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,
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);
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)) {
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))
1436 * THP may get swapped out in a whole, need account
1439 nr_reclaimed += nr_pages;
1442 * Is there need to periodically free_page_list? It would
1443 * appear not as the counts should be low
1445 if (unlikely(PageTransHuge(page)))
1446 (*get_compound_page_dtor(page))(page);
1448 list_add(&page->lru, &free_pages);
1451 activate_locked_split:
1453 * The tail pages that are failed to add into swap cache
1454 * reach here. Fixup nr_scanned and nr_pages.
1457 sc->nr_scanned -= (nr_pages - 1);
1461 /* Not a candidate for swapping, so reclaim swap space. */
1462 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1464 try_to_free_swap(page);
1465 VM_BUG_ON_PAGE(PageActive(page), page);
1466 if (!PageMlocked(page)) {
1467 int type = page_is_file_cache(page);
1468 SetPageActive(page);
1469 stat->nr_activate[type] += nr_pages;
1470 count_memcg_page_event(page, PGACTIVATE);
1475 list_add(&page->lru, &ret_pages);
1476 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1479 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1481 mem_cgroup_uncharge_list(&free_pages);
1482 try_to_unmap_flush();
1483 free_unref_page_list(&free_pages);
1485 list_splice(&ret_pages, page_list);
1486 count_vm_events(PGACTIVATE, pgactivate);
1488 return nr_reclaimed;
1491 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1492 struct list_head *page_list)
1494 struct scan_control sc = {
1495 .gfp_mask = GFP_KERNEL,
1496 .priority = DEF_PRIORITY,
1499 struct reclaim_stat dummy_stat;
1501 struct page *page, *next;
1502 LIST_HEAD(clean_pages);
1504 list_for_each_entry_safe(page, next, page_list, lru) {
1505 if (page_is_file_cache(page) && !PageDirty(page) &&
1506 !__PageMovable(page) && !PageUnevictable(page)) {
1507 ClearPageActive(page);
1508 list_move(&page->lru, &clean_pages);
1512 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1513 TTU_IGNORE_ACCESS, &dummy_stat, true);
1514 list_splice(&clean_pages, page_list);
1515 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1520 * Attempt to remove the specified page from its LRU. Only take this page
1521 * if it is of the appropriate PageActive status. Pages which are being
1522 * freed elsewhere are also ignored.
1524 * page: page to consider
1525 * mode: one of the LRU isolation modes defined above
1527 * returns 0 on success, -ve errno on failure.
1529 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1533 /* Only take pages on the LRU. */
1537 /* Compaction should not handle unevictable pages but CMA can do so */
1538 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1544 * To minimise LRU disruption, the caller can indicate that it only
1545 * wants to isolate pages it will be able to operate on without
1546 * blocking - clean pages for the most part.
1548 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1549 * that it is possible to migrate without blocking
1551 if (mode & ISOLATE_ASYNC_MIGRATE) {
1552 /* All the caller can do on PageWriteback is block */
1553 if (PageWriteback(page))
1556 if (PageDirty(page)) {
1557 struct address_space *mapping;
1561 * Only pages without mappings or that have a
1562 * ->migratepage callback are possible to migrate
1563 * without blocking. However, we can be racing with
1564 * truncation so it's necessary to lock the page
1565 * to stabilise the mapping as truncation holds
1566 * the page lock until after the page is removed
1567 * from the page cache.
1569 if (!trylock_page(page))
1572 mapping = page_mapping(page);
1573 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1580 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1583 if (likely(get_page_unless_zero(page))) {
1585 * Be careful not to clear PageLRU until after we're
1586 * sure the page is not being freed elsewhere -- the
1587 * page release code relies on it.
1598 * Update LRU sizes after isolating pages. The LRU size updates must
1599 * be complete before mem_cgroup_update_lru_size due to a santity check.
1601 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1602 enum lru_list lru, unsigned long *nr_zone_taken)
1606 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1607 if (!nr_zone_taken[zid])
1610 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1612 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1619 * pgdat->lru_lock is heavily contended. Some of the functions that
1620 * shrink the lists perform better by taking out a batch of pages
1621 * and working on them outside the LRU lock.
1623 * For pagecache intensive workloads, this function is the hottest
1624 * spot in the kernel (apart from copy_*_user functions).
1626 * Appropriate locks must be held before calling this function.
1628 * @nr_to_scan: The number of eligible pages to look through on the list.
1629 * @lruvec: The LRU vector to pull pages from.
1630 * @dst: The temp list to put pages on to.
1631 * @nr_scanned: The number of pages that were scanned.
1632 * @sc: The scan_control struct for this reclaim session
1633 * @mode: One of the LRU isolation modes
1634 * @lru: LRU list id for isolating
1636 * returns how many pages were moved onto *@dst.
1638 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1639 struct lruvec *lruvec, struct list_head *dst,
1640 unsigned long *nr_scanned, struct scan_control *sc,
1643 struct list_head *src = &lruvec->lists[lru];
1644 unsigned long nr_taken = 0;
1645 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1646 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1647 unsigned long skipped = 0;
1648 unsigned long scan, total_scan, nr_pages;
1649 LIST_HEAD(pages_skipped);
1650 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1654 while (scan < nr_to_scan && !list_empty(src)) {
1657 page = lru_to_page(src);
1658 prefetchw_prev_lru_page(page, src, flags);
1660 VM_BUG_ON_PAGE(!PageLRU(page), page);
1662 nr_pages = compound_nr(page);
1663 total_scan += nr_pages;
1665 if (page_zonenum(page) > sc->reclaim_idx) {
1666 list_move(&page->lru, &pages_skipped);
1667 nr_skipped[page_zonenum(page)] += nr_pages;
1672 * Do not count skipped pages because that makes the function
1673 * return with no isolated pages if the LRU mostly contains
1674 * ineligible pages. This causes the VM to not reclaim any
1675 * pages, triggering a premature OOM.
1677 * Account all tail pages of THP. This would not cause
1678 * premature OOM since __isolate_lru_page() returns -EBUSY
1679 * only when the page is being freed somewhere else.
1682 switch (__isolate_lru_page(page, mode)) {
1684 nr_taken += nr_pages;
1685 nr_zone_taken[page_zonenum(page)] += nr_pages;
1686 list_move(&page->lru, dst);
1690 /* else it is being freed elsewhere */
1691 list_move(&page->lru, src);
1700 * Splice any skipped pages to the start of the LRU list. Note that
1701 * this disrupts the LRU order when reclaiming for lower zones but
1702 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1703 * scanning would soon rescan the same pages to skip and put the
1704 * system at risk of premature OOM.
1706 if (!list_empty(&pages_skipped)) {
1709 list_splice(&pages_skipped, src);
1710 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1711 if (!nr_skipped[zid])
1714 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1715 skipped += nr_skipped[zid];
1718 *nr_scanned = total_scan;
1719 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1720 total_scan, skipped, nr_taken, mode, lru);
1721 update_lru_sizes(lruvec, lru, nr_zone_taken);
1726 * isolate_lru_page - tries to isolate a page from its LRU list
1727 * @page: page to isolate from its LRU list
1729 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1730 * vmstat statistic corresponding to whatever LRU list the page was on.
1732 * Returns 0 if the page was removed from an LRU list.
1733 * Returns -EBUSY if the page was not on an LRU list.
1735 * The returned page will have PageLRU() cleared. If it was found on
1736 * the active list, it will have PageActive set. If it was found on
1737 * the unevictable list, it will have the PageUnevictable bit set. That flag
1738 * may need to be cleared by the caller before letting the page go.
1740 * The vmstat statistic corresponding to the list on which the page was
1741 * found will be decremented.
1745 * (1) Must be called with an elevated refcount on the page. This is a
1746 * fundamentnal difference from isolate_lru_pages (which is called
1747 * without a stable reference).
1748 * (2) the lru_lock must not be held.
1749 * (3) interrupts must be enabled.
1751 int isolate_lru_page(struct page *page)
1755 VM_BUG_ON_PAGE(!page_count(page), page);
1756 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1758 if (PageLRU(page)) {
1759 pg_data_t *pgdat = page_pgdat(page);
1760 struct lruvec *lruvec;
1762 spin_lock_irq(&pgdat->lru_lock);
1763 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1764 if (PageLRU(page)) {
1765 int lru = page_lru(page);
1768 del_page_from_lru_list(page, lruvec, lru);
1771 spin_unlock_irq(&pgdat->lru_lock);
1777 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1778 * then get resheduled. When there are massive number of tasks doing page
1779 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1780 * the LRU list will go small and be scanned faster than necessary, leading to
1781 * unnecessary swapping, thrashing and OOM.
1783 static int too_many_isolated(struct pglist_data *pgdat, int file,
1784 struct scan_control *sc)
1786 unsigned long inactive, isolated;
1788 if (current_is_kswapd())
1791 if (!writeback_throttling_sane(sc))
1795 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1796 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1798 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1799 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1803 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1804 * won't get blocked by normal direct-reclaimers, forming a circular
1807 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1810 return isolated > inactive;
1814 * This moves pages from @list to corresponding LRU list.
1816 * We move them the other way if the page is referenced by one or more
1817 * processes, from rmap.
1819 * If the pages are mostly unmapped, the processing is fast and it is
1820 * appropriate to hold zone_lru_lock across the whole operation. But if
1821 * the pages are mapped, the processing is slow (page_referenced()) so we
1822 * should drop zone_lru_lock around each page. It's impossible to balance
1823 * this, so instead we remove the pages from the LRU while processing them.
1824 * It is safe to rely on PG_active against the non-LRU pages in here because
1825 * nobody will play with that bit on a non-LRU page.
1827 * The downside is that we have to touch page->_refcount against each page.
1828 * But we had to alter page->flags anyway.
1830 * Returns the number of pages moved to the given lruvec.
1833 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1834 struct list_head *list)
1836 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1837 int nr_pages, nr_moved = 0;
1838 LIST_HEAD(pages_to_free);
1842 while (!list_empty(list)) {
1843 page = lru_to_page(list);
1844 VM_BUG_ON_PAGE(PageLRU(page), page);
1845 if (unlikely(!page_evictable(page))) {
1846 list_del(&page->lru);
1847 spin_unlock_irq(&pgdat->lru_lock);
1848 putback_lru_page(page);
1849 spin_lock_irq(&pgdat->lru_lock);
1852 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1855 lru = page_lru(page);
1857 nr_pages = hpage_nr_pages(page);
1858 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1859 list_move(&page->lru, &lruvec->lists[lru]);
1861 if (put_page_testzero(page)) {
1862 __ClearPageLRU(page);
1863 __ClearPageActive(page);
1864 del_page_from_lru_list(page, lruvec, lru);
1866 if (unlikely(PageCompound(page))) {
1867 spin_unlock_irq(&pgdat->lru_lock);
1868 (*get_compound_page_dtor(page))(page);
1869 spin_lock_irq(&pgdat->lru_lock);
1871 list_add(&page->lru, &pages_to_free);
1873 nr_moved += nr_pages;
1878 * To save our caller's stack, now use input list for pages to free.
1880 list_splice(&pages_to_free, list);
1886 * If a kernel thread (such as nfsd for loop-back mounts) services
1887 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1888 * In that case we should only throttle if the backing device it is
1889 * writing to is congested. In other cases it is safe to throttle.
1891 static int current_may_throttle(void)
1893 return !(current->flags & PF_LESS_THROTTLE) ||
1894 current->backing_dev_info == NULL ||
1895 bdi_write_congested(current->backing_dev_info);
1899 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1900 * of reclaimed pages
1902 static noinline_for_stack unsigned long
1903 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1904 struct scan_control *sc, enum lru_list lru)
1906 LIST_HEAD(page_list);
1907 unsigned long nr_scanned;
1908 unsigned long nr_reclaimed = 0;
1909 unsigned long nr_taken;
1910 struct reclaim_stat stat;
1911 int file = is_file_lru(lru);
1912 enum vm_event_item item;
1913 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1914 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1915 bool stalled = false;
1917 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1921 /* wait a bit for the reclaimer. */
1925 /* We are about to die and free our memory. Return now. */
1926 if (fatal_signal_pending(current))
1927 return SWAP_CLUSTER_MAX;
1932 spin_lock_irq(&pgdat->lru_lock);
1934 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1935 &nr_scanned, sc, lru);
1937 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1938 reclaim_stat->recent_scanned[file] += nr_taken;
1940 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1941 if (!cgroup_reclaim(sc))
1942 __count_vm_events(item, nr_scanned);
1943 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1944 spin_unlock_irq(&pgdat->lru_lock);
1949 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1952 spin_lock_irq(&pgdat->lru_lock);
1954 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1955 if (!cgroup_reclaim(sc))
1956 __count_vm_events(item, nr_reclaimed);
1957 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1958 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1959 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1961 move_pages_to_lru(lruvec, &page_list);
1963 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1965 spin_unlock_irq(&pgdat->lru_lock);
1967 mem_cgroup_uncharge_list(&page_list);
1968 free_unref_page_list(&page_list);
1971 * If dirty pages are scanned that are not queued for IO, it
1972 * implies that flushers are not doing their job. This can
1973 * happen when memory pressure pushes dirty pages to the end of
1974 * the LRU before the dirty limits are breached and the dirty
1975 * data has expired. It can also happen when the proportion of
1976 * dirty pages grows not through writes but through memory
1977 * pressure reclaiming all the clean cache. And in some cases,
1978 * the flushers simply cannot keep up with the allocation
1979 * rate. Nudge the flusher threads in case they are asleep.
1981 if (stat.nr_unqueued_dirty == nr_taken)
1982 wakeup_flusher_threads(WB_REASON_VMSCAN);
1984 sc->nr.dirty += stat.nr_dirty;
1985 sc->nr.congested += stat.nr_congested;
1986 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1987 sc->nr.writeback += stat.nr_writeback;
1988 sc->nr.immediate += stat.nr_immediate;
1989 sc->nr.taken += nr_taken;
1991 sc->nr.file_taken += nr_taken;
1993 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1994 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1995 return nr_reclaimed;
1998 static void shrink_active_list(unsigned long nr_to_scan,
1999 struct lruvec *lruvec,
2000 struct scan_control *sc,
2003 unsigned long nr_taken;
2004 unsigned long nr_scanned;
2005 unsigned long vm_flags;
2006 LIST_HEAD(l_hold); /* The pages which were snipped off */
2007 LIST_HEAD(l_active);
2008 LIST_HEAD(l_inactive);
2010 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2011 unsigned nr_deactivate, nr_activate;
2012 unsigned nr_rotated = 0;
2013 int file = is_file_lru(lru);
2014 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2018 spin_lock_irq(&pgdat->lru_lock);
2020 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2021 &nr_scanned, sc, lru);
2023 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2024 reclaim_stat->recent_scanned[file] += nr_taken;
2026 __count_vm_events(PGREFILL, nr_scanned);
2027 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2029 spin_unlock_irq(&pgdat->lru_lock);
2031 while (!list_empty(&l_hold)) {
2033 page = lru_to_page(&l_hold);
2034 list_del(&page->lru);
2036 if (unlikely(!page_evictable(page))) {
2037 putback_lru_page(page);
2041 if (unlikely(buffer_heads_over_limit)) {
2042 if (page_has_private(page) && trylock_page(page)) {
2043 if (page_has_private(page))
2044 try_to_release_page(page, 0);
2049 if (page_referenced(page, 0, sc->target_mem_cgroup,
2051 nr_rotated += hpage_nr_pages(page);
2053 * Identify referenced, file-backed active pages and
2054 * give them one more trip around the active list. So
2055 * that executable code get better chances to stay in
2056 * memory under moderate memory pressure. Anon pages
2057 * are not likely to be evicted by use-once streaming
2058 * IO, plus JVM can create lots of anon VM_EXEC pages,
2059 * so we ignore them here.
2061 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2062 list_add(&page->lru, &l_active);
2067 ClearPageActive(page); /* we are de-activating */
2068 SetPageWorkingset(page);
2069 list_add(&page->lru, &l_inactive);
2073 * Move pages back to the lru list.
2075 spin_lock_irq(&pgdat->lru_lock);
2077 * Count referenced pages from currently used mappings as rotated,
2078 * even though only some of them are actually re-activated. This
2079 * helps balance scan pressure between file and anonymous pages in
2082 reclaim_stat->recent_rotated[file] += nr_rotated;
2084 nr_activate = move_pages_to_lru(lruvec, &l_active);
2085 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2086 /* Keep all free pages in l_active list */
2087 list_splice(&l_inactive, &l_active);
2089 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2090 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2092 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2093 spin_unlock_irq(&pgdat->lru_lock);
2095 mem_cgroup_uncharge_list(&l_active);
2096 free_unref_page_list(&l_active);
2097 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2098 nr_deactivate, nr_rotated, sc->priority, file);
2101 unsigned long reclaim_pages(struct list_head *page_list)
2104 unsigned long nr_reclaimed = 0;
2105 LIST_HEAD(node_page_list);
2106 struct reclaim_stat dummy_stat;
2108 struct scan_control sc = {
2109 .gfp_mask = GFP_KERNEL,
2110 .priority = DEF_PRIORITY,
2116 while (!list_empty(page_list)) {
2117 page = lru_to_page(page_list);
2119 nid = page_to_nid(page);
2120 INIT_LIST_HEAD(&node_page_list);
2123 if (nid == page_to_nid(page)) {
2124 ClearPageActive(page);
2125 list_move(&page->lru, &node_page_list);
2129 nr_reclaimed += shrink_page_list(&node_page_list,
2132 &dummy_stat, false);
2133 while (!list_empty(&node_page_list)) {
2134 page = lru_to_page(&node_page_list);
2135 list_del(&page->lru);
2136 putback_lru_page(page);
2142 if (!list_empty(&node_page_list)) {
2143 nr_reclaimed += shrink_page_list(&node_page_list,
2146 &dummy_stat, false);
2147 while (!list_empty(&node_page_list)) {
2148 page = lru_to_page(&node_page_list);
2149 list_del(&page->lru);
2150 putback_lru_page(page);
2154 return nr_reclaimed;
2158 * The inactive anon list should be small enough that the VM never has
2159 * to do too much work.
2161 * The inactive file list should be small enough to leave most memory
2162 * to the established workingset on the scan-resistant active list,
2163 * but large enough to avoid thrashing the aggregate readahead window.
2165 * Both inactive lists should also be large enough that each inactive
2166 * page has a chance to be referenced again before it is reclaimed.
2168 * If that fails and refaulting is observed, the inactive list grows.
2170 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2171 * on this LRU, maintained by the pageout code. An inactive_ratio
2172 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2175 * memory ratio inactive
2176 * -------------------------------------
2185 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2186 struct scan_control *sc, bool trace)
2188 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2189 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2190 enum lru_list inactive_lru = file * LRU_FILE;
2191 unsigned long inactive, active;
2192 unsigned long inactive_ratio;
2193 unsigned long refaults;
2196 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2197 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2200 * When refaults are being observed, it means a new workingset
2201 * is being established. Disable active list protection to get
2202 * rid of the stale workingset quickly.
2204 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2205 if (file && lruvec->refaults != refaults) {
2208 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2210 inactive_ratio = int_sqrt(10 * gb);
2216 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2217 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2218 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2219 inactive_ratio, file);
2221 return inactive * inactive_ratio < active;
2224 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2225 struct lruvec *lruvec, struct scan_control *sc)
2227 if (is_active_lru(lru)) {
2228 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2229 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2233 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2244 * Determine how aggressively the anon and file LRU lists should be
2245 * scanned. The relative value of each set of LRU lists is determined
2246 * by looking at the fraction of the pages scanned we did rotate back
2247 * onto the active list instead of evict.
2249 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2250 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2252 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2255 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2256 int swappiness = mem_cgroup_swappiness(memcg);
2257 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2259 u64 denominator = 0; /* gcc */
2260 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2261 unsigned long anon_prio, file_prio;
2262 enum scan_balance scan_balance;
2263 unsigned long anon, file;
2264 unsigned long ap, fp;
2267 /* If we have no swap space, do not bother scanning anon pages. */
2268 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2269 scan_balance = SCAN_FILE;
2274 * Global reclaim will swap to prevent OOM even with no
2275 * swappiness, but memcg users want to use this knob to
2276 * disable swapping for individual groups completely when
2277 * using the memory controller's swap limit feature would be
2280 if (cgroup_reclaim(sc) && !swappiness) {
2281 scan_balance = SCAN_FILE;
2286 * Do not apply any pressure balancing cleverness when the
2287 * system is close to OOM, scan both anon and file equally
2288 * (unless the swappiness setting disagrees with swapping).
2290 if (!sc->priority && swappiness) {
2291 scan_balance = SCAN_EQUAL;
2296 * If the system is almost out of file pages, force-scan anon.
2297 * But only if there are enough inactive anonymous pages on
2298 * the LRU. Otherwise, the small LRU gets thrashed.
2300 if (sc->file_is_tiny &&
2301 !inactive_list_is_low(lruvec, false, sc, false) &&
2302 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON,
2303 sc->reclaim_idx) >> sc->priority) {
2304 scan_balance = SCAN_ANON;
2309 * If there is enough inactive page cache, i.e. if the size of the
2310 * inactive list is greater than that of the active list *and* the
2311 * inactive list actually has some pages to scan on this priority, we
2312 * do not reclaim anything from the anonymous working set right now.
2313 * Without the second condition we could end up never scanning an
2314 * lruvec even if it has plenty of old anonymous pages unless the
2315 * system is under heavy pressure.
2317 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2318 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2319 scan_balance = SCAN_FILE;
2323 scan_balance = SCAN_FRACT;
2326 * With swappiness at 100, anonymous and file have the same priority.
2327 * This scanning priority is essentially the inverse of IO cost.
2329 anon_prio = swappiness;
2330 file_prio = 200 - anon_prio;
2333 * OK, so we have swap space and a fair amount of page cache
2334 * pages. We use the recently rotated / recently scanned
2335 * ratios to determine how valuable each cache is.
2337 * Because workloads change over time (and to avoid overflow)
2338 * we keep these statistics as a floating average, which ends
2339 * up weighing recent references more than old ones.
2341 * anon in [0], file in [1]
2344 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2345 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2346 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2347 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2349 spin_lock_irq(&pgdat->lru_lock);
2350 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2351 reclaim_stat->recent_scanned[0] /= 2;
2352 reclaim_stat->recent_rotated[0] /= 2;
2355 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2356 reclaim_stat->recent_scanned[1] /= 2;
2357 reclaim_stat->recent_rotated[1] /= 2;
2361 * The amount of pressure on anon vs file pages is inversely
2362 * proportional to the fraction of recently scanned pages on
2363 * each list that were recently referenced and in active use.
2365 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2366 ap /= reclaim_stat->recent_rotated[0] + 1;
2368 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2369 fp /= reclaim_stat->recent_rotated[1] + 1;
2370 spin_unlock_irq(&pgdat->lru_lock);
2374 denominator = ap + fp + 1;
2376 for_each_evictable_lru(lru) {
2377 int file = is_file_lru(lru);
2378 unsigned long lruvec_size;
2380 unsigned long protection;
2382 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2383 protection = mem_cgroup_protection(memcg,
2384 sc->memcg_low_reclaim);
2388 * Scale a cgroup's reclaim pressure by proportioning
2389 * its current usage to its memory.low or memory.min
2392 * This is important, as otherwise scanning aggression
2393 * becomes extremely binary -- from nothing as we
2394 * approach the memory protection threshold, to totally
2395 * nominal as we exceed it. This results in requiring
2396 * setting extremely liberal protection thresholds. It
2397 * also means we simply get no protection at all if we
2398 * set it too low, which is not ideal.
2400 * If there is any protection in place, we reduce scan
2401 * pressure by how much of the total memory used is
2402 * within protection thresholds.
2404 * There is one special case: in the first reclaim pass,
2405 * we skip over all groups that are within their low
2406 * protection. If that fails to reclaim enough pages to
2407 * satisfy the reclaim goal, we come back and override
2408 * the best-effort low protection. However, we still
2409 * ideally want to honor how well-behaved groups are in
2410 * that case instead of simply punishing them all
2411 * equally. As such, we reclaim them based on how much
2412 * memory they are using, reducing the scan pressure
2413 * again by how much of the total memory used is under
2416 unsigned long cgroup_size = mem_cgroup_size(memcg);
2418 /* Avoid TOCTOU with earlier protection check */
2419 cgroup_size = max(cgroup_size, protection);
2421 scan = lruvec_size - lruvec_size * protection /
2425 * Minimally target SWAP_CLUSTER_MAX pages to keep
2426 * reclaim moving forwards, avoiding decremeting
2427 * sc->priority further than desirable.
2429 scan = max(scan, SWAP_CLUSTER_MAX);
2434 scan >>= sc->priority;
2437 * If the cgroup's already been deleted, make sure to
2438 * scrape out the remaining cache.
2440 if (!scan && !mem_cgroup_online(memcg))
2441 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2443 switch (scan_balance) {
2445 /* Scan lists relative to size */
2449 * Scan types proportional to swappiness and
2450 * their relative recent reclaim efficiency.
2451 * Make sure we don't miss the last page on
2452 * the offlined memory cgroups because of a
2455 scan = mem_cgroup_online(memcg) ?
2456 div64_u64(scan * fraction[file], denominator) :
2457 DIV64_U64_ROUND_UP(scan * fraction[file],
2462 /* Scan one type exclusively */
2463 if ((scan_balance == SCAN_FILE) != file) {
2469 /* Look ma, no brain */
2477 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2479 unsigned long nr[NR_LRU_LISTS];
2480 unsigned long targets[NR_LRU_LISTS];
2481 unsigned long nr_to_scan;
2483 unsigned long nr_reclaimed = 0;
2484 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2485 struct blk_plug plug;
2488 get_scan_count(lruvec, sc, nr);
2490 /* Record the original scan target for proportional adjustments later */
2491 memcpy(targets, nr, sizeof(nr));
2494 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2495 * event that can occur when there is little memory pressure e.g.
2496 * multiple streaming readers/writers. Hence, we do not abort scanning
2497 * when the requested number of pages are reclaimed when scanning at
2498 * DEF_PRIORITY on the assumption that the fact we are direct
2499 * reclaiming implies that kswapd is not keeping up and it is best to
2500 * do a batch of work at once. For memcg reclaim one check is made to
2501 * abort proportional reclaim if either the file or anon lru has already
2502 * dropped to zero at the first pass.
2504 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2505 sc->priority == DEF_PRIORITY);
2507 blk_start_plug(&plug);
2508 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2509 nr[LRU_INACTIVE_FILE]) {
2510 unsigned long nr_anon, nr_file, percentage;
2511 unsigned long nr_scanned;
2513 for_each_evictable_lru(lru) {
2515 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2516 nr[lru] -= nr_to_scan;
2518 nr_reclaimed += shrink_list(lru, nr_to_scan,
2525 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2529 * For kswapd and memcg, reclaim at least the number of pages
2530 * requested. Ensure that the anon and file LRUs are scanned
2531 * proportionally what was requested by get_scan_count(). We
2532 * stop reclaiming one LRU and reduce the amount scanning
2533 * proportional to the original scan target.
2535 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2536 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2539 * It's just vindictive to attack the larger once the smaller
2540 * has gone to zero. And given the way we stop scanning the
2541 * smaller below, this makes sure that we only make one nudge
2542 * towards proportionality once we've got nr_to_reclaim.
2544 if (!nr_file || !nr_anon)
2547 if (nr_file > nr_anon) {
2548 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2549 targets[LRU_ACTIVE_ANON] + 1;
2551 percentage = nr_anon * 100 / scan_target;
2553 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2554 targets[LRU_ACTIVE_FILE] + 1;
2556 percentage = nr_file * 100 / scan_target;
2559 /* Stop scanning the smaller of the LRU */
2561 nr[lru + LRU_ACTIVE] = 0;
2564 * Recalculate the other LRU scan count based on its original
2565 * scan target and the percentage scanning already complete
2567 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2568 nr_scanned = targets[lru] - nr[lru];
2569 nr[lru] = targets[lru] * (100 - percentage) / 100;
2570 nr[lru] -= min(nr[lru], nr_scanned);
2573 nr_scanned = targets[lru] - nr[lru];
2574 nr[lru] = targets[lru] * (100 - percentage) / 100;
2575 nr[lru] -= min(nr[lru], nr_scanned);
2577 scan_adjusted = true;
2579 blk_finish_plug(&plug);
2580 sc->nr_reclaimed += nr_reclaimed;
2583 * Even if we did not try to evict anon pages at all, we want to
2584 * rebalance the anon lru active/inactive ratio.
2586 if (total_swap_pages && inactive_list_is_low(lruvec, false, sc, true))
2587 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2588 sc, LRU_ACTIVE_ANON);
2591 /* Use reclaim/compaction for costly allocs or under memory pressure */
2592 static bool in_reclaim_compaction(struct scan_control *sc)
2594 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2595 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2596 sc->priority < DEF_PRIORITY - 2))
2603 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2604 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2605 * true if more pages should be reclaimed such that when the page allocator
2606 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2607 * It will give up earlier than that if there is difficulty reclaiming pages.
2609 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2610 unsigned long nr_reclaimed,
2611 struct scan_control *sc)
2613 unsigned long pages_for_compaction;
2614 unsigned long inactive_lru_pages;
2617 /* If not in reclaim/compaction mode, stop */
2618 if (!in_reclaim_compaction(sc))
2622 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2623 * number of pages that were scanned. This will return to the caller
2624 * with the risk reclaim/compaction and the resulting allocation attempt
2625 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2626 * allocations through requiring that the full LRU list has been scanned
2627 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2628 * scan, but that approximation was wrong, and there were corner cases
2629 * where always a non-zero amount of pages were scanned.
2634 /* If compaction would go ahead or the allocation would succeed, stop */
2635 for (z = 0; z <= sc->reclaim_idx; z++) {
2636 struct zone *zone = &pgdat->node_zones[z];
2637 if (!managed_zone(zone))
2640 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2641 case COMPACT_SUCCESS:
2642 case COMPACT_CONTINUE:
2645 /* check next zone */
2651 * If we have not reclaimed enough pages for compaction and the
2652 * inactive lists are large enough, continue reclaiming
2654 pages_for_compaction = compact_gap(sc->order);
2655 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2656 if (get_nr_swap_pages() > 0)
2657 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2659 return inactive_lru_pages > pages_for_compaction;
2662 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2664 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2665 struct mem_cgroup *memcg;
2667 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2669 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2670 unsigned long reclaimed;
2671 unsigned long scanned;
2673 switch (mem_cgroup_protected(target_memcg, memcg)) {
2674 case MEMCG_PROT_MIN:
2677 * If there is no reclaimable memory, OOM.
2680 case MEMCG_PROT_LOW:
2683 * Respect the protection only as long as
2684 * there is an unprotected supply
2685 * of reclaimable memory from other cgroups.
2687 if (!sc->memcg_low_reclaim) {
2688 sc->memcg_low_skipped = 1;
2691 memcg_memory_event(memcg, MEMCG_LOW);
2693 case MEMCG_PROT_NONE:
2695 * All protection thresholds breached. We may
2696 * still choose to vary the scan pressure
2697 * applied based on by how much the cgroup in
2698 * question has exceeded its protection
2699 * thresholds (see get_scan_count).
2704 reclaimed = sc->nr_reclaimed;
2705 scanned = sc->nr_scanned;
2707 shrink_lruvec(lruvec, sc);
2709 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2712 /* Record the group's reclaim efficiency */
2713 vmpressure(sc->gfp_mask, memcg, false,
2714 sc->nr_scanned - scanned,
2715 sc->nr_reclaimed - reclaimed);
2717 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2720 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2722 struct reclaim_state *reclaim_state = current->reclaim_state;
2723 unsigned long nr_reclaimed, nr_scanned;
2724 struct lruvec *target_lruvec;
2725 bool reclaimable = false;
2727 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2730 memset(&sc->nr, 0, sizeof(sc->nr));
2732 nr_reclaimed = sc->nr_reclaimed;
2733 nr_scanned = sc->nr_scanned;
2736 * Prevent the reclaimer from falling into the cache trap: as
2737 * cache pages start out inactive, every cache fault will tip
2738 * the scan balance towards the file LRU. And as the file LRU
2739 * shrinks, so does the window for rotation from references.
2740 * This means we have a runaway feedback loop where a tiny
2741 * thrashing file LRU becomes infinitely more attractive than
2742 * anon pages. Try to detect this based on file LRU size.
2744 if (!cgroup_reclaim(sc)) {
2748 unsigned long total_high_wmark = 0;
2750 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2751 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2752 node_page_state(pgdat, NR_INACTIVE_FILE);
2754 for (z = 0; z < MAX_NR_ZONES; z++) {
2755 struct zone *zone = &pgdat->node_zones[z];
2756 if (!managed_zone(zone))
2759 total_high_wmark += high_wmark_pages(zone);
2762 sc->file_is_tiny = file + free <= total_high_wmark;
2765 shrink_node_memcgs(pgdat, sc);
2767 if (reclaim_state) {
2768 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2769 reclaim_state->reclaimed_slab = 0;
2772 /* Record the subtree's reclaim efficiency */
2773 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2774 sc->nr_scanned - nr_scanned,
2775 sc->nr_reclaimed - nr_reclaimed);
2777 if (sc->nr_reclaimed - nr_reclaimed)
2780 if (current_is_kswapd()) {
2782 * If reclaim is isolating dirty pages under writeback,
2783 * it implies that the long-lived page allocation rate
2784 * is exceeding the page laundering rate. Either the
2785 * global limits are not being effective at throttling
2786 * processes due to the page distribution throughout
2787 * zones or there is heavy usage of a slow backing
2788 * device. The only option is to throttle from reclaim
2789 * context which is not ideal as there is no guarantee
2790 * the dirtying process is throttled in the same way
2791 * balance_dirty_pages() manages.
2793 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2794 * count the number of pages under pages flagged for
2795 * immediate reclaim and stall if any are encountered
2796 * in the nr_immediate check below.
2798 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2799 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2801 /* Allow kswapd to start writing pages during reclaim.*/
2802 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2803 set_bit(PGDAT_DIRTY, &pgdat->flags);
2806 * If kswapd scans pages marked marked for immediate
2807 * reclaim and under writeback (nr_immediate), it
2808 * implies that pages are cycling through the LRU
2809 * faster than they are written so also forcibly stall.
2811 if (sc->nr.immediate)
2812 congestion_wait(BLK_RW_ASYNC, HZ/10);
2816 * Tag a node/memcg as congested if all the dirty pages
2817 * scanned were backed by a congested BDI and
2818 * wait_iff_congested will stall.
2820 * Legacy memcg will stall in page writeback so avoid forcibly
2821 * stalling in wait_iff_congested().
2823 if ((current_is_kswapd() ||
2824 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2825 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2826 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2829 * Stall direct reclaim for IO completions if underlying BDIs
2830 * and node is congested. Allow kswapd to continue until it
2831 * starts encountering unqueued dirty pages or cycling through
2832 * the LRU too quickly.
2834 if (!current_is_kswapd() && current_may_throttle() &&
2835 !sc->hibernation_mode &&
2836 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2837 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2839 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2844 * Kswapd gives up on balancing particular nodes after too
2845 * many failures to reclaim anything from them and goes to
2846 * sleep. On reclaim progress, reset the failure counter. A
2847 * successful direct reclaim run will revive a dormant kswapd.
2850 pgdat->kswapd_failures = 0;
2856 * Returns true if compaction should go ahead for a costly-order request, or
2857 * the allocation would already succeed without compaction. Return false if we
2858 * should reclaim first.
2860 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2862 unsigned long watermark;
2863 enum compact_result suitable;
2865 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2866 if (suitable == COMPACT_SUCCESS)
2867 /* Allocation should succeed already. Don't reclaim. */
2869 if (suitable == COMPACT_SKIPPED)
2870 /* Compaction cannot yet proceed. Do reclaim. */
2874 * Compaction is already possible, but it takes time to run and there
2875 * are potentially other callers using the pages just freed. So proceed
2876 * with reclaim to make a buffer of free pages available to give
2877 * compaction a reasonable chance of completing and allocating the page.
2878 * Note that we won't actually reclaim the whole buffer in one attempt
2879 * as the target watermark in should_continue_reclaim() is lower. But if
2880 * we are already above the high+gap watermark, don't reclaim at all.
2882 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2884 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2888 * This is the direct reclaim path, for page-allocating processes. We only
2889 * try to reclaim pages from zones which will satisfy the caller's allocation
2892 * If a zone is deemed to be full of pinned pages then just give it a light
2893 * scan then give up on it.
2895 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2899 unsigned long nr_soft_reclaimed;
2900 unsigned long nr_soft_scanned;
2902 pg_data_t *last_pgdat = NULL;
2905 * If the number of buffer_heads in the machine exceeds the maximum
2906 * allowed level, force direct reclaim to scan the highmem zone as
2907 * highmem pages could be pinning lowmem pages storing buffer_heads
2909 orig_mask = sc->gfp_mask;
2910 if (buffer_heads_over_limit) {
2911 sc->gfp_mask |= __GFP_HIGHMEM;
2912 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2915 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2916 sc->reclaim_idx, sc->nodemask) {
2918 * Take care memory controller reclaiming has small influence
2921 if (!cgroup_reclaim(sc)) {
2922 if (!cpuset_zone_allowed(zone,
2923 GFP_KERNEL | __GFP_HARDWALL))
2927 * If we already have plenty of memory free for
2928 * compaction in this zone, don't free any more.
2929 * Even though compaction is invoked for any
2930 * non-zero order, only frequent costly order
2931 * reclamation is disruptive enough to become a
2932 * noticeable problem, like transparent huge
2935 if (IS_ENABLED(CONFIG_COMPACTION) &&
2936 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2937 compaction_ready(zone, sc)) {
2938 sc->compaction_ready = true;
2943 * Shrink each node in the zonelist once. If the
2944 * zonelist is ordered by zone (not the default) then a
2945 * node may be shrunk multiple times but in that case
2946 * the user prefers lower zones being preserved.
2948 if (zone->zone_pgdat == last_pgdat)
2952 * This steals pages from memory cgroups over softlimit
2953 * and returns the number of reclaimed pages and
2954 * scanned pages. This works for global memory pressure
2955 * and balancing, not for a memcg's limit.
2957 nr_soft_scanned = 0;
2958 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2959 sc->order, sc->gfp_mask,
2961 sc->nr_reclaimed += nr_soft_reclaimed;
2962 sc->nr_scanned += nr_soft_scanned;
2963 /* need some check for avoid more shrink_zone() */
2966 /* See comment about same check for global reclaim above */
2967 if (zone->zone_pgdat == last_pgdat)
2969 last_pgdat = zone->zone_pgdat;
2970 shrink_node(zone->zone_pgdat, sc);
2974 * Restore to original mask to avoid the impact on the caller if we
2975 * promoted it to __GFP_HIGHMEM.
2977 sc->gfp_mask = orig_mask;
2980 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2982 struct mem_cgroup *memcg;
2984 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2986 unsigned long refaults;
2987 struct lruvec *lruvec;
2989 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2990 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2991 lruvec->refaults = refaults;
2992 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2996 * This is the main entry point to direct page reclaim.
2998 * If a full scan of the inactive list fails to free enough memory then we
2999 * are "out of memory" and something needs to be killed.
3001 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3002 * high - the zone may be full of dirty or under-writeback pages, which this
3003 * caller can't do much about. We kick the writeback threads and take explicit
3004 * naps in the hope that some of these pages can be written. But if the
3005 * allocating task holds filesystem locks which prevent writeout this might not
3006 * work, and the allocation attempt will fail.
3008 * returns: 0, if no pages reclaimed
3009 * else, the number of pages reclaimed
3011 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3012 struct scan_control *sc)
3014 int initial_priority = sc->priority;
3015 pg_data_t *last_pgdat;
3019 delayacct_freepages_start();
3021 if (!cgroup_reclaim(sc))
3022 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3025 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3028 shrink_zones(zonelist, sc);
3030 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3033 if (sc->compaction_ready)
3037 * If we're getting trouble reclaiming, start doing
3038 * writepage even in laptop mode.
3040 if (sc->priority < DEF_PRIORITY - 2)
3041 sc->may_writepage = 1;
3042 } while (--sc->priority >= 0);
3045 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3047 if (zone->zone_pgdat == last_pgdat)
3049 last_pgdat = zone->zone_pgdat;
3051 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3053 if (cgroup_reclaim(sc)) {
3054 struct lruvec *lruvec;
3056 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3058 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3062 delayacct_freepages_end();
3064 if (sc->nr_reclaimed)
3065 return sc->nr_reclaimed;
3067 /* Aborted reclaim to try compaction? don't OOM, then */
3068 if (sc->compaction_ready)
3071 /* Untapped cgroup reserves? Don't OOM, retry. */
3072 if (sc->memcg_low_skipped) {
3073 sc->priority = initial_priority;
3074 sc->memcg_low_reclaim = 1;
3075 sc->memcg_low_skipped = 0;
3082 static bool allow_direct_reclaim(pg_data_t *pgdat)
3085 unsigned long pfmemalloc_reserve = 0;
3086 unsigned long free_pages = 0;
3090 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3093 for (i = 0; i <= ZONE_NORMAL; i++) {
3094 zone = &pgdat->node_zones[i];
3095 if (!managed_zone(zone))
3098 if (!zone_reclaimable_pages(zone))
3101 pfmemalloc_reserve += min_wmark_pages(zone);
3102 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3105 /* If there are no reserves (unexpected config) then do not throttle */
3106 if (!pfmemalloc_reserve)
3109 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3111 /* kswapd must be awake if processes are being throttled */
3112 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3113 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3114 (enum zone_type)ZONE_NORMAL);
3115 wake_up_interruptible(&pgdat->kswapd_wait);
3122 * Throttle direct reclaimers if backing storage is backed by the network
3123 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3124 * depleted. kswapd will continue to make progress and wake the processes
3125 * when the low watermark is reached.
3127 * Returns true if a fatal signal was delivered during throttling. If this
3128 * happens, the page allocator should not consider triggering the OOM killer.
3130 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3131 nodemask_t *nodemask)
3135 pg_data_t *pgdat = NULL;
3138 * Kernel threads should not be throttled as they may be indirectly
3139 * responsible for cleaning pages necessary for reclaim to make forward
3140 * progress. kjournald for example may enter direct reclaim while
3141 * committing a transaction where throttling it could forcing other
3142 * processes to block on log_wait_commit().
3144 if (current->flags & PF_KTHREAD)
3148 * If a fatal signal is pending, this process should not throttle.
3149 * It should return quickly so it can exit and free its memory
3151 if (fatal_signal_pending(current))
3155 * Check if the pfmemalloc reserves are ok by finding the first node
3156 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3157 * GFP_KERNEL will be required for allocating network buffers when
3158 * swapping over the network so ZONE_HIGHMEM is unusable.
3160 * Throttling is based on the first usable node and throttled processes
3161 * wait on a queue until kswapd makes progress and wakes them. There
3162 * is an affinity then between processes waking up and where reclaim
3163 * progress has been made assuming the process wakes on the same node.
3164 * More importantly, processes running on remote nodes will not compete
3165 * for remote pfmemalloc reserves and processes on different nodes
3166 * should make reasonable progress.
3168 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3169 gfp_zone(gfp_mask), nodemask) {
3170 if (zone_idx(zone) > ZONE_NORMAL)
3173 /* Throttle based on the first usable node */
3174 pgdat = zone->zone_pgdat;
3175 if (allow_direct_reclaim(pgdat))
3180 /* If no zone was usable by the allocation flags then do not throttle */
3184 /* Account for the throttling */
3185 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3188 * If the caller cannot enter the filesystem, it's possible that it
3189 * is due to the caller holding an FS lock or performing a journal
3190 * transaction in the case of a filesystem like ext[3|4]. In this case,
3191 * it is not safe to block on pfmemalloc_wait as kswapd could be
3192 * blocked waiting on the same lock. Instead, throttle for up to a
3193 * second before continuing.
3195 if (!(gfp_mask & __GFP_FS)) {
3196 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3197 allow_direct_reclaim(pgdat), HZ);
3202 /* Throttle until kswapd wakes the process */
3203 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3204 allow_direct_reclaim(pgdat));
3207 if (fatal_signal_pending(current))
3214 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3215 gfp_t gfp_mask, nodemask_t *nodemask)
3217 unsigned long nr_reclaimed;
3218 struct scan_control sc = {
3219 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3220 .gfp_mask = current_gfp_context(gfp_mask),
3221 .reclaim_idx = gfp_zone(gfp_mask),
3223 .nodemask = nodemask,
3224 .priority = DEF_PRIORITY,
3225 .may_writepage = !laptop_mode,
3231 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3232 * Confirm they are large enough for max values.
3234 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3235 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3236 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3239 * Do not enter reclaim if fatal signal was delivered while throttled.
3240 * 1 is returned so that the page allocator does not OOM kill at this
3243 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3246 set_task_reclaim_state(current, &sc.reclaim_state);
3247 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3249 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3251 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3252 set_task_reclaim_state(current, NULL);
3254 return nr_reclaimed;
3259 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3260 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3261 gfp_t gfp_mask, bool noswap,
3263 unsigned long *nr_scanned)
3265 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3266 struct scan_control sc = {
3267 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3268 .target_mem_cgroup = memcg,
3269 .may_writepage = !laptop_mode,
3271 .reclaim_idx = MAX_NR_ZONES - 1,
3272 .may_swap = !noswap,
3275 WARN_ON_ONCE(!current->reclaim_state);
3277 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3278 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3280 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3284 * NOTE: Although we can get the priority field, using it
3285 * here is not a good idea, since it limits the pages we can scan.
3286 * if we don't reclaim here, the shrink_node from balance_pgdat
3287 * will pick up pages from other mem cgroup's as well. We hack
3288 * the priority and make it zero.
3290 shrink_lruvec(lruvec, &sc);
3292 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3294 *nr_scanned = sc.nr_scanned;
3296 return sc.nr_reclaimed;
3299 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3300 unsigned long nr_pages,
3304 struct zonelist *zonelist;
3305 unsigned long nr_reclaimed;
3306 unsigned long pflags;
3308 unsigned int noreclaim_flag;
3309 struct scan_control sc = {
3310 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3311 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3312 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3313 .reclaim_idx = MAX_NR_ZONES - 1,
3314 .target_mem_cgroup = memcg,
3315 .priority = DEF_PRIORITY,
3316 .may_writepage = !laptop_mode,
3318 .may_swap = may_swap,
3321 set_task_reclaim_state(current, &sc.reclaim_state);
3323 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3324 * take care of from where we get pages. So the node where we start the
3325 * scan does not need to be the current node.
3327 nid = mem_cgroup_select_victim_node(memcg);
3329 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3331 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3333 psi_memstall_enter(&pflags);
3334 noreclaim_flag = memalloc_noreclaim_save();
3336 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3338 memalloc_noreclaim_restore(noreclaim_flag);
3339 psi_memstall_leave(&pflags);
3341 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3342 set_task_reclaim_state(current, NULL);
3344 return nr_reclaimed;
3348 static void age_active_anon(struct pglist_data *pgdat,
3349 struct scan_control *sc)
3351 struct mem_cgroup *memcg;
3353 if (!total_swap_pages)
3356 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3358 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3360 if (inactive_list_is_low(lruvec, false, sc, true))
3361 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3362 sc, LRU_ACTIVE_ANON);
3364 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3368 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3374 * Check for watermark boosts top-down as the higher zones
3375 * are more likely to be boosted. Both watermarks and boosts
3376 * should not be checked at the time time as reclaim would
3377 * start prematurely when there is no boosting and a lower
3380 for (i = classzone_idx; i >= 0; i--) {
3381 zone = pgdat->node_zones + i;
3382 if (!managed_zone(zone))
3385 if (zone->watermark_boost)
3393 * Returns true if there is an eligible zone balanced for the request order
3396 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3399 unsigned long mark = -1;
3403 * Check watermarks bottom-up as lower zones are more likely to
3406 for (i = 0; i <= classzone_idx; i++) {
3407 zone = pgdat->node_zones + i;
3409 if (!managed_zone(zone))
3412 mark = high_wmark_pages(zone);
3413 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3418 * If a node has no populated zone within classzone_idx, it does not
3419 * need balancing by definition. This can happen if a zone-restricted
3420 * allocation tries to wake a remote kswapd.
3428 /* Clear pgdat state for congested, dirty or under writeback. */
3429 static void clear_pgdat_congested(pg_data_t *pgdat)
3431 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3433 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3434 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3435 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3439 * Prepare kswapd for sleeping. This verifies that there are no processes
3440 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3442 * Returns true if kswapd is ready to sleep
3444 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3447 * The throttled processes are normally woken up in balance_pgdat() as
3448 * soon as allow_direct_reclaim() is true. But there is a potential
3449 * race between when kswapd checks the watermarks and a process gets
3450 * throttled. There is also a potential race if processes get
3451 * throttled, kswapd wakes, a large process exits thereby balancing the
3452 * zones, which causes kswapd to exit balance_pgdat() before reaching
3453 * the wake up checks. If kswapd is going to sleep, no process should
3454 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3455 * the wake up is premature, processes will wake kswapd and get
3456 * throttled again. The difference from wake ups in balance_pgdat() is
3457 * that here we are under prepare_to_wait().
3459 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3460 wake_up_all(&pgdat->pfmemalloc_wait);
3462 /* Hopeless node, leave it to direct reclaim */
3463 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3466 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3467 clear_pgdat_congested(pgdat);
3475 * kswapd shrinks a node of pages that are at or below the highest usable
3476 * zone that is currently unbalanced.
3478 * Returns true if kswapd scanned at least the requested number of pages to
3479 * reclaim or if the lack of progress was due to pages under writeback.
3480 * This is used to determine if the scanning priority needs to be raised.
3482 static bool kswapd_shrink_node(pg_data_t *pgdat,
3483 struct scan_control *sc)
3488 /* Reclaim a number of pages proportional to the number of zones */
3489 sc->nr_to_reclaim = 0;
3490 for (z = 0; z <= sc->reclaim_idx; z++) {
3491 zone = pgdat->node_zones + z;
3492 if (!managed_zone(zone))
3495 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3499 * Historically care was taken to put equal pressure on all zones but
3500 * now pressure is applied based on node LRU order.
3502 shrink_node(pgdat, sc);
3505 * Fragmentation may mean that the system cannot be rebalanced for
3506 * high-order allocations. If twice the allocation size has been
3507 * reclaimed then recheck watermarks only at order-0 to prevent
3508 * excessive reclaim. Assume that a process requested a high-order
3509 * can direct reclaim/compact.
3511 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3514 return sc->nr_scanned >= sc->nr_to_reclaim;
3518 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3519 * that are eligible for use by the caller until at least one zone is
3522 * Returns the order kswapd finished reclaiming at.
3524 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3525 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3526 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3527 * or lower is eligible for reclaim until at least one usable zone is
3530 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3533 unsigned long nr_soft_reclaimed;
3534 unsigned long nr_soft_scanned;
3535 unsigned long pflags;
3536 unsigned long nr_boost_reclaim;
3537 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3540 struct scan_control sc = {
3541 .gfp_mask = GFP_KERNEL,
3546 set_task_reclaim_state(current, &sc.reclaim_state);
3547 psi_memstall_enter(&pflags);
3548 __fs_reclaim_acquire();
3550 count_vm_event(PAGEOUTRUN);
3553 * Account for the reclaim boost. Note that the zone boost is left in
3554 * place so that parallel allocations that are near the watermark will
3555 * stall or direct reclaim until kswapd is finished.
3557 nr_boost_reclaim = 0;
3558 for (i = 0; i <= classzone_idx; i++) {
3559 zone = pgdat->node_zones + i;
3560 if (!managed_zone(zone))
3563 nr_boost_reclaim += zone->watermark_boost;
3564 zone_boosts[i] = zone->watermark_boost;
3566 boosted = nr_boost_reclaim;
3569 sc.priority = DEF_PRIORITY;
3571 unsigned long nr_reclaimed = sc.nr_reclaimed;
3572 bool raise_priority = true;
3576 sc.reclaim_idx = classzone_idx;
3579 * If the number of buffer_heads exceeds the maximum allowed
3580 * then consider reclaiming from all zones. This has a dual
3581 * purpose -- on 64-bit systems it is expected that
3582 * buffer_heads are stripped during active rotation. On 32-bit
3583 * systems, highmem pages can pin lowmem memory and shrinking
3584 * buffers can relieve lowmem pressure. Reclaim may still not
3585 * go ahead if all eligible zones for the original allocation
3586 * request are balanced to avoid excessive reclaim from kswapd.
3588 if (buffer_heads_over_limit) {
3589 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3590 zone = pgdat->node_zones + i;
3591 if (!managed_zone(zone))
3600 * If the pgdat is imbalanced then ignore boosting and preserve
3601 * the watermarks for a later time and restart. Note that the
3602 * zone watermarks will be still reset at the end of balancing
3603 * on the grounds that the normal reclaim should be enough to
3604 * re-evaluate if boosting is required when kswapd next wakes.
3606 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3607 if (!balanced && nr_boost_reclaim) {
3608 nr_boost_reclaim = 0;
3613 * If boosting is not active then only reclaim if there are no
3614 * eligible zones. Note that sc.reclaim_idx is not used as
3615 * buffer_heads_over_limit may have adjusted it.
3617 if (!nr_boost_reclaim && balanced)
3620 /* Limit the priority of boosting to avoid reclaim writeback */
3621 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3622 raise_priority = false;
3625 * Do not writeback or swap pages for boosted reclaim. The
3626 * intent is to relieve pressure not issue sub-optimal IO
3627 * from reclaim context. If no pages are reclaimed, the
3628 * reclaim will be aborted.
3630 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3631 sc.may_swap = !nr_boost_reclaim;
3634 * Do some background aging of the anon list, to give
3635 * pages a chance to be referenced before reclaiming. All
3636 * pages are rotated regardless of classzone as this is
3637 * about consistent aging.
3639 age_active_anon(pgdat, &sc);
3642 * If we're getting trouble reclaiming, start doing writepage
3643 * even in laptop mode.
3645 if (sc.priority < DEF_PRIORITY - 2)
3646 sc.may_writepage = 1;
3648 /* Call soft limit reclaim before calling shrink_node. */
3650 nr_soft_scanned = 0;
3651 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3652 sc.gfp_mask, &nr_soft_scanned);
3653 sc.nr_reclaimed += nr_soft_reclaimed;
3656 * There should be no need to raise the scanning priority if
3657 * enough pages are already being scanned that that high
3658 * watermark would be met at 100% efficiency.
3660 if (kswapd_shrink_node(pgdat, &sc))
3661 raise_priority = false;
3664 * If the low watermark is met there is no need for processes
3665 * to be throttled on pfmemalloc_wait as they should not be
3666 * able to safely make forward progress. Wake them
3668 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3669 allow_direct_reclaim(pgdat))
3670 wake_up_all(&pgdat->pfmemalloc_wait);
3672 /* Check if kswapd should be suspending */
3673 __fs_reclaim_release();
3674 ret = try_to_freeze();
3675 __fs_reclaim_acquire();
3676 if (ret || kthread_should_stop())
3680 * Raise priority if scanning rate is too low or there was no
3681 * progress in reclaiming pages
3683 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3684 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3687 * If reclaim made no progress for a boost, stop reclaim as
3688 * IO cannot be queued and it could be an infinite loop in
3689 * extreme circumstances.
3691 if (nr_boost_reclaim && !nr_reclaimed)
3694 if (raise_priority || !nr_reclaimed)
3696 } while (sc.priority >= 1);
3698 if (!sc.nr_reclaimed)
3699 pgdat->kswapd_failures++;
3702 /* If reclaim was boosted, account for the reclaim done in this pass */
3704 unsigned long flags;
3706 for (i = 0; i <= classzone_idx; i++) {
3707 if (!zone_boosts[i])
3710 /* Increments are under the zone lock */
3711 zone = pgdat->node_zones + i;
3712 spin_lock_irqsave(&zone->lock, flags);
3713 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3714 spin_unlock_irqrestore(&zone->lock, flags);
3718 * As there is now likely space, wakeup kcompact to defragment
3721 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3724 snapshot_refaults(NULL, pgdat);
3725 __fs_reclaim_release();
3726 psi_memstall_leave(&pflags);
3727 set_task_reclaim_state(current, NULL);
3730 * Return the order kswapd stopped reclaiming at as
3731 * prepare_kswapd_sleep() takes it into account. If another caller
3732 * entered the allocator slow path while kswapd was awake, order will
3733 * remain at the higher level.
3739 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3740 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3741 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3742 * after previous reclaim attempt (node is still unbalanced). In that case
3743 * return the zone index of the previous kswapd reclaim cycle.
3745 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3746 enum zone_type prev_classzone_idx)
3748 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3749 return prev_classzone_idx;
3750 return pgdat->kswapd_classzone_idx;
3753 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3754 unsigned int classzone_idx)
3759 if (freezing(current) || kthread_should_stop())
3762 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3765 * Try to sleep for a short interval. Note that kcompactd will only be
3766 * woken if it is possible to sleep for a short interval. This is
3767 * deliberate on the assumption that if reclaim cannot keep an
3768 * eligible zone balanced that it's also unlikely that compaction will
3771 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3773 * Compaction records what page blocks it recently failed to
3774 * isolate pages from and skips them in the future scanning.
3775 * When kswapd is going to sleep, it is reasonable to assume
3776 * that pages and compaction may succeed so reset the cache.
3778 reset_isolation_suitable(pgdat);
3781 * We have freed the memory, now we should compact it to make
3782 * allocation of the requested order possible.
3784 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3786 remaining = schedule_timeout(HZ/10);
3789 * If woken prematurely then reset kswapd_classzone_idx and
3790 * order. The values will either be from a wakeup request or
3791 * the previous request that slept prematurely.
3794 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3795 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3798 finish_wait(&pgdat->kswapd_wait, &wait);
3799 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3803 * After a short sleep, check if it was a premature sleep. If not, then
3804 * go fully to sleep until explicitly woken up.
3807 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3808 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3811 * vmstat counters are not perfectly accurate and the estimated
3812 * value for counters such as NR_FREE_PAGES can deviate from the
3813 * true value by nr_online_cpus * threshold. To avoid the zone
3814 * watermarks being breached while under pressure, we reduce the
3815 * per-cpu vmstat threshold while kswapd is awake and restore
3816 * them before going back to sleep.
3818 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3820 if (!kthread_should_stop())
3823 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3826 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3828 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3830 finish_wait(&pgdat->kswapd_wait, &wait);
3834 * The background pageout daemon, started as a kernel thread
3835 * from the init process.
3837 * This basically trickles out pages so that we have _some_
3838 * free memory available even if there is no other activity
3839 * that frees anything up. This is needed for things like routing
3840 * etc, where we otherwise might have all activity going on in
3841 * asynchronous contexts that cannot page things out.
3843 * If there are applications that are active memory-allocators
3844 * (most normal use), this basically shouldn't matter.
3846 static int kswapd(void *p)
3848 unsigned int alloc_order, reclaim_order;
3849 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3850 pg_data_t *pgdat = (pg_data_t*)p;
3851 struct task_struct *tsk = current;
3852 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3854 if (!cpumask_empty(cpumask))
3855 set_cpus_allowed_ptr(tsk, cpumask);
3858 * Tell the memory management that we're a "memory allocator",
3859 * and that if we need more memory we should get access to it
3860 * regardless (see "__alloc_pages()"). "kswapd" should
3861 * never get caught in the normal page freeing logic.
3863 * (Kswapd normally doesn't need memory anyway, but sometimes
3864 * you need a small amount of memory in order to be able to
3865 * page out something else, and this flag essentially protects
3866 * us from recursively trying to free more memory as we're
3867 * trying to free the first piece of memory in the first place).
3869 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3872 pgdat->kswapd_order = 0;
3873 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3877 alloc_order = reclaim_order = pgdat->kswapd_order;
3878 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3881 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3884 /* Read the new order and classzone_idx */
3885 alloc_order = reclaim_order = pgdat->kswapd_order;
3886 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3887 pgdat->kswapd_order = 0;
3888 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3890 ret = try_to_freeze();
3891 if (kthread_should_stop())
3895 * We can speed up thawing tasks if we don't call balance_pgdat
3896 * after returning from the refrigerator
3902 * Reclaim begins at the requested order but if a high-order
3903 * reclaim fails then kswapd falls back to reclaiming for
3904 * order-0. If that happens, kswapd will consider sleeping
3905 * for the order it finished reclaiming at (reclaim_order)
3906 * but kcompactd is woken to compact for the original
3907 * request (alloc_order).
3909 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3911 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3912 if (reclaim_order < alloc_order)
3913 goto kswapd_try_sleep;
3916 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3922 * A zone is low on free memory or too fragmented for high-order memory. If
3923 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3924 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3925 * has failed or is not needed, still wake up kcompactd if only compaction is
3928 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3929 enum zone_type classzone_idx)
3933 if (!managed_zone(zone))
3936 if (!cpuset_zone_allowed(zone, gfp_flags))
3938 pgdat = zone->zone_pgdat;
3940 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3941 pgdat->kswapd_classzone_idx = classzone_idx;
3943 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3945 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3946 if (!waitqueue_active(&pgdat->kswapd_wait))
3949 /* Hopeless node, leave it to direct reclaim if possible */
3950 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3951 (pgdat_balanced(pgdat, order, classzone_idx) &&
3952 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3954 * There may be plenty of free memory available, but it's too
3955 * fragmented for high-order allocations. Wake up kcompactd
3956 * and rely on compaction_suitable() to determine if it's
3957 * needed. If it fails, it will defer subsequent attempts to
3958 * ratelimit its work.
3960 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3961 wakeup_kcompactd(pgdat, order, classzone_idx);
3965 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3967 wake_up_interruptible(&pgdat->kswapd_wait);
3970 #ifdef CONFIG_HIBERNATION
3972 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3975 * Rather than trying to age LRUs the aim is to preserve the overall
3976 * LRU order by reclaiming preferentially
3977 * inactive > active > active referenced > active mapped
3979 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3981 struct scan_control sc = {
3982 .nr_to_reclaim = nr_to_reclaim,
3983 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3984 .reclaim_idx = MAX_NR_ZONES - 1,
3985 .priority = DEF_PRIORITY,
3989 .hibernation_mode = 1,
3991 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3992 unsigned long nr_reclaimed;
3993 unsigned int noreclaim_flag;
3995 fs_reclaim_acquire(sc.gfp_mask);
3996 noreclaim_flag = memalloc_noreclaim_save();
3997 set_task_reclaim_state(current, &sc.reclaim_state);
3999 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4001 set_task_reclaim_state(current, NULL);
4002 memalloc_noreclaim_restore(noreclaim_flag);
4003 fs_reclaim_release(sc.gfp_mask);
4005 return nr_reclaimed;
4007 #endif /* CONFIG_HIBERNATION */
4009 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4010 not required for correctness. So if the last cpu in a node goes
4011 away, we get changed to run anywhere: as the first one comes back,
4012 restore their cpu bindings. */
4013 static int kswapd_cpu_online(unsigned int cpu)
4017 for_each_node_state(nid, N_MEMORY) {
4018 pg_data_t *pgdat = NODE_DATA(nid);
4019 const struct cpumask *mask;
4021 mask = cpumask_of_node(pgdat->node_id);
4023 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4024 /* One of our CPUs online: restore mask */
4025 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4031 * This kswapd start function will be called by init and node-hot-add.
4032 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4034 int kswapd_run(int nid)
4036 pg_data_t *pgdat = NODE_DATA(nid);
4042 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4043 if (IS_ERR(pgdat->kswapd)) {
4044 /* failure at boot is fatal */
4045 BUG_ON(system_state < SYSTEM_RUNNING);
4046 pr_err("Failed to start kswapd on node %d\n", nid);
4047 ret = PTR_ERR(pgdat->kswapd);
4048 pgdat->kswapd = NULL;
4054 * Called by memory hotplug when all memory in a node is offlined. Caller must
4055 * hold mem_hotplug_begin/end().
4057 void kswapd_stop(int nid)
4059 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4062 kthread_stop(kswapd);
4063 NODE_DATA(nid)->kswapd = NULL;
4067 static int __init kswapd_init(void)
4072 for_each_node_state(nid, N_MEMORY)
4074 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4075 "mm/vmscan:online", kswapd_cpu_online,
4081 module_init(kswapd_init)
4087 * If non-zero call node_reclaim when the number of free pages falls below
4090 int node_reclaim_mode __read_mostly;
4092 #define RECLAIM_OFF 0
4093 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4094 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4095 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4098 * Priority for NODE_RECLAIM. This determines the fraction of pages
4099 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4102 #define NODE_RECLAIM_PRIORITY 4
4105 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4108 int sysctl_min_unmapped_ratio = 1;
4111 * If the number of slab pages in a zone grows beyond this percentage then
4112 * slab reclaim needs to occur.
4114 int sysctl_min_slab_ratio = 5;
4116 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4118 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4119 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4120 node_page_state(pgdat, NR_ACTIVE_FILE);
4123 * It's possible for there to be more file mapped pages than
4124 * accounted for by the pages on the file LRU lists because
4125 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4127 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4130 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4131 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4133 unsigned long nr_pagecache_reclaimable;
4134 unsigned long delta = 0;
4137 * If RECLAIM_UNMAP is set, then all file pages are considered
4138 * potentially reclaimable. Otherwise, we have to worry about
4139 * pages like swapcache and node_unmapped_file_pages() provides
4142 if (node_reclaim_mode & RECLAIM_UNMAP)
4143 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4145 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4147 /* If we can't clean pages, remove dirty pages from consideration */
4148 if (!(node_reclaim_mode & RECLAIM_WRITE))
4149 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4151 /* Watch for any possible underflows due to delta */
4152 if (unlikely(delta > nr_pagecache_reclaimable))
4153 delta = nr_pagecache_reclaimable;
4155 return nr_pagecache_reclaimable - delta;
4159 * Try to free up some pages from this node through reclaim.
4161 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4163 /* Minimum pages needed in order to stay on node */
4164 const unsigned long nr_pages = 1 << order;
4165 struct task_struct *p = current;
4166 unsigned int noreclaim_flag;
4167 struct scan_control sc = {
4168 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4169 .gfp_mask = current_gfp_context(gfp_mask),
4171 .priority = NODE_RECLAIM_PRIORITY,
4172 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4173 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4175 .reclaim_idx = gfp_zone(gfp_mask),
4178 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4182 fs_reclaim_acquire(sc.gfp_mask);
4184 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4185 * and we also need to be able to write out pages for RECLAIM_WRITE
4186 * and RECLAIM_UNMAP.
4188 noreclaim_flag = memalloc_noreclaim_save();
4189 p->flags |= PF_SWAPWRITE;
4190 set_task_reclaim_state(p, &sc.reclaim_state);
4192 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4194 * Free memory by calling shrink node with increasing
4195 * priorities until we have enough memory freed.
4198 shrink_node(pgdat, &sc);
4199 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4202 set_task_reclaim_state(p, NULL);
4203 current->flags &= ~PF_SWAPWRITE;
4204 memalloc_noreclaim_restore(noreclaim_flag);
4205 fs_reclaim_release(sc.gfp_mask);
4207 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4209 return sc.nr_reclaimed >= nr_pages;
4212 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4217 * Node reclaim reclaims unmapped file backed pages and
4218 * slab pages if we are over the defined limits.
4220 * A small portion of unmapped file backed pages is needed for
4221 * file I/O otherwise pages read by file I/O will be immediately
4222 * thrown out if the node is overallocated. So we do not reclaim
4223 * if less than a specified percentage of the node is used by
4224 * unmapped file backed pages.
4226 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4227 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4228 return NODE_RECLAIM_FULL;
4231 * Do not scan if the allocation should not be delayed.
4233 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4234 return NODE_RECLAIM_NOSCAN;
4237 * Only run node reclaim on the local node or on nodes that do not
4238 * have associated processors. This will favor the local processor
4239 * over remote processors and spread off node memory allocations
4240 * as wide as possible.
4242 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4243 return NODE_RECLAIM_NOSCAN;
4245 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4246 return NODE_RECLAIM_NOSCAN;
4248 ret = __node_reclaim(pgdat, gfp_mask, order);
4249 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4252 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4259 * page_evictable - test whether a page is evictable
4260 * @page: the page to test
4262 * Test whether page is evictable--i.e., should be placed on active/inactive
4263 * lists vs unevictable list.
4265 * Reasons page might not be evictable:
4266 * (1) page's mapping marked unevictable
4267 * (2) page is part of an mlocked VMA
4270 int page_evictable(struct page *page)
4274 /* Prevent address_space of inode and swap cache from being freed */
4276 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4282 * check_move_unevictable_pages - check pages for evictability and move to
4283 * appropriate zone lru list
4284 * @pvec: pagevec with lru pages to check
4286 * Checks pages for evictability, if an evictable page is in the unevictable
4287 * lru list, moves it to the appropriate evictable lru list. This function
4288 * should be only used for lru pages.
4290 void check_move_unevictable_pages(struct pagevec *pvec)
4292 struct lruvec *lruvec;
4293 struct pglist_data *pgdat = NULL;
4298 for (i = 0; i < pvec->nr; i++) {
4299 struct page *page = pvec->pages[i];
4300 struct pglist_data *pagepgdat = page_pgdat(page);
4303 if (pagepgdat != pgdat) {
4305 spin_unlock_irq(&pgdat->lru_lock);
4307 spin_lock_irq(&pgdat->lru_lock);
4309 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4311 if (!PageLRU(page) || !PageUnevictable(page))
4314 if (page_evictable(page)) {
4315 enum lru_list lru = page_lru_base_type(page);
4317 VM_BUG_ON_PAGE(PageActive(page), page);
4318 ClearPageUnevictable(page);
4319 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4320 add_page_to_lru_list(page, lruvec, lru);
4326 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4327 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4328 spin_unlock_irq(&pgdat->lru_lock);
4331 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);