = 0;
2098 int file = is_file_lru(lru);
2099 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2104 isolate_mode |= ISOLATE_UNMAPPED;
2106 spin_lock_irq(&pgdat->lru_lock);
2108 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2109 &nr_scanned, sc, isolate_mode, lru);
2111 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2112 reclaim_stat->recent_scanned[file] += nr_taken;
2114 __count_vm_events(PGREFILL, nr_scanned);
2115 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2117 spin_unlock_irq(&pgdat->lru_lock);
2119 while (!list_empty(&l_hold)) {
2121 page = lru_to_page(&l_hold);
2122 list_del(&page->lru);
2124 if (unlikely(!page_evictable(page))) {
2125 putback_lru_page(page);
2129 if (unlikely(buffer_heads_over_limit)) {
2130 if (page_has_private(page) && trylock_page(page)) {
2131 if (page_has_private(page))
2132 try_to_release_page(page, 0);
2137 if (page_referenced(page, 0, sc->target_mem_cgroup,
2139 nr_rotated += hpage_nr_pages(page);
2141 * Identify referenced, file-backed active pages and
2142 * give them one more trip around the active list. So
2143 * that executable code get better chances to stay in
2144 * memory under moderate memory pressure. Anon pages
2145 * are not likely to be evicted by use-once streaming
2146 * IO, plus JVM can create lots of anon VM_EXEC pages,
2147 * so we ignore them here.
2149 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2150 list_add(&page->lru, &l_active);
2155 ClearPageActive(page); /* we are de-activating */
2156 SetPageWorkingset(page);
2157 list_add(&page->lru, &l_inactive);
2161 * Move pages back to the lru list.
2163 spin_lock_irq(&pgdat->lru_lock);
2165 * Count referenced pages from currently used mappings as rotated,
2166 * even though only some of them are actually re-activated. This
2167 * helps balance scan pressure between file and anonymous pages in
2170 reclaim_stat->recent_rotated[file] += nr_rotated;
2172 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2173 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2174 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2175 spin_unlock_irq(&pgdat->lru_lock);
2177 mem_cgroup_uncharge_list(&l_hold);
2178 free_unref_page_list(&l_hold);
2179 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2180 nr_deactivate, nr_rotated, sc->priority, file);
2184 * The inactive anon list should be small enough that the VM never has
2185 * to do too much work.
2187 * The inactive file list should be small enough to leave most memory
2188 * to the established workingset on the scan-resistant active list,
2189 * but large enough to avoid thrashing the aggregate readahead window.
2191 * Both inactive lists should also be large enough that each inactive
2192 * page has a chance to be referenced again before it is reclaimed.
2194 * If that fails and refaulting is observed, the inactive list grows.
2196 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2197 * on this LRU, maintained by the pageout code. An inactive_ratio
2198 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2201 * memory ratio inactive
2202 * -------------------------------------
2211 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2212 struct mem_cgroup *memcg,
2213 struct scan_control *sc, bool actual_reclaim)
2215 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2216 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2217 enum lru_list inactive_lru = file * LRU_FILE;
2218 unsigned long inactive, active;
2219 unsigned long inactive_ratio;
2220 unsigned long refaults;
2224 * If we don't have swap space, anonymous page deactivation
2227 if (!file && !total_swap_pages)
2230 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2231 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2234 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2236 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2239 * When refaults are being observed, it means a new workingset
2240 * is being established. Disable active list protection to get
2241 * rid of the stale workingset quickly.
2243 if (file && actual_reclaim && lruvec->refaults != refaults) {
2246 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2248 inactive_ratio = int_sqrt(10 * gb);
2254 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2255 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2256 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2257 inactive_ratio, file);
2259 return inactive * inactive_ratio < active;
2262 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2263 struct lruvec *lruvec, struct mem_cgroup *memcg,
2264 struct scan_control *sc)
2266 if (is_active_lru(lru)) {
2267 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2269 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2273 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2284 * Determine how aggressively the anon and file LRU lists should be
2285 * scanned. The relative value of each set of LRU lists is determined
2286 * by looking at the fraction of the pages scanned we did rotate back
2287 * onto the active list instead of evict.
2289 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2290 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2292 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2293 struct scan_control *sc, unsigned long *nr,
2294 unsigned long *lru_pages)
2296 int swappiness = mem_cgroup_swappiness(memcg);
2297 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2299 u64 denominator = 0; /* gcc */
2300 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2301 unsigned long anon_prio, file_prio;
2302 enum scan_balance scan_balance;
2303 unsigned long anon, file;
2304 unsigned long ap, fp;
2307 /* If we have no swap space, do not bother scanning anon pages. */
2308 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2309 scan_balance = SCAN_FILE;
2314 * Global reclaim will swap to prevent OOM even with no
2315 * swappiness, but memcg users want to use this knob to
2316 * disable swapping for individual groups completely when
2317 * using the memory controller's swap limit feature would be
2320 if (!global_reclaim(sc) && !swappiness) {
2321 scan_balance = SCAN_FILE;
2326 * Do not apply any pressure balancing cleverness when the
2327 * system is close to OOM, scan both anon and file equally
2328 * (unless the swappiness setting disagrees with swapping).
2330 if (!sc->priority && swappiness) {
2331 scan_balance = SCAN_EQUAL;
2336 * Prevent the reclaimer from falling into the cache trap: as
2337 * cache pages start out inactive, every cache fault will tip
2338 * the scan balance towards the file LRU. And as the file LRU
2339 * shrinks, so does the window for rotation from references.
2340 * This means we have a runaway feedback loop where a tiny
2341 * thrashing file LRU becomes infinitely more attractive than
2342 * anon pages. Try to detect this based on file LRU size.
2344 if (global_reclaim(sc)) {
2345 unsigned long pgdatfile;
2346 unsigned long pgdatfree;
2348 unsigned long total_high_wmark = 0;
2350 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2351 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2352 node_page_state(pgdat, NR_INACTIVE_FILE);
2354 for (z = 0; z < MAX_NR_ZONES; z++) {
2355 struct zone *zone = &pgdat->node_zones[z];
2356 if (!managed_zone(zone))
2359 total_high_wmark += high_wmark_pages(zone);
2362 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2364 * Force SCAN_ANON if there are enough inactive
2365 * anonymous pages on the LRU in eligible zones.
2366 * Otherwise, the small LRU gets thrashed.
2368 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2369 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2371 scan_balance = SCAN_ANON;
2378 * If there is enough inactive page cache, i.e. if the size of the
2379 * inactive list is greater than that of the active list *and* the
2380 * inactive list actually has some pages to scan on this priority, we
2381 * do not reclaim anything from the anonymous working set right now.
2382 * Without the second condition we could end up never scanning an
2383 * lruvec even if it has plenty of old anonymous pages unless the
2384 * system is under heavy pressure.
2386 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2387 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2388 scan_balance = SCAN_FILE;
2392 scan_balance = SCAN_FRACT;
2395 * With swappiness at 100, anonymous and file have the same priority.
2396 * This scanning priority is essentially the inverse of IO cost.
2398 anon_prio = swappiness;
2399 file_prio = 200 - anon_prio;
2402 * OK, so we have swap space and a fair amount of page cache
2403 * pages. We use the recently rotated / recently scanned
2404 * ratios to determine how valuable each cache is.
2406 * Because workloads change over time (and to avoid overflow)
2407 * we keep these statistics as a floating average, which ends
2408 * up weighing recent references more than old ones.
2410 * anon in [0], file in [1]
2413 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2414 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2415 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2416 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2418 spin_lock_irq(&pgdat->lru_lock);
2419 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2420 reclaim_stat->recent_scanned[0] /= 2;
2421 reclaim_stat->recent_rotated[0] /= 2;
2424 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2425 reclaim_stat->recent_scanned[1] /= 2;
2426 reclaim_stat->recent_rotated[1] /= 2;
2430 * The amount of pressure on anon vs file pages is inversely
2431 * proportional to the fraction of recently scanned pages on
2432 * each list that were recently referenced and in active use.
2434 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2435 ap /= reclaim_stat->recent_rotated[0] + 1;
2437 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2438 fp /= reclaim_stat->recent_rotated[1] + 1;
2439 spin_unlock_irq(&pgdat->lru_lock);
2443 denominator = ap + fp + 1;
2446 for_each_evictable_lru(lru) {
2447 int file = is_file_lru(lru);
2451 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2452 scan = size >> sc->priority;
2454 * If the cgroup's already been deleted, make sure to
2455 * scrape out the remaining cache.
2457 if (!scan && !mem_cgroup_online(memcg))
2458 scan = min(size, SWAP_CLUSTER_MAX);
2460 switch (scan_balance) {
2462 /* Scan lists relative to size */
2466 * Scan types proportional to swappiness and
2467 * their relative recent reclaim efficiency.
2468 * Make sure we don't miss the last page
2469 * because of a round-off error.
2471 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2476 /* Scan one type exclusively */
2477 if ((scan_balance == SCAN_FILE) != file) {
2483 /* Look ma, no brain */
2493 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2495 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2496 struct scan_control *sc, unsigned long *lru_pages)
2498 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2499 unsigned long nr[NR_LRU_LISTS];
2500 unsigned long targets[NR_LRU_LISTS];
2501 unsigned long nr_to_scan;
2503 unsigned long nr_reclaimed = 0;
2504 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2505 struct blk_plug plug;
2508 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2510 /* Record the original scan target for proportional adjustments later */
2511 memcpy(targets, nr, sizeof(nr));
2514 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2515 * event that can occur when there is little memory pressure e.g.
2516 * multiple streaming readers/writers. Hence, we do not abort scanning
2517 * when the requested number of pages are reclaimed when scanning at
2518 * DEF_PRIORITY on the assumption that the fact we are direct
2519 * reclaiming implies that kswapd is not keeping up and it is best to
2520 * do a batch of work at once. For memcg reclaim one check is made to
2521 * abort proportional reclaim if either the file or anon lru has already
2522 * dropped to zero at the first pass.
2524 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2525 sc->priority == DEF_PRIORITY);
2527 blk_start_plug(&plug);
2528 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2529 nr[LRU_INACTIVE_FILE]) {
2530 unsigned long nr_anon, nr_file, percentage;
2531 unsigned long nr_scanned;
2533 for_each_evictable_lru(lru) {
2535 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2536 nr[lru] -= nr_to_scan;
2538 nr_reclaimed += shrink_list(lru, nr_to_scan,
2545 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2549 * For kswapd and memcg, reclaim at least the number of pages
2550 * requested. Ensure that the anon and file LRUs are scanned
2551 * proportionally what was requested by get_scan_count(). We
2552 * stop reclaiming one LRU and reduce the amount scanning
2553 * proportional to the original scan target.
2555 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2556 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2559 * It's just vindictive to attack the larger once the smaller
2560 * has gone to zero. And given the way we stop scanning the
2561 * smaller below, this makes sure that we only make one nudge
2562 * towards proportionality once we've got nr_to_reclaim.
2564 if (!nr_file || !nr_anon)
2567 if (nr_file > nr_anon) {
2568 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2569 targets[LRU_ACTIVE_ANON] + 1;
2571 percentage = nr_anon * 100 / scan_target;
2573 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2574 targets[LRU_ACTIVE_FILE] + 1;
2576 percentage = nr_file * 100 / scan_target;
2579 /* Stop scanning the smaller of the LRU */
2581 nr[lru + LRU_ACTIVE] = 0;
2584 * Recalculate the other LRU scan count based on its original
2585 * scan target and the percentage scanning already complete
2587 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2588 nr_scanned = targets[lru] - nr[lru];
2589 nr[lru] = targets[lru] * (100 - percentage) / 100;
2590 nr[lru] -= min(nr[lru], nr_scanned);
2593 nr_scanned = targets[lru] - nr[lru];
2594 nr[lru] = targets[lru] * (100 - percentage) / 100;
2595 nr[lru] -= min(nr[lru], nr_scanned);
2597 scan_adjusted = true;
2599 blk_finish_plug(&plug);
2600 sc->nr_reclaimed += nr_reclaimed;
2603 * Even if we did not try to evict anon pages at all, we want to
2604 * rebalance the anon lru active/inactive ratio.
2606 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2607 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2608 sc, LRU_ACTIVE_ANON);
2611 /* Use reclaim/compaction for costly allocs or under memory pressure */
2612 static bool in_reclaim_compaction(struct scan_control *sc)
2614 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2615 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2616 sc->priority < DEF_PRIORITY - 2))
2623 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2624 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2625 * true if more pages should be reclaimed such that when the page allocator
2626 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2627 * It will give up earlier than that if there is difficulty reclaiming pages.
2629 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2630 unsigned long nr_reclaimed,
2631 unsigned long nr_scanned,
2632 struct scan_control *sc)
2634 unsigned long pages_for_compaction;
2635 unsigned long inactive_lru_pages;
2638 /* If not in reclaim/compaction mode, stop */
2639 if (!in_reclaim_compaction(sc))
2642 /* Consider stopping depending on scan and reclaim activity */
2643 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2645 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2646 * full LRU list has been scanned and we are still failing
2647 * to reclaim pages. This full LRU scan is potentially
2648 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2650 if (!nr_reclaimed && !nr_scanned)
2654 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2655 * fail without consequence, stop if we failed to reclaim
2656 * any pages from the last SWAP_CLUSTER_MAX number of
2657 * pages that were scanned. This will return to the
2658 * caller faster at the risk reclaim/compaction and
2659 * the resulting allocation attempt fails
2666 * If we have not reclaimed enough pages for compaction and the
2667 * inactive lists are large enough, continue reclaiming
2669 pages_for_compaction = compact_gap(sc->order);
2670 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2671 if (get_nr_swap_pages() > 0)
2672 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2673 if (sc->nr_reclaimed < pages_for_compaction &&
2674 inactive_lru_pages > pages_for_compaction)
2677 /* If compaction would go ahead or the allocation would succeed, stop */
2678 for (z = 0; z <= sc->reclaim_idx; z++) {
2679 struct zone *zone = &pgdat->node_zones[z];
2680 if (!managed_zone(zone))
2683 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2684 case COMPACT_SUCCESS:
2685 case COMPACT_CONTINUE:
2688 /* check next zone */
2695 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2697 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2698 (memcg && memcg_congested(pgdat, memcg));
2701 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2703 struct reclaim_state *reclaim_state = current->reclaim_state;
2704 unsigned long nr_reclaimed, nr_scanned;
2705 bool reclaimable = false;
2708 struct mem_cgroup *root = sc->target_mem_cgroup;
2709 struct mem_cgroup_reclaim_cookie reclaim = {
2711 .priority = sc->priority,
2713 unsigned long node_lru_pages = 0;
2714 struct mem_cgroup *memcg;
2716 memset(&sc->nr, 0, sizeof(sc->nr));
2718 nr_reclaimed = sc->nr_reclaimed;
2719 nr_scanned = sc->nr_scanned;
2721 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2723 unsigned long lru_pages;
2724 unsigned long reclaimed;
2725 unsigned long scanned;
2727 switch (mem_cgroup_protected(root, memcg)) {
2728 case MEMCG_PROT_MIN:
2731 * If there is no reclaimable memory, OOM.
2734 case MEMCG_PROT_LOW:
2737 * Respect the protection only as long as
2738 * there is an unprotected supply
2739 * of reclaimable memory from other cgroups.
2741 if (!sc->memcg_low_reclaim) {
2742 sc->memcg_low_skipped = 1;
2745 memcg_memory_event(memcg, MEMCG_LOW);
2747 case MEMCG_PROT_NONE:
2751 reclaimed = sc->nr_reclaimed;
2752 scanned = sc->nr_scanned;
2753 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2754 node_lru_pages += lru_pages;
2756 if (sc->may_shrinkslab) {
2757 shrink_slab(sc->gfp_mask, pgdat->node_id,
2758 memcg, sc->priority);
2761 /* Record the group's reclaim efficiency */
2762 vmpressure(sc->gfp_mask, memcg, false,
2763 sc->nr_scanned - scanned,
2764 sc->nr_reclaimed - reclaimed);
2767 * Direct reclaim and kswapd have to scan all memory
2768 * cgroups to fulfill the overall scan target for the
2771 * Limit reclaim, on the other hand, only cares about
2772 * nr_to_reclaim pages to be reclaimed and it will
2773 * retry with decreasing priority if one round over the
2774 * whole hierarchy is not sufficient.
2776 if (!global_reclaim(sc) &&
2777 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2778 mem_cgroup_iter_break(root, memcg);
2781 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2783 if (reclaim_state) {
2784 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2785 reclaim_state->reclaimed_slab = 0;
2788 /* Record the subtree's reclaim efficiency */
2789 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2790 sc->nr_scanned - nr_scanned,
2791 sc->nr_reclaimed - nr_reclaimed);
2793 if (sc->nr_reclaimed - nr_reclaimed)
2796 if (current_is_kswapd()) {
2798 * If reclaim is isolating dirty pages under writeback,
2799 * it implies that the long-lived page allocation rate
2800 * is exceeding the page laundering rate. Either the
2801 * global limits are not being effective at throttling
2802 * processes due to the page distribution throughout
2803 * zones or there is heavy usage of a slow backing
2804 * device. The only option is to throttle from reclaim
2805 * context which is not ideal as there is no guarantee
2806 * the dirtying process is throttled in the same way
2807 * balance_dirty_pages() manages.
2809 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2810 * count the number of pages under pages flagged for
2811 * immediate reclaim and stall if any are encountered
2812 * in the nr_immediate check below.
2814 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2815 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2818 * Tag a node as congested if all the dirty pages
2819 * scanned were backed by a congested BDI and
2820 * wait_iff_congested will stall.
2822 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2823 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2825 /* Allow kswapd to start writing pages during reclaim.*/
2826 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2827 set_bit(PGDAT_DIRTY, &pgdat->flags);
2830 * If kswapd scans pages marked marked for immediate
2831 * reclaim and under writeback (nr_immediate), it
2832 * implies that pages are cycling through the LRU
2833 * faster than they are written so also forcibly stall.
2835 if (sc->nr.immediate)
2836 congestion_wait(BLK_RW_ASYNC, HZ/10);
2840 * Legacy memcg will stall in page writeback so avoid forcibly
2841 * stalling in wait_iff_congested().
2843 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2844 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2845 set_memcg_congestion(pgdat, root, true);
2848 * Stall direct reclaim for IO completions if underlying BDIs
2849 * and node is congested. Allow kswapd to continue until it
2850 * starts encountering unqueued dirty pages or cycling through
2851 * the LRU too quickly.
2853 if (!sc->hibernation_mode && !current_is_kswapd() &&
2854 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2855 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2857 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2858 sc->nr_scanned - nr_scanned, sc));
2861 * Kswapd gives up on balancing particular nodes after too
2862 * many failures to reclaim anything from them and goes to
2863 * sleep. On reclaim progress, reset the failure counter. A
2864 * successful direct reclaim run will revive a dormant kswapd.
2867 pgdat->kswapd_failures = 0;
2873 * Returns true if compaction should go ahead for a costly-order request, or
2874 * the allocation would already succeed without compaction. Return false if we
2875 * should reclaim first.
2877 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2879 unsigned long watermark;
2880 enum compact_result suitable;
2882 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2883 if (suitable == COMPACT_SUCCESS)
2884 /* Allocation should succeed already. Don't reclaim. */
2886 if (suitable == COMPACT_SKIPPED)
2887 /* Compaction cannot yet proceed. Do reclaim. */
2891 * Compaction is already possible, but it takes time to run and there
2892 * are potentially other callers using the pages just freed. So proceed
2893 * with reclaim to make a buffer of free pages available to give
2894 * compaction a reasonable chance of completing and allocating the page.
2895 * Note that we won't actually reclaim the whole buffer in one attempt
2896 * as the target watermark in should_continue_reclaim() is lower. But if
2897 * we are already above the high+gap watermark, don't reclaim at all.
2899 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2901 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2905 * This is the direct reclaim path, for page-allocating processes. We only
2906 * try to reclaim pages from zones which will satisfy the caller's allocation
2909 * If a zone is deemed to be full of pinned pages then just give it a light
2910 * scan then give up on it.
2912 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2916 unsigned long nr_soft_reclaimed;
2917 unsigned long nr_soft_scanned;
2919 pg_data_t *last_pgdat = NULL;
2922 * If the number of buffer_heads in the machine exceeds the maximum
2923 * allowed level, force direct reclaim to scan the highmem zone as
2924 * highmem pages could be pinning lowmem pages storing buffer_heads
2926 orig_mask = sc->gfp_mask;
2927 if (buffer_heads_over_limit) {
2928 sc->gfp_mask |= __GFP_HIGHMEM;
2929 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2932 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2933 sc->reclaim_idx, sc->nodemask) {
2935 * Take care memory controller reclaiming has small influence
2938 if (global_reclaim(sc)) {
2939 if (!cpuset_zone_allowed(zone,
2940 GFP_KERNEL | __GFP_HARDWALL))
2944 * If we already have plenty of memory free for
2945 * compaction in this zone, don't free any more.
2946 * Even though compaction is invoked for any
2947 * non-zero order, only frequent costly order
2948 * reclamation is disruptive enough to become a
2949 * noticeable problem, like transparent huge
2952 if (IS_ENABLED(CONFIG_COMPACTION) &&
2953 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2954 compaction_ready(zone, sc)) {
2955 sc->compaction_ready = true;
2960 * Shrink each node in the zonelist once. If the
2961 * zonelist is ordered by zone (not the default) then a
2962 * node may be shrunk multiple times but in that case
2963 * the user prefers lower zones being preserved.
2965 if (zone->zone_pgdat == last_pgdat)
2969 * This steals pages from memory cgroups over softlimit
2970 * and returns the number of reclaimed pages and
2971 * scanned pages. This works for global memory pressure
2972 * and balancing, not for a memcg's limit.
2974 nr_soft_scanned = 0;
2975 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2976 sc->order, sc->gfp_mask,
2978 sc->nr_reclaimed += nr_soft_reclaimed;
2979 sc->nr_scanned += nr_soft_scanned;
2980 /* need some check for avoid more shrink_zone() */
2983 /* See comment about same check for global reclaim above */
2984 if (zone->zone_pgdat == last_pgdat)
2986 last_pgdat = zone->zone_pgdat;
2987 shrink_node(zone->zone_pgdat, sc);
2991 * Restore to original mask to avoid the impact on the caller if we
2992 * promoted it to __GFP_HIGHMEM.
2994 sc->gfp_mask = orig_mask;
2997 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2999 struct mem_cgroup *memcg;
3001 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3003 unsigned long refaults;
3004 struct lruvec *lruvec;
3007 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
3009 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
3011 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3012 lruvec->refaults = refaults;
3013 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3017 * This is the main entry point to direct page reclaim.
3019 * If a full scan of the inactive list fails to free enough memory then we
3020 * are "out of memory" and something needs to be killed.
3022 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3023 * high - the zone may be full of dirty or under-writeback pages, which this
3024 * caller can't do much about. We kick the writeback threads and take explicit
3025 * naps in the hope that some of these pages can be written. But if the
3026 * allocating task holds filesystem locks which prevent writeout this might not
3027 * work, and the allocation attempt will fail.
3029 * returns: 0, if no pages reclaimed
3030 * else, the number of pages reclaimed
3032 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3033 struct scan_control *sc)
3035 int initial_priority = sc->priority;
3036 pg_data_t *last_pgdat;
3040 delayacct_freepages_start();
3042 if (global_reclaim(sc))
3043 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3046 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3049 shrink_zones(zonelist, sc);
3051 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3054 if (sc->compaction_ready)
3058 * If we're getting trouble reclaiming, start doing
3059 * writepage even in laptop mode.
3061 if (sc->priority < DEF_PRIORITY - 2)
3062 sc->may_writepage = 1;
3063 } while (--sc->priority >= 0);
3066 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3068 if (zone->zone_pgdat == last_pgdat)
3070 last_pgdat = zone->zone_pgdat;
3071 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3072 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3075 delayacct_freepages_end();
3077 if (sc->nr_reclaimed)
3078 return sc->nr_reclaimed;
3080 /* Aborted reclaim to try compaction? don't OOM, then */
3081 if (sc->compaction_ready)
3084 /* Untapped cgroup reserves? Don't OOM, retry. */
3085 if (sc->memcg_low_skipped) {
3086 sc->priority = initial_priority;
3087 sc->memcg_low_reclaim = 1;
3088 sc->memcg_low_skipped = 0;
3095 static bool allow_direct_reclaim(pg_data_t *pgdat)
3098 unsigned long pfmemalloc_reserve = 0;
3099 unsigned long free_pages = 0;
3103 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3106 for (i = 0; i <= ZONE_NORMAL; i++) {
3107 zone = &pgdat->node_zones[i];
3108 if (!managed_zone(zone))
3111 if (!zone_reclaimable_pages(zone))
3114 pfmemalloc_reserve += min_wmark_pages(zone);
3115 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3118 /* If there are no reserves (unexpected config) then do not throttle */
3119 if (!pfmemalloc_reserve)
3122 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3124 /* kswapd must be awake if processes are being throttled */
3125 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3126 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3127 (enum zone_type)ZONE_NORMAL);
3128 wake_up_interruptible(&pgdat->kswapd_wait);
3135 * Throttle direct reclaimers if backing storage is backed by the network
3136 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3137 * depleted. kswapd will continue to make progress and wake the processes
3138 * when the low watermark is reached.
3140 * Returns true if a fatal signal was delivered during throttling. If this
3141 * happens, the page allocator should not consider triggering the OOM killer.
3143 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3144 nodemask_t *nodemask)
3148 pg_data_t *pgdat = NULL;
3151 * Kernel threads should not be throttled as they may be indirectly
3152 * responsible for cleaning pages necessary for reclaim to make forward
3153 * progress. kjournald for example may enter direct reclaim while
3154 * committing a transaction where throttling it could forcing other
3155 * processes to block on log_wait_commit().
3157 if (current->flags & PF_KTHREAD)
3161 * If a fatal signal is pending, this process should not throttle.
3162 * It should return quickly so it can exit and free its memory
3164 if (fatal_signal_pending(current))
3168 * Check if the pfmemalloc reserves are ok by finding the first node
3169 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3170 * GFP_KERNEL will be required for allocating network buffers when
3171 * swapping over the network so ZONE_HIGHMEM is unusable.
3173 * Throttling is based on the first usable node and throttled processes
3174 * wait on a queue until kswapd makes progress and wakes them. There
3175 * is an affinity then between processes waking up and where reclaim
3176 * progress has been made assuming the process wakes on the same node.
3177 * More importantly, processes running on remote nodes will not compete
3178 * for remote pfmemalloc reserves and processes on different nodes
3179 * should make reasonable progress.
3181 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3182 gfp_zone(gfp_mask), nodemask) {
3183 if (zone_idx(zone) > ZONE_NORMAL)
3186 /* Throttle based on the first usable node */
3187 pgdat = zone->zone_pgdat;
3188 if (allow_direct_reclaim(pgdat))
3193 /* If no zone was usable by the allocation flags then do not throttle */
3197 /* Account for the throttling */
3198 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3201 * If the caller cannot enter the filesystem, it's possible that it
3202 * is due to the caller holding an FS lock or performing a journal
3203 * transaction in the case of a filesystem like ext[3|4]. In this case,
3204 * it is not safe to block on pfmemalloc_wait as kswapd could be
3205 * blocked waiting on the same lock. Instead, throttle for up to a
3206 * second before continuing.
3208 if (!(gfp_mask & __GFP_FS)) {
3209 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3210 allow_direct_reclaim(pgdat), HZ);
3215 /* Throttle until kswapd wakes the process */
3216 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3217 allow_direct_reclaim(pgdat));
3220 if (fatal_signal_pending(current))
3227 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3228 gfp_t gfp_mask, nodemask_t *nodemask)
3230 unsigned long nr_reclaimed;
3231 struct scan_control sc = {
3232 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3233 .gfp_mask = current_gfp_context(gfp_mask),
3234 .reclaim_idx = gfp_zone(gfp_mask),
3236 .nodemask = nodemask,
3237 .priority = DEF_PRIORITY,
3238 .may_writepage = !laptop_mode,
3241 .may_shrinkslab = 1,
3245 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3246 * Confirm they are large enough for max values.
3248 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3249 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3250 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3253 * Do not enter reclaim if fatal signal was delivered while throttled.
3254 * 1 is returned so that the page allocator does not OOM kill at this
3257 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3260 trace_mm_vmscan_direct_reclaim_begin(order,
3265 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3267 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3269 return nr_reclaimed;
3274 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3275 gfp_t gfp_mask, bool noswap,
3277 unsigned long *nr_scanned)
3279 struct scan_control sc = {
3280 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3281 .target_mem_cgroup = memcg,
3282 .may_writepage = !laptop_mode,
3284 .reclaim_idx = MAX_NR_ZONES - 1,
3285 .may_swap = !noswap,
3286 .may_shrinkslab = 1,
3288 unsigned long lru_pages;
3290 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3291 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3293 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3299 * NOTE: Although we can get the priority field, using it
3300 * here is not a good idea, since it limits the pages we can scan.
3301 * if we don't reclaim here, the shrink_node from balance_pgdat
3302 * will pick up pages from other mem cgroup's as well. We hack
3303 * the priority and make it zero.
3305 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3307 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3309 *nr_scanned = sc.nr_scanned;
3310 return sc.nr_reclaimed;
3313 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3314 unsigned long nr_pages,
3318 struct zonelist *zonelist;
3319 unsigned long nr_reclaimed;
3320 unsigned long pflags;
3322 unsigned int noreclaim_flag;
3323 struct scan_control sc = {
3324 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3325 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3326 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3327 .reclaim_idx = MAX_NR_ZONES - 1,
3328 .target_mem_cgroup = memcg,
3329 .priority = DEF_PRIORITY,
3330 .may_writepage = !laptop_mode,
3332 .may_swap = may_swap,
3333 .may_shrinkslab = 1,
3337 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3338 * take care of from where we get pages. So the node where we start the
3339 * scan does not need to be the current node.
3341 nid = mem_cgroup_select_victim_node(memcg);
3343 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3345 trace_mm_vmscan_memcg_reclaim_begin(0,
3350 psi_memstall_enter(&pflags);
3351 noreclaim_flag = memalloc_noreclaim_save();
3353 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3355 memalloc_noreclaim_restore(noreclaim_flag);
3356 psi_memstall_leave(&pflags);
3358 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3360 return nr_reclaimed;
3364 static void age_active_anon(struct pglist_data *pgdat,
3365 struct scan_control *sc)
3367 struct mem_cgroup *memcg;
3369 if (!total_swap_pages)
3372 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3374 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3376 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3377 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3378 sc, LRU_ACTIVE_ANON);
3380 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3384 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3390 * Check for watermark boosts top-down as the higher zones
3391 * are more likely to be boosted. Both watermarks and boosts
3392 * should not be checked at the time time as reclaim would
3393 * start prematurely when there is no boosting and a lower
3396 for (i = classzone_idx; i >= 0; i--) {
3397 zone = pgdat->node_zones + i;
3398 if (!managed_zone(zone))
3401 if (zone->watermark_boost)
3409 * Returns true if there is an eligible zone balanced for the request order
3412 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3415 unsigned long mark = -1;
3419 * Check watermarks bottom-up as lower zones are more likely to
3422 for (i = 0; i <= classzone_idx; i++) {
3423 zone = pgdat->node_zones + i;
3425 if (!managed_zone(zone))
3428 mark = high_wmark_pages(zone);
3429 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3434 * If a node has no populated zone within classzone_idx, it does not
3435 * need balancing by definition. This can happen if a zone-restricted
3436 * allocation tries to wake a remote kswapd.
3444 /* Clear pgdat state for congested, dirty or under writeback. */
3445 static void clear_pgdat_congested(pg_data_t *pgdat)
3447 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3448 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3449 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3453 * Prepare kswapd for sleeping. This verifies that there are no processes
3454 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3456 * Returns true if kswapd is ready to sleep
3458 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3461 * The throttled processes are normally woken up in balance_pgdat() as
3462 * soon as allow_direct_reclaim() is true. But there is a potential
3463 * race between when kswapd checks the watermarks and a process gets
3464 * throttled. There is also a potential race if processes get
3465 * throttled, kswapd wakes, a large process exits thereby balancing the
3466 * zones, which causes kswapd to exit balance_pgdat() before reaching
3467 * the wake up checks. If kswapd is going to sleep, no process should
3468 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3469 * the wake up is premature, processes will wake kswapd and get
3470 * throttled again. The difference from wake ups in balance_pgdat() is
3471 * that here we are under prepare_to_wait().
3473 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3474 wake_up_all(&pgdat->pfmemalloc_wait);
3476 /* Hopeless node, leave it to direct reclaim */
3477 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3480 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3481 clear_pgdat_congested(pgdat);
3489 * kswapd shrinks a node of pages that are at or below the highest usable
3490 * zone that is currently unbalanced.
3492 * Returns true if kswapd scanned at least the requested number of pages to
3493 * reclaim or if the lack of progress was due to pages under writeback.
3494 * This is used to determine if the scanning priority needs to be raised.
3496 static bool kswapd_shrink_node(pg_data_t *pgdat,
3497 struct scan_control *sc)
3502 /* Reclaim a number of pages proportional to the number of zones */
3503 sc->nr_to_reclaim = 0;
3504 for (z = 0; z <= sc->reclaim_idx; z++) {
3505 zone = pgdat->node_zones + z;
3506 if (!managed_zone(zone))
3509 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3513 * Historically care was taken to put equal pressure on all zones but
3514 * now pressure is applied based on node LRU order.
3516 shrink_node(pgdat, sc);
3519 * Fragmentation may mean that the system cannot be rebalanced for
3520 * high-order allocations. If twice the allocation size has been
3521 * reclaimed then recheck watermarks only at order-0 to prevent
3522 * excessive reclaim. Assume that a process requested a high-order
3523 * can direct reclaim/compact.
3525 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3528 return sc->nr_scanned >= sc->nr_to_reclaim;
3532 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3533 * that are eligible for use by the caller until at least one zone is
3536 * Returns the order kswapd finished reclaiming at.
3538 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3539 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3540 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3541 * or lower is eligible for reclaim until at least one usable zone is
3544 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3547 unsigned long nr_soft_reclaimed;
3548 unsigned long nr_soft_scanned;
3549 unsigned long pflags;
3550 unsigned long nr_boost_reclaim;
3551 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3554 struct scan_control sc = {
3555 .gfp_mask = GFP_KERNEL,
3560 psi_memstall_enter(&pflags);
3561 __fs_reclaim_acquire();
3563 count_vm_event(PAGEOUTRUN);
3566 * Account for the reclaim boost. Note that the zone boost is left in
3567 * place so that parallel allocations that are near the watermark will
3568 * stall or direct reclaim until kswapd is finished.
3570 nr_boost_reclaim = 0;
3571 for (i = 0; i <= classzone_idx; i++) {
3572 zone = pgdat->node_zones + i;
3573 if (!managed_zone(zone))
3576 nr_boost_reclaim += zone->watermark_boost;
3577 zone_boosts[i] = zone->watermark_boost;
3579 boosted = nr_boost_reclaim;
3582 sc.priority = DEF_PRIORITY;
3584 unsigned long nr_reclaimed = sc.nr_reclaimed;
3585 bool raise_priority = true;
3589 sc.reclaim_idx = classzone_idx;
3592 * If the number of buffer_heads exceeds the maximum allowed
3593 * then consider reclaiming from all zones. This has a dual
3594 * purpose -- on 64-bit systems it is expected that
3595 * buffer_heads are stripped during active rotation. On 32-bit
3596 * systems, highmem pages can pin lowmem memory and shrinking
3597 * buffers can relieve lowmem pressure. Reclaim may still not
3598 * go ahead if all eligible zones for the original allocation
3599 * request are balanced to avoid excessive reclaim from kswapd.
3601 if (buffer_heads_over_limit) {
3602 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3603 zone = pgdat->node_zones + i;
3604 if (!managed_zone(zone))
3613 * If the pgdat is imbalanced then ignore boosting and preserve
3614 * the watermarks for a later time and restart. Note that the
3615 * zone watermarks will be still reset at the end of balancing
3616 * on the grounds that the normal reclaim should be enough to
3617 * re-evaluate if boosting is required when kswapd next wakes.
3619 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3620 if (!balanced && nr_boost_reclaim) {
3621 nr_boost_reclaim = 0;
3626 * If boosting is not active then only reclaim if there are no
3627 * eligible zones. Note that sc.reclaim_idx is not used as
3628 * buffer_heads_over_limit may have adjusted it.
3630 if (!nr_boost_reclaim && balanced)
3633 /* Limit the priority of boosting to avoid reclaim writeback */
3634 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3635 raise_priority = false;
3638 * Do not writeback or swap pages for boosted reclaim. The
3639 * intent is to relieve pressure not issue sub-optimal IO
3640 * from reclaim context. If no pages are reclaimed, the
3641 * reclaim will be aborted.
3643 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3644 sc.may_swap = !nr_boost_reclaim;
3645 sc.may_shrinkslab = !nr_boost_reclaim;
3648 * Do some background aging of the anon list, to give
3649 * pages a chance to be referenced before reclaiming. All
3650 * pages are rotated regardless of classzone as this is
3651 * about consistent aging.
3653 age_active_anon(pgdat, &sc);
3656 * If we're getting trouble reclaiming, start doing writepage
3657 * even in laptop mode.
3659 if (sc.priority < DEF_PRIORITY - 2)
3660 sc.may_writepage = 1;
3662 /* Call soft limit reclaim before calling shrink_node. */
3664 nr_soft_scanned = 0;
3665 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3666 sc.gfp_mask, &nr_soft_scanned);
3667 sc.nr_reclaimed += nr_soft_reclaimed;
3670 * There should be no need to raise the scanning priority if
3671 * enough pages are already being scanned that that high
3672 * watermark would be met at 100% efficiency.
3674 if (kswapd_shrink_node(pgdat, &sc))
3675 raise_priority = false;
3678 * If the low watermark is met there is no need for processes
3679 * to be throttled on pfmemalloc_wait as they should not be
3680 * able to safely make forward progress. Wake them
3682 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3683 allow_direct_reclaim(pgdat))
3684 wake_up_all(&pgdat->pfmemalloc_wait);
3686 /* Check if kswapd should be suspending */
3687 __fs_reclaim_release();
3688 ret = try_to_freeze();
3689 __fs_reclaim_acquire();
3690 if (ret || kthread_should_stop())
3694 * Raise priority if scanning rate is too low or there was no
3695 * progress in reclaiming pages
3697 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3698 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3701 * If reclaim made no progress for a boost, stop reclaim as
3702 * IO cannot be queued and it could be an infinite loop in
3703 * extreme circumstances.
3705 if (nr_boost_reclaim && !nr_reclaimed)
3708 if (raise_priority || !nr_reclaimed)
3710 } while (sc.priority >= 1);
3712 if (!sc.nr_reclaimed)
3713 pgdat->kswapd_failures++;
3716 /* If reclaim was boosted, account for the reclaim done in this pass */
3718 unsigned long flags;
3720 for (i = 0; i <= classzone_idx; i++) {
3721 if (!zone_boosts[i])
3724 /* Increments are under the zone lock */
3725 zone = pgdat->node_zones + i;
3726 spin_lock_irqsave(&zone->lock, flags);
3727 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3728 spin_unlock_irqrestore(&zone->lock, flags);
3732 * As there is now likely space, wakeup kcompact to defragment
3735 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3738 snapshot_refaults(NULL, pgdat);
3739 __fs_reclaim_release();
3740 psi_memstall_leave(&pflags);
3742 * Return the order kswapd stopped reclaiming at as
3743 * prepare_kswapd_sleep() takes it into account. If another caller
3744 * entered the allocator slow path while kswapd was awake, order will
3745 * remain at the higher level.
3751 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3752 * allocation request woke kswapd for. When kswapd has not woken recently,
3753 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3754 * given classzone and returns it or the highest classzone index kswapd
3755 * was recently woke for.
3757 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3758 enum zone_type classzone_idx)
3760 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3761 return classzone_idx;
3763 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3766 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3767 unsigned int classzone_idx)
3772 if (freezing(current) || kthread_should_stop())
3775 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3778 * Try to sleep for a short interval. Note that kcompactd will only be
3779 * woken if it is possible to sleep for a short interval. This is
3780 * deliberate on the assumption that if reclaim cannot keep an
3781 * eligible zone balanced that it's also unlikely that compaction will
3784 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3786 * Compaction records what page blocks it recently failed to
3787 * isolate pages from and skips them in the future scanning.
3788 * When kswapd is going to sleep, it is reasonable to assume
3789 * that pages and compaction may succeed so reset the cache.
3791 reset_isolation_suitable(pgdat);
3794 * We have freed the memory, now we should compact it to make
3795 * allocation of the requested order possible.
3797 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3799 remaining = schedule_timeout(HZ/10);
3802 * If woken prematurely then reset kswapd_classzone_idx and
3803 * order. The values will either be from a wakeup request or
3804 * the previous request that slept prematurely.
3807 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3808 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3811 finish_wait(&pgdat->kswapd_wait, &wait);
3812 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3816 * After a short sleep, check if it was a premature sleep. If not, then
3817 * go fully to sleep until explicitly woken up.
3820 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3821 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3824 * vmstat counters are not perfectly accurate and the estimated
3825 * value for counters such as NR_FREE_PAGES can deviate from the
3826 * true value by nr_online_cpus * threshold. To avoid the zone
3827 * watermarks being breached while under pressure, we reduce the
3828 * per-cpu vmstat threshold while kswapd is awake and restore
3829 * them before going back to sleep.
3831 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3833 if (!kthread_should_stop())
3836 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3839 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3841 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3843 finish_wait(&pgdat->kswapd_wait, &wait);
3847 * The background pageout daemon, started as a kernel thread
3848 * from the init process.
3850 * This basically trickles out pages so that we have _some_
3851 * free memory available even if there is no other activity
3852 * that frees anything up. This is needed for things like routing
3853 * etc, where we otherwise might have all activity going on in
3854 * asynchronous contexts that cannot page things out.
3856 * If there are applications that are active memory-allocators
3857 * (most normal use), this basically shouldn't matter.
3859 static int kswapd(void *p)
3861 unsigned int alloc_order, reclaim_order;
3862 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3863 pg_data_t *pgdat = (pg_data_t*)p;
3864 struct task_struct *tsk = current;
3866 struct reclaim_state reclaim_state = {
3867 .reclaimed_slab = 0,
3869 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3871 if (!cpumask_empty(cpumask))
3872 set_cpus_allowed_ptr(tsk, cpumask);
3873 current->reclaim_state = &reclaim_state;
3876 * Tell the memory management that we're a "memory allocator",
3877 * and that if we need more memory we should get access to it
3878 * regardless (see "__alloc_pages()"). "kswapd" should
3879 * never get caught in the normal page freeing logic.
3881 * (Kswapd normally doesn't need memory anyway, but sometimes
3882 * you need a small amount of memory in order to be able to
3883 * page out something else, and this flag essentially protects
3884 * us from recursively trying to free more memory as we're
3885 * trying to free the first piece of memory in the first place).
3887 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3890 pgdat->kswapd_order = 0;
3891 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3895 alloc_order = reclaim_order = pgdat->kswapd_order;
3896 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3899 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3902 /* Read the new order and classzone_idx */
3903 alloc_order = reclaim_order = pgdat->kswapd_order;
3904 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3905 pgdat->kswapd_order = 0;
3906 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3908 ret = try_to_freeze();
3909 if (kthread_should_stop())
3913 * We can speed up thawing tasks if we don't call balance_pgdat
3914 * after returning from the refrigerator
3920 * Reclaim begins at the requested order but if a high-order
3921 * reclaim fails then kswapd falls back to reclaiming for
3922 * order-0. If that happens, kswapd will consider sleeping
3923 * for the order it finished reclaiming at (reclaim_order)
3924 * but kcompactd is woken to compact for the original
3925 * request (alloc_order).
3927 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3929 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3930 if (reclaim_order < alloc_order)
3931 goto kswapd_try_sleep;
3934 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3935 current->reclaim_state = NULL;
3941 * A zone is low on free memory or too fragmented for high-order memory. If
3942 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3943 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3944 * has failed or is not needed, still wake up kcompactd if only compaction is
3947 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3948 enum zone_type classzone_idx)
3952 if (!managed_zone(zone))
3955 if (!cpuset_zone_allowed(zone, gfp_flags))
3957 pgdat = zone->zone_pgdat;
3958 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3960 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3961 if (!waitqueue_active(&pgdat->kswapd_wait))
3964 /* Hopeless node, leave it to direct reclaim if possible */
3965 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3966 (pgdat_balanced(pgdat, order, classzone_idx) &&
3967 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3969 * There may be plenty of free memory available, but it's too
3970 * fragmented for high-order allocations. Wake up kcompactd
3971 * and rely on compaction_suitable() to determine if it's
3972 * needed. If it fails, it will defer subsequent attempts to
3973 * ratelimit its work.
3975 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3976 wakeup_kcompactd(pgdat, order, classzone_idx);
3980 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3982 wake_up_interruptible(&pgdat->kswapd_wait);
3985 #ifdef CONFIG_HIBERNATION
3987 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3990 * Rather than trying to age LRUs the aim is to preserve the overall
3991 * LRU order by reclaiming preferentially
3992 * inactive > active > active referenced > active mapped
3994 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3996 struct reclaim_state reclaim_state;
3997 struct scan_control sc = {
3998 .nr_to_reclaim = nr_to_reclaim,
3999 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4000 .reclaim_idx = MAX_NR_ZONES - 1,
4001 .priority = DEF_PRIORITY,
4005 .hibernation_mode = 1,
4007 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4008 struct task_struct *p = current;
4009 unsigned long nr_reclaimed;
4010 unsigned int noreclaim_flag;
4012 fs_reclaim_acquire(sc.gfp_mask);
4013 noreclaim_flag = memalloc_noreclaim_save();
4014 reclaim_state.reclaimed_slab = 0;
4015 p->reclaim_state = &reclaim_state;
4017 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4019 p->reclaim_state = NULL;
4020 memalloc_noreclaim_restore(noreclaim_flag);
4021 fs_reclaim_release(sc.gfp_mask);
4023 return nr_reclaimed;
4025 #endif /* CONFIG_HIBERNATION */
4027 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4028 not required for correctness. So if the last cpu in a node goes
4029 away, we get changed to run anywhere: as the first one comes back,
4030 restore their cpu bindings. */
4031 static int kswapd_cpu_online(unsigned int cpu)
4035 for_each_node_state(nid, N_MEMORY) {
4036 pg_data_t *pgdat = NODE_DATA(nid);
4037 const struct cpumask *mask;
4039 mask = cpumask_of_node(pgdat->node_id);
4041 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4042 /* One of our CPUs online: restore mask */
4043 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4049 * This kswapd start function will be called by init and node-hot-add.
4050 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4052 int kswapd_run(int nid)
4054 pg_data_t *pgdat = NODE_DATA(nid);
4060 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4061 if (IS_ERR(pgdat->kswapd)) {
4062 /* failure at boot is fatal */
4063 BUG_ON(system_state < SYSTEM_RUNNING);
4064 pr_err("Failed to start kswapd on node %d\n", nid);
4065 ret = PTR_ERR(pgdat->kswapd);
4066 pgdat->kswapd = NULL;
4072 * Called by memory hotplug when all memory in a node is offlined. Caller must
4073 * hold mem_hotplug_begin/end().
4075 void kswapd_stop(int nid)
4077 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4080 kthread_stop(kswapd);
4081 NODE_DATA(nid)->kswapd = NULL;
4085 static int __init kswapd_init(void)
4090 for_each_node_state(nid, N_MEMORY)
4092 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4093 "mm/vmscan:online", kswapd_cpu_online,
4099 module_init(kswapd_init)
4105 * If non-zero call node_reclaim when the number of free pages falls below
4108 int node_reclaim_mode __read_mostly;
4110 #define RECLAIM_OFF 0
4111 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4112 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4113 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4116 * Priority for NODE_RECLAIM. This determines the fraction of pages
4117 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4120 #define NODE_RECLAIM_PRIORITY 4
4123 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4126 int sysctl_min_unmapped_ratio = 1;
4129 * If the number of slab pages in a zone grows beyond this percentage then
4130 * slab reclaim needs to occur.
4132 int sysctl_min_slab_ratio = 5;
4134 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4136 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4137 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4138 node_page_state(pgdat, NR_ACTIVE_FILE);
4141 * It's possible for there to be more file mapped pages than
4142 * accounted for by the pages on the file LRU lists because
4143 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4145 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4148 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4149 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4151 unsigned long nr_pagecache_reclaimable;
4152 unsigned long delta = 0;
4155 * If RECLAIM_UNMAP is set, then all file pages are considered
4156 * potentially reclaimable. Otherwise, we have to worry about
4157 * pages like swapcache and node_unmapped_file_pages() provides
4160 if (node_reclaim_mode & RECLAIM_UNMAP)
4161 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4163 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4165 /* If we can't clean pages, remove dirty pages from consideration */
4166 if (!(node_reclaim_mode & RECLAIM_WRITE))
4167 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4169 /* Watch for any possible underflows due to delta */
4170 if (unlikely(delta > nr_pagecache_reclaimable))
4171 delta = nr_pagecache_reclaimable;
4173 return nr_pagecache_reclaimable - delta;
4177 * Try to free up some pages from this node through reclaim.
4179 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4181 /* Minimum pages needed in order to stay on node */
4182 const unsigned long nr_pages = 1 << order;
4183 struct task_struct *p = current;
4184 struct reclaim_state reclaim_state;
4185 unsigned int noreclaim_flag;
4186 struct scan_control sc = {
4187 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4188 .gfp_mask = current_gfp_context(gfp_mask),
4190 .priority = NODE_RECLAIM_PRIORITY,
4191 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4192 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4194 .reclaim_idx = gfp_zone(gfp_mask),
4198 fs_reclaim_acquire(sc.gfp_mask);
4200 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4201 * and we also need to be able to write out pages for RECLAIM_WRITE
4202 * and RECLAIM_UNMAP.
4204 noreclaim_flag = memalloc_noreclaim_save();
4205 p->flags |= PF_SWAPWRITE;
4206 reclaim_state.reclaimed_slab = 0;
4207 p->reclaim_state = &reclaim_state;
4209 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4211 * Free memory by calling shrink node with increasing
4212 * priorities until we have enough memory freed.
4215 shrink_node(pgdat, &sc);
4216 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4219 p->reclaim_state = NULL;
4220 current->flags &= ~PF_SWAPWRITE;
4221 memalloc_noreclaim_restore(noreclaim_flag);
4222 fs_reclaim_release(sc.gfp_mask);
4223 return sc.nr_reclaimed >= nr_pages;
4226 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4231 * Node reclaim reclaims unmapped file backed pages and
4232 * slab pages if we are over the defined limits.
4234 * A small portion of unmapped file backed pages is needed for
4235 * file I/O otherwise pages read by file I/O will be immediately
4236 * thrown out if the node is overallocated. So we do not reclaim
4237 * if less than a specified percentage of the node is used by
4238 * unmapped file backed pages.
4240 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4241 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4242 return NODE_RECLAIM_FULL;
4245 * Do not scan if the allocation should not be delayed.
4247 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4248 return NODE_RECLAIM_NOSCAN;
4251 * Only run node reclaim on the local node or on nodes that do not
4252 * have associated processors. This will favor the local processor
4253 * over remote processors and spread off node memory allocations
4254 * as wide as possible.
4256 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4257 return NODE_RECLAIM_NOSCAN;
4259 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4260 return NODE_RECLAIM_NOSCAN;
4262 ret = __node_reclaim(pgdat, gfp_mask, order);
4263 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4266 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4273 * page_evictable - test whether a page is evictable
4274 * @page: the page to test
4276 * Test whether page is evictable--i.e., should be placed on active/inactive
4277 * lists vs unevictable list.
4279 * Reasons page might not be evictable:
4280 * (1) page's mapping marked unevictable
4281 * (2) page is part of an mlocked VMA
4284 int page_evictable(struct page *page)
4288 /* Prevent address_space of inode and swap cache from being freed */
4290 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4296 * check_move_unevictable_pages - check pages for evictability and move to
4297 * appropriate zone lru list
4298 * @pvec: pagevec with lru pages to check
4300 * Checks pages for evictability, if an evictable page is in the unevictable
4301 * lru list, moves it to the appropriate evictable lru list. This function
4302 * should be only used for lru pages.
4304 void check_move_unevictable_pages(struct pagevec *pvec)
4306 struct lruvec *lruvec;
4307 struct pglist_data *pgdat = NULL;
4312 for (i = 0; i < pvec->nr; i++) {
4313 struct page *page = pvec->pages[i];
4314 struct pglist_data *pagepgdat = page_pgdat(page);
4317 if (pagepgdat != pgdat) {
4319 spin_unlock_irq(&pgdat->lru_lock);
4321 spin_lock_irq(&pgdat->lru_lock);
4323 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4325 if (!PageLRU(page) || !PageUnevictable(page))
4328 if (page_evictable(page)) {
4329 enum lru_list lru = page_lru_base_type(page);
4331 VM_BUG_ON_PAGE(PageActive(page), page);
4332 ClearPageUnevictable(page);
4333 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4334 add_page_to_lru_list(page, lruvec, lru);
4340 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4341 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4342 spin_unlock_irq(&pgdat->lru_lock);
4345 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);