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 /* Can active pages be deactivated as part of reclaim? */
83 #define DEACTIVATE_ANON 1
84 #define DEACTIVATE_FILE 2
85 unsigned int may_deactivate:2;
86 unsigned int force_deactivate:1;
87 unsigned int skipped_deactivate:1;
89 /* Writepage batching in laptop mode; RECLAIM_WRITE */
90 unsigned int may_writepage:1;
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
99 * Cgroups are not reclaimed below their configured memory.low,
100 * unless we threaten to OOM. If any cgroups are skipped due to
101 * memory.low and nothing was reclaimed, go back for memory.low.
103 unsigned int memcg_low_reclaim:1;
104 unsigned int memcg_low_skipped:1;
106 unsigned int hibernation_mode:1;
108 /* One of the zones is ready for compaction */
109 unsigned int compaction_ready:1;
111 /* There is easily reclaimable cold cache in the current node */
112 unsigned int cache_trim_mode:1;
114 /* The file pages on the current node are dangerously low */
115 unsigned int file_is_tiny:1;
117 /* Allocation order */
120 /* Scan (total_size >> priority) pages at once */
123 /* The highest zone to isolate pages for reclaim from */
126 /* This context's GFP mask */
129 /* Incremented by the number of inactive pages that were scanned */
130 unsigned long nr_scanned;
132 /* Number of pages freed so far during a call to shrink_zones() */
133 unsigned long nr_reclaimed;
137 unsigned int unqueued_dirty;
138 unsigned int congested;
139 unsigned int writeback;
140 unsigned int immediate;
141 unsigned int file_taken;
145 /* for recording the reclaimed slab by now */
146 struct reclaim_state reclaim_state;
149 #ifdef ARCH_HAS_PREFETCHW
150 #define prefetchw_prev_lru_page(_page, _base, _field) \
152 if ((_page)->lru.prev != _base) { \
155 prev = lru_to_page(&(_page->lru)); \
156 prefetchw(&prev->_field); \
160 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
164 * From 0 .. 100. Higher means more swappy.
166 int vm_swappiness = 60;
168 * The total number of pages which are beyond the high watermark within all
171 unsigned long vm_total_pages;
173 static void set_task_reclaim_state(struct task_struct *task,
174 struct reclaim_state *rs)
176 /* Check for an overwrite */
177 WARN_ON_ONCE(rs && task->reclaim_state);
179 /* Check for the nulling of an already-nulled member */
180 WARN_ON_ONCE(!rs && !task->reclaim_state);
182 task->reclaim_state = rs;
185 static LIST_HEAD(shrinker_list);
186 static DECLARE_RWSEM(shrinker_rwsem);
190 * We allow subsystems to populate their shrinker-related
191 * LRU lists before register_shrinker_prepared() is called
192 * for the shrinker, since we don't want to impose
193 * restrictions on their internal registration order.
194 * In this case shrink_slab_memcg() may find corresponding
195 * bit is set in the shrinkers map.
197 * This value is used by the function to detect registering
198 * shrinkers and to skip do_shrink_slab() calls for them.
200 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202 static DEFINE_IDR(shrinker_idr);
203 static int shrinker_nr_max;
205 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
207 int id, ret = -ENOMEM;
209 down_write(&shrinker_rwsem);
210 /* This may call shrinker, so it must use down_read_trylock() */
211 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
215 if (id >= shrinker_nr_max) {
216 if (memcg_expand_shrinker_maps(id)) {
217 idr_remove(&shrinker_idr, id);
221 shrinker_nr_max = id + 1;
226 up_write(&shrinker_rwsem);
230 static void unregister_memcg_shrinker(struct shrinker *shrinker)
232 int id = shrinker->id;
236 down_write(&shrinker_rwsem);
237 idr_remove(&shrinker_idr, id);
238 up_write(&shrinker_rwsem);
241 static bool cgroup_reclaim(struct scan_control *sc)
243 return sc->target_mem_cgroup;
247 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
248 * @sc: scan_control in question
250 * The normal page dirty throttling mechanism in balance_dirty_pages() is
251 * completely broken with the legacy memcg and direct stalling in
252 * shrink_page_list() is used for throttling instead, which lacks all the
253 * niceties such as fairness, adaptive pausing, bandwidth proportional
254 * allocation and configurability.
256 * This function tests whether the vmscan currently in progress can assume
257 * that the normal dirty throttling mechanism is operational.
259 static bool writeback_throttling_sane(struct scan_control *sc)
261 if (!cgroup_reclaim(sc))
263 #ifdef CONFIG_CGROUP_WRITEBACK
264 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
270 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
275 static void unregister_memcg_shrinker(struct shrinker *shrinker)
279 static bool cgroup_reclaim(struct scan_control *sc)
284 static bool writeback_throttling_sane(struct scan_control *sc)
291 * This misses isolated pages which are not accounted for to save counters.
292 * As the data only determines if reclaim or compaction continues, it is
293 * not expected that isolated pages will be a dominating factor.
295 unsigned long zone_reclaimable_pages(struct zone *zone)
299 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
300 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
301 if (get_nr_swap_pages() > 0)
302 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
303 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
309 * lruvec_lru_size - Returns the number of pages on the given LRU list.
310 * @lruvec: lru vector
312 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
314 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
316 unsigned long size = 0;
319 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
320 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
322 if (!managed_zone(zone))
325 if (!mem_cgroup_disabled())
326 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
328 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
334 * Add a shrinker callback to be called from the vm.
336 int prealloc_shrinker(struct shrinker *shrinker)
338 unsigned int size = sizeof(*shrinker->nr_deferred);
340 if (shrinker->flags & SHRINKER_NUMA_AWARE)
343 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
344 if (!shrinker->nr_deferred)
347 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
348 if (prealloc_memcg_shrinker(shrinker))
355 kfree(shrinker->nr_deferred);
356 shrinker->nr_deferred = NULL;
360 void free_prealloced_shrinker(struct shrinker *shrinker)
362 if (!shrinker->nr_deferred)
365 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
366 unregister_memcg_shrinker(shrinker);
368 kfree(shrinker->nr_deferred);
369 shrinker->nr_deferred = NULL;
372 void register_shrinker_prepared(struct shrinker *shrinker)
374 down_write(&shrinker_rwsem);
375 list_add_tail(&shrinker->list, &shrinker_list);
377 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
378 idr_replace(&shrinker_idr, shrinker, shrinker->id);
380 up_write(&shrinker_rwsem);
383 int register_shrinker(struct shrinker *shrinker)
385 int err = prealloc_shrinker(shrinker);
389 register_shrinker_prepared(shrinker);
392 EXPORT_SYMBOL(register_shrinker);
397 void unregister_shrinker(struct shrinker *shrinker)
399 if (!shrinker->nr_deferred)
401 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
402 unregister_memcg_shrinker(shrinker);
403 down_write(&shrinker_rwsem);
404 list_del(&shrinker->list);
405 up_write(&shrinker_rwsem);
406 kfree(shrinker->nr_deferred);
407 shrinker->nr_deferred = NULL;
409 EXPORT_SYMBOL(unregister_shrinker);
411 #define SHRINK_BATCH 128
413 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
414 struct shrinker *shrinker, int priority)
416 unsigned long freed = 0;
417 unsigned long long delta;
422 int nid = shrinkctl->nid;
423 long batch_size = shrinker->batch ? shrinker->batch
425 long scanned = 0, next_deferred;
427 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
430 freeable = shrinker->count_objects(shrinker, shrinkctl);
431 if (freeable == 0 || freeable == SHRINK_EMPTY)
435 * copy the current shrinker scan count into a local variable
436 * and zero it so that other concurrent shrinker invocations
437 * don't also do this scanning work.
439 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
442 if (shrinker->seeks) {
443 delta = freeable >> priority;
445 do_div(delta, shrinker->seeks);
448 * These objects don't require any IO to create. Trim
449 * them aggressively under memory pressure to keep
450 * them from causing refetches in the IO caches.
452 delta = freeable / 2;
456 if (total_scan < 0) {
457 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
458 shrinker->scan_objects, total_scan);
459 total_scan = freeable;
462 next_deferred = total_scan;
465 * We need to avoid excessive windup on filesystem shrinkers
466 * due to large numbers of GFP_NOFS allocations causing the
467 * shrinkers to return -1 all the time. This results in a large
468 * nr being built up so when a shrink that can do some work
469 * comes along it empties the entire cache due to nr >>>
470 * freeable. This is bad for sustaining a working set in
473 * Hence only allow the shrinker to scan the entire cache when
474 * a large delta change is calculated directly.
476 if (delta < freeable / 4)
477 total_scan = min(total_scan, freeable / 2);
480 * Avoid risking looping forever due to too large nr value:
481 * never try to free more than twice the estimate number of
484 if (total_scan > freeable * 2)
485 total_scan = freeable * 2;
487 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
488 freeable, delta, total_scan, priority);
491 * Normally, we should not scan less than batch_size objects in one
492 * pass to avoid too frequent shrinker calls, but if the slab has less
493 * than batch_size objects in total and we are really tight on memory,
494 * we will try to reclaim all available objects, otherwise we can end
495 * up failing allocations although there are plenty of reclaimable
496 * objects spread over several slabs with usage less than the
499 * We detect the "tight on memory" situations by looking at the total
500 * number of objects we want to scan (total_scan). If it is greater
501 * than the total number of objects on slab (freeable), we must be
502 * scanning at high prio and therefore should try to reclaim as much as
505 while (total_scan >= batch_size ||
506 total_scan >= freeable) {
508 unsigned long nr_to_scan = min(batch_size, total_scan);
510 shrinkctl->nr_to_scan = nr_to_scan;
511 shrinkctl->nr_scanned = nr_to_scan;
512 ret = shrinker->scan_objects(shrinker, shrinkctl);
513 if (ret == SHRINK_STOP)
517 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
518 total_scan -= shrinkctl->nr_scanned;
519 scanned += shrinkctl->nr_scanned;
524 if (next_deferred >= scanned)
525 next_deferred -= scanned;
529 * move the unused scan count back into the shrinker in a
530 * manner that handles concurrent updates. If we exhausted the
531 * scan, there is no need to do an update.
533 if (next_deferred > 0)
534 new_nr = atomic_long_add_return(next_deferred,
535 &shrinker->nr_deferred[nid]);
537 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
539 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
544 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
545 struct mem_cgroup *memcg, int priority)
547 struct memcg_shrinker_map *map;
548 unsigned long ret, freed = 0;
551 if (!mem_cgroup_online(memcg))
554 if (!down_read_trylock(&shrinker_rwsem))
557 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
562 for_each_set_bit(i, map->map, shrinker_nr_max) {
563 struct shrink_control sc = {
564 .gfp_mask = gfp_mask,
568 struct shrinker *shrinker;
570 shrinker = idr_find(&shrinker_idr, i);
571 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
573 clear_bit(i, map->map);
577 /* Call non-slab shrinkers even though kmem is disabled */
578 if (!memcg_kmem_enabled() &&
579 !(shrinker->flags & SHRINKER_NONSLAB))
582 ret = do_shrink_slab(&sc, shrinker, priority);
583 if (ret == SHRINK_EMPTY) {
584 clear_bit(i, map->map);
586 * After the shrinker reported that it had no objects to
587 * free, but before we cleared the corresponding bit in
588 * the memcg shrinker map, a new object might have been
589 * added. To make sure, we have the bit set in this
590 * case, we invoke the shrinker one more time and reset
591 * the bit if it reports that it is not empty anymore.
592 * The memory barrier here pairs with the barrier in
593 * memcg_set_shrinker_bit():
595 * list_lru_add() shrink_slab_memcg()
596 * list_add_tail() clear_bit()
598 * set_bit() do_shrink_slab()
600 smp_mb__after_atomic();
601 ret = do_shrink_slab(&sc, shrinker, priority);
602 if (ret == SHRINK_EMPTY)
605 memcg_set_shrinker_bit(memcg, nid, i);
609 if (rwsem_is_contended(&shrinker_rwsem)) {
615 up_read(&shrinker_rwsem);
618 #else /* CONFIG_MEMCG */
619 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
620 struct mem_cgroup *memcg, int priority)
624 #endif /* CONFIG_MEMCG */
627 * shrink_slab - shrink slab caches
628 * @gfp_mask: allocation context
629 * @nid: node whose slab caches to target
630 * @memcg: memory cgroup whose slab caches to target
631 * @priority: the reclaim priority
633 * Call the shrink functions to age shrinkable caches.
635 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
636 * unaware shrinkers will receive a node id of 0 instead.
638 * @memcg specifies the memory cgroup to target. Unaware shrinkers
639 * are called only if it is the root cgroup.
641 * @priority is sc->priority, we take the number of objects and >> by priority
642 * in order to get the scan target.
644 * Returns the number of reclaimed slab objects.
646 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
647 struct mem_cgroup *memcg,
650 unsigned long ret, freed = 0;
651 struct shrinker *shrinker;
654 * The root memcg might be allocated even though memcg is disabled
655 * via "cgroup_disable=memory" boot parameter. This could make
656 * mem_cgroup_is_root() return false, then just run memcg slab
657 * shrink, but skip global shrink. This may result in premature
660 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
661 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
663 if (!down_read_trylock(&shrinker_rwsem))
666 list_for_each_entry(shrinker, &shrinker_list, list) {
667 struct shrink_control sc = {
668 .gfp_mask = gfp_mask,
673 ret = do_shrink_slab(&sc, shrinker, priority);
674 if (ret == SHRINK_EMPTY)
678 * Bail out if someone want to register a new shrinker to
679 * prevent the regsitration from being stalled for long periods
680 * by parallel ongoing shrinking.
682 if (rwsem_is_contended(&shrinker_rwsem)) {
688 up_read(&shrinker_rwsem);
694 void drop_slab_node(int nid)
699 struct mem_cgroup *memcg = NULL;
702 memcg = mem_cgroup_iter(NULL, NULL, NULL);
704 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
705 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
706 } while (freed > 10);
713 for_each_online_node(nid)
717 static inline int is_page_cache_freeable(struct page *page)
720 * A freeable page cache page is referenced only by the caller
721 * that isolated the page, the page cache and optional buffer
722 * heads at page->private.
724 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
726 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
729 static int may_write_to_inode(struct inode *inode)
731 if (current->flags & PF_SWAPWRITE)
733 if (!inode_write_congested(inode))
735 if (inode_to_bdi(inode) == current->backing_dev_info)
741 * We detected a synchronous write error writing a page out. Probably
742 * -ENOSPC. We need to propagate that into the address_space for a subsequent
743 * fsync(), msync() or close().
745 * The tricky part is that after writepage we cannot touch the mapping: nothing
746 * prevents it from being freed up. But we have a ref on the page and once
747 * that page is locked, the mapping is pinned.
749 * We're allowed to run sleeping lock_page() here because we know the caller has
752 static void handle_write_error(struct address_space *mapping,
753 struct page *page, int error)
756 if (page_mapping(page) == mapping)
757 mapping_set_error(mapping, error);
761 /* possible outcome of pageout() */
763 /* failed to write page out, page is locked */
765 /* move page to the active list, page is locked */
767 /* page has been sent to the disk successfully, page is unlocked */
769 /* page is clean and locked */
774 * pageout is called by shrink_page_list() for each dirty page.
775 * Calls ->writepage().
777 static pageout_t pageout(struct page *page, struct address_space *mapping)
780 * If the page is dirty, only perform writeback if that write
781 * will be non-blocking. To prevent this allocation from being
782 * stalled by pagecache activity. But note that there may be
783 * stalls if we need to run get_block(). We could test
784 * PagePrivate for that.
786 * If this process is currently in __generic_file_write_iter() against
787 * this page's queue, we can perform writeback even if that
790 * If the page is swapcache, write it back even if that would
791 * block, for some throttling. This happens by accident, because
792 * swap_backing_dev_info is bust: it doesn't reflect the
793 * congestion state of the swapdevs. Easy to fix, if needed.
795 if (!is_page_cache_freeable(page))
799 * Some data journaling orphaned pages can have
800 * page->mapping == NULL while being dirty with clean buffers.
802 if (page_has_private(page)) {
803 if (try_to_free_buffers(page)) {
804 ClearPageDirty(page);
805 pr_info("%s: orphaned page\n", __func__);
811 if (mapping->a_ops->writepage == NULL)
812 return PAGE_ACTIVATE;
813 if (!may_write_to_inode(mapping->host))
816 if (clear_page_dirty_for_io(page)) {
818 struct writeback_control wbc = {
819 .sync_mode = WB_SYNC_NONE,
820 .nr_to_write = SWAP_CLUSTER_MAX,
822 .range_end = LLONG_MAX,
826 SetPageReclaim(page);
827 res = mapping->a_ops->writepage(page, &wbc);
829 handle_write_error(mapping, page, res);
830 if (res == AOP_WRITEPAGE_ACTIVATE) {
831 ClearPageReclaim(page);
832 return PAGE_ACTIVATE;
835 if (!PageWriteback(page)) {
836 /* synchronous write or broken a_ops? */
837 ClearPageReclaim(page);
839 trace_mm_vmscan_writepage(page);
840 inc_node_page_state(page, NR_VMSCAN_WRITE);
848 * Same as remove_mapping, but if the page is removed from the mapping, it
849 * gets returned with a refcount of 0.
851 static int __remove_mapping(struct address_space *mapping, struct page *page,
852 bool reclaimed, struct mem_cgroup *target_memcg)
857 BUG_ON(!PageLocked(page));
858 BUG_ON(mapping != page_mapping(page));
860 xa_lock_irqsave(&mapping->i_pages, flags);
862 * The non racy check for a busy page.
864 * Must be careful with the order of the tests. When someone has
865 * a ref to the page, it may be possible that they dirty it then
866 * drop the reference. So if PageDirty is tested before page_count
867 * here, then the following race may occur:
869 * get_user_pages(&page);
870 * [user mapping goes away]
872 * !PageDirty(page) [good]
873 * SetPageDirty(page);
875 * !page_count(page) [good, discard it]
877 * [oops, our write_to data is lost]
879 * Reversing the order of the tests ensures such a situation cannot
880 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
881 * load is not satisfied before that of page->_refcount.
883 * Note that if SetPageDirty is always performed via set_page_dirty,
884 * and thus under the i_pages lock, then this ordering is not required.
886 refcount = 1 + compound_nr(page);
887 if (!page_ref_freeze(page, refcount))
889 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
890 if (unlikely(PageDirty(page))) {
891 page_ref_unfreeze(page, refcount);
895 if (PageSwapCache(page)) {
896 swp_entry_t swap = { .val = page_private(page) };
897 mem_cgroup_swapout(page, swap);
898 __delete_from_swap_cache(page, swap);
899 xa_unlock_irqrestore(&mapping->i_pages, flags);
900 put_swap_page(page, swap);
902 void (*freepage)(struct page *);
905 freepage = mapping->a_ops->freepage;
907 * Remember a shadow entry for reclaimed file cache in
908 * order to detect refaults, thus thrashing, later on.
910 * But don't store shadows in an address space that is
911 * already exiting. This is not just an optizimation,
912 * inode reclaim needs to empty out the radix tree or
913 * the nodes are lost. Don't plant shadows behind its
916 * We also don't store shadows for DAX mappings because the
917 * only page cache pages found in these are zero pages
918 * covering holes, and because we don't want to mix DAX
919 * exceptional entries and shadow exceptional entries in the
920 * same address_space.
922 if (reclaimed && page_is_file_cache(page) &&
923 !mapping_exiting(mapping) && !dax_mapping(mapping))
924 shadow = workingset_eviction(page, target_memcg);
925 __delete_from_page_cache(page, shadow);
926 xa_unlock_irqrestore(&mapping->i_pages, flags);
928 if (freepage != NULL)
935 xa_unlock_irqrestore(&mapping->i_pages, flags);
940 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
941 * someone else has a ref on the page, abort and return 0. If it was
942 * successfully detached, return 1. Assumes the caller has a single ref on
945 int remove_mapping(struct address_space *mapping, struct page *page)
947 if (__remove_mapping(mapping, page, false, NULL)) {
949 * Unfreezing the refcount with 1 rather than 2 effectively
950 * drops the pagecache ref for us without requiring another
953 page_ref_unfreeze(page, 1);
960 * putback_lru_page - put previously isolated page onto appropriate LRU list
961 * @page: page to be put back to appropriate lru list
963 * Add previously isolated @page to appropriate LRU list.
964 * Page may still be unevictable for other reasons.
966 * lru_lock must not be held, interrupts must be enabled.
968 void putback_lru_page(struct page *page)
971 put_page(page); /* drop ref from isolate */
974 enum page_references {
976 PAGEREF_RECLAIM_CLEAN,
981 static enum page_references page_check_references(struct page *page,
982 struct scan_control *sc)
984 int referenced_ptes, referenced_page;
985 unsigned long vm_flags;
987 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
989 referenced_page = TestClearPageReferenced(page);
992 * Mlock lost the isolation race with us. Let try_to_unmap()
993 * move the page to the unevictable list.
995 if (vm_flags & VM_LOCKED)
996 return PAGEREF_RECLAIM;
998 if (referenced_ptes) {
999 if (PageSwapBacked(page))
1000 return PAGEREF_ACTIVATE;
1002 * All mapped pages start out with page table
1003 * references from the instantiating fault, so we need
1004 * to look twice if a mapped file page is used more
1007 * Mark it and spare it for another trip around the
1008 * inactive list. Another page table reference will
1009 * lead to its activation.
1011 * Note: the mark is set for activated pages as well
1012 * so that recently deactivated but used pages are
1013 * quickly recovered.
1015 SetPageReferenced(page);
1017 if (referenced_page || referenced_ptes > 1)
1018 return PAGEREF_ACTIVATE;
1021 * Activate file-backed executable pages after first usage.
1023 if (vm_flags & VM_EXEC)
1024 return PAGEREF_ACTIVATE;
1026 return PAGEREF_KEEP;
1029 /* Reclaim if clean, defer dirty pages to writeback */
1030 if (referenced_page && !PageSwapBacked(page))
1031 return PAGEREF_RECLAIM_CLEAN;
1033 return PAGEREF_RECLAIM;
1036 /* Check if a page is dirty or under writeback */
1037 static void page_check_dirty_writeback(struct page *page,
1038 bool *dirty, bool *writeback)
1040 struct address_space *mapping;
1043 * Anonymous pages are not handled by flushers and must be written
1044 * from reclaim context. Do not stall reclaim based on them
1046 if (!page_is_file_cache(page) ||
1047 (PageAnon(page) && !PageSwapBacked(page))) {
1053 /* By default assume that the page flags are accurate */
1054 *dirty = PageDirty(page);
1055 *writeback = PageWriteback(page);
1057 /* Verify dirty/writeback state if the filesystem supports it */
1058 if (!page_has_private(page))
1061 mapping = page_mapping(page);
1062 if (mapping && mapping->a_ops->is_dirty_writeback)
1063 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1067 * shrink_page_list() returns the number of reclaimed pages
1069 static unsigned long shrink_page_list(struct list_head *page_list,
1070 struct pglist_data *pgdat,
1071 struct scan_control *sc,
1072 enum ttu_flags ttu_flags,
1073 struct reclaim_stat *stat,
1074 bool ignore_references)
1076 LIST_HEAD(ret_pages);
1077 LIST_HEAD(free_pages);
1078 unsigned nr_reclaimed = 0;
1079 unsigned pgactivate = 0;
1081 memset(stat, 0, sizeof(*stat));
1084 while (!list_empty(page_list)) {
1085 struct address_space *mapping;
1088 enum page_references references = PAGEREF_RECLAIM;
1089 bool dirty, writeback;
1090 unsigned int nr_pages;
1094 page = lru_to_page(page_list);
1095 list_del(&page->lru);
1097 if (!trylock_page(page))
1100 VM_BUG_ON_PAGE(PageActive(page), page);
1102 nr_pages = compound_nr(page);
1104 /* Account the number of base pages even though THP */
1105 sc->nr_scanned += nr_pages;
1107 if (unlikely(!page_evictable(page)))
1108 goto activate_locked;
1110 if (!sc->may_unmap && page_mapped(page))
1113 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1114 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1117 * The number of dirty pages determines if a node is marked
1118 * reclaim_congested which affects wait_iff_congested. kswapd
1119 * will stall and start writing pages if the tail of the LRU
1120 * is all dirty unqueued pages.
1122 page_check_dirty_writeback(page, &dirty, &writeback);
1123 if (dirty || writeback)
1126 if (dirty && !writeback)
1127 stat->nr_unqueued_dirty++;
1130 * Treat this page as congested if the underlying BDI is or if
1131 * pages are cycling through the LRU so quickly that the
1132 * pages marked for immediate reclaim are making it to the
1133 * end of the LRU a second time.
1135 mapping = page_mapping(page);
1136 if (((dirty || writeback) && mapping &&
1137 inode_write_congested(mapping->host)) ||
1138 (writeback && PageReclaim(page)))
1139 stat->nr_congested++;
1142 * If a page at the tail of the LRU is under writeback, there
1143 * are three cases to consider.
1145 * 1) If reclaim is encountering an excessive number of pages
1146 * under writeback and this page is both under writeback and
1147 * PageReclaim then it indicates that pages are being queued
1148 * for IO but are being recycled through the LRU before the
1149 * IO can complete. Waiting on the page itself risks an
1150 * indefinite stall if it is impossible to writeback the
1151 * page due to IO error or disconnected storage so instead
1152 * note that the LRU is being scanned too quickly and the
1153 * caller can stall after page list has been processed.
1155 * 2) Global or new memcg reclaim encounters a page that is
1156 * not marked for immediate reclaim, or the caller does not
1157 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1158 * not to fs). In this case mark the page for immediate
1159 * reclaim and continue scanning.
1161 * Require may_enter_fs because we would wait on fs, which
1162 * may not have submitted IO yet. And the loop driver might
1163 * enter reclaim, and deadlock if it waits on a page for
1164 * which it is needed to do the write (loop masks off
1165 * __GFP_IO|__GFP_FS for this reason); but more thought
1166 * would probably show more reasons.
1168 * 3) Legacy memcg encounters a page that is already marked
1169 * PageReclaim. memcg does not have any dirty pages
1170 * throttling so we could easily OOM just because too many
1171 * pages are in writeback and there is nothing else to
1172 * reclaim. Wait for the writeback to complete.
1174 * In cases 1) and 2) we activate the pages to get them out of
1175 * the way while we continue scanning for clean pages on the
1176 * inactive list and refilling from the active list. The
1177 * observation here is that waiting for disk writes is more
1178 * expensive than potentially causing reloads down the line.
1179 * Since they're marked for immediate reclaim, they won't put
1180 * memory pressure on the cache working set any longer than it
1181 * takes to write them to disk.
1183 if (PageWriteback(page)) {
1185 if (current_is_kswapd() &&
1186 PageReclaim(page) &&
1187 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1188 stat->nr_immediate++;
1189 goto activate_locked;
1192 } else if (writeback_throttling_sane(sc) ||
1193 !PageReclaim(page) || !may_enter_fs) {
1195 * This is slightly racy - end_page_writeback()
1196 * might have just cleared PageReclaim, then
1197 * setting PageReclaim here end up interpreted
1198 * as PageReadahead - but that does not matter
1199 * enough to care. What we do want is for this
1200 * page to have PageReclaim set next time memcg
1201 * reclaim reaches the tests above, so it will
1202 * then wait_on_page_writeback() to avoid OOM;
1203 * and it's also appropriate in global reclaim.
1205 SetPageReclaim(page);
1206 stat->nr_writeback++;
1207 goto activate_locked;
1212 wait_on_page_writeback(page);
1213 /* then go back and try same page again */
1214 list_add_tail(&page->lru, page_list);
1219 if (!ignore_references)
1220 references = page_check_references(page, sc);
1222 switch (references) {
1223 case PAGEREF_ACTIVATE:
1224 goto activate_locked;
1226 stat->nr_ref_keep += nr_pages;
1228 case PAGEREF_RECLAIM:
1229 case PAGEREF_RECLAIM_CLEAN:
1230 ; /* try to reclaim the page below */
1234 * Anonymous process memory has backing store?
1235 * Try to allocate it some swap space here.
1236 * Lazyfree page could be freed directly
1238 if (PageAnon(page) && PageSwapBacked(page)) {
1239 if (!PageSwapCache(page)) {
1240 if (!(sc->gfp_mask & __GFP_IO))
1242 if (PageTransHuge(page)) {
1243 /* cannot split THP, skip it */
1244 if (!can_split_huge_page(page, NULL))
1245 goto activate_locked;
1247 * Split pages without a PMD map right
1248 * away. Chances are some or all of the
1249 * tail pages can be freed without IO.
1251 if (!compound_mapcount(page) &&
1252 split_huge_page_to_list(page,
1254 goto activate_locked;
1256 if (!add_to_swap(page)) {
1257 if (!PageTransHuge(page))
1258 goto activate_locked_split;
1259 /* Fallback to swap normal pages */
1260 if (split_huge_page_to_list(page,
1262 goto activate_locked;
1263 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1264 count_vm_event(THP_SWPOUT_FALLBACK);
1266 if (!add_to_swap(page))
1267 goto activate_locked_split;
1272 /* Adding to swap updated mapping */
1273 mapping = page_mapping(page);
1275 } else if (unlikely(PageTransHuge(page))) {
1276 /* Split file THP */
1277 if (split_huge_page_to_list(page, page_list))
1282 * THP may get split above, need minus tail pages and update
1283 * nr_pages to avoid accounting tail pages twice.
1285 * The tail pages that are added into swap cache successfully
1288 if ((nr_pages > 1) && !PageTransHuge(page)) {
1289 sc->nr_scanned -= (nr_pages - 1);
1294 * The page is mapped into the page tables of one or more
1295 * processes. Try to unmap it here.
1297 if (page_mapped(page)) {
1298 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1300 if (unlikely(PageTransHuge(page)))
1301 flags |= TTU_SPLIT_HUGE_PMD;
1302 if (!try_to_unmap(page, flags)) {
1303 stat->nr_unmap_fail += nr_pages;
1304 goto activate_locked;
1308 if (PageDirty(page)) {
1310 * Only kswapd can writeback filesystem pages
1311 * to avoid risk of stack overflow. But avoid
1312 * injecting inefficient single-page IO into
1313 * flusher writeback as much as possible: only
1314 * write pages when we've encountered many
1315 * dirty pages, and when we've already scanned
1316 * the rest of the LRU for clean pages and see
1317 * the same dirty pages again (PageReclaim).
1319 if (page_is_file_cache(page) &&
1320 (!current_is_kswapd() || !PageReclaim(page) ||
1321 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1323 * Immediately reclaim when written back.
1324 * Similar in principal to deactivate_page()
1325 * except we already have the page isolated
1326 * and know it's dirty
1328 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1329 SetPageReclaim(page);
1331 goto activate_locked;
1334 if (references == PAGEREF_RECLAIM_CLEAN)
1338 if (!sc->may_writepage)
1342 * Page is dirty. Flush the TLB if a writable entry
1343 * potentially exists to avoid CPU writes after IO
1344 * starts and then write it out here.
1346 try_to_unmap_flush_dirty();
1347 switch (pageout(page, mapping)) {
1351 goto activate_locked;
1353 if (PageWriteback(page))
1355 if (PageDirty(page))
1359 * A synchronous write - probably a ramdisk. Go
1360 * ahead and try to reclaim the page.
1362 if (!trylock_page(page))
1364 if (PageDirty(page) || PageWriteback(page))
1366 mapping = page_mapping(page);
1368 ; /* try to free the page below */
1373 * If the page has buffers, try to free the buffer mappings
1374 * associated with this page. If we succeed we try to free
1377 * We do this even if the page is PageDirty().
1378 * try_to_release_page() does not perform I/O, but it is
1379 * possible for a page to have PageDirty set, but it is actually
1380 * clean (all its buffers are clean). This happens if the
1381 * buffers were written out directly, with submit_bh(). ext3
1382 * will do this, as well as the blockdev mapping.
1383 * try_to_release_page() will discover that cleanness and will
1384 * drop the buffers and mark the page clean - it can be freed.
1386 * Rarely, pages can have buffers and no ->mapping. These are
1387 * the pages which were not successfully invalidated in
1388 * truncate_complete_page(). We try to drop those buffers here
1389 * and if that worked, and the page is no longer mapped into
1390 * process address space (page_count == 1) it can be freed.
1391 * Otherwise, leave the page on the LRU so it is swappable.
1393 if (page_has_private(page)) {
1394 if (!try_to_release_page(page, sc->gfp_mask))
1395 goto activate_locked;
1396 if (!mapping && page_count(page) == 1) {
1398 if (put_page_testzero(page))
1402 * rare race with speculative reference.
1403 * the speculative reference will free
1404 * this page shortly, so we may
1405 * increment nr_reclaimed here (and
1406 * leave it off the LRU).
1414 if (PageAnon(page) && !PageSwapBacked(page)) {
1415 /* follow __remove_mapping for reference */
1416 if (!page_ref_freeze(page, 1))
1418 if (PageDirty(page)) {
1419 page_ref_unfreeze(page, 1);
1423 count_vm_event(PGLAZYFREED);
1424 count_memcg_page_event(page, PGLAZYFREED);
1425 } else if (!mapping || !__remove_mapping(mapping, page, true,
1426 sc->target_mem_cgroup))
1432 * THP may get swapped out in a whole, need account
1435 nr_reclaimed += nr_pages;
1438 * Is there need to periodically free_page_list? It would
1439 * appear not as the counts should be low
1441 if (unlikely(PageTransHuge(page)))
1442 (*get_compound_page_dtor(page))(page);
1444 list_add(&page->lru, &free_pages);
1447 activate_locked_split:
1449 * The tail pages that are failed to add into swap cache
1450 * reach here. Fixup nr_scanned and nr_pages.
1453 sc->nr_scanned -= (nr_pages - 1);
1457 /* Not a candidate for swapping, so reclaim swap space. */
1458 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1460 try_to_free_swap(page);
1461 VM_BUG_ON_PAGE(PageActive(page), page);
1462 if (!PageMlocked(page)) {
1463 int type = page_is_file_cache(page);
1464 SetPageActive(page);
1465 stat->nr_activate[type] += nr_pages;
1466 count_memcg_page_event(page, PGACTIVATE);
1471 list_add(&page->lru, &ret_pages);
1472 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1475 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1477 mem_cgroup_uncharge_list(&free_pages);
1478 try_to_unmap_flush();
1479 free_unref_page_list(&free_pages);
1481 list_splice(&ret_pages, page_list);
1482 count_vm_events(PGACTIVATE, pgactivate);
1484 return nr_reclaimed;
1487 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1488 struct list_head *page_list)
1490 struct scan_control sc = {
1491 .gfp_mask = GFP_KERNEL,
1492 .priority = DEF_PRIORITY,
1495 struct reclaim_stat dummy_stat;
1497 struct page *page, *next;
1498 LIST_HEAD(clean_pages);
1500 list_for_each_entry_safe(page, next, page_list, lru) {
1501 if (page_is_file_cache(page) && !PageDirty(page) &&
1502 !__PageMovable(page) && !PageUnevictable(page)) {
1503 ClearPageActive(page);
1504 list_move(&page->lru, &clean_pages);
1508 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1509 TTU_IGNORE_ACCESS, &dummy_stat, true);
1510 list_splice(&clean_pages, page_list);
1511 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1516 * Attempt to remove the specified page from its LRU. Only take this page
1517 * if it is of the appropriate PageActive status. Pages which are being
1518 * freed elsewhere are also ignored.
1520 * page: page to consider
1521 * mode: one of the LRU isolation modes defined above
1523 * returns 0 on success, -ve errno on failure.
1525 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1529 /* Only take pages on the LRU. */
1533 /* Compaction should not handle unevictable pages but CMA can do so */
1534 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1540 * To minimise LRU disruption, the caller can indicate that it only
1541 * wants to isolate pages it will be able to operate on without
1542 * blocking - clean pages for the most part.
1544 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1545 * that it is possible to migrate without blocking
1547 if (mode & ISOLATE_ASYNC_MIGRATE) {
1548 /* All the caller can do on PageWriteback is block */
1549 if (PageWriteback(page))
1552 if (PageDirty(page)) {
1553 struct address_space *mapping;
1557 * Only pages without mappings or that have a
1558 * ->migratepage callback are possible to migrate
1559 * without blocking. However, we can be racing with
1560 * truncation so it's necessary to lock the page
1561 * to stabilise the mapping as truncation holds
1562 * the page lock until after the page is removed
1563 * from the page cache.
1565 if (!trylock_page(page))
1568 mapping = page_mapping(page);
1569 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1576 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1579 if (likely(get_page_unless_zero(page))) {
1581 * Be careful not to clear PageLRU until after we're
1582 * sure the page is not being freed elsewhere -- the
1583 * page release code relies on it.
1594 * Update LRU sizes after isolating pages. The LRU size updates must
1595 * be complete before mem_cgroup_update_lru_size due to a santity check.
1597 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1598 enum lru_list lru, unsigned long *nr_zone_taken)
1602 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1603 if (!nr_zone_taken[zid])
1606 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1608 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1615 * pgdat->lru_lock is heavily contended. Some of the functions that
1616 * shrink the lists perform better by taking out a batch of pages
1617 * and working on them outside the LRU lock.
1619 * For pagecache intensive workloads, this function is the hottest
1620 * spot in the kernel (apart from copy_*_user functions).
1622 * Appropriate locks must be held before calling this function.
1624 * @nr_to_scan: The number of eligible pages to look through on the list.
1625 * @lruvec: The LRU vector to pull pages from.
1626 * @dst: The temp list to put pages on to.
1627 * @nr_scanned: The number of pages that were scanned.
1628 * @sc: The scan_control struct for this reclaim session
1629 * @mode: One of the LRU isolation modes
1630 * @lru: LRU list id for isolating
1632 * returns how many pages were moved onto *@dst.
1634 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1635 struct lruvec *lruvec, struct list_head *dst,
1636 unsigned long *nr_scanned, struct scan_control *sc,
1639 struct list_head *src = &lruvec->lists[lru];
1640 unsigned long nr_taken = 0;
1641 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1642 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1643 unsigned long skipped = 0;
1644 unsigned long scan, total_scan, nr_pages;
1645 LIST_HEAD(pages_skipped);
1646 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1650 while (scan < nr_to_scan && !list_empty(src)) {
1653 page = lru_to_page(src);
1654 prefetchw_prev_lru_page(page, src, flags);
1656 VM_BUG_ON_PAGE(!PageLRU(page), page);
1658 nr_pages = compound_nr(page);
1659 total_scan += nr_pages;
1661 if (page_zonenum(page) > sc->reclaim_idx) {
1662 list_move(&page->lru, &pages_skipped);
1663 nr_skipped[page_zonenum(page)] += nr_pages;
1668 * Do not count skipped pages because that makes the function
1669 * return with no isolated pages if the LRU mostly contains
1670 * ineligible pages. This causes the VM to not reclaim any
1671 * pages, triggering a premature OOM.
1673 * Account all tail pages of THP. This would not cause
1674 * premature OOM since __isolate_lru_page() returns -EBUSY
1675 * only when the page is being freed somewhere else.
1678 switch (__isolate_lru_page(page, mode)) {
1680 nr_taken += nr_pages;
1681 nr_zone_taken[page_zonenum(page)] += nr_pages;
1682 list_move(&page->lru, dst);
1686 /* else it is being freed elsewhere */
1687 list_move(&page->lru, src);
1696 * Splice any skipped pages to the start of the LRU list. Note that
1697 * this disrupts the LRU order when reclaiming for lower zones but
1698 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1699 * scanning would soon rescan the same pages to skip and put the
1700 * system at risk of premature OOM.
1702 if (!list_empty(&pages_skipped)) {
1705 list_splice(&pages_skipped, src);
1706 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1707 if (!nr_skipped[zid])
1710 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1711 skipped += nr_skipped[zid];
1714 *nr_scanned = total_scan;
1715 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1716 total_scan, skipped, nr_taken, mode, lru);
1717 update_lru_sizes(lruvec, lru, nr_zone_taken);
1722 * isolate_lru_page - tries to isolate a page from its LRU list
1723 * @page: page to isolate from its LRU list
1725 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1726 * vmstat statistic corresponding to whatever LRU list the page was on.
1728 * Returns 0 if the page was removed from an LRU list.
1729 * Returns -EBUSY if the page was not on an LRU list.
1731 * The returned page will have PageLRU() cleared. If it was found on
1732 * the active list, it will have PageActive set. If it was found on
1733 * the unevictable list, it will have the PageUnevictable bit set. That flag
1734 * may need to be cleared by the caller before letting the page go.
1736 * The vmstat statistic corresponding to the list on which the page was
1737 * found will be decremented.
1741 * (1) Must be called with an elevated refcount on the page. This is a
1742 * fundamentnal difference from isolate_lru_pages (which is called
1743 * without a stable reference).
1744 * (2) the lru_lock must not be held.
1745 * (3) interrupts must be enabled.
1747 int isolate_lru_page(struct page *page)
1751 VM_BUG_ON_PAGE(!page_count(page), page);
1752 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1754 if (PageLRU(page)) {
1755 pg_data_t *pgdat = page_pgdat(page);
1756 struct lruvec *lruvec;
1758 spin_lock_irq(&pgdat->lru_lock);
1759 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1760 if (PageLRU(page)) {
1761 int lru = page_lru(page);
1764 del_page_from_lru_list(page, lruvec, lru);
1767 spin_unlock_irq(&pgdat->lru_lock);
1773 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1774 * then get rescheduled. When there are massive number of tasks doing page
1775 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1776 * the LRU list will go small and be scanned faster than necessary, leading to
1777 * unnecessary swapping, thrashing and OOM.
1779 static int too_many_isolated(struct pglist_data *pgdat, int file,
1780 struct scan_control *sc)
1782 unsigned long inactive, isolated;
1784 if (current_is_kswapd())
1787 if (!writeback_throttling_sane(sc))
1791 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1792 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1794 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1795 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1799 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1800 * won't get blocked by normal direct-reclaimers, forming a circular
1803 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1806 return isolated > inactive;
1810 * This moves pages from @list to corresponding LRU list.
1812 * We move them the other way if the page is referenced by one or more
1813 * processes, from rmap.
1815 * If the pages are mostly unmapped, the processing is fast and it is
1816 * appropriate to hold zone_lru_lock across the whole operation. But if
1817 * the pages are mapped, the processing is slow (page_referenced()) so we
1818 * should drop zone_lru_lock around each page. It's impossible to balance
1819 * this, so instead we remove the pages from the LRU while processing them.
1820 * It is safe to rely on PG_active against the non-LRU pages in here because
1821 * nobody will play with that bit on a non-LRU page.
1823 * The downside is that we have to touch page->_refcount against each page.
1824 * But we had to alter page->flags anyway.
1826 * Returns the number of pages moved to the given lruvec.
1829 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1830 struct list_head *list)
1832 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1833 int nr_pages, nr_moved = 0;
1834 LIST_HEAD(pages_to_free);
1838 while (!list_empty(list)) {
1839 page = lru_to_page(list);
1840 VM_BUG_ON_PAGE(PageLRU(page), page);
1841 if (unlikely(!page_evictable(page))) {
1842 list_del(&page->lru);
1843 spin_unlock_irq(&pgdat->lru_lock);
1844 putback_lru_page(page);
1845 spin_lock_irq(&pgdat->lru_lock);
1848 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1851 lru = page_lru(page);
1853 nr_pages = hpage_nr_pages(page);
1854 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1855 list_move(&page->lru, &lruvec->lists[lru]);
1857 if (put_page_testzero(page)) {
1858 __ClearPageLRU(page);
1859 __ClearPageActive(page);
1860 del_page_from_lru_list(page, lruvec, lru);
1862 if (unlikely(PageCompound(page))) {
1863 spin_unlock_irq(&pgdat->lru_lock);
1864 (*get_compound_page_dtor(page))(page);
1865 spin_lock_irq(&pgdat->lru_lock);
1867 list_add(&page->lru, &pages_to_free);
1869 nr_moved += nr_pages;
1874 * To save our caller's stack, now use input list for pages to free.
1876 list_splice(&pages_to_free, list);
1882 * If a kernel thread (such as nfsd for loop-back mounts) services
1883 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1884 * In that case we should only throttle if the backing device it is
1885 * writing to is congested. In other cases it is safe to throttle.
1887 static int current_may_throttle(void)
1889 return !(current->flags & PF_LESS_THROTTLE) ||
1890 current->backing_dev_info == NULL ||
1891 bdi_write_congested(current->backing_dev_info);
1895 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1896 * of reclaimed pages
1898 static noinline_for_stack unsigned long
1899 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1900 struct scan_control *sc, enum lru_list lru)
1902 LIST_HEAD(page_list);
1903 unsigned long nr_scanned;
1904 unsigned long nr_reclaimed = 0;
1905 unsigned long nr_taken;
1906 struct reclaim_stat stat;
1907 int file = is_file_lru(lru);
1908 enum vm_event_item item;
1909 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1910 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1911 bool stalled = false;
1913 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1917 /* wait a bit for the reclaimer. */
1921 /* We are about to die and free our memory. Return now. */
1922 if (fatal_signal_pending(current))
1923 return SWAP_CLUSTER_MAX;
1928 spin_lock_irq(&pgdat->lru_lock);
1930 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1931 &nr_scanned, sc, lru);
1933 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1934 reclaim_stat->recent_scanned[file] += nr_taken;
1936 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1937 if (!cgroup_reclaim(sc))
1938 __count_vm_events(item, nr_scanned);
1939 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1940 spin_unlock_irq(&pgdat->lru_lock);
1945 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1948 spin_lock_irq(&pgdat->lru_lock);
1950 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1951 if (!cgroup_reclaim(sc))
1952 __count_vm_events(item, nr_reclaimed);
1953 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1954 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1955 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1957 move_pages_to_lru(lruvec, &page_list);
1959 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1961 spin_unlock_irq(&pgdat->lru_lock);
1963 mem_cgroup_uncharge_list(&page_list);
1964 free_unref_page_list(&page_list);
1967 * If dirty pages are scanned that are not queued for IO, it
1968 * implies that flushers are not doing their job. This can
1969 * happen when memory pressure pushes dirty pages to the end of
1970 * the LRU before the dirty limits are breached and the dirty
1971 * data has expired. It can also happen when the proportion of
1972 * dirty pages grows not through writes but through memory
1973 * pressure reclaiming all the clean cache. And in some cases,
1974 * the flushers simply cannot keep up with the allocation
1975 * rate. Nudge the flusher threads in case they are asleep.
1977 if (stat.nr_unqueued_dirty == nr_taken)
1978 wakeup_flusher_threads(WB_REASON_VMSCAN);
1980 sc->nr.dirty += stat.nr_dirty;
1981 sc->nr.congested += stat.nr_congested;
1982 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1983 sc->nr.writeback += stat.nr_writeback;
1984 sc->nr.immediate += stat.nr_immediate;
1985 sc->nr.taken += nr_taken;
1987 sc->nr.file_taken += nr_taken;
1989 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1990 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1991 return nr_reclaimed;
1994 static void shrink_active_list(unsigned long nr_to_scan,
1995 struct lruvec *lruvec,
1996 struct scan_control *sc,
1999 unsigned long nr_taken;
2000 unsigned long nr_scanned;
2001 unsigned long vm_flags;
2002 LIST_HEAD(l_hold); /* The pages which were snipped off */
2003 LIST_HEAD(l_active);
2004 LIST_HEAD(l_inactive);
2006 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2007 unsigned nr_deactivate, nr_activate;
2008 unsigned nr_rotated = 0;
2009 int file = is_file_lru(lru);
2010 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2014 spin_lock_irq(&pgdat->lru_lock);
2016 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2017 &nr_scanned, sc, lru);
2019 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2020 reclaim_stat->recent_scanned[file] += nr_taken;
2022 __count_vm_events(PGREFILL, nr_scanned);
2023 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2025 spin_unlock_irq(&pgdat->lru_lock);
2027 while (!list_empty(&l_hold)) {
2029 page = lru_to_page(&l_hold);
2030 list_del(&page->lru);
2032 if (unlikely(!page_evictable(page))) {
2033 putback_lru_page(page);
2037 if (unlikely(buffer_heads_over_limit)) {
2038 if (page_has_private(page) && trylock_page(page)) {
2039 if (page_has_private(page))
2040 try_to_release_page(page, 0);
2045 if (page_referenced(page, 0, sc->target_mem_cgroup,
2047 nr_rotated += hpage_nr_pages(page);
2049 * Identify referenced, file-backed active pages and
2050 * give them one more trip around the active list. So
2051 * that executable code get better chances to stay in
2052 * memory under moderate memory pressure. Anon pages
2053 * are not likely to be evicted by use-once streaming
2054 * IO, plus JVM can create lots of anon VM_EXEC pages,
2055 * so we ignore them here.
2057 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2058 list_add(&page->lru, &l_active);
2063 ClearPageActive(page); /* we are de-activating */
2064 SetPageWorkingset(page);
2065 list_add(&page->lru, &l_inactive);
2069 * Move pages back to the lru list.
2071 spin_lock_irq(&pgdat->lru_lock);
2073 * Count referenced pages from currently used mappings as rotated,
2074 * even though only some of them are actually re-activated. This
2075 * helps balance scan pressure between file and anonymous pages in
2078 reclaim_stat->recent_rotated[file] += nr_rotated;
2080 nr_activate = move_pages_to_lru(lruvec, &l_active);
2081 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2082 /* Keep all free pages in l_active list */
2083 list_splice(&l_inactive, &l_active);
2085 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2086 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2088 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2089 spin_unlock_irq(&pgdat->lru_lock);
2091 mem_cgroup_uncharge_list(&l_active);
2092 free_unref_page_list(&l_active);
2093 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2094 nr_deactivate, nr_rotated, sc->priority, file);
2097 unsigned long reclaim_pages(struct list_head *page_list)
2099 int nid = NUMA_NO_NODE;
2100 unsigned long nr_reclaimed = 0;
2101 LIST_HEAD(node_page_list);
2102 struct reclaim_stat dummy_stat;
2104 struct scan_control sc = {
2105 .gfp_mask = GFP_KERNEL,
2106 .priority = DEF_PRIORITY,
2112 while (!list_empty(page_list)) {
2113 page = lru_to_page(page_list);
2114 if (nid == NUMA_NO_NODE) {
2115 nid = page_to_nid(page);
2116 INIT_LIST_HEAD(&node_page_list);
2119 if (nid == page_to_nid(page)) {
2120 ClearPageActive(page);
2121 list_move(&page->lru, &node_page_list);
2125 nr_reclaimed += shrink_page_list(&node_page_list,
2128 &dummy_stat, false);
2129 while (!list_empty(&node_page_list)) {
2130 page = lru_to_page(&node_page_list);
2131 list_del(&page->lru);
2132 putback_lru_page(page);
2138 if (!list_empty(&node_page_list)) {
2139 nr_reclaimed += shrink_page_list(&node_page_list,
2142 &dummy_stat, false);
2143 while (!list_empty(&node_page_list)) {
2144 page = lru_to_page(&node_page_list);
2145 list_del(&page->lru);
2146 putback_lru_page(page);
2150 return nr_reclaimed;
2153 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2154 struct lruvec *lruvec, struct scan_control *sc)
2156 if (is_active_lru(lru)) {
2157 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2158 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2160 sc->skipped_deactivate = 1;
2164 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2168 * The inactive anon list should be small enough that the VM never has
2169 * to do too much work.
2171 * The inactive file list should be small enough to leave most memory
2172 * to the established workingset on the scan-resistant active list,
2173 * but large enough to avoid thrashing the aggregate readahead window.
2175 * Both inactive lists should also be large enough that each inactive
2176 * page has a chance to be referenced again before it is reclaimed.
2178 * If that fails and refaulting is observed, the inactive list grows.
2180 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2181 * on this LRU, maintained by the pageout code. An inactive_ratio
2182 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2185 * memory ratio inactive
2186 * -------------------------------------
2195 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2197 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2198 unsigned long inactive, active;
2199 unsigned long inactive_ratio;
2202 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2203 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2205 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2207 inactive_ratio = int_sqrt(10 * gb);
2211 return inactive * inactive_ratio < active;
2222 * Determine how aggressively the anon and file LRU lists should be
2223 * scanned. The relative value of each set of LRU lists is determined
2224 * by looking at the fraction of the pages scanned we did rotate back
2225 * onto the active list instead of evict.
2227 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2228 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2230 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2233 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2234 int swappiness = mem_cgroup_swappiness(memcg);
2235 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2237 u64 denominator = 0; /* gcc */
2238 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2239 unsigned long anon_prio, file_prio;
2240 enum scan_balance scan_balance;
2241 unsigned long anon, file;
2242 unsigned long ap, fp;
2245 /* If we have no swap space, do not bother scanning anon pages. */
2246 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2247 scan_balance = SCAN_FILE;
2252 * Global reclaim will swap to prevent OOM even with no
2253 * swappiness, but memcg users want to use this knob to
2254 * disable swapping for individual groups completely when
2255 * using the memory controller's swap limit feature would be
2258 if (cgroup_reclaim(sc) && !swappiness) {
2259 scan_balance = SCAN_FILE;
2264 * Do not apply any pressure balancing cleverness when the
2265 * system is close to OOM, scan both anon and file equally
2266 * (unless the swappiness setting disagrees with swapping).
2268 if (!sc->priority && swappiness) {
2269 scan_balance = SCAN_EQUAL;
2274 * If the system is almost out of file pages, force-scan anon.
2276 if (sc->file_is_tiny) {
2277 scan_balance = SCAN_ANON;
2282 * If there is enough inactive page cache, we do not reclaim
2283 * anything from the anonymous working right now.
2285 if (sc->cache_trim_mode) {
2286 scan_balance = SCAN_FILE;
2290 scan_balance = SCAN_FRACT;
2293 * With swappiness at 100, anonymous and file have the same priority.
2294 * This scanning priority is essentially the inverse of IO cost.
2296 anon_prio = swappiness;
2297 file_prio = 200 - anon_prio;
2300 * OK, so we have swap space and a fair amount of page cache
2301 * pages. We use the recently rotated / recently scanned
2302 * ratios to determine how valuable each cache is.
2304 * Because workloads change over time (and to avoid overflow)
2305 * we keep these statistics as a floating average, which ends
2306 * up weighing recent references more than old ones.
2308 * anon in [0], file in [1]
2311 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2312 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2313 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2314 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2316 spin_lock_irq(&pgdat->lru_lock);
2317 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2318 reclaim_stat->recent_scanned[0] /= 2;
2319 reclaim_stat->recent_rotated[0] /= 2;
2322 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2323 reclaim_stat->recent_scanned[1] /= 2;
2324 reclaim_stat->recent_rotated[1] /= 2;
2328 * The amount of pressure on anon vs file pages is inversely
2329 * proportional to the fraction of recently scanned pages on
2330 * each list that were recently referenced and in active use.
2332 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2333 ap /= reclaim_stat->recent_rotated[0] + 1;
2335 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2336 fp /= reclaim_stat->recent_rotated[1] + 1;
2337 spin_unlock_irq(&pgdat->lru_lock);
2341 denominator = ap + fp + 1;
2343 for_each_evictable_lru(lru) {
2344 int file = is_file_lru(lru);
2345 unsigned long lruvec_size;
2347 unsigned long protection;
2349 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2350 protection = mem_cgroup_protection(memcg,
2351 sc->memcg_low_reclaim);
2355 * Scale a cgroup's reclaim pressure by proportioning
2356 * its current usage to its memory.low or memory.min
2359 * This is important, as otherwise scanning aggression
2360 * becomes extremely binary -- from nothing as we
2361 * approach the memory protection threshold, to totally
2362 * nominal as we exceed it. This results in requiring
2363 * setting extremely liberal protection thresholds. It
2364 * also means we simply get no protection at all if we
2365 * set it too low, which is not ideal.
2367 * If there is any protection in place, we reduce scan
2368 * pressure by how much of the total memory used is
2369 * within protection thresholds.
2371 * There is one special case: in the first reclaim pass,
2372 * we skip over all groups that are within their low
2373 * protection. If that fails to reclaim enough pages to
2374 * satisfy the reclaim goal, we come back and override
2375 * the best-effort low protection. However, we still
2376 * ideally want to honor how well-behaved groups are in
2377 * that case instead of simply punishing them all
2378 * equally. As such, we reclaim them based on how much
2379 * memory they are using, reducing the scan pressure
2380 * again by how much of the total memory used is under
2383 unsigned long cgroup_size = mem_cgroup_size(memcg);
2385 /* Avoid TOCTOU with earlier protection check */
2386 cgroup_size = max(cgroup_size, protection);
2388 scan = lruvec_size - lruvec_size * protection /
2392 * Minimally target SWAP_CLUSTER_MAX pages to keep
2393 * reclaim moving forwards, avoiding decremeting
2394 * sc->priority further than desirable.
2396 scan = max(scan, SWAP_CLUSTER_MAX);
2401 scan >>= sc->priority;
2404 * If the cgroup's already been deleted, make sure to
2405 * scrape out the remaining cache.
2407 if (!scan && !mem_cgroup_online(memcg))
2408 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2410 switch (scan_balance) {
2412 /* Scan lists relative to size */
2416 * Scan types proportional to swappiness and
2417 * their relative recent reclaim efficiency.
2418 * Make sure we don't miss the last page on
2419 * the offlined memory cgroups because of a
2422 scan = mem_cgroup_online(memcg) ?
2423 div64_u64(scan * fraction[file], denominator) :
2424 DIV64_U64_ROUND_UP(scan * fraction[file],
2429 /* Scan one type exclusively */
2430 if ((scan_balance == SCAN_FILE) != file)
2434 /* Look ma, no brain */
2442 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2444 unsigned long nr[NR_LRU_LISTS];
2445 unsigned long targets[NR_LRU_LISTS];
2446 unsigned long nr_to_scan;
2448 unsigned long nr_reclaimed = 0;
2449 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2450 struct blk_plug plug;
2453 get_scan_count(lruvec, sc, nr);
2455 /* Record the original scan target for proportional adjustments later */
2456 memcpy(targets, nr, sizeof(nr));
2459 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2460 * event that can occur when there is little memory pressure e.g.
2461 * multiple streaming readers/writers. Hence, we do not abort scanning
2462 * when the requested number of pages are reclaimed when scanning at
2463 * DEF_PRIORITY on the assumption that the fact we are direct
2464 * reclaiming implies that kswapd is not keeping up and it is best to
2465 * do a batch of work at once. For memcg reclaim one check is made to
2466 * abort proportional reclaim if either the file or anon lru has already
2467 * dropped to zero at the first pass.
2469 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2470 sc->priority == DEF_PRIORITY);
2472 blk_start_plug(&plug);
2473 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2474 nr[LRU_INACTIVE_FILE]) {
2475 unsigned long nr_anon, nr_file, percentage;
2476 unsigned long nr_scanned;
2478 for_each_evictable_lru(lru) {
2480 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2481 nr[lru] -= nr_to_scan;
2483 nr_reclaimed += shrink_list(lru, nr_to_scan,
2490 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2494 * For kswapd and memcg, reclaim at least the number of pages
2495 * requested. Ensure that the anon and file LRUs are scanned
2496 * proportionally what was requested by get_scan_count(). We
2497 * stop reclaiming one LRU and reduce the amount scanning
2498 * proportional to the original scan target.
2500 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2501 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2504 * It's just vindictive to attack the larger once the smaller
2505 * has gone to zero. And given the way we stop scanning the
2506 * smaller below, this makes sure that we only make one nudge
2507 * towards proportionality once we've got nr_to_reclaim.
2509 if (!nr_file || !nr_anon)
2512 if (nr_file > nr_anon) {
2513 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2514 targets[LRU_ACTIVE_ANON] + 1;
2516 percentage = nr_anon * 100 / scan_target;
2518 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2519 targets[LRU_ACTIVE_FILE] + 1;
2521 percentage = nr_file * 100 / scan_target;
2524 /* Stop scanning the smaller of the LRU */
2526 nr[lru + LRU_ACTIVE] = 0;
2529 * Recalculate the other LRU scan count based on its original
2530 * scan target and the percentage scanning already complete
2532 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2533 nr_scanned = targets[lru] - nr[lru];
2534 nr[lru] = targets[lru] * (100 - percentage) / 100;
2535 nr[lru] -= min(nr[lru], nr_scanned);
2538 nr_scanned = targets[lru] - nr[lru];
2539 nr[lru] = targets[lru] * (100 - percentage) / 100;
2540 nr[lru] -= min(nr[lru], nr_scanned);
2542 scan_adjusted = true;
2544 blk_finish_plug(&plug);
2545 sc->nr_reclaimed += nr_reclaimed;
2548 * Even if we did not try to evict anon pages at all, we want to
2549 * rebalance the anon lru active/inactive ratio.
2551 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2552 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2553 sc, LRU_ACTIVE_ANON);
2556 /* Use reclaim/compaction for costly allocs or under memory pressure */
2557 static bool in_reclaim_compaction(struct scan_control *sc)
2559 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2560 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2561 sc->priority < DEF_PRIORITY - 2))
2568 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2569 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2570 * true if more pages should be reclaimed such that when the page allocator
2571 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2572 * It will give up earlier than that if there is difficulty reclaiming pages.
2574 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2575 unsigned long nr_reclaimed,
2576 struct scan_control *sc)
2578 unsigned long pages_for_compaction;
2579 unsigned long inactive_lru_pages;
2582 /* If not in reclaim/compaction mode, stop */
2583 if (!in_reclaim_compaction(sc))
2587 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2588 * number of pages that were scanned. This will return to the caller
2589 * with the risk reclaim/compaction and the resulting allocation attempt
2590 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2591 * allocations through requiring that the full LRU list has been scanned
2592 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2593 * scan, but that approximation was wrong, and there were corner cases
2594 * where always a non-zero amount of pages were scanned.
2599 /* If compaction would go ahead or the allocation would succeed, stop */
2600 for (z = 0; z <= sc->reclaim_idx; z++) {
2601 struct zone *zone = &pgdat->node_zones[z];
2602 if (!managed_zone(zone))
2605 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2606 case COMPACT_SUCCESS:
2607 case COMPACT_CONTINUE:
2610 /* check next zone */
2616 * If we have not reclaimed enough pages for compaction and the
2617 * inactive lists are large enough, continue reclaiming
2619 pages_for_compaction = compact_gap(sc->order);
2620 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2621 if (get_nr_swap_pages() > 0)
2622 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2624 return inactive_lru_pages > pages_for_compaction;
2627 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2629 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2630 struct mem_cgroup *memcg;
2632 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2634 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2635 unsigned long reclaimed;
2636 unsigned long scanned;
2638 switch (mem_cgroup_protected(target_memcg, memcg)) {
2639 case MEMCG_PROT_MIN:
2642 * If there is no reclaimable memory, OOM.
2645 case MEMCG_PROT_LOW:
2648 * Respect the protection only as long as
2649 * there is an unprotected supply
2650 * of reclaimable memory from other cgroups.
2652 if (!sc->memcg_low_reclaim) {
2653 sc->memcg_low_skipped = 1;
2656 memcg_memory_event(memcg, MEMCG_LOW);
2658 case MEMCG_PROT_NONE:
2660 * All protection thresholds breached. We may
2661 * still choose to vary the scan pressure
2662 * applied based on by how much the cgroup in
2663 * question has exceeded its protection
2664 * thresholds (see get_scan_count).
2669 reclaimed = sc->nr_reclaimed;
2670 scanned = sc->nr_scanned;
2672 shrink_lruvec(lruvec, sc);
2674 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2677 /* Record the group's reclaim efficiency */
2678 vmpressure(sc->gfp_mask, memcg, false,
2679 sc->nr_scanned - scanned,
2680 sc->nr_reclaimed - reclaimed);
2682 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2685 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2687 struct reclaim_state *reclaim_state = current->reclaim_state;
2688 unsigned long nr_reclaimed, nr_scanned;
2689 struct lruvec *target_lruvec;
2690 bool reclaimable = false;
2693 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2696 memset(&sc->nr, 0, sizeof(sc->nr));
2698 nr_reclaimed = sc->nr_reclaimed;
2699 nr_scanned = sc->nr_scanned;
2702 * Target desirable inactive:active list ratios for the anon
2703 * and file LRU lists.
2705 if (!sc->force_deactivate) {
2706 unsigned long refaults;
2708 if (inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2709 sc->may_deactivate |= DEACTIVATE_ANON;
2711 sc->may_deactivate &= ~DEACTIVATE_ANON;
2714 * When refaults are being observed, it means a new
2715 * workingset is being established. Deactivate to get
2716 * rid of any stale active pages quickly.
2718 refaults = lruvec_page_state(target_lruvec,
2719 WORKINGSET_ACTIVATE);
2720 if (refaults != target_lruvec->refaults ||
2721 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2722 sc->may_deactivate |= DEACTIVATE_FILE;
2724 sc->may_deactivate &= ~DEACTIVATE_FILE;
2726 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2729 * If we have plenty of inactive file pages that aren't
2730 * thrashing, try to reclaim those first before touching
2733 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2734 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2735 sc->cache_trim_mode = 1;
2737 sc->cache_trim_mode = 0;
2740 * Prevent the reclaimer from falling into the cache trap: as
2741 * cache pages start out inactive, every cache fault will tip
2742 * the scan balance towards the file LRU. And as the file LRU
2743 * shrinks, so does the window for rotation from references.
2744 * This means we have a runaway feedback loop where a tiny
2745 * thrashing file LRU becomes infinitely more attractive than
2746 * anon pages. Try to detect this based on file LRU size.
2748 if (!cgroup_reclaim(sc)) {
2749 unsigned long total_high_wmark = 0;
2750 unsigned long free, anon;
2753 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2754 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2755 node_page_state(pgdat, NR_INACTIVE_FILE);
2757 for (z = 0; z < MAX_NR_ZONES; z++) {
2758 struct zone *zone = &pgdat->node_zones[z];
2759 if (!managed_zone(zone))
2762 total_high_wmark += high_wmark_pages(zone);
2766 * Consider anon: if that's low too, this isn't a
2767 * runaway file reclaim problem, but rather just
2768 * extreme pressure. Reclaim as per usual then.
2770 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2773 file + free <= total_high_wmark &&
2774 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2775 anon >> sc->priority;
2778 shrink_node_memcgs(pgdat, sc);
2780 if (reclaim_state) {
2781 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2782 reclaim_state->reclaimed_slab = 0;
2785 /* Record the subtree's reclaim efficiency */
2786 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2787 sc->nr_scanned - nr_scanned,
2788 sc->nr_reclaimed - nr_reclaimed);
2790 if (sc->nr_reclaimed - nr_reclaimed)
2793 if (current_is_kswapd()) {
2795 * If reclaim is isolating dirty pages under writeback,
2796 * it implies that the long-lived page allocation rate
2797 * is exceeding the page laundering rate. Either the
2798 * global limits are not being effective at throttling
2799 * processes due to the page distribution throughout
2800 * zones or there is heavy usage of a slow backing
2801 * device. The only option is to throttle from reclaim
2802 * context which is not ideal as there is no guarantee
2803 * the dirtying process is throttled in the same way
2804 * balance_dirty_pages() manages.
2806 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2807 * count the number of pages under pages flagged for
2808 * immediate reclaim and stall if any are encountered
2809 * in the nr_immediate check below.
2811 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2812 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2814 /* Allow kswapd to start writing pages during reclaim.*/
2815 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2816 set_bit(PGDAT_DIRTY, &pgdat->flags);
2819 * If kswapd scans pages marked marked for immediate
2820 * reclaim and under writeback (nr_immediate), it
2821 * implies that pages are cycling through the LRU
2822 * faster than they are written so also forcibly stall.
2824 if (sc->nr.immediate)
2825 congestion_wait(BLK_RW_ASYNC, HZ/10);
2829 * Tag a node/memcg as congested if all the dirty pages
2830 * scanned were backed by a congested BDI and
2831 * wait_iff_congested will stall.
2833 * Legacy memcg will stall in page writeback so avoid forcibly
2834 * stalling in wait_iff_congested().
2836 if ((current_is_kswapd() ||
2837 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2838 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2839 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2842 * Stall direct reclaim for IO completions if underlying BDIs
2843 * and node is congested. Allow kswapd to continue until it
2844 * starts encountering unqueued dirty pages or cycling through
2845 * the LRU too quickly.
2847 if (!current_is_kswapd() && current_may_throttle() &&
2848 !sc->hibernation_mode &&
2849 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2850 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2852 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2857 * Kswapd gives up on balancing particular nodes after too
2858 * many failures to reclaim anything from them and goes to
2859 * sleep. On reclaim progress, reset the failure counter. A
2860 * successful direct reclaim run will revive a dormant kswapd.
2863 pgdat->kswapd_failures = 0;
2867 * Returns true if compaction should go ahead for a costly-order request, or
2868 * the allocation would already succeed without compaction. Return false if we
2869 * should reclaim first.
2871 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2873 unsigned long watermark;
2874 enum compact_result suitable;
2876 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2877 if (suitable == COMPACT_SUCCESS)
2878 /* Allocation should succeed already. Don't reclaim. */
2880 if (suitable == COMPACT_SKIPPED)
2881 /* Compaction cannot yet proceed. Do reclaim. */
2885 * Compaction is already possible, but it takes time to run and there
2886 * are potentially other callers using the pages just freed. So proceed
2887 * with reclaim to make a buffer of free pages available to give
2888 * compaction a reasonable chance of completing and allocating the page.
2889 * Note that we won't actually reclaim the whole buffer in one attempt
2890 * as the target watermark in should_continue_reclaim() is lower. But if
2891 * we are already above the high+gap watermark, don't reclaim at all.
2893 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2895 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2899 * This is the direct reclaim path, for page-allocating processes. We only
2900 * try to reclaim pages from zones which will satisfy the caller's allocation
2903 * If a zone is deemed to be full of pinned pages then just give it a light
2904 * scan then give up on it.
2906 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2910 unsigned long nr_soft_reclaimed;
2911 unsigned long nr_soft_scanned;
2913 pg_data_t *last_pgdat = NULL;
2916 * If the number of buffer_heads in the machine exceeds the maximum
2917 * allowed level, force direct reclaim to scan the highmem zone as
2918 * highmem pages could be pinning lowmem pages storing buffer_heads
2920 orig_mask = sc->gfp_mask;
2921 if (buffer_heads_over_limit) {
2922 sc->gfp_mask |= __GFP_HIGHMEM;
2923 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2926 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2927 sc->reclaim_idx, sc->nodemask) {
2929 * Take care memory controller reclaiming has small influence
2932 if (!cgroup_reclaim(sc)) {
2933 if (!cpuset_zone_allowed(zone,
2934 GFP_KERNEL | __GFP_HARDWALL))
2938 * If we already have plenty of memory free for
2939 * compaction in this zone, don't free any more.
2940 * Even though compaction is invoked for any
2941 * non-zero order, only frequent costly order
2942 * reclamation is disruptive enough to become a
2943 * noticeable problem, like transparent huge
2946 if (IS_ENABLED(CONFIG_COMPACTION) &&
2947 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2948 compaction_ready(zone, sc)) {
2949 sc->compaction_ready = true;
2954 * Shrink each node in the zonelist once. If the
2955 * zonelist is ordered by zone (not the default) then a
2956 * node may be shrunk multiple times but in that case
2957 * the user prefers lower zones being preserved.
2959 if (zone->zone_pgdat == last_pgdat)
2963 * This steals pages from memory cgroups over softlimit
2964 * and returns the number of reclaimed pages and
2965 * scanned pages. This works for global memory pressure
2966 * and balancing, not for a memcg's limit.
2968 nr_soft_scanned = 0;
2969 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2970 sc->order, sc->gfp_mask,
2972 sc->nr_reclaimed += nr_soft_reclaimed;
2973 sc->nr_scanned += nr_soft_scanned;
2974 /* need some check for avoid more shrink_zone() */
2977 /* See comment about same check for global reclaim above */
2978 if (zone->zone_pgdat == last_pgdat)
2980 last_pgdat = zone->zone_pgdat;
2981 shrink_node(zone->zone_pgdat, sc);
2985 * Restore to original mask to avoid the impact on the caller if we
2986 * promoted it to __GFP_HIGHMEM.
2988 sc->gfp_mask = orig_mask;
2991 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2993 struct lruvec *target_lruvec;
2994 unsigned long refaults;
2996 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2997 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE);
2998 target_lruvec->refaults = refaults;
3002 * This is the main entry point to direct page reclaim.
3004 * If a full scan of the inactive list fails to free enough memory then we
3005 * are "out of memory" and something needs to be killed.
3007 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3008 * high - the zone may be full of dirty or under-writeback pages, which this
3009 * caller can't do much about. We kick the writeback threads and take explicit
3010 * naps in the hope that some of these pages can be written. But if the
3011 * allocating task holds filesystem locks which prevent writeout this might not
3012 * work, and the allocation attempt will fail.
3014 * returns: 0, if no pages reclaimed
3015 * else, the number of pages reclaimed
3017 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3018 struct scan_control *sc)
3020 int initial_priority = sc->priority;
3021 pg_data_t *last_pgdat;
3025 delayacct_freepages_start();
3027 if (!cgroup_reclaim(sc))
3028 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3031 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3034 shrink_zones(zonelist, sc);
3036 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3039 if (sc->compaction_ready)
3043 * If we're getting trouble reclaiming, start doing
3044 * writepage even in laptop mode.
3046 if (sc->priority < DEF_PRIORITY - 2)
3047 sc->may_writepage = 1;
3048 } while (--sc->priority >= 0);
3051 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3053 if (zone->zone_pgdat == last_pgdat)
3055 last_pgdat = zone->zone_pgdat;
3057 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3059 if (cgroup_reclaim(sc)) {
3060 struct lruvec *lruvec;
3062 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3064 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3068 delayacct_freepages_end();
3070 if (sc->nr_reclaimed)
3071 return sc->nr_reclaimed;
3073 /* Aborted reclaim to try compaction? don't OOM, then */
3074 if (sc->compaction_ready)
3078 * We make inactive:active ratio decisions based on the node's
3079 * composition of memory, but a restrictive reclaim_idx or a
3080 * memory.low cgroup setting can exempt large amounts of
3081 * memory from reclaim. Neither of which are very common, so
3082 * instead of doing costly eligibility calculations of the
3083 * entire cgroup subtree up front, we assume the estimates are
3084 * good, and retry with forcible deactivation if that fails.
3086 if (sc->skipped_deactivate) {
3087 sc->priority = initial_priority;
3088 sc->force_deactivate = 1;
3089 sc->skipped_deactivate = 0;
3093 /* Untapped cgroup reserves? Don't OOM, retry. */
3094 if (sc->memcg_low_skipped) {
3095 sc->priority = initial_priority;
3096 sc->force_deactivate = 0;
3097 sc->skipped_deactivate = 0;
3098 sc->memcg_low_reclaim = 1;
3099 sc->memcg_low_skipped = 0;
3106 static bool allow_direct_reclaim(pg_data_t *pgdat)
3109 unsigned long pfmemalloc_reserve = 0;
3110 unsigned long free_pages = 0;
3114 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3117 for (i = 0; i <= ZONE_NORMAL; i++) {
3118 zone = &pgdat->node_zones[i];
3119 if (!managed_zone(zone))
3122 if (!zone_reclaimable_pages(zone))
3125 pfmemalloc_reserve += min_wmark_pages(zone);
3126 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3129 /* If there are no reserves (unexpected config) then do not throttle */
3130 if (!pfmemalloc_reserve)
3133 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3135 /* kswapd must be awake if processes are being throttled */
3136 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3137 if (READ_ONCE(pgdat->kswapd_classzone_idx) > ZONE_NORMAL)
3138 WRITE_ONCE(pgdat->kswapd_classzone_idx, ZONE_NORMAL);
3140 wake_up_interruptible(&pgdat->kswapd_wait);
3147 * Throttle direct reclaimers if backing storage is backed by the network
3148 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3149 * depleted. kswapd will continue to make progress and wake the processes
3150 * when the low watermark is reached.
3152 * Returns true if a fatal signal was delivered during throttling. If this
3153 * happens, the page allocator should not consider triggering the OOM killer.
3155 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3156 nodemask_t *nodemask)
3160 pg_data_t *pgdat = NULL;
3163 * Kernel threads should not be throttled as they may be indirectly
3164 * responsible for cleaning pages necessary for reclaim to make forward
3165 * progress. kjournald for example may enter direct reclaim while
3166 * committing a transaction where throttling it could forcing other
3167 * processes to block on log_wait_commit().
3169 if (current->flags & PF_KTHREAD)
3173 * If a fatal signal is pending, this process should not throttle.
3174 * It should return quickly so it can exit and free its memory
3176 if (fatal_signal_pending(current))
3180 * Check if the pfmemalloc reserves are ok by finding the first node
3181 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3182 * GFP_KERNEL will be required for allocating network buffers when
3183 * swapping over the network so ZONE_HIGHMEM is unusable.
3185 * Throttling is based on the first usable node and throttled processes
3186 * wait on a queue until kswapd makes progress and wakes them. There
3187 * is an affinity then between processes waking up and where reclaim
3188 * progress has been made assuming the process wakes on the same node.
3189 * More importantly, processes running on remote nodes will not compete
3190 * for remote pfmemalloc reserves and processes on different nodes
3191 * should make reasonable progress.
3193 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3194 gfp_zone(gfp_mask), nodemask) {
3195 if (zone_idx(zone) > ZONE_NORMAL)
3198 /* Throttle based on the first usable node */
3199 pgdat = zone->zone_pgdat;
3200 if (allow_direct_reclaim(pgdat))
3205 /* If no zone was usable by the allocation flags then do not throttle */
3209 /* Account for the throttling */
3210 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3213 * If the caller cannot enter the filesystem, it's possible that it
3214 * is due to the caller holding an FS lock or performing a journal
3215 * transaction in the case of a filesystem like ext[3|4]. In this case,
3216 * it is not safe to block on pfmemalloc_wait as kswapd could be
3217 * blocked waiting on the same lock. Instead, throttle for up to a
3218 * second before continuing.
3220 if (!(gfp_mask & __GFP_FS)) {
3221 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3222 allow_direct_reclaim(pgdat), HZ);
3227 /* Throttle until kswapd wakes the process */
3228 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3229 allow_direct_reclaim(pgdat));
3232 if (fatal_signal_pending(current))
3239 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3240 gfp_t gfp_mask, nodemask_t *nodemask)
3242 unsigned long nr_reclaimed;
3243 struct scan_control sc = {
3244 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3245 .gfp_mask = current_gfp_context(gfp_mask),
3246 .reclaim_idx = gfp_zone(gfp_mask),
3248 .nodemask = nodemask,
3249 .priority = DEF_PRIORITY,
3250 .may_writepage = !laptop_mode,
3256 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3257 * Confirm they are large enough for max values.
3259 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3260 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3261 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3264 * Do not enter reclaim if fatal signal was delivered while throttled.
3265 * 1 is returned so that the page allocator does not OOM kill at this
3268 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3271 set_task_reclaim_state(current, &sc.reclaim_state);
3272 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3274 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3276 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3277 set_task_reclaim_state(current, NULL);
3279 return nr_reclaimed;
3284 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3285 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3286 gfp_t gfp_mask, bool noswap,
3288 unsigned long *nr_scanned)
3290 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3291 struct scan_control sc = {
3292 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3293 .target_mem_cgroup = memcg,
3294 .may_writepage = !laptop_mode,
3296 .reclaim_idx = MAX_NR_ZONES - 1,
3297 .may_swap = !noswap,
3300 WARN_ON_ONCE(!current->reclaim_state);
3302 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3303 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3305 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3309 * NOTE: Although we can get the priority field, using it
3310 * here is not a good idea, since it limits the pages we can scan.
3311 * if we don't reclaim here, the shrink_node from balance_pgdat
3312 * will pick up pages from other mem cgroup's as well. We hack
3313 * the priority and make it zero.
3315 shrink_lruvec(lruvec, &sc);
3317 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3319 *nr_scanned = sc.nr_scanned;
3321 return sc.nr_reclaimed;
3324 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3325 unsigned long nr_pages,
3329 unsigned long nr_reclaimed;
3330 unsigned long pflags;
3331 unsigned int noreclaim_flag;
3332 struct scan_control sc = {
3333 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3334 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3335 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3336 .reclaim_idx = MAX_NR_ZONES - 1,
3337 .target_mem_cgroup = memcg,
3338 .priority = DEF_PRIORITY,
3339 .may_writepage = !laptop_mode,
3341 .may_swap = may_swap,
3344 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3345 * equal pressure on all the nodes. This is based on the assumption that
3346 * the reclaim does not bail out early.
3348 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3350 set_task_reclaim_state(current, &sc.reclaim_state);
3352 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3354 psi_memstall_enter(&pflags);
3355 noreclaim_flag = memalloc_noreclaim_save();
3357 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3359 memalloc_noreclaim_restore(noreclaim_flag);
3360 psi_memstall_leave(&pflags);
3362 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3363 set_task_reclaim_state(current, NULL);
3365 return nr_reclaimed;
3369 static void age_active_anon(struct pglist_data *pgdat,
3370 struct scan_control *sc)
3372 struct mem_cgroup *memcg;
3373 struct lruvec *lruvec;
3375 if (!total_swap_pages)
3378 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3379 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3382 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3384 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3385 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3386 sc, LRU_ACTIVE_ANON);
3387 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3391 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3397 * Check for watermark boosts top-down as the higher zones
3398 * are more likely to be boosted. Both watermarks and boosts
3399 * should not be checked at the time time as reclaim would
3400 * start prematurely when there is no boosting and a lower
3403 for (i = classzone_idx; i >= 0; i--) {
3404 zone = pgdat->node_zones + i;
3405 if (!managed_zone(zone))
3408 if (zone->watermark_boost)
3416 * Returns true if there is an eligible zone balanced for the request order
3419 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3422 unsigned long mark = -1;
3426 * Check watermarks bottom-up as lower zones are more likely to
3429 for (i = 0; i <= classzone_idx; i++) {
3430 zone = pgdat->node_zones + i;
3432 if (!managed_zone(zone))
3435 mark = high_wmark_pages(zone);
3436 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3441 * If a node has no populated zone within classzone_idx, it does not
3442 * need balancing by definition. This can happen if a zone-restricted
3443 * allocation tries to wake a remote kswapd.
3451 /* Clear pgdat state for congested, dirty or under writeback. */
3452 static void clear_pgdat_congested(pg_data_t *pgdat)
3454 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3456 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3457 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3458 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3462 * Prepare kswapd for sleeping. This verifies that there are no processes
3463 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3465 * Returns true if kswapd is ready to sleep
3467 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3470 * The throttled processes are normally woken up in balance_pgdat() as
3471 * soon as allow_direct_reclaim() is true. But there is a potential
3472 * race between when kswapd checks the watermarks and a process gets
3473 * throttled. There is also a potential race if processes get
3474 * throttled, kswapd wakes, a large process exits thereby balancing the
3475 * zones, which causes kswapd to exit balance_pgdat() before reaching
3476 * the wake up checks. If kswapd is going to sleep, no process should
3477 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3478 * the wake up is premature, processes will wake kswapd and get
3479 * throttled again. The difference from wake ups in balance_pgdat() is
3480 * that here we are under prepare_to_wait().
3482 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3483 wake_up_all(&pgdat->pfmemalloc_wait);
3485 /* Hopeless node, leave it to direct reclaim */
3486 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3489 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3490 clear_pgdat_congested(pgdat);
3498 * kswapd shrinks a node of pages that are at or below the highest usable
3499 * zone that is currently unbalanced.
3501 * Returns true if kswapd scanned at least the requested number of pages to
3502 * reclaim or if the lack of progress was due to pages under writeback.
3503 * This is used to determine if the scanning priority needs to be raised.
3505 static bool kswapd_shrink_node(pg_data_t *pgdat,
3506 struct scan_control *sc)
3511 /* Reclaim a number of pages proportional to the number of zones */
3512 sc->nr_to_reclaim = 0;
3513 for (z = 0; z <= sc->reclaim_idx; z++) {
3514 zone = pgdat->node_zones + z;
3515 if (!managed_zone(zone))
3518 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3522 * Historically care was taken to put equal pressure on all zones but
3523 * now pressure is applied based on node LRU order.
3525 shrink_node(pgdat, sc);
3528 * Fragmentation may mean that the system cannot be rebalanced for
3529 * high-order allocations. If twice the allocation size has been
3530 * reclaimed then recheck watermarks only at order-0 to prevent
3531 * excessive reclaim. Assume that a process requested a high-order
3532 * can direct reclaim/compact.
3534 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3537 return sc->nr_scanned >= sc->nr_to_reclaim;
3541 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3542 * that are eligible for use by the caller until at least one zone is
3545 * Returns the order kswapd finished reclaiming at.
3547 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3548 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3549 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3550 * or lower is eligible for reclaim until at least one usable zone is
3553 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3556 unsigned long nr_soft_reclaimed;
3557 unsigned long nr_soft_scanned;
3558 unsigned long pflags;
3559 unsigned long nr_boost_reclaim;
3560 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3563 struct scan_control sc = {
3564 .gfp_mask = GFP_KERNEL,
3569 set_task_reclaim_state(current, &sc.reclaim_state);
3570 psi_memstall_enter(&pflags);
3571 __fs_reclaim_acquire();
3573 count_vm_event(PAGEOUTRUN);
3576 * Account for the reclaim boost. Note that the zone boost is left in
3577 * place so that parallel allocations that are near the watermark will
3578 * stall or direct reclaim until kswapd is finished.
3580 nr_boost_reclaim = 0;
3581 for (i = 0; i <= classzone_idx; i++) {
3582 zone = pgdat->node_zones + i;
3583 if (!managed_zone(zone))
3586 nr_boost_reclaim += zone->watermark_boost;
3587 zone_boosts[i] = zone->watermark_boost;
3589 boosted = nr_boost_reclaim;
3592 sc.priority = DEF_PRIORITY;
3594 unsigned long nr_reclaimed = sc.nr_reclaimed;
3595 bool raise_priority = true;
3599 sc.reclaim_idx = classzone_idx;
3602 * If the number of buffer_heads exceeds the maximum allowed
3603 * then consider reclaiming from all zones. This has a dual
3604 * purpose -- on 64-bit systems it is expected that
3605 * buffer_heads are stripped during active rotation. On 32-bit
3606 * systems, highmem pages can pin lowmem memory and shrinking
3607 * buffers can relieve lowmem pressure. Reclaim may still not
3608 * go ahead if all eligible zones for the original allocation
3609 * request are balanced to avoid excessive reclaim from kswapd.
3611 if (buffer_heads_over_limit) {
3612 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3613 zone = pgdat->node_zones + i;
3614 if (!managed_zone(zone))
3623 * If the pgdat is imbalanced then ignore boosting and preserve
3624 * the watermarks for a later time and restart. Note that the
3625 * zone watermarks will be still reset at the end of balancing
3626 * on the grounds that the normal reclaim should be enough to
3627 * re-evaluate if boosting is required when kswapd next wakes.
3629 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3630 if (!balanced && nr_boost_reclaim) {
3631 nr_boost_reclaim = 0;
3636 * If boosting is not active then only reclaim if there are no
3637 * eligible zones. Note that sc.reclaim_idx is not used as
3638 * buffer_heads_over_limit may have adjusted it.
3640 if (!nr_boost_reclaim && balanced)
3643 /* Limit the priority of boosting to avoid reclaim writeback */
3644 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3645 raise_priority = false;
3648 * Do not writeback or swap pages for boosted reclaim. The
3649 * intent is to relieve pressure not issue sub-optimal IO
3650 * from reclaim context. If no pages are reclaimed, the
3651 * reclaim will be aborted.
3653 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3654 sc.may_swap = !nr_boost_reclaim;
3657 * Do some background aging of the anon list, to give
3658 * pages a chance to be referenced before reclaiming. All
3659 * pages are rotated regardless of classzone as this is
3660 * about consistent aging.
3662 age_active_anon(pgdat, &sc);
3665 * If we're getting trouble reclaiming, start doing writepage
3666 * even in laptop mode.
3668 if (sc.priority < DEF_PRIORITY - 2)
3669 sc.may_writepage = 1;
3671 /* Call soft limit reclaim before calling shrink_node. */
3673 nr_soft_scanned = 0;
3674 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3675 sc.gfp_mask, &nr_soft_scanned);
3676 sc.nr_reclaimed += nr_soft_reclaimed;
3679 * There should be no need to raise the scanning priority if
3680 * enough pages are already being scanned that that high
3681 * watermark would be met at 100% efficiency.
3683 if (kswapd_shrink_node(pgdat, &sc))
3684 raise_priority = false;
3687 * If the low watermark is met there is no need for processes
3688 * to be throttled on pfmemalloc_wait as they should not be
3689 * able to safely make forward progress. Wake them
3691 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3692 allow_direct_reclaim(pgdat))
3693 wake_up_all(&pgdat->pfmemalloc_wait);
3695 /* Check if kswapd should be suspending */
3696 __fs_reclaim_release();
3697 ret = try_to_freeze();
3698 __fs_reclaim_acquire();
3699 if (ret || kthread_should_stop())
3703 * Raise priority if scanning rate is too low or there was no
3704 * progress in reclaiming pages
3706 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3707 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3710 * If reclaim made no progress for a boost, stop reclaim as
3711 * IO cannot be queued and it could be an infinite loop in
3712 * extreme circumstances.
3714 if (nr_boost_reclaim && !nr_reclaimed)
3717 if (raise_priority || !nr_reclaimed)
3719 } while (sc.priority >= 1);
3721 if (!sc.nr_reclaimed)
3722 pgdat->kswapd_failures++;
3725 /* If reclaim was boosted, account for the reclaim done in this pass */
3727 unsigned long flags;
3729 for (i = 0; i <= classzone_idx; i++) {
3730 if (!zone_boosts[i])
3733 /* Increments are under the zone lock */
3734 zone = pgdat->node_zones + i;
3735 spin_lock_irqsave(&zone->lock, flags);
3736 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3737 spin_unlock_irqrestore(&zone->lock, flags);
3741 * As there is now likely space, wakeup kcompact to defragment
3744 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3747 snapshot_refaults(NULL, pgdat);
3748 __fs_reclaim_release();
3749 psi_memstall_leave(&pflags);
3750 set_task_reclaim_state(current, NULL);
3753 * Return the order kswapd stopped reclaiming at as
3754 * prepare_kswapd_sleep() takes it into account. If another caller
3755 * entered the allocator slow path while kswapd was awake, order will
3756 * remain at the higher level.
3762 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3763 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3764 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3765 * after previous reclaim attempt (node is still unbalanced). In that case
3766 * return the zone index of the previous kswapd reclaim cycle.
3768 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3769 enum zone_type prev_classzone_idx)
3771 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_classzone_idx);
3773 return curr_idx == MAX_NR_ZONES ? prev_classzone_idx : curr_idx;
3776 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3777 unsigned int classzone_idx)
3782 if (freezing(current) || kthread_should_stop())
3785 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3788 * Try to sleep for a short interval. Note that kcompactd will only be
3789 * woken if it is possible to sleep for a short interval. This is
3790 * deliberate on the assumption that if reclaim cannot keep an
3791 * eligible zone balanced that it's also unlikely that compaction will
3794 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3796 * Compaction records what page blocks it recently failed to
3797 * isolate pages from and skips them in the future scanning.
3798 * When kswapd is going to sleep, it is reasonable to assume
3799 * that pages and compaction may succeed so reset the cache.
3801 reset_isolation_suitable(pgdat);
3804 * We have freed the memory, now we should compact it to make
3805 * allocation of the requested order possible.
3807 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3809 remaining = schedule_timeout(HZ/10);
3812 * If woken prematurely then reset kswapd_classzone_idx and
3813 * order. The values will either be from a wakeup request or
3814 * the previous request that slept prematurely.
3817 WRITE_ONCE(pgdat->kswapd_classzone_idx,
3818 kswapd_classzone_idx(pgdat, classzone_idx));
3820 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3821 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3824 finish_wait(&pgdat->kswapd_wait, &wait);
3825 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3829 * After a short sleep, check if it was a premature sleep. If not, then
3830 * go fully to sleep until explicitly woken up.
3833 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3834 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3837 * vmstat counters are not perfectly accurate and the estimated
3838 * value for counters such as NR_FREE_PAGES can deviate from the
3839 * true value by nr_online_cpus * threshold. To avoid the zone
3840 * watermarks being breached while under pressure, we reduce the
3841 * per-cpu vmstat threshold while kswapd is awake and restore
3842 * them before going back to sleep.
3844 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3846 if (!kthread_should_stop())
3849 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3852 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3854 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3856 finish_wait(&pgdat->kswapd_wait, &wait);
3860 * The background pageout daemon, started as a kernel thread
3861 * from the init process.
3863 * This basically trickles out pages so that we have _some_
3864 * free memory available even if there is no other activity
3865 * that frees anything up. This is needed for things like routing
3866 * etc, where we otherwise might have all activity going on in
3867 * asynchronous contexts that cannot page things out.
3869 * If there are applications that are active memory-allocators
3870 * (most normal use), this basically shouldn't matter.
3872 static int kswapd(void *p)
3874 unsigned int alloc_order, reclaim_order;
3875 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3876 pg_data_t *pgdat = (pg_data_t*)p;
3877 struct task_struct *tsk = current;
3878 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3880 if (!cpumask_empty(cpumask))
3881 set_cpus_allowed_ptr(tsk, cpumask);
3884 * Tell the memory management that we're a "memory allocator",
3885 * and that if we need more memory we should get access to it
3886 * regardless (see "__alloc_pages()"). "kswapd" should
3887 * never get caught in the normal page freeing logic.
3889 * (Kswapd normally doesn't need memory anyway, but sometimes
3890 * you need a small amount of memory in order to be able to
3891 * page out something else, and this flag essentially protects
3892 * us from recursively trying to free more memory as we're
3893 * trying to free the first piece of memory in the first place).
3895 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3898 WRITE_ONCE(pgdat->kswapd_order, 0);
3899 WRITE_ONCE(pgdat->kswapd_classzone_idx, MAX_NR_ZONES);
3903 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3904 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3907 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3910 /* Read the new order and classzone_idx */
3911 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3912 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3913 WRITE_ONCE(pgdat->kswapd_order, 0);
3914 WRITE_ONCE(pgdat->kswapd_classzone_idx, MAX_NR_ZONES);
3916 ret = try_to_freeze();
3917 if (kthread_should_stop())
3921 * We can speed up thawing tasks if we don't call balance_pgdat
3922 * after returning from the refrigerator
3928 * Reclaim begins at the requested order but if a high-order
3929 * reclaim fails then kswapd falls back to reclaiming for
3930 * order-0. If that happens, kswapd will consider sleeping
3931 * for the order it finished reclaiming at (reclaim_order)
3932 * but kcompactd is woken to compact for the original
3933 * request (alloc_order).
3935 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3937 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3938 if (reclaim_order < alloc_order)
3939 goto kswapd_try_sleep;
3942 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3948 * A zone is low on free memory or too fragmented for high-order memory. If
3949 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3950 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3951 * has failed or is not needed, still wake up kcompactd if only compaction is
3954 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3955 enum zone_type classzone_idx)
3958 enum zone_type curr_idx;
3960 if (!managed_zone(zone))
3963 if (!cpuset_zone_allowed(zone, gfp_flags))
3966 pgdat = zone->zone_pgdat;
3967 curr_idx = READ_ONCE(pgdat->kswapd_classzone_idx);
3969 if (curr_idx == MAX_NR_ZONES || curr_idx < classzone_idx)
3970 WRITE_ONCE(pgdat->kswapd_classzone_idx, classzone_idx);
3972 if (READ_ONCE(pgdat->kswapd_order) < order)
3973 WRITE_ONCE(pgdat->kswapd_order, order);
3975 if (!waitqueue_active(&pgdat->kswapd_wait))
3978 /* Hopeless node, leave it to direct reclaim if possible */
3979 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3980 (pgdat_balanced(pgdat, order, classzone_idx) &&
3981 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3983 * There may be plenty of free memory available, but it's too
3984 * fragmented for high-order allocations. Wake up kcompactd
3985 * and rely on compaction_suitable() to determine if it's
3986 * needed. If it fails, it will defer subsequent attempts to
3987 * ratelimit its work.
3989 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3990 wakeup_kcompactd(pgdat, order, classzone_idx);
3994 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3996 wake_up_interruptible(&pgdat->kswapd_wait);
3999 #ifdef CONFIG_HIBERNATION
4001 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4004 * Rather than trying to age LRUs the aim is to preserve the overall
4005 * LRU order by reclaiming preferentially
4006 * inactive > active > active referenced > active mapped
4008 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4010 struct scan_control sc = {
4011 .nr_to_reclaim = nr_to_reclaim,
4012 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4013 .reclaim_idx = MAX_NR_ZONES - 1,
4014 .priority = DEF_PRIORITY,
4018 .hibernation_mode = 1,
4020 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4021 unsigned long nr_reclaimed;
4022 unsigned int noreclaim_flag;
4024 fs_reclaim_acquire(sc.gfp_mask);
4025 noreclaim_flag = memalloc_noreclaim_save();
4026 set_task_reclaim_state(current, &sc.reclaim_state);
4028 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4030 set_task_reclaim_state(current, NULL);
4031 memalloc_noreclaim_restore(noreclaim_flag);
4032 fs_reclaim_release(sc.gfp_mask);
4034 return nr_reclaimed;
4036 #endif /* CONFIG_HIBERNATION */
4039 * This kswapd start function will be called by init and node-hot-add.
4040 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4042 int kswapd_run(int nid)
4044 pg_data_t *pgdat = NODE_DATA(nid);
4050 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4051 if (IS_ERR(pgdat->kswapd)) {
4052 /* failure at boot is fatal */
4053 BUG_ON(system_state < SYSTEM_RUNNING);
4054 pr_err("Failed to start kswapd on node %d\n", nid);
4055 ret = PTR_ERR(pgdat->kswapd);
4056 pgdat->kswapd = NULL;
4062 * Called by memory hotplug when all memory in a node is offlined. Caller must
4063 * hold mem_hotplug_begin/end().
4065 void kswapd_stop(int nid)
4067 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4070 kthread_stop(kswapd);
4071 NODE_DATA(nid)->kswapd = NULL;
4075 static int __init kswapd_init(void)
4080 for_each_node_state(nid, N_MEMORY)
4085 module_init(kswapd_init)
4091 * If non-zero call node_reclaim when the number of free pages falls below
4094 int node_reclaim_mode __read_mostly;
4096 #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4097 #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4100 * Priority for NODE_RECLAIM. This determines the fraction of pages
4101 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4104 #define NODE_RECLAIM_PRIORITY 4
4107 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4110 int sysctl_min_unmapped_ratio = 1;
4113 * If the number of slab pages in a zone grows beyond this percentage then
4114 * slab reclaim needs to occur.
4116 int sysctl_min_slab_ratio = 5;
4118 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4120 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4121 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4122 node_page_state(pgdat, NR_ACTIVE_FILE);
4125 * It's possible for there to be more file mapped pages than
4126 * accounted for by the pages on the file LRU lists because
4127 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4129 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4132 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4133 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4135 unsigned long nr_pagecache_reclaimable;
4136 unsigned long delta = 0;
4139 * If RECLAIM_UNMAP is set, then all file pages are considered
4140 * potentially reclaimable. Otherwise, we have to worry about
4141 * pages like swapcache and node_unmapped_file_pages() provides
4144 if (node_reclaim_mode & RECLAIM_UNMAP)
4145 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4147 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4149 /* If we can't clean pages, remove dirty pages from consideration */
4150 if (!(node_reclaim_mode & RECLAIM_WRITE))
4151 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4153 /* Watch for any possible underflows due to delta */
4154 if (unlikely(delta > nr_pagecache_reclaimable))
4155 delta = nr_pagecache_reclaimable;
4157 return nr_pagecache_reclaimable - delta;
4161 * Try to free up some pages from this node through reclaim.
4163 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4165 /* Minimum pages needed in order to stay on node */
4166 const unsigned long nr_pages = 1 << order;
4167 struct task_struct *p = current;
4168 unsigned int noreclaim_flag;
4169 struct scan_control sc = {
4170 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4171 .gfp_mask = current_gfp_context(gfp_mask),
4173 .priority = NODE_RECLAIM_PRIORITY,
4174 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4175 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4177 .reclaim_idx = gfp_zone(gfp_mask),
4180 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4184 fs_reclaim_acquire(sc.gfp_mask);
4186 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4187 * and we also need to be able to write out pages for RECLAIM_WRITE
4188 * and RECLAIM_UNMAP.
4190 noreclaim_flag = memalloc_noreclaim_save();
4191 p->flags |= PF_SWAPWRITE;
4192 set_task_reclaim_state(p, &sc.reclaim_state);
4194 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4196 * Free memory by calling shrink node with increasing
4197 * priorities until we have enough memory freed.
4200 shrink_node(pgdat, &sc);
4201 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4204 set_task_reclaim_state(p, NULL);
4205 current->flags &= ~PF_SWAPWRITE;
4206 memalloc_noreclaim_restore(noreclaim_flag);
4207 fs_reclaim_release(sc.gfp_mask);
4209 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4211 return sc.nr_reclaimed >= nr_pages;
4214 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4219 * Node reclaim reclaims unmapped file backed pages and
4220 * slab pages if we are over the defined limits.
4222 * A small portion of unmapped file backed pages is needed for
4223 * file I/O otherwise pages read by file I/O will be immediately
4224 * thrown out if the node is overallocated. So we do not reclaim
4225 * if less than a specified percentage of the node is used by
4226 * unmapped file backed pages.
4228 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4229 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4230 return NODE_RECLAIM_FULL;
4233 * Do not scan if the allocation should not be delayed.
4235 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4236 return NODE_RECLAIM_NOSCAN;
4239 * Only run node reclaim on the local node or on nodes that do not
4240 * have associated processors. This will favor the local processor
4241 * over remote processors and spread off node memory allocations
4242 * as wide as possible.
4244 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4245 return NODE_RECLAIM_NOSCAN;
4247 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4248 return NODE_RECLAIM_NOSCAN;
4250 ret = __node_reclaim(pgdat, gfp_mask, order);
4251 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4254 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4261 * check_move_unevictable_pages - check pages for evictability and move to
4262 * appropriate zone lru list
4263 * @pvec: pagevec with lru pages to check
4265 * Checks pages for evictability, if an evictable page is in the unevictable
4266 * lru list, moves it to the appropriate evictable lru list. This function
4267 * should be only used for lru pages.
4269 void check_move_unevictable_pages(struct pagevec *pvec)
4271 struct lruvec *lruvec;
4272 struct pglist_data *pgdat = NULL;
4277 for (i = 0; i < pvec->nr; i++) {
4278 struct page *page = pvec->pages[i];
4279 struct pglist_data *pagepgdat = page_pgdat(page);
4282 if (pagepgdat != pgdat) {
4284 spin_unlock_irq(&pgdat->lru_lock);
4286 spin_lock_irq(&pgdat->lru_lock);
4288 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4290 if (!PageLRU(page) || !PageUnevictable(page))
4293 if (page_evictable(page)) {
4294 enum lru_list lru = page_lru_base_type(page);
4296 VM_BUG_ON_PAGE(PageActive(page), page);
4297 ClearPageUnevictable(page);
4298 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4299 add_page_to_lru_list(page, lruvec, lru);
4305 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4306 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4307 spin_unlock_irq(&pgdat->lru_lock);
4310 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);