1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup *target_mem_cgroup;
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
92 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
96 unsigned int memcg_low_reclaim:1;
97 unsigned int memcg_low_skipped:1;
99 unsigned int hibernation_mode:1;
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready:1;
104 /* The file pages on the current node are dangerously low */
105 unsigned int file_is_tiny:1;
107 /* Allocation order */
110 /* Scan (total_size >> priority) pages at once */
113 /* The highest zone to isolate pages for reclaim from */
116 /* This context's GFP mask */
119 /* Incremented by the number of inactive pages that were scanned */
120 unsigned long nr_scanned;
122 /* Number of pages freed so far during a call to shrink_zones() */
123 unsigned long nr_reclaimed;
127 unsigned int unqueued_dirty;
128 unsigned int congested;
129 unsigned int writeback;
130 unsigned int immediate;
131 unsigned int file_taken;
135 /* for recording the reclaimed slab by now */
136 struct reclaim_state reclaim_state;
139 #ifdef ARCH_HAS_PREFETCH
140 #define prefetch_prev_lru_page(_page, _base, _field) \
142 if ((_page)->lru.prev != _base) { \
145 prev = lru_to_page(&(_page->lru)); \
146 prefetch(&prev->_field); \
150 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
153 #ifdef ARCH_HAS_PREFETCHW
154 #define prefetchw_prev_lru_page(_page, _base, _field) \
156 if ((_page)->lru.prev != _base) { \
159 prev = lru_to_page(&(_page->lru)); \
160 prefetchw(&prev->_field); \
164 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168 * From 0 .. 100. Higher means more swappy.
170 int vm_swappiness = 60;
172 * The total number of pages which are beyond the high watermark within all
175 unsigned long vm_total_pages;
177 static void set_task_reclaim_state(struct task_struct *task,
178 struct reclaim_state *rs)
180 /* Check for an overwrite */
181 WARN_ON_ONCE(rs && task->reclaim_state);
183 /* Check for the nulling of an already-nulled member */
184 WARN_ON_ONCE(!rs && !task->reclaim_state);
186 task->reclaim_state = rs;
189 static LIST_HEAD(shrinker_list);
190 static DECLARE_RWSEM(shrinker_rwsem);
194 * We allow subsystems to populate their shrinker-related
195 * LRU lists before register_shrinker_prepared() is called
196 * for the shrinker, since we don't want to impose
197 * restrictions on their internal registration order.
198 * In this case shrink_slab_memcg() may find corresponding
199 * bit is set in the shrinkers map.
201 * This value is used by the function to detect registering
202 * shrinkers and to skip do_shrink_slab() calls for them.
204 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
206 static DEFINE_IDR(shrinker_idr);
207 static int shrinker_nr_max;
209 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
211 int id, ret = -ENOMEM;
213 down_write(&shrinker_rwsem);
214 /* This may call shrinker, so it must use down_read_trylock() */
215 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
219 if (id >= shrinker_nr_max) {
220 if (memcg_expand_shrinker_maps(id)) {
221 idr_remove(&shrinker_idr, id);
225 shrinker_nr_max = id + 1;
230 up_write(&shrinker_rwsem);
234 static void unregister_memcg_shrinker(struct shrinker *shrinker)
236 int id = shrinker->id;
240 down_write(&shrinker_rwsem);
241 idr_remove(&shrinker_idr, id);
242 up_write(&shrinker_rwsem);
245 static bool cgroup_reclaim(struct scan_control *sc)
247 return sc->target_mem_cgroup;
251 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
252 * @sc: scan_control in question
254 * The normal page dirty throttling mechanism in balance_dirty_pages() is
255 * completely broken with the legacy memcg and direct stalling in
256 * shrink_page_list() is used for throttling instead, which lacks all the
257 * niceties such as fairness, adaptive pausing, bandwidth proportional
258 * allocation and configurability.
260 * This function tests whether the vmscan currently in progress can assume
261 * that the normal dirty throttling mechanism is operational.
263 static bool writeback_throttling_sane(struct scan_control *sc)
265 if (!cgroup_reclaim(sc))
267 #ifdef CONFIG_CGROUP_WRITEBACK
268 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
274 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
279 static void unregister_memcg_shrinker(struct shrinker *shrinker)
283 static bool cgroup_reclaim(struct scan_control *sc)
288 static bool writeback_throttling_sane(struct scan_control *sc)
295 * This misses isolated pages which are not accounted for to save counters.
296 * As the data only determines if reclaim or compaction continues, it is
297 * not expected that isolated pages will be a dominating factor.
299 unsigned long zone_reclaimable_pages(struct zone *zone)
303 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
304 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
305 if (get_nr_swap_pages() > 0)
306 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
307 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
313 * lruvec_lru_size - Returns the number of pages on the given LRU list.
314 * @lruvec: lru vector
316 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
318 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
320 unsigned long size = 0;
323 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
324 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
326 if (!managed_zone(zone))
329 if (!mem_cgroup_disabled())
330 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
332 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
338 * Add a shrinker callback to be called from the vm.
340 int prealloc_shrinker(struct shrinker *shrinker)
342 unsigned int size = sizeof(*shrinker->nr_deferred);
344 if (shrinker->flags & SHRINKER_NUMA_AWARE)
347 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
348 if (!shrinker->nr_deferred)
351 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
352 if (prealloc_memcg_shrinker(shrinker))
359 kfree(shrinker->nr_deferred);
360 shrinker->nr_deferred = NULL;
364 void free_prealloced_shrinker(struct shrinker *shrinker)
366 if (!shrinker->nr_deferred)
369 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
370 unregister_memcg_shrinker(shrinker);
372 kfree(shrinker->nr_deferred);
373 shrinker->nr_deferred = NULL;
376 void register_shrinker_prepared(struct shrinker *shrinker)
378 down_write(&shrinker_rwsem);
379 list_add_tail(&shrinker->list, &shrinker_list);
381 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
382 idr_replace(&shrinker_idr, shrinker, shrinker->id);
384 up_write(&shrinker_rwsem);
387 int register_shrinker(struct shrinker *shrinker)
389 int err = prealloc_shrinker(shrinker);
393 register_shrinker_prepared(shrinker);
396 EXPORT_SYMBOL(register_shrinker);
401 void unregister_shrinker(struct shrinker *shrinker)
403 if (!shrinker->nr_deferred)
405 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
406 unregister_memcg_shrinker(shrinker);
407 down_write(&shrinker_rwsem);
408 list_del(&shrinker->list);
409 up_write(&shrinker_rwsem);
410 kfree(shrinker->nr_deferred);
411 shrinker->nr_deferred = NULL;
413 EXPORT_SYMBOL(unregister_shrinker);
415 #define SHRINK_BATCH 128
417 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
418 struct shrinker *shrinker, int priority)
420 unsigned long freed = 0;
421 unsigned long long delta;
426 int nid = shrinkctl->nid;
427 long batch_size = shrinker->batch ? shrinker->batch
429 long scanned = 0, next_deferred;
431 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
434 freeable = shrinker->count_objects(shrinker, shrinkctl);
435 if (freeable == 0 || freeable == SHRINK_EMPTY)
439 * copy the current shrinker scan count into a local variable
440 * and zero it so that other concurrent shrinker invocations
441 * don't also do this scanning work.
443 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
446 if (shrinker->seeks) {
447 delta = freeable >> priority;
449 do_div(delta, shrinker->seeks);
452 * These objects don't require any IO to create. Trim
453 * them aggressively under memory pressure to keep
454 * them from causing refetches in the IO caches.
456 delta = freeable / 2;
460 if (total_scan < 0) {
461 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
462 shrinker->scan_objects, total_scan);
463 total_scan = freeable;
466 next_deferred = total_scan;
469 * We need to avoid excessive windup on filesystem shrinkers
470 * due to large numbers of GFP_NOFS allocations causing the
471 * shrinkers to return -1 all the time. This results in a large
472 * nr being built up so when a shrink that can do some work
473 * comes along it empties the entire cache due to nr >>>
474 * freeable. This is bad for sustaining a working set in
477 * Hence only allow the shrinker to scan the entire cache when
478 * a large delta change is calculated directly.
480 if (delta < freeable / 4)
481 total_scan = min(total_scan, freeable / 2);
484 * Avoid risking looping forever due to too large nr value:
485 * never try to free more than twice the estimate number of
488 if (total_scan > freeable * 2)
489 total_scan = freeable * 2;
491 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
492 freeable, delta, total_scan, priority);
495 * Normally, we should not scan less than batch_size objects in one
496 * pass to avoid too frequent shrinker calls, but if the slab has less
497 * than batch_size objects in total and we are really tight on memory,
498 * we will try to reclaim all available objects, otherwise we can end
499 * up failing allocations although there are plenty of reclaimable
500 * objects spread over several slabs with usage less than the
503 * We detect the "tight on memory" situations by looking at the total
504 * number of objects we want to scan (total_scan). If it is greater
505 * than the total number of objects on slab (freeable), we must be
506 * scanning at high prio and therefore should try to reclaim as much as
509 while (total_scan >= batch_size ||
510 total_scan >= freeable) {
512 unsigned long nr_to_scan = min(batch_size, total_scan);
514 shrinkctl->nr_to_scan = nr_to_scan;
515 shrinkctl->nr_scanned = nr_to_scan;
516 ret = shrinker->scan_objects(shrinker, shrinkctl);
517 if (ret == SHRINK_STOP)
521 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
522 total_scan -= shrinkctl->nr_scanned;
523 scanned += shrinkctl->nr_scanned;
528 if (next_deferred >= scanned)
529 next_deferred -= scanned;
533 * move the unused scan count back into the shrinker in a
534 * manner that handles concurrent updates. If we exhausted the
535 * scan, there is no need to do an update.
537 if (next_deferred > 0)
538 new_nr = atomic_long_add_return(next_deferred,
539 &shrinker->nr_deferred[nid]);
541 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
543 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
548 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
549 struct mem_cgroup *memcg, int priority)
551 struct memcg_shrinker_map *map;
552 unsigned long ret, freed = 0;
555 if (!mem_cgroup_online(memcg))
558 if (!down_read_trylock(&shrinker_rwsem))
561 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
566 for_each_set_bit(i, map->map, shrinker_nr_max) {
567 struct shrink_control sc = {
568 .gfp_mask = gfp_mask,
572 struct shrinker *shrinker;
574 shrinker = idr_find(&shrinker_idr, i);
575 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
577 clear_bit(i, map->map);
581 /* Call non-slab shrinkers even though kmem is disabled */
582 if (!memcg_kmem_enabled() &&
583 !(shrinker->flags & SHRINKER_NONSLAB))
586 ret = do_shrink_slab(&sc, shrinker, priority);
587 if (ret == SHRINK_EMPTY) {
588 clear_bit(i, map->map);
590 * After the shrinker reported that it had no objects to
591 * free, but before we cleared the corresponding bit in
592 * the memcg shrinker map, a new object might have been
593 * added. To make sure, we have the bit set in this
594 * case, we invoke the shrinker one more time and reset
595 * the bit if it reports that it is not empty anymore.
596 * The memory barrier here pairs with the barrier in
597 * memcg_set_shrinker_bit():
599 * list_lru_add() shrink_slab_memcg()
600 * list_add_tail() clear_bit()
602 * set_bit() do_shrink_slab()
604 smp_mb__after_atomic();
605 ret = do_shrink_slab(&sc, shrinker, priority);
606 if (ret == SHRINK_EMPTY)
609 memcg_set_shrinker_bit(memcg, nid, i);
613 if (rwsem_is_contended(&shrinker_rwsem)) {
619 up_read(&shrinker_rwsem);
622 #else /* CONFIG_MEMCG */
623 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
624 struct mem_cgroup *memcg, int priority)
628 #endif /* CONFIG_MEMCG */
631 * shrink_slab - shrink slab caches
632 * @gfp_mask: allocation context
633 * @nid: node whose slab caches to target
634 * @memcg: memory cgroup whose slab caches to target
635 * @priority: the reclaim priority
637 * Call the shrink functions to age shrinkable caches.
639 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
640 * unaware shrinkers will receive a node id of 0 instead.
642 * @memcg specifies the memory cgroup to target. Unaware shrinkers
643 * are called only if it is the root cgroup.
645 * @priority is sc->priority, we take the number of objects and >> by priority
646 * in order to get the scan target.
648 * Returns the number of reclaimed slab objects.
650 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
651 struct mem_cgroup *memcg,
654 unsigned long ret, freed = 0;
655 struct shrinker *shrinker;
658 * The root memcg might be allocated even though memcg is disabled
659 * via "cgroup_disable=memory" boot parameter. This could make
660 * mem_cgroup_is_root() return false, then just run memcg slab
661 * shrink, but skip global shrink. This may result in premature
664 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
665 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
667 if (!down_read_trylock(&shrinker_rwsem))
670 list_for_each_entry(shrinker, &shrinker_list, list) {
671 struct shrink_control sc = {
672 .gfp_mask = gfp_mask,
677 ret = do_shrink_slab(&sc, shrinker, priority);
678 if (ret == SHRINK_EMPTY)
682 * Bail out if someone want to register a new shrinker to
683 * prevent the regsitration from being stalled for long periods
684 * by parallel ongoing shrinking.
686 if (rwsem_is_contended(&shrinker_rwsem)) {
692 up_read(&shrinker_rwsem);
698 void drop_slab_node(int nid)
703 struct mem_cgroup *memcg = NULL;
706 memcg = mem_cgroup_iter(NULL, NULL, NULL);
708 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
709 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
710 } while (freed > 10);
717 for_each_online_node(nid)
721 static inline int is_page_cache_freeable(struct page *page)
724 * A freeable page cache page is referenced only by the caller
725 * that isolated the page, the page cache and optional buffer
726 * heads at page->private.
728 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
730 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
733 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
735 if (current->flags & PF_SWAPWRITE)
737 if (!inode_write_congested(inode))
739 if (inode_to_bdi(inode) == current->backing_dev_info)
745 * We detected a synchronous write error writing a page out. Probably
746 * -ENOSPC. We need to propagate that into the address_space for a subsequent
747 * fsync(), msync() or close().
749 * The tricky part is that after writepage we cannot touch the mapping: nothing
750 * prevents it from being freed up. But we have a ref on the page and once
751 * that page is locked, the mapping is pinned.
753 * We're allowed to run sleeping lock_page() here because we know the caller has
756 static void handle_write_error(struct address_space *mapping,
757 struct page *page, int error)
760 if (page_mapping(page) == mapping)
761 mapping_set_error(mapping, error);
765 /* possible outcome of pageout() */
767 /* failed to write page out, page is locked */
769 /* move page to the active list, page is locked */
771 /* page has been sent to the disk successfully, page is unlocked */
773 /* page is clean and locked */
778 * pageout is called by shrink_page_list() for each dirty page.
779 * Calls ->writepage().
781 static pageout_t pageout(struct page *page, struct address_space *mapping,
782 struct scan_control *sc)
785 * If the page is dirty, only perform writeback if that write
786 * will be non-blocking. To prevent this allocation from being
787 * stalled by pagecache activity. But note that there may be
788 * stalls if we need to run get_block(). We could test
789 * PagePrivate for that.
791 * If this process is currently in __generic_file_write_iter() against
792 * this page's queue, we can perform writeback even if that
795 * If the page is swapcache, write it back even if that would
796 * block, for some throttling. This happens by accident, because
797 * swap_backing_dev_info is bust: it doesn't reflect the
798 * congestion state of the swapdevs. Easy to fix, if needed.
800 if (!is_page_cache_freeable(page))
804 * Some data journaling orphaned pages can have
805 * page->mapping == NULL while being dirty with clean buffers.
807 if (page_has_private(page)) {
808 if (try_to_free_buffers(page)) {
809 ClearPageDirty(page);
810 pr_info("%s: orphaned page\n", __func__);
816 if (mapping->a_ops->writepage == NULL)
817 return PAGE_ACTIVATE;
818 if (!may_write_to_inode(mapping->host, sc))
821 if (clear_page_dirty_for_io(page)) {
823 struct writeback_control wbc = {
824 .sync_mode = WB_SYNC_NONE,
825 .nr_to_write = SWAP_CLUSTER_MAX,
827 .range_end = LLONG_MAX,
831 SetPageReclaim(page);
832 res = mapping->a_ops->writepage(page, &wbc);
834 handle_write_error(mapping, page, res);
835 if (res == AOP_WRITEPAGE_ACTIVATE) {
836 ClearPageReclaim(page);
837 return PAGE_ACTIVATE;
840 if (!PageWriteback(page)) {
841 /* synchronous write or broken a_ops? */
842 ClearPageReclaim(page);
844 trace_mm_vmscan_writepage(page);
845 inc_node_page_state(page, NR_VMSCAN_WRITE);
853 * Same as remove_mapping, but if the page is removed from the mapping, it
854 * gets returned with a refcount of 0.
856 static int __remove_mapping(struct address_space *mapping, struct page *page,
857 bool reclaimed, struct mem_cgroup *target_memcg)
862 BUG_ON(!PageLocked(page));
863 BUG_ON(mapping != page_mapping(page));
865 xa_lock_irqsave(&mapping->i_pages, flags);
867 * The non racy check for a busy page.
869 * Must be careful with the order of the tests. When someone has
870 * a ref to the page, it may be possible that they dirty it then
871 * drop the reference. So if PageDirty is tested before page_count
872 * here, then the following race may occur:
874 * get_user_pages(&page);
875 * [user mapping goes away]
877 * !PageDirty(page) [good]
878 * SetPageDirty(page);
880 * !page_count(page) [good, discard it]
882 * [oops, our write_to data is lost]
884 * Reversing the order of the tests ensures such a situation cannot
885 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
886 * load is not satisfied before that of page->_refcount.
888 * Note that if SetPageDirty is always performed via set_page_dirty,
889 * and thus under the i_pages lock, then this ordering is not required.
891 refcount = 1 + compound_nr(page);
892 if (!page_ref_freeze(page, refcount))
894 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
895 if (unlikely(PageDirty(page))) {
896 page_ref_unfreeze(page, refcount);
900 if (PageSwapCache(page)) {
901 swp_entry_t swap = { .val = page_private(page) };
902 mem_cgroup_swapout(page, swap);
903 __delete_from_swap_cache(page, swap);
904 xa_unlock_irqrestore(&mapping->i_pages, flags);
905 put_swap_page(page, swap);
907 void (*freepage)(struct page *);
910 freepage = mapping->a_ops->freepage;
912 * Remember a shadow entry for reclaimed file cache in
913 * order to detect refaults, thus thrashing, later on.
915 * But don't store shadows in an address space that is
916 * already exiting. This is not just an optizimation,
917 * inode reclaim needs to empty out the radix tree or
918 * the nodes are lost. Don't plant shadows behind its
921 * We also don't store shadows for DAX mappings because the
922 * only page cache pages found in these are zero pages
923 * covering holes, and because we don't want to mix DAX
924 * exceptional entries and shadow exceptional entries in the
925 * same address_space.
927 if (reclaimed && page_is_file_cache(page) &&
928 !mapping_exiting(mapping) && !dax_mapping(mapping))
929 shadow = workingset_eviction(page, target_memcg);
930 __delete_from_page_cache(page, shadow);
931 xa_unlock_irqrestore(&mapping->i_pages, flags);
933 if (freepage != NULL)
940 xa_unlock_irqrestore(&mapping->i_pages, flags);
945 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
946 * someone else has a ref on the page, abort and return 0. If it was
947 * successfully detached, return 1. Assumes the caller has a single ref on
950 int remove_mapping(struct address_space *mapping, struct page *page)
952 if (__remove_mapping(mapping, page, false, NULL)) {
954 * Unfreezing the refcount with 1 rather than 2 effectively
955 * drops the pagecache ref for us without requiring another
958 page_ref_unfreeze(page, 1);
965 * putback_lru_page - put previously isolated page onto appropriate LRU list
966 * @page: page to be put back to appropriate lru list
968 * Add previously isolated @page to appropriate LRU list.
969 * Page may still be unevictable for other reasons.
971 * lru_lock must not be held, interrupts must be enabled.
973 void putback_lru_page(struct page *page)
976 put_page(page); /* drop ref from isolate */
979 enum page_references {
981 PAGEREF_RECLAIM_CLEAN,
986 static enum page_references page_check_references(struct page *page,
987 struct scan_control *sc)
989 int referenced_ptes, referenced_page;
990 unsigned long vm_flags;
992 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
994 referenced_page = TestClearPageReferenced(page);
997 * Mlock lost the isolation race with us. Let try_to_unmap()
998 * move the page to the unevictable list.
1000 if (vm_flags & VM_LOCKED)
1001 return PAGEREF_RECLAIM;
1003 if (referenced_ptes) {
1004 if (PageSwapBacked(page))
1005 return PAGEREF_ACTIVATE;
1007 * All mapped pages start out with page table
1008 * references from the instantiating fault, so we need
1009 * to look twice if a mapped file page is used more
1012 * Mark it and spare it for another trip around the
1013 * inactive list. Another page table reference will
1014 * lead to its activation.
1016 * Note: the mark is set for activated pages as well
1017 * so that recently deactivated but used pages are
1018 * quickly recovered.
1020 SetPageReferenced(page);
1022 if (referenced_page || referenced_ptes > 1)
1023 return PAGEREF_ACTIVATE;
1026 * Activate file-backed executable pages after first usage.
1028 if (vm_flags & VM_EXEC)
1029 return PAGEREF_ACTIVATE;
1031 return PAGEREF_KEEP;
1034 /* Reclaim if clean, defer dirty pages to writeback */
1035 if (referenced_page && !PageSwapBacked(page))
1036 return PAGEREF_RECLAIM_CLEAN;
1038 return PAGEREF_RECLAIM;
1041 /* Check if a page is dirty or under writeback */
1042 static void page_check_dirty_writeback(struct page *page,
1043 bool *dirty, bool *writeback)
1045 struct address_space *mapping;
1048 * Anonymous pages are not handled by flushers and must be written
1049 * from reclaim context. Do not stall reclaim based on them
1051 if (!page_is_file_cache(page) ||
1052 (PageAnon(page) && !PageSwapBacked(page))) {
1058 /* By default assume that the page flags are accurate */
1059 *dirty = PageDirty(page);
1060 *writeback = PageWriteback(page);
1062 /* Verify dirty/writeback state if the filesystem supports it */
1063 if (!page_has_private(page))
1066 mapping = page_mapping(page);
1067 if (mapping && mapping->a_ops->is_dirty_writeback)
1068 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1072 * shrink_page_list() returns the number of reclaimed pages
1074 static unsigned long shrink_page_list(struct list_head *page_list,
1075 struct pglist_data *pgdat,
1076 struct scan_control *sc,
1077 enum ttu_flags ttu_flags,
1078 struct reclaim_stat *stat,
1079 bool ignore_references)
1081 LIST_HEAD(ret_pages);
1082 LIST_HEAD(free_pages);
1083 unsigned nr_reclaimed = 0;
1084 unsigned pgactivate = 0;
1086 memset(stat, 0, sizeof(*stat));
1089 while (!list_empty(page_list)) {
1090 struct address_space *mapping;
1093 enum page_references references = PAGEREF_RECLAIM;
1094 bool dirty, writeback;
1095 unsigned int nr_pages;
1099 page = lru_to_page(page_list);
1100 list_del(&page->lru);
1102 if (!trylock_page(page))
1105 VM_BUG_ON_PAGE(PageActive(page), page);
1107 nr_pages = compound_nr(page);
1109 /* Account the number of base pages even though THP */
1110 sc->nr_scanned += nr_pages;
1112 if (unlikely(!page_evictable(page)))
1113 goto activate_locked;
1115 if (!sc->may_unmap && page_mapped(page))
1118 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1119 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1122 * The number of dirty pages determines if a node is marked
1123 * reclaim_congested which affects wait_iff_congested. kswapd
1124 * will stall and start writing pages if the tail of the LRU
1125 * is all dirty unqueued pages.
1127 page_check_dirty_writeback(page, &dirty, &writeback);
1128 if (dirty || writeback)
1131 if (dirty && !writeback)
1132 stat->nr_unqueued_dirty++;
1135 * Treat this page as congested if the underlying BDI is or if
1136 * pages are cycling through the LRU so quickly that the
1137 * pages marked for immediate reclaim are making it to the
1138 * end of the LRU a second time.
1140 mapping = page_mapping(page);
1141 if (((dirty || writeback) && mapping &&
1142 inode_write_congested(mapping->host)) ||
1143 (writeback && PageReclaim(page)))
1144 stat->nr_congested++;
1147 * If a page at the tail of the LRU is under writeback, there
1148 * are three cases to consider.
1150 * 1) If reclaim is encountering an excessive number of pages
1151 * under writeback and this page is both under writeback and
1152 * PageReclaim then it indicates that pages are being queued
1153 * for IO but are being recycled through the LRU before the
1154 * IO can complete. Waiting on the page itself risks an
1155 * indefinite stall if it is impossible to writeback the
1156 * page due to IO error or disconnected storage so instead
1157 * note that the LRU is being scanned too quickly and the
1158 * caller can stall after page list has been processed.
1160 * 2) Global or new memcg reclaim encounters a page that is
1161 * not marked for immediate reclaim, or the caller does not
1162 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1163 * not to fs). In this case mark the page for immediate
1164 * reclaim and continue scanning.
1166 * Require may_enter_fs because we would wait on fs, which
1167 * may not have submitted IO yet. And the loop driver might
1168 * enter reclaim, and deadlock if it waits on a page for
1169 * which it is needed to do the write (loop masks off
1170 * __GFP_IO|__GFP_FS for this reason); but more thought
1171 * would probably show more reasons.
1173 * 3) Legacy memcg encounters a page that is already marked
1174 * PageReclaim. memcg does not have any dirty pages
1175 * throttling so we could easily OOM just because too many
1176 * pages are in writeback and there is nothing else to
1177 * reclaim. Wait for the writeback to complete.
1179 * In cases 1) and 2) we activate the pages to get them out of
1180 * the way while we continue scanning for clean pages on the
1181 * inactive list and refilling from the active list. The
1182 * observation here is that waiting for disk writes is more
1183 * expensive than potentially causing reloads down the line.
1184 * Since they're marked for immediate reclaim, they won't put
1185 * memory pressure on the cache working set any longer than it
1186 * takes to write them to disk.
1188 if (PageWriteback(page)) {
1190 if (current_is_kswapd() &&
1191 PageReclaim(page) &&
1192 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1193 stat->nr_immediate++;
1194 goto activate_locked;
1197 } else if (writeback_throttling_sane(sc) ||
1198 !PageReclaim(page) || !may_enter_fs) {
1200 * This is slightly racy - end_page_writeback()
1201 * might have just cleared PageReclaim, then
1202 * setting PageReclaim here end up interpreted
1203 * as PageReadahead - but that does not matter
1204 * enough to care. What we do want is for this
1205 * page to have PageReclaim set next time memcg
1206 * reclaim reaches the tests above, so it will
1207 * then wait_on_page_writeback() to avoid OOM;
1208 * and it's also appropriate in global reclaim.
1210 SetPageReclaim(page);
1211 stat->nr_writeback++;
1212 goto activate_locked;
1217 wait_on_page_writeback(page);
1218 /* then go back and try same page again */
1219 list_add_tail(&page->lru, page_list);
1224 if (!ignore_references)
1225 references = page_check_references(page, sc);
1227 switch (references) {
1228 case PAGEREF_ACTIVATE:
1229 goto activate_locked;
1231 stat->nr_ref_keep += nr_pages;
1233 case PAGEREF_RECLAIM:
1234 case PAGEREF_RECLAIM_CLEAN:
1235 ; /* try to reclaim the page below */
1239 * Anonymous process memory has backing store?
1240 * Try to allocate it some swap space here.
1241 * Lazyfree page could be freed directly
1243 if (PageAnon(page) && PageSwapBacked(page)) {
1244 if (!PageSwapCache(page)) {
1245 if (!(sc->gfp_mask & __GFP_IO))
1247 if (PageTransHuge(page)) {
1248 /* cannot split THP, skip it */
1249 if (!can_split_huge_page(page, NULL))
1250 goto activate_locked;
1252 * Split pages without a PMD map right
1253 * away. Chances are some or all of the
1254 * tail pages can be freed without IO.
1256 if (!compound_mapcount(page) &&
1257 split_huge_page_to_list(page,
1259 goto activate_locked;
1261 if (!add_to_swap(page)) {
1262 if (!PageTransHuge(page))
1263 goto activate_locked_split;
1264 /* Fallback to swap normal pages */
1265 if (split_huge_page_to_list(page,
1267 goto activate_locked;
1268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1269 count_vm_event(THP_SWPOUT_FALLBACK);
1271 if (!add_to_swap(page))
1272 goto activate_locked_split;
1277 /* Adding to swap updated mapping */
1278 mapping = page_mapping(page);
1280 } else if (unlikely(PageTransHuge(page))) {
1281 /* Split file THP */
1282 if (split_huge_page_to_list(page, page_list))
1287 * THP may get split above, need minus tail pages and update
1288 * nr_pages to avoid accounting tail pages twice.
1290 * The tail pages that are added into swap cache successfully
1293 if ((nr_pages > 1) && !PageTransHuge(page)) {
1294 sc->nr_scanned -= (nr_pages - 1);
1299 * The page is mapped into the page tables of one or more
1300 * processes. Try to unmap it here.
1302 if (page_mapped(page)) {
1303 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1305 if (unlikely(PageTransHuge(page)))
1306 flags |= TTU_SPLIT_HUGE_PMD;
1307 if (!try_to_unmap(page, flags)) {
1308 stat->nr_unmap_fail += nr_pages;
1309 goto activate_locked;
1313 if (PageDirty(page)) {
1315 * Only kswapd can writeback filesystem pages
1316 * to avoid risk of stack overflow. But avoid
1317 * injecting inefficient single-page IO into
1318 * flusher writeback as much as possible: only
1319 * write pages when we've encountered many
1320 * dirty pages, and when we've already scanned
1321 * the rest of the LRU for clean pages and see
1322 * the same dirty pages again (PageReclaim).
1324 if (page_is_file_cache(page) &&
1325 (!current_is_kswapd() || !PageReclaim(page) ||
1326 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1328 * Immediately reclaim when written back.
1329 * Similar in principal to deactivate_page()
1330 * except we already have the page isolated
1331 * and know it's dirty
1333 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1334 SetPageReclaim(page);
1336 goto activate_locked;
1339 if (references == PAGEREF_RECLAIM_CLEAN)
1343 if (!sc->may_writepage)
1347 * Page is dirty. Flush the TLB if a writable entry
1348 * potentially exists to avoid CPU writes after IO
1349 * starts and then write it out here.
1351 try_to_unmap_flush_dirty();
1352 switch (pageout(page, mapping, sc)) {
1356 goto activate_locked;
1358 if (PageWriteback(page))
1360 if (PageDirty(page))
1364 * A synchronous write - probably a ramdisk. Go
1365 * ahead and try to reclaim the page.
1367 if (!trylock_page(page))
1369 if (PageDirty(page) || PageWriteback(page))
1371 mapping = page_mapping(page);
1373 ; /* try to free the page below */
1378 * If the page has buffers, try to free the buffer mappings
1379 * associated with this page. If we succeed we try to free
1382 * We do this even if the page is PageDirty().
1383 * try_to_release_page() does not perform I/O, but it is
1384 * possible for a page to have PageDirty set, but it is actually
1385 * clean (all its buffers are clean). This happens if the
1386 * buffers were written out directly, with submit_bh(). ext3
1387 * will do this, as well as the blockdev mapping.
1388 * try_to_release_page() will discover that cleanness and will
1389 * drop the buffers and mark the page clean - it can be freed.
1391 * Rarely, pages can have buffers and no ->mapping. These are
1392 * the pages which were not successfully invalidated in
1393 * truncate_complete_page(). We try to drop those buffers here
1394 * and if that worked, and the page is no longer mapped into
1395 * process address space (page_count == 1) it can be freed.
1396 * Otherwise, leave the page on the LRU so it is swappable.
1398 if (page_has_private(page)) {
1399 if (!try_to_release_page(page, sc->gfp_mask))
1400 goto activate_locked;
1401 if (!mapping && page_count(page) == 1) {
1403 if (put_page_testzero(page))
1407 * rare race with speculative reference.
1408 * the speculative reference will free
1409 * this page shortly, so we may
1410 * increment nr_reclaimed here (and
1411 * leave it off the LRU).
1419 if (PageAnon(page) && !PageSwapBacked(page)) {
1420 /* follow __remove_mapping for reference */
1421 if (!page_ref_freeze(page, 1))
1423 if (PageDirty(page)) {
1424 page_ref_unfreeze(page, 1);
1428 count_vm_event(PGLAZYFREED);
1429 count_memcg_page_event(page, PGLAZYFREED);
1430 } else if (!mapping || !__remove_mapping(mapping, page, true,
1431 sc->target_mem_cgroup))
1437 * THP may get swapped out in a whole, need account
1440 nr_reclaimed += nr_pages;
1443 * Is there need to periodically free_page_list? It would
1444 * appear not as the counts should be low
1446 if (unlikely(PageTransHuge(page)))
1447 (*get_compound_page_dtor(page))(page);
1449 list_add(&page->lru, &free_pages);
1452 activate_locked_split:
1454 * The tail pages that are failed to add into swap cache
1455 * reach here. Fixup nr_scanned and nr_pages.
1458 sc->nr_scanned -= (nr_pages - 1);
1462 /* Not a candidate for swapping, so reclaim swap space. */
1463 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1465 try_to_free_swap(page);
1466 VM_BUG_ON_PAGE(PageActive(page), page);
1467 if (!PageMlocked(page)) {
1468 int type = page_is_file_cache(page);
1469 SetPageActive(page);
1470 stat->nr_activate[type] += nr_pages;
1471 count_memcg_page_event(page, PGACTIVATE);
1476 list_add(&page->lru, &ret_pages);
1477 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1480 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1482 mem_cgroup_uncharge_list(&free_pages);
1483 try_to_unmap_flush();
1484 free_unref_page_list(&free_pages);
1486 list_splice(&ret_pages, page_list);
1487 count_vm_events(PGACTIVATE, pgactivate);
1489 return nr_reclaimed;
1492 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1493 struct list_head *page_list)
1495 struct scan_control sc = {
1496 .gfp_mask = GFP_KERNEL,
1497 .priority = DEF_PRIORITY,
1500 struct reclaim_stat dummy_stat;
1502 struct page *page, *next;
1503 LIST_HEAD(clean_pages);
1505 list_for_each_entry_safe(page, next, page_list, lru) {
1506 if (page_is_file_cache(page) && !PageDirty(page) &&
1507 !__PageMovable(page) && !PageUnevictable(page)) {
1508 ClearPageActive(page);
1509 list_move(&page->lru, &clean_pages);
1513 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1514 TTU_IGNORE_ACCESS, &dummy_stat, true);
1515 list_splice(&clean_pages, page_list);
1516 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1521 * Attempt to remove the specified page from its LRU. Only take this page
1522 * if it is of the appropriate PageActive status. Pages which are being
1523 * freed elsewhere are also ignored.
1525 * page: page to consider
1526 * mode: one of the LRU isolation modes defined above
1528 * returns 0 on success, -ve errno on failure.
1530 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1534 /* Only take pages on the LRU. */
1538 /* Compaction should not handle unevictable pages but CMA can do so */
1539 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1545 * To minimise LRU disruption, the caller can indicate that it only
1546 * wants to isolate pages it will be able to operate on without
1547 * blocking - clean pages for the most part.
1549 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1550 * that it is possible to migrate without blocking
1552 if (mode & ISOLATE_ASYNC_MIGRATE) {
1553 /* All the caller can do on PageWriteback is block */
1554 if (PageWriteback(page))
1557 if (PageDirty(page)) {
1558 struct address_space *mapping;
1562 * Only pages without mappings or that have a
1563 * ->migratepage callback are possible to migrate
1564 * without blocking. However, we can be racing with
1565 * truncation so it's necessary to lock the page
1566 * to stabilise the mapping as truncation holds
1567 * the page lock until after the page is removed
1568 * from the page cache.
1570 if (!trylock_page(page))
1573 mapping = page_mapping(page);
1574 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1581 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1584 if (likely(get_page_unless_zero(page))) {
1586 * Be careful not to clear PageLRU until after we're
1587 * sure the page is not being freed elsewhere -- the
1588 * page release code relies on it.
1599 * Update LRU sizes after isolating pages. The LRU size updates must
1600 * be complete before mem_cgroup_update_lru_size due to a santity check.
1602 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1603 enum lru_list lru, unsigned long *nr_zone_taken)
1607 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1608 if (!nr_zone_taken[zid])
1611 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1613 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1620 * pgdat->lru_lock is heavily contended. Some of the functions that
1621 * shrink the lists perform better by taking out a batch of pages
1622 * and working on them outside the LRU lock.
1624 * For pagecache intensive workloads, this function is the hottest
1625 * spot in the kernel (apart from copy_*_user functions).
1627 * Appropriate locks must be held before calling this function.
1629 * @nr_to_scan: The number of eligible pages to look through on the list.
1630 * @lruvec: The LRU vector to pull pages from.
1631 * @dst: The temp list to put pages on to.
1632 * @nr_scanned: The number of pages that were scanned.
1633 * @sc: The scan_control struct for this reclaim session
1634 * @mode: One of the LRU isolation modes
1635 * @lru: LRU list id for isolating
1637 * returns how many pages were moved onto *@dst.
1639 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1640 struct lruvec *lruvec, struct list_head *dst,
1641 unsigned long *nr_scanned, struct scan_control *sc,
1644 struct list_head *src = &lruvec->lists[lru];
1645 unsigned long nr_taken = 0;
1646 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1647 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1648 unsigned long skipped = 0;
1649 unsigned long scan, total_scan, nr_pages;
1650 LIST_HEAD(pages_skipped);
1651 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1655 while (scan < nr_to_scan && !list_empty(src)) {
1658 page = lru_to_page(src);
1659 prefetchw_prev_lru_page(page, src, flags);
1661 VM_BUG_ON_PAGE(!PageLRU(page), page);
1663 nr_pages = compound_nr(page);
1664 total_scan += nr_pages;
1666 if (page_zonenum(page) > sc->reclaim_idx) {
1667 list_move(&page->lru, &pages_skipped);
1668 nr_skipped[page_zonenum(page)] += nr_pages;
1673 * Do not count skipped pages because that makes the function
1674 * return with no isolated pages if the LRU mostly contains
1675 * ineligible pages. This causes the VM to not reclaim any
1676 * pages, triggering a premature OOM.
1678 * Account all tail pages of THP. This would not cause
1679 * premature OOM since __isolate_lru_page() returns -EBUSY
1680 * only when the page is being freed somewhere else.
1683 switch (__isolate_lru_page(page, mode)) {
1685 nr_taken += nr_pages;
1686 nr_zone_taken[page_zonenum(page)] += nr_pages;
1687 list_move(&page->lru, dst);
1691 /* else it is being freed elsewhere */
1692 list_move(&page->lru, src);
1701 * Splice any skipped pages to the start of the LRU list. Note that
1702 * this disrupts the LRU order when reclaiming for lower zones but
1703 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1704 * scanning would soon rescan the same pages to skip and put the
1705 * system at risk of premature OOM.
1707 if (!list_empty(&pages_skipped)) {
1710 list_splice(&pages_skipped, src);
1711 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1712 if (!nr_skipped[zid])
1715 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1716 skipped += nr_skipped[zid];
1719 *nr_scanned = total_scan;
1720 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1721 total_scan, skipped, nr_taken, mode, lru);
1722 update_lru_sizes(lruvec, lru, nr_zone_taken);
1727 * isolate_lru_page - tries to isolate a page from its LRU list
1728 * @page: page to isolate from its LRU list
1730 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1731 * vmstat statistic corresponding to whatever LRU list the page was on.
1733 * Returns 0 if the page was removed from an LRU list.
1734 * Returns -EBUSY if the page was not on an LRU list.
1736 * The returned page will have PageLRU() cleared. If it was found on
1737 * the active list, it will have PageActive set. If it was found on
1738 * the unevictable list, it will have the PageUnevictable bit set. That flag
1739 * may need to be cleared by the caller before letting the page go.
1741 * The vmstat statistic corresponding to the list on which the page was
1742 * found will be decremented.
1746 * (1) Must be called with an elevated refcount on the page. This is a
1747 * fundamentnal difference from isolate_lru_pages (which is called
1748 * without a stable reference).
1749 * (2) the lru_lock must not be held.
1750 * (3) interrupts must be enabled.
1752 int isolate_lru_page(struct page *page)
1756 VM_BUG_ON_PAGE(!page_count(page), page);
1757 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1759 if (PageLRU(page)) {
1760 pg_data_t *pgdat = page_pgdat(page);
1761 struct lruvec *lruvec;
1763 spin_lock_irq(&pgdat->lru_lock);
1764 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1765 if (PageLRU(page)) {
1766 int lru = page_lru(page);
1769 del_page_from_lru_list(page, lruvec, lru);
1772 spin_unlock_irq(&pgdat->lru_lock);
1778 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1779 * then get resheduled. When there are massive number of tasks doing page
1780 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1781 * the LRU list will go small and be scanned faster than necessary, leading to
1782 * unnecessary swapping, thrashing and OOM.
1784 static int too_many_isolated(struct pglist_data *pgdat, int file,
1785 struct scan_control *sc)
1787 unsigned long inactive, isolated;
1789 if (current_is_kswapd())
1792 if (!writeback_throttling_sane(sc))
1796 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1797 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1799 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1800 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1804 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1805 * won't get blocked by normal direct-reclaimers, forming a circular
1808 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1811 return isolated > inactive;
1815 * This moves pages from @list to corresponding LRU list.
1817 * We move them the other way if the page is referenced by one or more
1818 * processes, from rmap.
1820 * If the pages are mostly unmapped, the processing is fast and it is
1821 * appropriate to hold zone_lru_lock across the whole operation. But if
1822 * the pages are mapped, the processing is slow (page_referenced()) so we
1823 * should drop zone_lru_lock around each page. It's impossible to balance
1824 * this, so instead we remove the pages from the LRU while processing them.
1825 * It is safe to rely on PG_active against the non-LRU pages in here because
1826 * nobody will play with that bit on a non-LRU page.
1828 * The downside is that we have to touch page->_refcount against each page.
1829 * But we had to alter page->flags anyway.
1831 * Returns the number of pages moved to the given lruvec.
1834 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1835 struct list_head *list)
1837 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1838 int nr_pages, nr_moved = 0;
1839 LIST_HEAD(pages_to_free);
1843 while (!list_empty(list)) {
1844 page = lru_to_page(list);
1845 VM_BUG_ON_PAGE(PageLRU(page), page);
1846 if (unlikely(!page_evictable(page))) {
1847 list_del(&page->lru);
1848 spin_unlock_irq(&pgdat->lru_lock);
1849 putback_lru_page(page);
1850 spin_lock_irq(&pgdat->lru_lock);
1853 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1856 lru = page_lru(page);
1858 nr_pages = hpage_nr_pages(page);
1859 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1860 list_move(&page->lru, &lruvec->lists[lru]);
1862 if (put_page_testzero(page)) {
1863 __ClearPageLRU(page);
1864 __ClearPageActive(page);
1865 del_page_from_lru_list(page, lruvec, lru);
1867 if (unlikely(PageCompound(page))) {
1868 spin_unlock_irq(&pgdat->lru_lock);
1869 (*get_compound_page_dtor(page))(page);
1870 spin_lock_irq(&pgdat->lru_lock);
1872 list_add(&page->lru, &pages_to_free);
1874 nr_moved += nr_pages;
1879 * To save our caller's stack, now use input list for pages to free.
1881 list_splice(&pages_to_free, list);
1887 * If a kernel thread (such as nfsd for loop-back mounts) services
1888 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1889 * In that case we should only throttle if the backing device it is
1890 * writing to is congested. In other cases it is safe to throttle.
1892 static int current_may_throttle(void)
1894 return !(current->flags & PF_LESS_THROTTLE) ||
1895 current->backing_dev_info == NULL ||
1896 bdi_write_congested(current->backing_dev_info);
1900 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1901 * of reclaimed pages
1903 static noinline_for_stack unsigned long
1904 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1905 struct scan_control *sc, enum lru_list lru)
1907 LIST_HEAD(page_list);
1908 unsigned long nr_scanned;
1909 unsigned long nr_reclaimed = 0;
1910 unsigned long nr_taken;
1911 struct reclaim_stat stat;
1912 int file = is_file_lru(lru);
1913 enum vm_event_item item;
1914 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1915 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1916 bool stalled = false;
1918 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1922 /* wait a bit for the reclaimer. */
1926 /* We are about to die and free our memory. Return now. */
1927 if (fatal_signal_pending(current))
1928 return SWAP_CLUSTER_MAX;
1933 spin_lock_irq(&pgdat->lru_lock);
1935 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1936 &nr_scanned, sc, lru);
1938 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1939 reclaim_stat->recent_scanned[file] += nr_taken;
1941 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1942 if (!cgroup_reclaim(sc))
1943 __count_vm_events(item, nr_scanned);
1944 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1945 spin_unlock_irq(&pgdat->lru_lock);
1950 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1953 spin_lock_irq(&pgdat->lru_lock);
1955 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1956 if (!cgroup_reclaim(sc))
1957 __count_vm_events(item, nr_reclaimed);
1958 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1959 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1960 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1962 move_pages_to_lru(lruvec, &page_list);
1964 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1966 spin_unlock_irq(&pgdat->lru_lock);
1968 mem_cgroup_uncharge_list(&page_list);
1969 free_unref_page_list(&page_list);
1972 * If dirty pages are scanned that are not queued for IO, it
1973 * implies that flushers are not doing their job. This can
1974 * happen when memory pressure pushes dirty pages to the end of
1975 * the LRU before the dirty limits are breached and the dirty
1976 * data has expired. It can also happen when the proportion of
1977 * dirty pages grows not through writes but through memory
1978 * pressure reclaiming all the clean cache. And in some cases,
1979 * the flushers simply cannot keep up with the allocation
1980 * rate. Nudge the flusher threads in case they are asleep.
1982 if (stat.nr_unqueued_dirty == nr_taken)
1983 wakeup_flusher_threads(WB_REASON_VMSCAN);
1985 sc->nr.dirty += stat.nr_dirty;
1986 sc->nr.congested += stat.nr_congested;
1987 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1988 sc->nr.writeback += stat.nr_writeback;
1989 sc->nr.immediate += stat.nr_immediate;
1990 sc->nr.taken += nr_taken;
1992 sc->nr.file_taken += nr_taken;
1994 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1995 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1996 return nr_reclaimed;
1999 static void shrink_active_list(unsigned long nr_to_scan,
2000 struct lruvec *lruvec,
2001 struct scan_control *sc,
2004 unsigned long nr_taken;
2005 unsigned long nr_scanned;
2006 unsigned long vm_flags;
2007 LIST_HEAD(l_hold); /* The pages which were snipped off */
2008 LIST_HEAD(l_active);
2009 LIST_HEAD(l_inactive);
2011 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2012 unsigned nr_deactivate, nr_activate;
2013 unsigned nr_rotated = 0;
2014 int file = is_file_lru(lru);
2015 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2019 spin_lock_irq(&pgdat->lru_lock);
2021 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2022 &nr_scanned, sc, lru);
2024 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2025 reclaim_stat->recent_scanned[file] += nr_taken;
2027 __count_vm_events(PGREFILL, nr_scanned);
2028 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2030 spin_unlock_irq(&pgdat->lru_lock);
2032 while (!list_empty(&l_hold)) {
2034 page = lru_to_page(&l_hold);
2035 list_del(&page->lru);
2037 if (unlikely(!page_evictable(page))) {
2038 putback_lru_page(page);
2042 if (unlikely(buffer_heads_over_limit)) {
2043 if (page_has_private(page) && trylock_page(page)) {
2044 if (page_has_private(page))
2045 try_to_release_page(page, 0);
2050 if (page_referenced(page, 0, sc->target_mem_cgroup,
2052 nr_rotated += hpage_nr_pages(page);
2054 * Identify referenced, file-backed active pages and
2055 * give them one more trip around the active list. So
2056 * that executable code get better chances to stay in
2057 * memory under moderate memory pressure. Anon pages
2058 * are not likely to be evicted by use-once streaming
2059 * IO, plus JVM can create lots of anon VM_EXEC pages,
2060 * so we ignore them here.
2062 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2063 list_add(&page->lru, &l_active);
2068 ClearPageActive(page); /* we are de-activating */
2069 SetPageWorkingset(page);
2070 list_add(&page->lru, &l_inactive);
2074 * Move pages back to the lru list.
2076 spin_lock_irq(&pgdat->lru_lock);
2078 * Count referenced pages from currently used mappings as rotated,
2079 * even though only some of them are actually re-activated. This
2080 * helps balance scan pressure between file and anonymous pages in
2083 reclaim_stat->recent_rotated[file] += nr_rotated;
2085 nr_activate = move_pages_to_lru(lruvec, &l_active);
2086 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2087 /* Keep all free pages in l_active list */
2088 list_splice(&l_inactive, &l_active);
2090 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2091 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2093 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2094 spin_unlock_irq(&pgdat->lru_lock);
2096 mem_cgroup_uncharge_list(&l_active);
2097 free_unref_page_list(&l_active);
2098 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2099 nr_deactivate, nr_rotated, sc->priority, file);
2102 unsigned long reclaim_pages(struct list_head *page_list)
2105 unsigned long nr_reclaimed = 0;
2106 LIST_HEAD(node_page_list);
2107 struct reclaim_stat dummy_stat;
2109 struct scan_control sc = {
2110 .gfp_mask = GFP_KERNEL,
2111 .priority = DEF_PRIORITY,
2117 while (!list_empty(page_list)) {
2118 page = lru_to_page(page_list);
2120 nid = page_to_nid(page);
2121 INIT_LIST_HEAD(&node_page_list);
2124 if (nid == page_to_nid(page)) {
2125 ClearPageActive(page);
2126 list_move(&page->lru, &node_page_list);
2130 nr_reclaimed += shrink_page_list(&node_page_list,
2133 &dummy_stat, false);
2134 while (!list_empty(&node_page_list)) {
2135 page = lru_to_page(&node_page_list);
2136 list_del(&page->lru);
2137 putback_lru_page(page);
2143 if (!list_empty(&node_page_list)) {
2144 nr_reclaimed += shrink_page_list(&node_page_list,
2147 &dummy_stat, false);
2148 while (!list_empty(&node_page_list)) {
2149 page = lru_to_page(&node_page_list);
2150 list_del(&page->lru);
2151 putback_lru_page(page);
2155 return nr_reclaimed;
2159 * The inactive anon list should be small enough that the VM never has
2160 * to do too much work.
2162 * The inactive file list should be small enough to leave most memory
2163 * to the established workingset on the scan-resistant active list,
2164 * but large enough to avoid thrashing the aggregate readahead window.
2166 * Both inactive lists should also be large enough that each inactive
2167 * page has a chance to be referenced again before it is reclaimed.
2169 * If that fails and refaulting is observed, the inactive list grows.
2171 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2172 * on this LRU, maintained by the pageout code. An inactive_ratio
2173 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2176 * memory ratio inactive
2177 * -------------------------------------
2186 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2187 struct scan_control *sc, bool trace)
2189 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2190 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2191 enum lru_list inactive_lru = file * LRU_FILE;
2192 unsigned long inactive, active;
2193 unsigned long inactive_ratio;
2194 struct lruvec *target_lruvec;
2195 unsigned long refaults;
2198 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2199 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2202 * When refaults are being observed, it means a new workingset
2203 * is being established. Disable active list protection to get
2204 * rid of the stale workingset quickly.
2206 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2207 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE);
2208 if (file && target_lruvec->refaults != refaults) {
2211 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2213 inactive_ratio = int_sqrt(10 * gb);
2219 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2220 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2221 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2222 inactive_ratio, file);
2224 return inactive * inactive_ratio < active;
2227 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2228 struct lruvec *lruvec, struct scan_control *sc)
2230 if (is_active_lru(lru)) {
2231 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2232 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2236 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2247 * Determine how aggressively the anon and file LRU lists should be
2248 * scanned. The relative value of each set of LRU lists is determined
2249 * by looking at the fraction of the pages scanned we did rotate back
2250 * onto the active list instead of evict.
2252 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2253 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2255 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2258 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2259 int swappiness = mem_cgroup_swappiness(memcg);
2260 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2262 u64 denominator = 0; /* gcc */
2263 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2264 unsigned long anon_prio, file_prio;
2265 enum scan_balance scan_balance;
2266 unsigned long anon, file;
2267 unsigned long ap, fp;
2270 /* If we have no swap space, do not bother scanning anon pages. */
2271 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2272 scan_balance = SCAN_FILE;
2277 * Global reclaim will swap to prevent OOM even with no
2278 * swappiness, but memcg users want to use this knob to
2279 * disable swapping for individual groups completely when
2280 * using the memory controller's swap limit feature would be
2283 if (cgroup_reclaim(sc) && !swappiness) {
2284 scan_balance = SCAN_FILE;
2289 * Do not apply any pressure balancing cleverness when the
2290 * system is close to OOM, scan both anon and file equally
2291 * (unless the swappiness setting disagrees with swapping).
2293 if (!sc->priority && swappiness) {
2294 scan_balance = SCAN_EQUAL;
2299 * If the system is almost out of file pages, force-scan anon.
2300 * But only if there are enough inactive anonymous pages on
2301 * the LRU. Otherwise, the small LRU gets thrashed.
2303 if (sc->file_is_tiny &&
2304 !inactive_list_is_low(lruvec, false, sc, false) &&
2305 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON,
2306 sc->reclaim_idx) >> sc->priority) {
2307 scan_balance = SCAN_ANON;
2312 * If there is enough inactive page cache, i.e. if the size of the
2313 * inactive list is greater than that of the active list *and* the
2314 * inactive list actually has some pages to scan on this priority, we
2315 * do not reclaim anything from the anonymous working set right now.
2316 * Without the second condition we could end up never scanning an
2317 * lruvec even if it has plenty of old anonymous pages unless the
2318 * system is under heavy pressure.
2320 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2321 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2322 scan_balance = SCAN_FILE;
2326 scan_balance = SCAN_FRACT;
2329 * With swappiness at 100, anonymous and file have the same priority.
2330 * This scanning priority is essentially the inverse of IO cost.
2332 anon_prio = swappiness;
2333 file_prio = 200 - anon_prio;
2336 * OK, so we have swap space and a fair amount of page cache
2337 * pages. We use the recently rotated / recently scanned
2338 * ratios to determine how valuable each cache is.
2340 * Because workloads change over time (and to avoid overflow)
2341 * we keep these statistics as a floating average, which ends
2342 * up weighing recent references more than old ones.
2344 * anon in [0], file in [1]
2347 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2348 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2349 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2350 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2352 spin_lock_irq(&pgdat->lru_lock);
2353 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2354 reclaim_stat->recent_scanned[0] /= 2;
2355 reclaim_stat->recent_rotated[0] /= 2;
2358 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2359 reclaim_stat->recent_scanned[1] /= 2;
2360 reclaim_stat->recent_rotated[1] /= 2;
2364 * The amount of pressure on anon vs file pages is inversely
2365 * proportional to the fraction of recently scanned pages on
2366 * each list that were recently referenced and in active use.
2368 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2369 ap /= reclaim_stat->recent_rotated[0] + 1;
2371 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2372 fp /= reclaim_stat->recent_rotated[1] + 1;
2373 spin_unlock_irq(&pgdat->lru_lock);
2377 denominator = ap + fp + 1;
2379 for_each_evictable_lru(lru) {
2380 int file = is_file_lru(lru);
2381 unsigned long lruvec_size;
2383 unsigned long protection;
2385 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2386 protection = mem_cgroup_protection(memcg,
2387 sc->memcg_low_reclaim);
2391 * Scale a cgroup's reclaim pressure by proportioning
2392 * its current usage to its memory.low or memory.min
2395 * This is important, as otherwise scanning aggression
2396 * becomes extremely binary -- from nothing as we
2397 * approach the memory protection threshold, to totally
2398 * nominal as we exceed it. This results in requiring
2399 * setting extremely liberal protection thresholds. It
2400 * also means we simply get no protection at all if we
2401 * set it too low, which is not ideal.
2403 * If there is any protection in place, we reduce scan
2404 * pressure by how much of the total memory used is
2405 * within protection thresholds.
2407 * There is one special case: in the first reclaim pass,
2408 * we skip over all groups that are within their low
2409 * protection. If that fails to reclaim enough pages to
2410 * satisfy the reclaim goal, we come back and override
2411 * the best-effort low protection. However, we still
2412 * ideally want to honor how well-behaved groups are in
2413 * that case instead of simply punishing them all
2414 * equally. As such, we reclaim them based on how much
2415 * memory they are using, reducing the scan pressure
2416 * again by how much of the total memory used is under
2419 unsigned long cgroup_size = mem_cgroup_size(memcg);
2421 /* Avoid TOCTOU with earlier protection check */
2422 cgroup_size = max(cgroup_size, protection);
2424 scan = lruvec_size - lruvec_size * protection /
2428 * Minimally target SWAP_CLUSTER_MAX pages to keep
2429 * reclaim moving forwards, avoiding decremeting
2430 * sc->priority further than desirable.
2432 scan = max(scan, SWAP_CLUSTER_MAX);
2437 scan >>= sc->priority;
2440 * If the cgroup's already been deleted, make sure to
2441 * scrape out the remaining cache.
2443 if (!scan && !mem_cgroup_online(memcg))
2444 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2446 switch (scan_balance) {
2448 /* Scan lists relative to size */
2452 * Scan types proportional to swappiness and
2453 * their relative recent reclaim efficiency.
2454 * Make sure we don't miss the last page on
2455 * the offlined memory cgroups because of a
2458 scan = mem_cgroup_online(memcg) ?
2459 div64_u64(scan * fraction[file], denominator) :
2460 DIV64_U64_ROUND_UP(scan * fraction[file],
2465 /* Scan one type exclusively */
2466 if ((scan_balance == SCAN_FILE) != file) {
2472 /* Look ma, no brain */
2480 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2482 unsigned long nr[NR_LRU_LISTS];
2483 unsigned long targets[NR_LRU_LISTS];
2484 unsigned long nr_to_scan;
2486 unsigned long nr_reclaimed = 0;
2487 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2488 struct blk_plug plug;
2491 get_scan_count(lruvec, sc, nr);
2493 /* Record the original scan target for proportional adjustments later */
2494 memcpy(targets, nr, sizeof(nr));
2497 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2498 * event that can occur when there is little memory pressure e.g.
2499 * multiple streaming readers/writers. Hence, we do not abort scanning
2500 * when the requested number of pages are reclaimed when scanning at
2501 * DEF_PRIORITY on the assumption that the fact we are direct
2502 * reclaiming implies that kswapd is not keeping up and it is best to
2503 * do a batch of work at once. For memcg reclaim one check is made to
2504 * abort proportional reclaim if either the file or anon lru has already
2505 * dropped to zero at the first pass.
2507 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2508 sc->priority == DEF_PRIORITY);
2510 blk_start_plug(&plug);
2511 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2512 nr[LRU_INACTIVE_FILE]) {
2513 unsigned long nr_anon, nr_file, percentage;
2514 unsigned long nr_scanned;
2516 for_each_evictable_lru(lru) {
2518 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2519 nr[lru] -= nr_to_scan;
2521 nr_reclaimed += shrink_list(lru, nr_to_scan,
2528 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2532 * For kswapd and memcg, reclaim at least the number of pages
2533 * requested. Ensure that the anon and file LRUs are scanned
2534 * proportionally what was requested by get_scan_count(). We
2535 * stop reclaiming one LRU and reduce the amount scanning
2536 * proportional to the original scan target.
2538 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2539 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2542 * It's just vindictive to attack the larger once the smaller
2543 * has gone to zero. And given the way we stop scanning the
2544 * smaller below, this makes sure that we only make one nudge
2545 * towards proportionality once we've got nr_to_reclaim.
2547 if (!nr_file || !nr_anon)
2550 if (nr_file > nr_anon) {
2551 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2552 targets[LRU_ACTIVE_ANON] + 1;
2554 percentage = nr_anon * 100 / scan_target;
2556 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2557 targets[LRU_ACTIVE_FILE] + 1;
2559 percentage = nr_file * 100 / scan_target;
2562 /* Stop scanning the smaller of the LRU */
2564 nr[lru + LRU_ACTIVE] = 0;
2567 * Recalculate the other LRU scan count based on its original
2568 * scan target and the percentage scanning already complete
2570 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2571 nr_scanned = targets[lru] - nr[lru];
2572 nr[lru] = targets[lru] * (100 - percentage) / 100;
2573 nr[lru] -= min(nr[lru], nr_scanned);
2576 nr_scanned = targets[lru] - nr[lru];
2577 nr[lru] = targets[lru] * (100 - percentage) / 100;
2578 nr[lru] -= min(nr[lru], nr_scanned);
2580 scan_adjusted = true;
2582 blk_finish_plug(&plug);
2583 sc->nr_reclaimed += nr_reclaimed;
2586 * Even if we did not try to evict anon pages at all, we want to
2587 * rebalance the anon lru active/inactive ratio.
2589 if (total_swap_pages && inactive_list_is_low(lruvec, false, sc, true))
2590 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2591 sc, LRU_ACTIVE_ANON);
2594 /* Use reclaim/compaction for costly allocs or under memory pressure */
2595 static bool in_reclaim_compaction(struct scan_control *sc)
2597 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2598 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2599 sc->priority < DEF_PRIORITY - 2))
2606 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2607 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2608 * true if more pages should be reclaimed such that when the page allocator
2609 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2610 * It will give up earlier than that if there is difficulty reclaiming pages.
2612 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2613 unsigned long nr_reclaimed,
2614 struct scan_control *sc)
2616 unsigned long pages_for_compaction;
2617 unsigned long inactive_lru_pages;
2620 /* If not in reclaim/compaction mode, stop */
2621 if (!in_reclaim_compaction(sc))
2625 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2626 * number of pages that were scanned. This will return to the caller
2627 * with the risk reclaim/compaction and the resulting allocation attempt
2628 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2629 * allocations through requiring that the full LRU list has been scanned
2630 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2631 * scan, but that approximation was wrong, and there were corner cases
2632 * where always a non-zero amount of pages were scanned.
2637 /* If compaction would go ahead or the allocation would succeed, stop */
2638 for (z = 0; z <= sc->reclaim_idx; z++) {
2639 struct zone *zone = &pgdat->node_zones[z];
2640 if (!managed_zone(zone))
2643 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2644 case COMPACT_SUCCESS:
2645 case COMPACT_CONTINUE:
2648 /* check next zone */
2654 * If we have not reclaimed enough pages for compaction and the
2655 * inactive lists are large enough, continue reclaiming
2657 pages_for_compaction = compact_gap(sc->order);
2658 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2659 if (get_nr_swap_pages() > 0)
2660 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2662 return inactive_lru_pages > pages_for_compaction;
2665 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2667 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2668 struct mem_cgroup *memcg;
2670 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2672 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2673 unsigned long reclaimed;
2674 unsigned long scanned;
2676 switch (mem_cgroup_protected(target_memcg, memcg)) {
2677 case MEMCG_PROT_MIN:
2680 * If there is no reclaimable memory, OOM.
2683 case MEMCG_PROT_LOW:
2686 * Respect the protection only as long as
2687 * there is an unprotected supply
2688 * of reclaimable memory from other cgroups.
2690 if (!sc->memcg_low_reclaim) {
2691 sc->memcg_low_skipped = 1;
2694 memcg_memory_event(memcg, MEMCG_LOW);
2696 case MEMCG_PROT_NONE:
2698 * All protection thresholds breached. We may
2699 * still choose to vary the scan pressure
2700 * applied based on by how much the cgroup in
2701 * question has exceeded its protection
2702 * thresholds (see get_scan_count).
2707 reclaimed = sc->nr_reclaimed;
2708 scanned = sc->nr_scanned;
2710 shrink_lruvec(lruvec, sc);
2712 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2715 /* Record the group's reclaim efficiency */
2716 vmpressure(sc->gfp_mask, memcg, false,
2717 sc->nr_scanned - scanned,
2718 sc->nr_reclaimed - reclaimed);
2720 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2723 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2725 struct reclaim_state *reclaim_state = current->reclaim_state;
2726 unsigned long nr_reclaimed, nr_scanned;
2727 struct lruvec *target_lruvec;
2728 bool reclaimable = false;
2730 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2733 memset(&sc->nr, 0, sizeof(sc->nr));
2735 nr_reclaimed = sc->nr_reclaimed;
2736 nr_scanned = sc->nr_scanned;
2739 * Prevent the reclaimer from falling into the cache trap: as
2740 * cache pages start out inactive, every cache fault will tip
2741 * the scan balance towards the file LRU. And as the file LRU
2742 * shrinks, so does the window for rotation from references.
2743 * This means we have a runaway feedback loop where a tiny
2744 * thrashing file LRU becomes infinitely more attractive than
2745 * anon pages. Try to detect this based on file LRU size.
2747 if (!cgroup_reclaim(sc)) {
2751 unsigned long total_high_wmark = 0;
2753 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2754 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2755 node_page_state(pgdat, NR_INACTIVE_FILE);
2757 for (z = 0; z < MAX_NR_ZONES; z++) {
2758 struct zone *zone = &pgdat->node_zones[z];
2759 if (!managed_zone(zone))
2762 total_high_wmark += high_wmark_pages(zone);
2765 sc->file_is_tiny = file + free <= total_high_wmark;
2768 shrink_node_memcgs(pgdat, sc);
2770 if (reclaim_state) {
2771 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2772 reclaim_state->reclaimed_slab = 0;
2775 /* Record the subtree's reclaim efficiency */
2776 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2777 sc->nr_scanned - nr_scanned,
2778 sc->nr_reclaimed - nr_reclaimed);
2780 if (sc->nr_reclaimed - nr_reclaimed)
2783 if (current_is_kswapd()) {
2785 * If reclaim is isolating dirty pages under writeback,
2786 * it implies that the long-lived page allocation rate
2787 * is exceeding the page laundering rate. Either the
2788 * global limits are not being effective at throttling
2789 * processes due to the page distribution throughout
2790 * zones or there is heavy usage of a slow backing
2791 * device. The only option is to throttle from reclaim
2792 * context which is not ideal as there is no guarantee
2793 * the dirtying process is throttled in the same way
2794 * balance_dirty_pages() manages.
2796 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2797 * count the number of pages under pages flagged for
2798 * immediate reclaim and stall if any are encountered
2799 * in the nr_immediate check below.
2801 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2802 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2804 /* Allow kswapd to start writing pages during reclaim.*/
2805 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2806 set_bit(PGDAT_DIRTY, &pgdat->flags);
2809 * If kswapd scans pages marked marked for immediate
2810 * reclaim and under writeback (nr_immediate), it
2811 * implies that pages are cycling through the LRU
2812 * faster than they are written so also forcibly stall.
2814 if (sc->nr.immediate)
2815 congestion_wait(BLK_RW_ASYNC, HZ/10);
2819 * Tag a node/memcg as congested if all the dirty pages
2820 * scanned were backed by a congested BDI and
2821 * wait_iff_congested will stall.
2823 * Legacy memcg will stall in page writeback so avoid forcibly
2824 * stalling in wait_iff_congested().
2826 if ((current_is_kswapd() ||
2827 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2828 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2829 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2832 * Stall direct reclaim for IO completions if underlying BDIs
2833 * and node is congested. Allow kswapd to continue until it
2834 * starts encountering unqueued dirty pages or cycling through
2835 * the LRU too quickly.
2837 if (!current_is_kswapd() && current_may_throttle() &&
2838 !sc->hibernation_mode &&
2839 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2840 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2842 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2847 * Kswapd gives up on balancing particular nodes after too
2848 * many failures to reclaim anything from them and goes to
2849 * sleep. On reclaim progress, reset the failure counter. A
2850 * successful direct reclaim run will revive a dormant kswapd.
2853 pgdat->kswapd_failures = 0;
2859 * Returns true if compaction should go ahead for a costly-order request, or
2860 * the allocation would already succeed without compaction. Return false if we
2861 * should reclaim first.
2863 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2865 unsigned long watermark;
2866 enum compact_result suitable;
2868 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2869 if (suitable == COMPACT_SUCCESS)
2870 /* Allocation should succeed already. Don't reclaim. */
2872 if (suitable == COMPACT_SKIPPED)
2873 /* Compaction cannot yet proceed. Do reclaim. */
2877 * Compaction is already possible, but it takes time to run and there
2878 * are potentially other callers using the pages just freed. So proceed
2879 * with reclaim to make a buffer of free pages available to give
2880 * compaction a reasonable chance of completing and allocating the page.
2881 * Note that we won't actually reclaim the whole buffer in one attempt
2882 * as the target watermark in should_continue_reclaim() is lower. But if
2883 * we are already above the high+gap watermark, don't reclaim at all.
2885 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2887 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2891 * This is the direct reclaim path, for page-allocating processes. We only
2892 * try to reclaim pages from zones which will satisfy the caller's allocation
2895 * If a zone is deemed to be full of pinned pages then just give it a light
2896 * scan then give up on it.
2898 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2902 unsigned long nr_soft_reclaimed;
2903 unsigned long nr_soft_scanned;
2905 pg_data_t *last_pgdat = NULL;
2908 * If the number of buffer_heads in the machine exceeds the maximum
2909 * allowed level, force direct reclaim to scan the highmem zone as
2910 * highmem pages could be pinning lowmem pages storing buffer_heads
2912 orig_mask = sc->gfp_mask;
2913 if (buffer_heads_over_limit) {
2914 sc->gfp_mask |= __GFP_HIGHMEM;
2915 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2918 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2919 sc->reclaim_idx, sc->nodemask) {
2921 * Take care memory controller reclaiming has small influence
2924 if (!cgroup_reclaim(sc)) {
2925 if (!cpuset_zone_allowed(zone,
2926 GFP_KERNEL | __GFP_HARDWALL))
2930 * If we already have plenty of memory free for
2931 * compaction in this zone, don't free any more.
2932 * Even though compaction is invoked for any
2933 * non-zero order, only frequent costly order
2934 * reclamation is disruptive enough to become a
2935 * noticeable problem, like transparent huge
2938 if (IS_ENABLED(CONFIG_COMPACTION) &&
2939 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2940 compaction_ready(zone, sc)) {
2941 sc->compaction_ready = true;
2946 * Shrink each node in the zonelist once. If the
2947 * zonelist is ordered by zone (not the default) then a
2948 * node may be shrunk multiple times but in that case
2949 * the user prefers lower zones being preserved.
2951 if (zone->zone_pgdat == last_pgdat)
2955 * This steals pages from memory cgroups over softlimit
2956 * and returns the number of reclaimed pages and
2957 * scanned pages. This works for global memory pressure
2958 * and balancing, not for a memcg's limit.
2960 nr_soft_scanned = 0;
2961 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2962 sc->order, sc->gfp_mask,
2964 sc->nr_reclaimed += nr_soft_reclaimed;
2965 sc->nr_scanned += nr_soft_scanned;
2966 /* need some check for avoid more shrink_zone() */
2969 /* See comment about same check for global reclaim above */
2970 if (zone->zone_pgdat == last_pgdat)
2972 last_pgdat = zone->zone_pgdat;
2973 shrink_node(zone->zone_pgdat, sc);
2977 * Restore to original mask to avoid the impact on the caller if we
2978 * promoted it to __GFP_HIGHMEM.
2980 sc->gfp_mask = orig_mask;
2983 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2985 struct lruvec *target_lruvec;
2986 unsigned long refaults;
2988 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2989 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE);
2990 target_lruvec->refaults = refaults;
2994 * This is the main entry point to direct page reclaim.
2996 * If a full scan of the inactive list fails to free enough memory then we
2997 * are "out of memory" and something needs to be killed.
2999 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3000 * high - the zone may be full of dirty or under-writeback pages, which this
3001 * caller can't do much about. We kick the writeback threads and take explicit
3002 * naps in the hope that some of these pages can be written. But if the
3003 * allocating task holds filesystem locks which prevent writeout this might not
3004 * work, and the allocation attempt will fail.
3006 * returns: 0, if no pages reclaimed
3007 * else, the number of pages reclaimed
3009 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3010 struct scan_control *sc)
3012 int initial_priority = sc->priority;
3013 pg_data_t *last_pgdat;
3017 delayacct_freepages_start();
3019 if (!cgroup_reclaim(sc))
3020 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3023 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3026 shrink_zones(zonelist, sc);
3028 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3031 if (sc->compaction_ready)
3035 * If we're getting trouble reclaiming, start doing
3036 * writepage even in laptop mode.
3038 if (sc->priority < DEF_PRIORITY - 2)
3039 sc->may_writepage = 1;
3040 } while (--sc->priority >= 0);
3043 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3045 if (zone->zone_pgdat == last_pgdat)
3047 last_pgdat = zone->zone_pgdat;
3049 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3051 if (cgroup_reclaim(sc)) {
3052 struct lruvec *lruvec;
3054 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3056 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3060 delayacct_freepages_end();
3062 if (sc->nr_reclaimed)
3063 return sc->nr_reclaimed;
3065 /* Aborted reclaim to try compaction? don't OOM, then */
3066 if (sc->compaction_ready)
3069 /* Untapped cgroup reserves? Don't OOM, retry. */
3070 if (sc->memcg_low_skipped) {
3071 sc->priority = initial_priority;
3072 sc->memcg_low_reclaim = 1;
3073 sc->memcg_low_skipped = 0;
3080 static bool allow_direct_reclaim(pg_data_t *pgdat)
3083 unsigned long pfmemalloc_reserve = 0;
3084 unsigned long free_pages = 0;
3088 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3091 for (i = 0; i <= ZONE_NORMAL; i++) {
3092 zone = &pgdat->node_zones[i];
3093 if (!managed_zone(zone))
3096 if (!zone_reclaimable_pages(zone))
3099 pfmemalloc_reserve += min_wmark_pages(zone);
3100 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3103 /* If there are no reserves (unexpected config) then do not throttle */
3104 if (!pfmemalloc_reserve)
3107 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3109 /* kswapd must be awake if processes are being throttled */
3110 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3111 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3112 (enum zone_type)ZONE_NORMAL);
3113 wake_up_interruptible(&pgdat->kswapd_wait);
3120 * Throttle direct reclaimers if backing storage is backed by the network
3121 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3122 * depleted. kswapd will continue to make progress and wake the processes
3123 * when the low watermark is reached.
3125 * Returns true if a fatal signal was delivered during throttling. If this
3126 * happens, the page allocator should not consider triggering the OOM killer.
3128 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3129 nodemask_t *nodemask)
3133 pg_data_t *pgdat = NULL;
3136 * Kernel threads should not be throttled as they may be indirectly
3137 * responsible for cleaning pages necessary for reclaim to make forward
3138 * progress. kjournald for example may enter direct reclaim while
3139 * committing a transaction where throttling it could forcing other
3140 * processes to block on log_wait_commit().
3142 if (current->flags & PF_KTHREAD)
3146 * If a fatal signal is pending, this process should not throttle.
3147 * It should return quickly so it can exit and free its memory
3149 if (fatal_signal_pending(current))
3153 * Check if the pfmemalloc reserves are ok by finding the first node
3154 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3155 * GFP_KERNEL will be required for allocating network buffers when
3156 * swapping over the network so ZONE_HIGHMEM is unusable.
3158 * Throttling is based on the first usable node and throttled processes
3159 * wait on a queue until kswapd makes progress and wakes them. There
3160 * is an affinity then between processes waking up and where reclaim
3161 * progress has been made assuming the process wakes on the same node.
3162 * More importantly, processes running on remote nodes will not compete
3163 * for remote pfmemalloc reserves and processes on different nodes
3164 * should make reasonable progress.
3166 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3167 gfp_zone(gfp_mask), nodemask) {
3168 if (zone_idx(zone) > ZONE_NORMAL)
3171 /* Throttle based on the first usable node */
3172 pgdat = zone->zone_pgdat;
3173 if (allow_direct_reclaim(pgdat))
3178 /* If no zone was usable by the allocation flags then do not throttle */
3182 /* Account for the throttling */
3183 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3186 * If the caller cannot enter the filesystem, it's possible that it
3187 * is due to the caller holding an FS lock or performing a journal
3188 * transaction in the case of a filesystem like ext[3|4]. In this case,
3189 * it is not safe to block on pfmemalloc_wait as kswapd could be
3190 * blocked waiting on the same lock. Instead, throttle for up to a
3191 * second before continuing.
3193 if (!(gfp_mask & __GFP_FS)) {
3194 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3195 allow_direct_reclaim(pgdat), HZ);
3200 /* Throttle until kswapd wakes the process */
3201 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3202 allow_direct_reclaim(pgdat));
3205 if (fatal_signal_pending(current))
3212 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3213 gfp_t gfp_mask, nodemask_t *nodemask)
3215 unsigned long nr_reclaimed;
3216 struct scan_control sc = {
3217 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3218 .gfp_mask = current_gfp_context(gfp_mask),
3219 .reclaim_idx = gfp_zone(gfp_mask),
3221 .nodemask = nodemask,
3222 .priority = DEF_PRIORITY,
3223 .may_writepage = !laptop_mode,
3229 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3230 * Confirm they are large enough for max values.
3232 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3233 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3234 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3237 * Do not enter reclaim if fatal signal was delivered while throttled.
3238 * 1 is returned so that the page allocator does not OOM kill at this
3241 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3244 set_task_reclaim_state(current, &sc.reclaim_state);
3245 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3247 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3249 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3250 set_task_reclaim_state(current, NULL);
3252 return nr_reclaimed;
3257 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3258 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3259 gfp_t gfp_mask, bool noswap,
3261 unsigned long *nr_scanned)
3263 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3264 struct scan_control sc = {
3265 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3266 .target_mem_cgroup = memcg,
3267 .may_writepage = !laptop_mode,
3269 .reclaim_idx = MAX_NR_ZONES - 1,
3270 .may_swap = !noswap,
3273 WARN_ON_ONCE(!current->reclaim_state);
3275 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3276 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3278 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3282 * NOTE: Although we can get the priority field, using it
3283 * here is not a good idea, since it limits the pages we can scan.
3284 * if we don't reclaim here, the shrink_node from balance_pgdat
3285 * will pick up pages from other mem cgroup's as well. We hack
3286 * the priority and make it zero.
3288 shrink_lruvec(lruvec, &sc);
3290 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3292 *nr_scanned = sc.nr_scanned;
3294 return sc.nr_reclaimed;
3297 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3298 unsigned long nr_pages,
3302 struct zonelist *zonelist;
3303 unsigned long nr_reclaimed;
3304 unsigned long pflags;
3306 unsigned int noreclaim_flag;
3307 struct scan_control sc = {
3308 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3309 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3310 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3311 .reclaim_idx = MAX_NR_ZONES - 1,
3312 .target_mem_cgroup = memcg,
3313 .priority = DEF_PRIORITY,
3314 .may_writepage = !laptop_mode,
3316 .may_swap = may_swap,
3319 set_task_reclaim_state(current, &sc.reclaim_state);
3321 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3322 * take care of from where we get pages. So the node where we start the
3323 * scan does not need to be the current node.
3325 nid = mem_cgroup_select_victim_node(memcg);
3327 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3329 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3331 psi_memstall_enter(&pflags);
3332 noreclaim_flag = memalloc_noreclaim_save();
3334 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3336 memalloc_noreclaim_restore(noreclaim_flag);
3337 psi_memstall_leave(&pflags);
3339 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3340 set_task_reclaim_state(current, NULL);
3342 return nr_reclaimed;
3346 static void age_active_anon(struct pglist_data *pgdat,
3347 struct scan_control *sc)
3349 struct mem_cgroup *memcg;
3351 if (!total_swap_pages)
3354 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3356 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3358 if (inactive_list_is_low(lruvec, false, sc, true))
3359 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3360 sc, LRU_ACTIVE_ANON);
3362 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3366 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3372 * Check for watermark boosts top-down as the higher zones
3373 * are more likely to be boosted. Both watermarks and boosts
3374 * should not be checked at the time time as reclaim would
3375 * start prematurely when there is no boosting and a lower
3378 for (i = classzone_idx; i >= 0; i--) {
3379 zone = pgdat->node_zones + i;
3380 if (!managed_zone(zone))
3383 if (zone->watermark_boost)
3391 * Returns true if there is an eligible zone balanced for the request order
3394 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3397 unsigned long mark = -1;
3401 * Check watermarks bottom-up as lower zones are more likely to
3404 for (i = 0; i <= classzone_idx; i++) {
3405 zone = pgdat->node_zones + i;
3407 if (!managed_zone(zone))
3410 mark = high_wmark_pages(zone);
3411 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3416 * If a node has no populated zone within classzone_idx, it does not
3417 * need balancing by definition. This can happen if a zone-restricted
3418 * allocation tries to wake a remote kswapd.
3426 /* Clear pgdat state for congested, dirty or under writeback. */
3427 static void clear_pgdat_congested(pg_data_t *pgdat)
3429 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3431 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3432 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3433 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3437 * Prepare kswapd for sleeping. This verifies that there are no processes
3438 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3440 * Returns true if kswapd is ready to sleep
3442 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3445 * The throttled processes are normally woken up in balance_pgdat() as
3446 * soon as allow_direct_reclaim() is true. But there is a potential
3447 * race between when kswapd checks the watermarks and a process gets
3448 * throttled. There is also a potential race if processes get
3449 * throttled, kswapd wakes, a large process exits thereby balancing the
3450 * zones, which causes kswapd to exit balance_pgdat() before reaching
3451 * the wake up checks. If kswapd is going to sleep, no process should
3452 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3453 * the wake up is premature, processes will wake kswapd and get
3454 * throttled again. The difference from wake ups in balance_pgdat() is
3455 * that here we are under prepare_to_wait().
3457 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3458 wake_up_all(&pgdat->pfmemalloc_wait);
3460 /* Hopeless node, leave it to direct reclaim */
3461 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3464 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3465 clear_pgdat_congested(pgdat);
3473 * kswapd shrinks a node of pages that are at or below the highest usable
3474 * zone that is currently unbalanced.
3476 * Returns true if kswapd scanned at least the requested number of pages to
3477 * reclaim or if the lack of progress was due to pages under writeback.
3478 * This is used to determine if the scanning priority needs to be raised.
3480 static bool kswapd_shrink_node(pg_data_t *pgdat,
3481 struct scan_control *sc)
3486 /* Reclaim a number of pages proportional to the number of zones */
3487 sc->nr_to_reclaim = 0;
3488 for (z = 0; z <= sc->reclaim_idx; z++) {
3489 zone = pgdat->node_zones + z;
3490 if (!managed_zone(zone))
3493 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3497 * Historically care was taken to put equal pressure on all zones but
3498 * now pressure is applied based on node LRU order.
3500 shrink_node(pgdat, sc);
3503 * Fragmentation may mean that the system cannot be rebalanced for
3504 * high-order allocations. If twice the allocation size has been
3505 * reclaimed then recheck watermarks only at order-0 to prevent
3506 * excessive reclaim. Assume that a process requested a high-order
3507 * can direct reclaim/compact.
3509 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3512 return sc->nr_scanned >= sc->nr_to_reclaim;
3516 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3517 * that are eligible for use by the caller until at least one zone is
3520 * Returns the order kswapd finished reclaiming at.
3522 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3523 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3524 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3525 * or lower is eligible for reclaim until at least one usable zone is
3528 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3531 unsigned long nr_soft_reclaimed;
3532 unsigned long nr_soft_scanned;
3533 unsigned long pflags;
3534 unsigned long nr_boost_reclaim;
3535 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3538 struct scan_control sc = {
3539 .gfp_mask = GFP_KERNEL,
3544 set_task_reclaim_state(current, &sc.reclaim_state);
3545 psi_memstall_enter(&pflags);
3546 __fs_reclaim_acquire();
3548 count_vm_event(PAGEOUTRUN);
3551 * Account for the reclaim boost. Note that the zone boost is left in
3552 * place so that parallel allocations that are near the watermark will
3553 * stall or direct reclaim until kswapd is finished.
3555 nr_boost_reclaim = 0;
3556 for (i = 0; i <= classzone_idx; i++) {
3557 zone = pgdat->node_zones + i;
3558 if (!managed_zone(zone))
3561 nr_boost_reclaim += zone->watermark_boost;
3562 zone_boosts[i] = zone->watermark_boost;
3564 boosted = nr_boost_reclaim;
3567 sc.priority = DEF_PRIORITY;
3569 unsigned long nr_reclaimed = sc.nr_reclaimed;
3570 bool raise_priority = true;
3574 sc.reclaim_idx = classzone_idx;
3577 * If the number of buffer_heads exceeds the maximum allowed
3578 * then consider reclaiming from all zones. This has a dual
3579 * purpose -- on 64-bit systems it is expected that
3580 * buffer_heads are stripped during active rotation. On 32-bit
3581 * systems, highmem pages can pin lowmem memory and shrinking
3582 * buffers can relieve lowmem pressure. Reclaim may still not
3583 * go ahead if all eligible zones for the original allocation
3584 * request are balanced to avoid excessive reclaim from kswapd.
3586 if (buffer_heads_over_limit) {
3587 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3588 zone = pgdat->node_zones + i;
3589 if (!managed_zone(zone))
3598 * If the pgdat is imbalanced then ignore boosting and preserve
3599 * the watermarks for a later time and restart. Note that the
3600 * zone watermarks will be still reset at the end of balancing
3601 * on the grounds that the normal reclaim should be enough to
3602 * re-evaluate if boosting is required when kswapd next wakes.
3604 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3605 if (!balanced && nr_boost_reclaim) {
3606 nr_boost_reclaim = 0;
3611 * If boosting is not active then only reclaim if there are no
3612 * eligible zones. Note that sc.reclaim_idx is not used as
3613 * buffer_heads_over_limit may have adjusted it.
3615 if (!nr_boost_reclaim && balanced)
3618 /* Limit the priority of boosting to avoid reclaim writeback */
3619 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3620 raise_priority = false;
3623 * Do not writeback or swap pages for boosted reclaim. The
3624 * intent is to relieve pressure not issue sub-optimal IO
3625 * from reclaim context. If no pages are reclaimed, the
3626 * reclaim will be aborted.
3628 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3629 sc.may_swap = !nr_boost_reclaim;
3632 * Do some background aging of the anon list, to give
3633 * pages a chance to be referenced before reclaiming. All
3634 * pages are rotated regardless of classzone as this is
3635 * about consistent aging.
3637 age_active_anon(pgdat, &sc);
3640 * If we're getting trouble reclaiming, start doing writepage
3641 * even in laptop mode.
3643 if (sc.priority < DEF_PRIORITY - 2)
3644 sc.may_writepage = 1;
3646 /* Call soft limit reclaim before calling shrink_node. */
3648 nr_soft_scanned = 0;
3649 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3650 sc.gfp_mask, &nr_soft_scanned);
3651 sc.nr_reclaimed += nr_soft_reclaimed;
3654 * There should be no need to raise the scanning priority if
3655 * enough pages are already being scanned that that high
3656 * watermark would be met at 100% efficiency.
3658 if (kswapd_shrink_node(pgdat, &sc))
3659 raise_priority = false;
3662 * If the low watermark is met there is no need for processes
3663 * to be throttled on pfmemalloc_wait as they should not be
3664 * able to safely make forward progress. Wake them
3666 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3667 allow_direct_reclaim(pgdat))
3668 wake_up_all(&pgdat->pfmemalloc_wait);
3670 /* Check if kswapd should be suspending */
3671 __fs_reclaim_release();
3672 ret = try_to_freeze();
3673 __fs_reclaim_acquire();
3674 if (ret || kthread_should_stop())
3678 * Raise priority if scanning rate is too low or there was no
3679 * progress in reclaiming pages
3681 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3682 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3685 * If reclaim made no progress for a boost, stop reclaim as
3686 * IO cannot be queued and it could be an infinite loop in
3687 * extreme circumstances.
3689 if (nr_boost_reclaim && !nr_reclaimed)
3692 if (raise_priority || !nr_reclaimed)
3694 } while (sc.priority >= 1);
3696 if (!sc.nr_reclaimed)
3697 pgdat->kswapd_failures++;
3700 /* If reclaim was boosted, account for the reclaim done in this pass */
3702 unsigned long flags;
3704 for (i = 0; i <= classzone_idx; i++) {
3705 if (!zone_boosts[i])
3708 /* Increments are under the zone lock */
3709 zone = pgdat->node_zones + i;
3710 spin_lock_irqsave(&zone->lock, flags);
3711 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3712 spin_unlock_irqrestore(&zone->lock, flags);
3716 * As there is now likely space, wakeup kcompact to defragment
3719 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3722 snapshot_refaults(NULL, pgdat);
3723 __fs_reclaim_release();
3724 psi_memstall_leave(&pflags);
3725 set_task_reclaim_state(current, NULL);
3728 * Return the order kswapd stopped reclaiming at as
3729 * prepare_kswapd_sleep() takes it into account. If another caller
3730 * entered the allocator slow path while kswapd was awake, order will
3731 * remain at the higher level.
3737 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3738 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3739 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3740 * after previous reclaim attempt (node is still unbalanced). In that case
3741 * return the zone index of the previous kswapd reclaim cycle.
3743 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3744 enum zone_type prev_classzone_idx)
3746 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3747 return prev_classzone_idx;
3748 return pgdat->kswapd_classzone_idx;
3751 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3752 unsigned int classzone_idx)
3757 if (freezing(current) || kthread_should_stop())
3760 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3763 * Try to sleep for a short interval. Note that kcompactd will only be
3764 * woken if it is possible to sleep for a short interval. This is
3765 * deliberate on the assumption that if reclaim cannot keep an
3766 * eligible zone balanced that it's also unlikely that compaction will
3769 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3771 * Compaction records what page blocks it recently failed to
3772 * isolate pages from and skips them in the future scanning.
3773 * When kswapd is going to sleep, it is reasonable to assume
3774 * that pages and compaction may succeed so reset the cache.
3776 reset_isolation_suitable(pgdat);
3779 * We have freed the memory, now we should compact it to make
3780 * allocation of the requested order possible.
3782 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3784 remaining = schedule_timeout(HZ/10);
3787 * If woken prematurely then reset kswapd_classzone_idx and
3788 * order. The values will either be from a wakeup request or
3789 * the previous request that slept prematurely.
3792 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3793 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3796 finish_wait(&pgdat->kswapd_wait, &wait);
3797 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3801 * After a short sleep, check if it was a premature sleep. If not, then
3802 * go fully to sleep until explicitly woken up.
3805 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3806 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3809 * vmstat counters are not perfectly accurate and the estimated
3810 * value for counters such as NR_FREE_PAGES can deviate from the
3811 * true value by nr_online_cpus * threshold. To avoid the zone
3812 * watermarks being breached while under pressure, we reduce the
3813 * per-cpu vmstat threshold while kswapd is awake and restore
3814 * them before going back to sleep.
3816 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3818 if (!kthread_should_stop())
3821 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3824 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3826 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3828 finish_wait(&pgdat->kswapd_wait, &wait);
3832 * The background pageout daemon, started as a kernel thread
3833 * from the init process.
3835 * This basically trickles out pages so that we have _some_
3836 * free memory available even if there is no other activity
3837 * that frees anything up. This is needed for things like routing
3838 * etc, where we otherwise might have all activity going on in
3839 * asynchronous contexts that cannot page things out.
3841 * If there are applications that are active memory-allocators
3842 * (most normal use), this basically shouldn't matter.
3844 static int kswapd(void *p)
3846 unsigned int alloc_order, reclaim_order;
3847 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3848 pg_data_t *pgdat = (pg_data_t*)p;
3849 struct task_struct *tsk = current;
3850 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3852 if (!cpumask_empty(cpumask))
3853 set_cpus_allowed_ptr(tsk, cpumask);
3856 * Tell the memory management that we're a "memory allocator",
3857 * and that if we need more memory we should get access to it
3858 * regardless (see "__alloc_pages()"). "kswapd" should
3859 * never get caught in the normal page freeing logic.
3861 * (Kswapd normally doesn't need memory anyway, but sometimes
3862 * you need a small amount of memory in order to be able to
3863 * page out something else, and this flag essentially protects
3864 * us from recursively trying to free more memory as we're
3865 * trying to free the first piece of memory in the first place).
3867 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3870 pgdat->kswapd_order = 0;
3871 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3875 alloc_order = reclaim_order = pgdat->kswapd_order;
3876 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3879 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3882 /* Read the new order and classzone_idx */
3883 alloc_order = reclaim_order = pgdat->kswapd_order;
3884 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3885 pgdat->kswapd_order = 0;
3886 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3888 ret = try_to_freeze();
3889 if (kthread_should_stop())
3893 * We can speed up thawing tasks if we don't call balance_pgdat
3894 * after returning from the refrigerator
3900 * Reclaim begins at the requested order but if a high-order
3901 * reclaim fails then kswapd falls back to reclaiming for
3902 * order-0. If that happens, kswapd will consider sleeping
3903 * for the order it finished reclaiming at (reclaim_order)
3904 * but kcompactd is woken to compact for the original
3905 * request (alloc_order).
3907 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3909 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3910 if (reclaim_order < alloc_order)
3911 goto kswapd_try_sleep;
3914 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3920 * A zone is low on free memory or too fragmented for high-order memory. If
3921 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3922 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3923 * has failed or is not needed, still wake up kcompactd if only compaction is
3926 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3927 enum zone_type classzone_idx)
3931 if (!managed_zone(zone))
3934 if (!cpuset_zone_allowed(zone, gfp_flags))
3936 pgdat = zone->zone_pgdat;
3938 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3939 pgdat->kswapd_classzone_idx = classzone_idx;
3941 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3943 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3944 if (!waitqueue_active(&pgdat->kswapd_wait))
3947 /* Hopeless node, leave it to direct reclaim if possible */
3948 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3949 (pgdat_balanced(pgdat, order, classzone_idx) &&
3950 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3952 * There may be plenty of free memory available, but it's too
3953 * fragmented for high-order allocations. Wake up kcompactd
3954 * and rely on compaction_suitable() to determine if it's
3955 * needed. If it fails, it will defer subsequent attempts to
3956 * ratelimit its work.
3958 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3959 wakeup_kcompactd(pgdat, order, classzone_idx);
3963 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3965 wake_up_interruptible(&pgdat->kswapd_wait);
3968 #ifdef CONFIG_HIBERNATION
3970 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3973 * Rather than trying to age LRUs the aim is to preserve the overall
3974 * LRU order by reclaiming preferentially
3975 * inactive > active > active referenced > active mapped
3977 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3979 struct scan_control sc = {
3980 .nr_to_reclaim = nr_to_reclaim,
3981 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3982 .reclaim_idx = MAX_NR_ZONES - 1,
3983 .priority = DEF_PRIORITY,
3987 .hibernation_mode = 1,
3989 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3990 unsigned long nr_reclaimed;
3991 unsigned int noreclaim_flag;
3993 fs_reclaim_acquire(sc.gfp_mask);
3994 noreclaim_flag = memalloc_noreclaim_save();
3995 set_task_reclaim_state(current, &sc.reclaim_state);
3997 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3999 set_task_reclaim_state(current, NULL);
4000 memalloc_noreclaim_restore(noreclaim_flag);
4001 fs_reclaim_release(sc.gfp_mask);
4003 return nr_reclaimed;
4005 #endif /* CONFIG_HIBERNATION */
4007 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4008 not required for correctness. So if the last cpu in a node goes
4009 away, we get changed to run anywhere: as the first one comes back,
4010 restore their cpu bindings. */
4011 static int kswapd_cpu_online(unsigned int cpu)
4015 for_each_node_state(nid, N_MEMORY) {
4016 pg_data_t *pgdat = NODE_DATA(nid);
4017 const struct cpumask *mask;
4019 mask = cpumask_of_node(pgdat->node_id);
4021 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4022 /* One of our CPUs online: restore mask */
4023 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4029 * This kswapd start function will be called by init and node-hot-add.
4030 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4032 int kswapd_run(int nid)
4034 pg_data_t *pgdat = NODE_DATA(nid);
4040 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4041 if (IS_ERR(pgdat->kswapd)) {
4042 /* failure at boot is fatal */
4043 BUG_ON(system_state < SYSTEM_RUNNING);
4044 pr_err("Failed to start kswapd on node %d\n", nid);
4045 ret = PTR_ERR(pgdat->kswapd);
4046 pgdat->kswapd = NULL;
4052 * Called by memory hotplug when all memory in a node is offlined. Caller must
4053 * hold mem_hotplug_begin/end().
4055 void kswapd_stop(int nid)
4057 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4060 kthread_stop(kswapd);
4061 NODE_DATA(nid)->kswapd = NULL;
4065 static int __init kswapd_init(void)
4070 for_each_node_state(nid, N_MEMORY)
4072 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4073 "mm/vmscan:online", kswapd_cpu_online,
4079 module_init(kswapd_init)
4085 * If non-zero call node_reclaim when the number of free pages falls below
4088 int node_reclaim_mode __read_mostly;
4090 #define RECLAIM_OFF 0
4091 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4092 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4093 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4096 * Priority for NODE_RECLAIM. This determines the fraction of pages
4097 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4100 #define NODE_RECLAIM_PRIORITY 4
4103 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4106 int sysctl_min_unmapped_ratio = 1;
4109 * If the number of slab pages in a zone grows beyond this percentage then
4110 * slab reclaim needs to occur.
4112 int sysctl_min_slab_ratio = 5;
4114 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4116 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4117 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4118 node_page_state(pgdat, NR_ACTIVE_FILE);
4121 * It's possible for there to be more file mapped pages than
4122 * accounted for by the pages on the file LRU lists because
4123 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4125 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4128 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4129 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4131 unsigned long nr_pagecache_reclaimable;
4132 unsigned long delta = 0;
4135 * If RECLAIM_UNMAP is set, then all file pages are considered
4136 * potentially reclaimable. Otherwise, we have to worry about
4137 * pages like swapcache and node_unmapped_file_pages() provides
4140 if (node_reclaim_mode & RECLAIM_UNMAP)
4141 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4143 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4145 /* If we can't clean pages, remove dirty pages from consideration */
4146 if (!(node_reclaim_mode & RECLAIM_WRITE))
4147 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4149 /* Watch for any possible underflows due to delta */
4150 if (unlikely(delta > nr_pagecache_reclaimable))
4151 delta = nr_pagecache_reclaimable;
4153 return nr_pagecache_reclaimable - delta;
4157 * Try to free up some pages from this node through reclaim.
4159 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4161 /* Minimum pages needed in order to stay on node */
4162 const unsigned long nr_pages = 1 << order;
4163 struct task_struct *p = current;
4164 unsigned int noreclaim_flag;
4165 struct scan_control sc = {
4166 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4167 .gfp_mask = current_gfp_context(gfp_mask),
4169 .priority = NODE_RECLAIM_PRIORITY,
4170 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4171 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4173 .reclaim_idx = gfp_zone(gfp_mask),
4176 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4180 fs_reclaim_acquire(sc.gfp_mask);
4182 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4183 * and we also need to be able to write out pages for RECLAIM_WRITE
4184 * and RECLAIM_UNMAP.
4186 noreclaim_flag = memalloc_noreclaim_save();
4187 p->flags |= PF_SWAPWRITE;
4188 set_task_reclaim_state(p, &sc.reclaim_state);
4190 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4192 * Free memory by calling shrink node with increasing
4193 * priorities until we have enough memory freed.
4196 shrink_node(pgdat, &sc);
4197 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4200 set_task_reclaim_state(p, NULL);
4201 current->flags &= ~PF_SWAPWRITE;
4202 memalloc_noreclaim_restore(noreclaim_flag);
4203 fs_reclaim_release(sc.gfp_mask);
4205 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4207 return sc.nr_reclaimed >= nr_pages;
4210 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4215 * Node reclaim reclaims unmapped file backed pages and
4216 * slab pages if we are over the defined limits.
4218 * A small portion of unmapped file backed pages is needed for
4219 * file I/O otherwise pages read by file I/O will be immediately
4220 * thrown out if the node is overallocated. So we do not reclaim
4221 * if less than a specified percentage of the node is used by
4222 * unmapped file backed pages.
4224 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4225 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4226 return NODE_RECLAIM_FULL;
4229 * Do not scan if the allocation should not be delayed.
4231 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4232 return NODE_RECLAIM_NOSCAN;
4235 * Only run node reclaim on the local node or on nodes that do not
4236 * have associated processors. This will favor the local processor
4237 * over remote processors and spread off node memory allocations
4238 * as wide as possible.
4240 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4241 return NODE_RECLAIM_NOSCAN;
4243 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4244 return NODE_RECLAIM_NOSCAN;
4246 ret = __node_reclaim(pgdat, gfp_mask, order);
4247 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4250 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4257 * page_evictable - test whether a page is evictable
4258 * @page: the page to test
4260 * Test whether page is evictable--i.e., should be placed on active/inactive
4261 * lists vs unevictable list.
4263 * Reasons page might not be evictable:
4264 * (1) page's mapping marked unevictable
4265 * (2) page is part of an mlocked VMA
4268 int page_evictable(struct page *page)
4272 /* Prevent address_space of inode and swap cache from being freed */
4274 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4280 * check_move_unevictable_pages - check pages for evictability and move to
4281 * appropriate zone lru list
4282 * @pvec: pagevec with lru pages to check
4284 * Checks pages for evictability, if an evictable page is in the unevictable
4285 * lru list, moves it to the appropriate evictable lru list. This function
4286 * should be only used for lru pages.
4288 void check_move_unevictable_pages(struct pagevec *pvec)
4290 struct lruvec *lruvec;
4291 struct pglist_data *pgdat = NULL;
4296 for (i = 0; i < pvec->nr; i++) {
4297 struct page *page = pvec->pages[i];
4298 struct pglist_data *pagepgdat = page_pgdat(page);
4301 if (pagepgdat != pgdat) {
4303 spin_unlock_irq(&pgdat->lru_lock);
4305 spin_lock_irq(&pgdat->lru_lock);
4307 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4309 if (!PageLRU(page) || !PageUnevictable(page))
4312 if (page_evictable(page)) {
4313 enum lru_list lru = page_lru_base_type(page);
4315 VM_BUG_ON_PAGE(PageActive(page), page);
4316 ClearPageUnevictable(page);
4317 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4318 add_page_to_lru_list(page, lruvec, lru);
4324 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4325 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4326 spin_unlock_irq(&pgdat->lru_lock);
4329 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);