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;
83 * Scan pressure balancing between anon and file LRUs
85 unsigned long anon_cost;
86 unsigned long file_cost;
88 /* Can active pages be deactivated as part of reclaim? */
89 #define DEACTIVATE_ANON 1
90 #define DEACTIVATE_FILE 2
91 unsigned int may_deactivate:2;
92 unsigned int force_deactivate:1;
93 unsigned int skipped_deactivate:1;
95 /* Writepage batching in laptop mode; RECLAIM_WRITE */
96 unsigned int may_writepage:1;
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap:1;
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap:1;
105 * Cgroups are not reclaimed below their configured memory.low,
106 * unless we threaten to OOM. If any cgroups are skipped due to
107 * memory.low and nothing was reclaimed, go back for memory.low.
109 unsigned int memcg_low_reclaim:1;
110 unsigned int memcg_low_skipped:1;
112 unsigned int hibernation_mode:1;
114 /* One of the zones is ready for compaction */
115 unsigned int compaction_ready:1;
117 /* There is easily reclaimable cold cache in the current node */
118 unsigned int cache_trim_mode:1;
120 /* The file pages on the current node are dangerously low */
121 unsigned int file_is_tiny:1;
123 /* Allocation order */
126 /* Scan (total_size >> priority) pages at once */
129 /* The highest zone to isolate pages for reclaim from */
132 /* This context's GFP mask */
135 /* Incremented by the number of inactive pages that were scanned */
136 unsigned long nr_scanned;
138 /* Number of pages freed so far during a call to shrink_zones() */
139 unsigned long nr_reclaimed;
143 unsigned int unqueued_dirty;
144 unsigned int congested;
145 unsigned int writeback;
146 unsigned int immediate;
147 unsigned int file_taken;
151 /* for recording the reclaimed slab by now */
152 struct reclaim_state reclaim_state;
155 #ifdef ARCH_HAS_PREFETCHW
156 #define prefetchw_prev_lru_page(_page, _base, _field) \
158 if ((_page)->lru.prev != _base) { \
161 prev = lru_to_page(&(_page->lru)); \
162 prefetchw(&prev->_field); \
166 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
170 * From 0 .. 200. Higher means more swappy.
172 int vm_swappiness = 60;
174 * The total number of pages which are beyond the high watermark within all
177 unsigned long vm_total_pages;
179 static void set_task_reclaim_state(struct task_struct *task,
180 struct reclaim_state *rs)
182 /* Check for an overwrite */
183 WARN_ON_ONCE(rs && task->reclaim_state);
185 /* Check for the nulling of an already-nulled member */
186 WARN_ON_ONCE(!rs && !task->reclaim_state);
188 task->reclaim_state = rs;
191 static LIST_HEAD(shrinker_list);
192 static DECLARE_RWSEM(shrinker_rwsem);
196 * We allow subsystems to populate their shrinker-related
197 * LRU lists before register_shrinker_prepared() is called
198 * for the shrinker, since we don't want to impose
199 * restrictions on their internal registration order.
200 * In this case shrink_slab_memcg() may find corresponding
201 * bit is set in the shrinkers map.
203 * This value is used by the function to detect registering
204 * shrinkers and to skip do_shrink_slab() calls for them.
206 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
208 static DEFINE_IDR(shrinker_idr);
209 static int shrinker_nr_max;
211 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
213 int id, ret = -ENOMEM;
215 down_write(&shrinker_rwsem);
216 /* This may call shrinker, so it must use down_read_trylock() */
217 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
221 if (id >= shrinker_nr_max) {
222 if (memcg_expand_shrinker_maps(id)) {
223 idr_remove(&shrinker_idr, id);
227 shrinker_nr_max = id + 1;
232 up_write(&shrinker_rwsem);
236 static void unregister_memcg_shrinker(struct shrinker *shrinker)
238 int id = shrinker->id;
242 down_write(&shrinker_rwsem);
243 idr_remove(&shrinker_idr, id);
244 up_write(&shrinker_rwsem);
247 static bool cgroup_reclaim(struct scan_control *sc)
249 return sc->target_mem_cgroup;
253 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
254 * @sc: scan_control in question
256 * The normal page dirty throttling mechanism in balance_dirty_pages() is
257 * completely broken with the legacy memcg and direct stalling in
258 * shrink_page_list() is used for throttling instead, which lacks all the
259 * niceties such as fairness, adaptive pausing, bandwidth proportional
260 * allocation and configurability.
262 * This function tests whether the vmscan currently in progress can assume
263 * that the normal dirty throttling mechanism is operational.
265 static bool writeback_throttling_sane(struct scan_control *sc)
267 if (!cgroup_reclaim(sc))
269 #ifdef CONFIG_CGROUP_WRITEBACK
270 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
276 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
281 static void unregister_memcg_shrinker(struct shrinker *shrinker)
285 static bool cgroup_reclaim(struct scan_control *sc)
290 static bool writeback_throttling_sane(struct scan_control *sc)
297 * This misses isolated pages which are not accounted for to save counters.
298 * As the data only determines if reclaim or compaction continues, it is
299 * not expected that isolated pages will be a dominating factor.
301 unsigned long zone_reclaimable_pages(struct zone *zone)
305 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
306 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
307 if (get_nr_swap_pages() > 0)
308 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
309 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
315 * lruvec_lru_size - Returns the number of pages on the given LRU list.
316 * @lruvec: lru vector
318 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
320 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
322 unsigned long size = 0;
325 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
326 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
328 if (!managed_zone(zone))
331 if (!mem_cgroup_disabled())
332 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
334 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
340 * Add a shrinker callback to be called from the vm.
342 int prealloc_shrinker(struct shrinker *shrinker)
344 unsigned int size = sizeof(*shrinker->nr_deferred);
346 if (shrinker->flags & SHRINKER_NUMA_AWARE)
349 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
350 if (!shrinker->nr_deferred)
353 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
354 if (prealloc_memcg_shrinker(shrinker))
361 kfree(shrinker->nr_deferred);
362 shrinker->nr_deferred = NULL;
366 void free_prealloced_shrinker(struct shrinker *shrinker)
368 if (!shrinker->nr_deferred)
371 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
372 unregister_memcg_shrinker(shrinker);
374 kfree(shrinker->nr_deferred);
375 shrinker->nr_deferred = NULL;
378 void register_shrinker_prepared(struct shrinker *shrinker)
380 down_write(&shrinker_rwsem);
381 list_add_tail(&shrinker->list, &shrinker_list);
383 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
384 idr_replace(&shrinker_idr, shrinker, shrinker->id);
386 up_write(&shrinker_rwsem);
389 int register_shrinker(struct shrinker *shrinker)
391 int err = prealloc_shrinker(shrinker);
395 register_shrinker_prepared(shrinker);
398 EXPORT_SYMBOL(register_shrinker);
403 void unregister_shrinker(struct shrinker *shrinker)
405 if (!shrinker->nr_deferred)
407 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
408 unregister_memcg_shrinker(shrinker);
409 down_write(&shrinker_rwsem);
410 list_del(&shrinker->list);
411 up_write(&shrinker_rwsem);
412 kfree(shrinker->nr_deferred);
413 shrinker->nr_deferred = NULL;
415 EXPORT_SYMBOL(unregister_shrinker);
417 #define SHRINK_BATCH 128
419 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
420 struct shrinker *shrinker, int priority)
422 unsigned long freed = 0;
423 unsigned long long delta;
428 int nid = shrinkctl->nid;
429 long batch_size = shrinker->batch ? shrinker->batch
431 long scanned = 0, next_deferred;
433 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
436 freeable = shrinker->count_objects(shrinker, shrinkctl);
437 if (freeable == 0 || freeable == SHRINK_EMPTY)
441 * copy the current shrinker scan count into a local variable
442 * and zero it so that other concurrent shrinker invocations
443 * don't also do this scanning work.
445 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
448 if (shrinker->seeks) {
449 delta = freeable >> priority;
451 do_div(delta, shrinker->seeks);
454 * These objects don't require any IO to create. Trim
455 * them aggressively under memory pressure to keep
456 * them from causing refetches in the IO caches.
458 delta = freeable / 2;
462 if (total_scan < 0) {
463 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
464 shrinker->scan_objects, total_scan);
465 total_scan = freeable;
468 next_deferred = total_scan;
471 * We need to avoid excessive windup on filesystem shrinkers
472 * due to large numbers of GFP_NOFS allocations causing the
473 * shrinkers to return -1 all the time. This results in a large
474 * nr being built up so when a shrink that can do some work
475 * comes along it empties the entire cache due to nr >>>
476 * freeable. This is bad for sustaining a working set in
479 * Hence only allow the shrinker to scan the entire cache when
480 * a large delta change is calculated directly.
482 if (delta < freeable / 4)
483 total_scan = min(total_scan, freeable / 2);
486 * Avoid risking looping forever due to too large nr value:
487 * never try to free more than twice the estimate number of
490 if (total_scan > freeable * 2)
491 total_scan = freeable * 2;
493 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
494 freeable, delta, total_scan, priority);
497 * Normally, we should not scan less than batch_size objects in one
498 * pass to avoid too frequent shrinker calls, but if the slab has less
499 * than batch_size objects in total and we are really tight on memory,
500 * we will try to reclaim all available objects, otherwise we can end
501 * up failing allocations although there are plenty of reclaimable
502 * objects spread over several slabs with usage less than the
505 * We detect the "tight on memory" situations by looking at the total
506 * number of objects we want to scan (total_scan). If it is greater
507 * than the total number of objects on slab (freeable), we must be
508 * scanning at high prio and therefore should try to reclaim as much as
511 while (total_scan >= batch_size ||
512 total_scan >= freeable) {
514 unsigned long nr_to_scan = min(batch_size, total_scan);
516 shrinkctl->nr_to_scan = nr_to_scan;
517 shrinkctl->nr_scanned = nr_to_scan;
518 ret = shrinker->scan_objects(shrinker, shrinkctl);
519 if (ret == SHRINK_STOP)
523 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
524 total_scan -= shrinkctl->nr_scanned;
525 scanned += shrinkctl->nr_scanned;
530 if (next_deferred >= scanned)
531 next_deferred -= scanned;
535 * move the unused scan count back into the shrinker in a
536 * manner that handles concurrent updates. If we exhausted the
537 * scan, there is no need to do an update.
539 if (next_deferred > 0)
540 new_nr = atomic_long_add_return(next_deferred,
541 &shrinker->nr_deferred[nid]);
543 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
545 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
550 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
551 struct mem_cgroup *memcg, int priority)
553 struct memcg_shrinker_map *map;
554 unsigned long ret, freed = 0;
557 if (!mem_cgroup_online(memcg))
560 if (!down_read_trylock(&shrinker_rwsem))
563 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
568 for_each_set_bit(i, map->map, shrinker_nr_max) {
569 struct shrink_control sc = {
570 .gfp_mask = gfp_mask,
574 struct shrinker *shrinker;
576 shrinker = idr_find(&shrinker_idr, i);
577 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
579 clear_bit(i, map->map);
583 /* Call non-slab shrinkers even though kmem is disabled */
584 if (!memcg_kmem_enabled() &&
585 !(shrinker->flags & SHRINKER_NONSLAB))
588 ret = do_shrink_slab(&sc, shrinker, priority);
589 if (ret == SHRINK_EMPTY) {
590 clear_bit(i, map->map);
592 * After the shrinker reported that it had no objects to
593 * free, but before we cleared the corresponding bit in
594 * the memcg shrinker map, a new object might have been
595 * added. To make sure, we have the bit set in this
596 * case, we invoke the shrinker one more time and reset
597 * the bit if it reports that it is not empty anymore.
598 * The memory barrier here pairs with the barrier in
599 * memcg_set_shrinker_bit():
601 * list_lru_add() shrink_slab_memcg()
602 * list_add_tail() clear_bit()
604 * set_bit() do_shrink_slab()
606 smp_mb__after_atomic();
607 ret = do_shrink_slab(&sc, shrinker, priority);
608 if (ret == SHRINK_EMPTY)
611 memcg_set_shrinker_bit(memcg, nid, i);
615 if (rwsem_is_contended(&shrinker_rwsem)) {
621 up_read(&shrinker_rwsem);
624 #else /* CONFIG_MEMCG */
625 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
626 struct mem_cgroup *memcg, int priority)
630 #endif /* CONFIG_MEMCG */
633 * shrink_slab - shrink slab caches
634 * @gfp_mask: allocation context
635 * @nid: node whose slab caches to target
636 * @memcg: memory cgroup whose slab caches to target
637 * @priority: the reclaim priority
639 * Call the shrink functions to age shrinkable caches.
641 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
642 * unaware shrinkers will receive a node id of 0 instead.
644 * @memcg specifies the memory cgroup to target. Unaware shrinkers
645 * are called only if it is the root cgroup.
647 * @priority is sc->priority, we take the number of objects and >> by priority
648 * in order to get the scan target.
650 * Returns the number of reclaimed slab objects.
652 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
653 struct mem_cgroup *memcg,
656 unsigned long ret, freed = 0;
657 struct shrinker *shrinker;
660 * The root memcg might be allocated even though memcg is disabled
661 * via "cgroup_disable=memory" boot parameter. This could make
662 * mem_cgroup_is_root() return false, then just run memcg slab
663 * shrink, but skip global shrink. This may result in premature
666 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
667 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
669 if (!down_read_trylock(&shrinker_rwsem))
672 list_for_each_entry(shrinker, &shrinker_list, list) {
673 struct shrink_control sc = {
674 .gfp_mask = gfp_mask,
679 ret = do_shrink_slab(&sc, shrinker, priority);
680 if (ret == SHRINK_EMPTY)
684 * Bail out if someone want to register a new shrinker to
685 * prevent the registration from being stalled for long periods
686 * by parallel ongoing shrinking.
688 if (rwsem_is_contended(&shrinker_rwsem)) {
694 up_read(&shrinker_rwsem);
700 void drop_slab_node(int nid)
705 struct mem_cgroup *memcg = NULL;
708 memcg = mem_cgroup_iter(NULL, NULL, NULL);
710 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
711 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
712 } while (freed > 10);
719 for_each_online_node(nid)
723 static inline int is_page_cache_freeable(struct page *page)
726 * A freeable page cache page is referenced only by the caller
727 * that isolated the page, the page cache and optional buffer
728 * heads at page->private.
730 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
732 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
735 static int may_write_to_inode(struct inode *inode)
737 if (current->flags & PF_SWAPWRITE)
739 if (!inode_write_congested(inode))
741 if (inode_to_bdi(inode) == current->backing_dev_info)
747 * We detected a synchronous write error writing a page out. Probably
748 * -ENOSPC. We need to propagate that into the address_space for a subsequent
749 * fsync(), msync() or close().
751 * The tricky part is that after writepage we cannot touch the mapping: nothing
752 * prevents it from being freed up. But we have a ref on the page and once
753 * that page is locked, the mapping is pinned.
755 * We're allowed to run sleeping lock_page() here because we know the caller has
758 static void handle_write_error(struct address_space *mapping,
759 struct page *page, int error)
762 if (page_mapping(page) == mapping)
763 mapping_set_error(mapping, error);
767 /* possible outcome of pageout() */
769 /* failed to write page out, page is locked */
771 /* move page to the active list, page is locked */
773 /* page has been sent to the disk successfully, page is unlocked */
775 /* page is clean and locked */
780 * pageout is called by shrink_page_list() for each dirty page.
781 * Calls ->writepage().
783 static pageout_t pageout(struct page *page, struct address_space *mapping)
786 * If the page is dirty, only perform writeback if that write
787 * will be non-blocking. To prevent this allocation from being
788 * stalled by pagecache activity. But note that there may be
789 * stalls if we need to run get_block(). We could test
790 * PagePrivate for that.
792 * If this process is currently in __generic_file_write_iter() against
793 * this page's queue, we can perform writeback even if that
796 * If the page is swapcache, write it back even if that would
797 * block, for some throttling. This happens by accident, because
798 * swap_backing_dev_info is bust: it doesn't reflect the
799 * congestion state of the swapdevs. Easy to fix, if needed.
801 if (!is_page_cache_freeable(page))
805 * Some data journaling orphaned pages can have
806 * page->mapping == NULL while being dirty with clean buffers.
808 if (page_has_private(page)) {
809 if (try_to_free_buffers(page)) {
810 ClearPageDirty(page);
811 pr_info("%s: orphaned page\n", __func__);
817 if (mapping->a_ops->writepage == NULL)
818 return PAGE_ACTIVATE;
819 if (!may_write_to_inode(mapping->host))
822 if (clear_page_dirty_for_io(page)) {
824 struct writeback_control wbc = {
825 .sync_mode = WB_SYNC_NONE,
826 .nr_to_write = SWAP_CLUSTER_MAX,
828 .range_end = LLONG_MAX,
832 SetPageReclaim(page);
833 res = mapping->a_ops->writepage(page, &wbc);
835 handle_write_error(mapping, page, res);
836 if (res == AOP_WRITEPAGE_ACTIVATE) {
837 ClearPageReclaim(page);
838 return PAGE_ACTIVATE;
841 if (!PageWriteback(page)) {
842 /* synchronous write or broken a_ops? */
843 ClearPageReclaim(page);
845 trace_mm_vmscan_writepage(page);
846 inc_node_page_state(page, NR_VMSCAN_WRITE);
854 * Same as remove_mapping, but if the page is removed from the mapping, it
855 * gets returned with a refcount of 0.
857 static int __remove_mapping(struct address_space *mapping, struct page *page,
858 bool reclaimed, struct mem_cgroup *target_memcg)
863 BUG_ON(!PageLocked(page));
864 BUG_ON(mapping != page_mapping(page));
866 xa_lock_irqsave(&mapping->i_pages, flags);
868 * The non racy check for a busy page.
870 * Must be careful with the order of the tests. When someone has
871 * a ref to the page, it may be possible that they dirty it then
872 * drop the reference. So if PageDirty is tested before page_count
873 * here, then the following race may occur:
875 * get_user_pages(&page);
876 * [user mapping goes away]
878 * !PageDirty(page) [good]
879 * SetPageDirty(page);
881 * !page_count(page) [good, discard it]
883 * [oops, our write_to data is lost]
885 * Reversing the order of the tests ensures such a situation cannot
886 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
887 * load is not satisfied before that of page->_refcount.
889 * Note that if SetPageDirty is always performed via set_page_dirty,
890 * and thus under the i_pages lock, then this ordering is not required.
892 refcount = 1 + compound_nr(page);
893 if (!page_ref_freeze(page, refcount))
895 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
896 if (unlikely(PageDirty(page))) {
897 page_ref_unfreeze(page, refcount);
901 if (PageSwapCache(page)) {
902 swp_entry_t swap = { .val = page_private(page) };
903 mem_cgroup_swapout(page, swap);
904 __delete_from_swap_cache(page, swap);
905 xa_unlock_irqrestore(&mapping->i_pages, flags);
906 put_swap_page(page, swap);
907 workingset_eviction(page, target_memcg);
909 void (*freepage)(struct page *);
912 freepage = mapping->a_ops->freepage;
914 * Remember a shadow entry for reclaimed file cache in
915 * order to detect refaults, thus thrashing, later on.
917 * But don't store shadows in an address space that is
918 * already exiting. This is not just an optizimation,
919 * inode reclaim needs to empty out the radix tree or
920 * the nodes are lost. Don't plant shadows behind its
923 * We also don't store shadows for DAX mappings because the
924 * only page cache pages found in these are zero pages
925 * covering holes, and because we don't want to mix DAX
926 * exceptional entries and shadow exceptional entries in the
927 * same address_space.
929 if (reclaimed && page_is_file_lru(page) &&
930 !mapping_exiting(mapping) && !dax_mapping(mapping))
931 shadow = workingset_eviction(page, target_memcg);
932 __delete_from_page_cache(page, shadow);
933 xa_unlock_irqrestore(&mapping->i_pages, flags);
935 if (freepage != NULL)
942 xa_unlock_irqrestore(&mapping->i_pages, flags);
947 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
948 * someone else has a ref on the page, abort and return 0. If it was
949 * successfully detached, return 1. Assumes the caller has a single ref on
952 int remove_mapping(struct address_space *mapping, struct page *page)
954 if (__remove_mapping(mapping, page, false, NULL)) {
956 * Unfreezing the refcount with 1 rather than 2 effectively
957 * drops the pagecache ref for us without requiring another
960 page_ref_unfreeze(page, 1);
967 * putback_lru_page - put previously isolated page onto appropriate LRU list
968 * @page: page to be put back to appropriate lru list
970 * Add previously isolated @page to appropriate LRU list.
971 * Page may still be unevictable for other reasons.
973 * lru_lock must not be held, interrupts must be enabled.
975 void putback_lru_page(struct page *page)
978 put_page(page); /* drop ref from isolate */
981 enum page_references {
983 PAGEREF_RECLAIM_CLEAN,
988 static enum page_references page_check_references(struct page *page,
989 struct scan_control *sc)
991 int referenced_ptes, referenced_page;
992 unsigned long vm_flags;
994 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
996 referenced_page = TestClearPageReferenced(page);
999 * Mlock lost the isolation race with us. Let try_to_unmap()
1000 * move the page to the unevictable list.
1002 if (vm_flags & VM_LOCKED)
1003 return PAGEREF_RECLAIM;
1005 if (referenced_ptes) {
1006 if (PageSwapBacked(page))
1007 return PAGEREF_ACTIVATE;
1009 * All mapped pages start out with page table
1010 * references from the instantiating fault, so we need
1011 * to look twice if a mapped file page is used more
1014 * Mark it and spare it for another trip around the
1015 * inactive list. Another page table reference will
1016 * lead to its activation.
1018 * Note: the mark is set for activated pages as well
1019 * so that recently deactivated but used pages are
1020 * quickly recovered.
1022 SetPageReferenced(page);
1024 if (referenced_page || referenced_ptes > 1)
1025 return PAGEREF_ACTIVATE;
1028 * Activate file-backed executable pages after first usage.
1030 if (vm_flags & VM_EXEC)
1031 return PAGEREF_ACTIVATE;
1033 return PAGEREF_KEEP;
1036 /* Reclaim if clean, defer dirty pages to writeback */
1037 if (referenced_page && !PageSwapBacked(page))
1038 return PAGEREF_RECLAIM_CLEAN;
1040 return PAGEREF_RECLAIM;
1043 /* Check if a page is dirty or under writeback */
1044 static void page_check_dirty_writeback(struct page *page,
1045 bool *dirty, bool *writeback)
1047 struct address_space *mapping;
1050 * Anonymous pages are not handled by flushers and must be written
1051 * from reclaim context. Do not stall reclaim based on them
1053 if (!page_is_file_lru(page) ||
1054 (PageAnon(page) && !PageSwapBacked(page))) {
1060 /* By default assume that the page flags are accurate */
1061 *dirty = PageDirty(page);
1062 *writeback = PageWriteback(page);
1064 /* Verify dirty/writeback state if the filesystem supports it */
1065 if (!page_has_private(page))
1068 mapping = page_mapping(page);
1069 if (mapping && mapping->a_ops->is_dirty_writeback)
1070 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1074 * shrink_page_list() returns the number of reclaimed pages
1076 static unsigned int shrink_page_list(struct list_head *page_list,
1077 struct pglist_data *pgdat,
1078 struct scan_control *sc,
1079 enum ttu_flags ttu_flags,
1080 struct reclaim_stat *stat,
1081 bool ignore_references)
1083 LIST_HEAD(ret_pages);
1084 LIST_HEAD(free_pages);
1085 unsigned int nr_reclaimed = 0;
1086 unsigned int pgactivate = 0;
1088 memset(stat, 0, sizeof(*stat));
1091 while (!list_empty(page_list)) {
1092 struct address_space *mapping;
1094 enum page_references references = PAGEREF_RECLAIM;
1095 bool dirty, writeback, may_enter_fs;
1096 unsigned int nr_pages;
1100 page = lru_to_page(page_list);
1101 list_del(&page->lru);
1103 if (!trylock_page(page))
1106 VM_BUG_ON_PAGE(PageActive(page), page);
1108 nr_pages = compound_nr(page);
1110 /* Account the number of base pages even though THP */
1111 sc->nr_scanned += nr_pages;
1113 if (unlikely(!page_evictable(page)))
1114 goto activate_locked;
1116 if (!sc->may_unmap && page_mapped(page))
1119 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1120 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1123 * The number of dirty pages determines if a node is marked
1124 * reclaim_congested which affects wait_iff_congested. kswapd
1125 * will stall and start writing pages if the tail of the LRU
1126 * is all dirty unqueued pages.
1128 page_check_dirty_writeback(page, &dirty, &writeback);
1129 if (dirty || writeback)
1132 if (dirty && !writeback)
1133 stat->nr_unqueued_dirty++;
1136 * Treat this page as congested if the underlying BDI is or if
1137 * pages are cycling through the LRU so quickly that the
1138 * pages marked for immediate reclaim are making it to the
1139 * end of the LRU a second time.
1141 mapping = page_mapping(page);
1142 if (((dirty || writeback) && mapping &&
1143 inode_write_congested(mapping->host)) ||
1144 (writeback && PageReclaim(page)))
1145 stat->nr_congested++;
1148 * If a page at the tail of the LRU is under writeback, there
1149 * are three cases to consider.
1151 * 1) If reclaim is encountering an excessive number of pages
1152 * under writeback and this page is both under writeback and
1153 * PageReclaim then it indicates that pages are being queued
1154 * for IO but are being recycled through the LRU before the
1155 * IO can complete. Waiting on the page itself risks an
1156 * indefinite stall if it is impossible to writeback the
1157 * page due to IO error or disconnected storage so instead
1158 * note that the LRU is being scanned too quickly and the
1159 * caller can stall after page list has been processed.
1161 * 2) Global or new memcg reclaim encounters a page that is
1162 * not marked for immediate reclaim, or the caller does not
1163 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1164 * not to fs). In this case mark the page for immediate
1165 * reclaim and continue scanning.
1167 * Require may_enter_fs because we would wait on fs, which
1168 * may not have submitted IO yet. And the loop driver might
1169 * enter reclaim, and deadlock if it waits on a page for
1170 * which it is needed to do the write (loop masks off
1171 * __GFP_IO|__GFP_FS for this reason); but more thought
1172 * would probably show more reasons.
1174 * 3) Legacy memcg encounters a page that is already marked
1175 * PageReclaim. memcg does not have any dirty pages
1176 * throttling so we could easily OOM just because too many
1177 * pages are in writeback and there is nothing else to
1178 * reclaim. Wait for the writeback to complete.
1180 * In cases 1) and 2) we activate the pages to get them out of
1181 * the way while we continue scanning for clean pages on the
1182 * inactive list and refilling from the active list. The
1183 * observation here is that waiting for disk writes is more
1184 * expensive than potentially causing reloads down the line.
1185 * Since they're marked for immediate reclaim, they won't put
1186 * memory pressure on the cache working set any longer than it
1187 * takes to write them to disk.
1189 if (PageWriteback(page)) {
1191 if (current_is_kswapd() &&
1192 PageReclaim(page) &&
1193 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1194 stat->nr_immediate++;
1195 goto activate_locked;
1198 } else if (writeback_throttling_sane(sc) ||
1199 !PageReclaim(page) || !may_enter_fs) {
1201 * This is slightly racy - end_page_writeback()
1202 * might have just cleared PageReclaim, then
1203 * setting PageReclaim here end up interpreted
1204 * as PageReadahead - but that does not matter
1205 * enough to care. What we do want is for this
1206 * page to have PageReclaim set next time memcg
1207 * reclaim reaches the tests above, so it will
1208 * then wait_on_page_writeback() to avoid OOM;
1209 * and it's also appropriate in global reclaim.
1211 SetPageReclaim(page);
1212 stat->nr_writeback++;
1213 goto activate_locked;
1218 wait_on_page_writeback(page);
1219 /* then go back and try same page again */
1220 list_add_tail(&page->lru, page_list);
1225 if (!ignore_references)
1226 references = page_check_references(page, sc);
1228 switch (references) {
1229 case PAGEREF_ACTIVATE:
1230 goto activate_locked;
1232 stat->nr_ref_keep += nr_pages;
1234 case PAGEREF_RECLAIM:
1235 case PAGEREF_RECLAIM_CLEAN:
1236 ; /* try to reclaim the page below */
1240 * Anonymous process memory has backing store?
1241 * Try to allocate it some swap space here.
1242 * Lazyfree page could be freed directly
1244 if (PageAnon(page) && PageSwapBacked(page)) {
1245 if (!PageSwapCache(page)) {
1246 if (!(sc->gfp_mask & __GFP_IO))
1248 if (PageTransHuge(page)) {
1249 /* cannot split THP, skip it */
1250 if (!can_split_huge_page(page, NULL))
1251 goto activate_locked;
1253 * Split pages without a PMD map right
1254 * away. Chances are some or all of the
1255 * tail pages can be freed without IO.
1257 if (!compound_mapcount(page) &&
1258 split_huge_page_to_list(page,
1260 goto activate_locked;
1262 if (!add_to_swap(page)) {
1263 if (!PageTransHuge(page))
1264 goto activate_locked_split;
1265 /* Fallback to swap normal pages */
1266 if (split_huge_page_to_list(page,
1268 goto activate_locked;
1269 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1270 count_vm_event(THP_SWPOUT_FALLBACK);
1272 if (!add_to_swap(page))
1273 goto activate_locked_split;
1276 may_enter_fs = true;
1278 /* Adding to swap updated mapping */
1279 mapping = page_mapping(page);
1281 } else if (unlikely(PageTransHuge(page))) {
1282 /* Split file THP */
1283 if (split_huge_page_to_list(page, page_list))
1288 * THP may get split above, need minus tail pages and update
1289 * nr_pages to avoid accounting tail pages twice.
1291 * The tail pages that are added into swap cache successfully
1294 if ((nr_pages > 1) && !PageTransHuge(page)) {
1295 sc->nr_scanned -= (nr_pages - 1);
1300 * The page is mapped into the page tables of one or more
1301 * processes. Try to unmap it here.
1303 if (page_mapped(page)) {
1304 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1305 bool was_swapbacked = PageSwapBacked(page);
1307 if (unlikely(PageTransHuge(page)))
1308 flags |= TTU_SPLIT_HUGE_PMD;
1310 if (!try_to_unmap(page, flags)) {
1311 stat->nr_unmap_fail += nr_pages;
1312 if (!was_swapbacked && PageSwapBacked(page))
1313 stat->nr_lazyfree_fail += nr_pages;
1314 goto activate_locked;
1318 if (PageDirty(page)) {
1320 * Only kswapd can writeback filesystem pages
1321 * to avoid risk of stack overflow. But avoid
1322 * injecting inefficient single-page IO into
1323 * flusher writeback as much as possible: only
1324 * write pages when we've encountered many
1325 * dirty pages, and when we've already scanned
1326 * the rest of the LRU for clean pages and see
1327 * the same dirty pages again (PageReclaim).
1329 if (page_is_file_lru(page) &&
1330 (!current_is_kswapd() || !PageReclaim(page) ||
1331 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1333 * Immediately reclaim when written back.
1334 * Similar in principal to deactivate_page()
1335 * except we already have the page isolated
1336 * and know it's dirty
1338 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1339 SetPageReclaim(page);
1341 goto activate_locked;
1344 if (references == PAGEREF_RECLAIM_CLEAN)
1348 if (!sc->may_writepage)
1352 * Page is dirty. Flush the TLB if a writable entry
1353 * potentially exists to avoid CPU writes after IO
1354 * starts and then write it out here.
1356 try_to_unmap_flush_dirty();
1357 switch (pageout(page, mapping)) {
1361 goto activate_locked;
1363 stat->nr_pageout += hpage_nr_pages(page);
1365 if (PageWriteback(page))
1367 if (PageDirty(page))
1371 * A synchronous write - probably a ramdisk. Go
1372 * ahead and try to reclaim the page.
1374 if (!trylock_page(page))
1376 if (PageDirty(page) || PageWriteback(page))
1378 mapping = page_mapping(page);
1380 ; /* try to free the page below */
1385 * If the page has buffers, try to free the buffer mappings
1386 * associated with this page. If we succeed we try to free
1389 * We do this even if the page is PageDirty().
1390 * try_to_release_page() does not perform I/O, but it is
1391 * possible for a page to have PageDirty set, but it is actually
1392 * clean (all its buffers are clean). This happens if the
1393 * buffers were written out directly, with submit_bh(). ext3
1394 * will do this, as well as the blockdev mapping.
1395 * try_to_release_page() will discover that cleanness and will
1396 * drop the buffers and mark the page clean - it can be freed.
1398 * Rarely, pages can have buffers and no ->mapping. These are
1399 * the pages which were not successfully invalidated in
1400 * truncate_complete_page(). We try to drop those buffers here
1401 * and if that worked, and the page is no longer mapped into
1402 * process address space (page_count == 1) it can be freed.
1403 * Otherwise, leave the page on the LRU so it is swappable.
1405 if (page_has_private(page)) {
1406 if (!try_to_release_page(page, sc->gfp_mask))
1407 goto activate_locked;
1408 if (!mapping && page_count(page) == 1) {
1410 if (put_page_testzero(page))
1414 * rare race with speculative reference.
1415 * the speculative reference will free
1416 * this page shortly, so we may
1417 * increment nr_reclaimed here (and
1418 * leave it off the LRU).
1426 if (PageAnon(page) && !PageSwapBacked(page)) {
1427 /* follow __remove_mapping for reference */
1428 if (!page_ref_freeze(page, 1))
1430 if (PageDirty(page)) {
1431 page_ref_unfreeze(page, 1);
1435 count_vm_event(PGLAZYFREED);
1436 count_memcg_page_event(page, PGLAZYFREED);
1437 } else if (!mapping || !__remove_mapping(mapping, page, true,
1438 sc->target_mem_cgroup))
1444 * THP may get swapped out in a whole, need account
1447 nr_reclaimed += nr_pages;
1450 * Is there need to periodically free_page_list? It would
1451 * appear not as the counts should be low
1453 if (unlikely(PageTransHuge(page)))
1454 destroy_compound_page(page);
1456 list_add(&page->lru, &free_pages);
1459 activate_locked_split:
1461 * The tail pages that are failed to add into swap cache
1462 * reach here. Fixup nr_scanned and nr_pages.
1465 sc->nr_scanned -= (nr_pages - 1);
1469 /* Not a candidate for swapping, so reclaim swap space. */
1470 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1472 try_to_free_swap(page);
1473 VM_BUG_ON_PAGE(PageActive(page), page);
1474 if (!PageMlocked(page)) {
1475 int type = page_is_file_lru(page);
1476 SetPageActive(page);
1477 stat->nr_activate[type] += nr_pages;
1478 count_memcg_page_event(page, PGACTIVATE);
1483 list_add(&page->lru, &ret_pages);
1484 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1487 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1489 mem_cgroup_uncharge_list(&free_pages);
1490 try_to_unmap_flush();
1491 free_unref_page_list(&free_pages);
1493 list_splice(&ret_pages, page_list);
1494 count_vm_events(PGACTIVATE, pgactivate);
1496 return nr_reclaimed;
1499 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1500 struct list_head *page_list)
1502 struct scan_control sc = {
1503 .gfp_mask = GFP_KERNEL,
1504 .priority = DEF_PRIORITY,
1507 struct reclaim_stat stat;
1508 unsigned int nr_reclaimed;
1509 struct page *page, *next;
1510 LIST_HEAD(clean_pages);
1512 list_for_each_entry_safe(page, next, page_list, lru) {
1513 if (page_is_file_lru(page) && !PageDirty(page) &&
1514 !__PageMovable(page) && !PageUnevictable(page)) {
1515 ClearPageActive(page);
1516 list_move(&page->lru, &clean_pages);
1520 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1521 TTU_IGNORE_ACCESS, &stat, true);
1522 list_splice(&clean_pages, page_list);
1523 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -nr_reclaimed);
1525 * Since lazyfree pages are isolated from file LRU from the beginning,
1526 * they will rotate back to anonymous LRU in the end if it failed to
1527 * discard so isolated count will be mismatched.
1528 * Compensate the isolated count for both LRU lists.
1530 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1531 stat.nr_lazyfree_fail);
1532 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1533 -stat.nr_lazyfree_fail);
1534 return nr_reclaimed;
1538 * Attempt to remove the specified page from its LRU. Only take this page
1539 * if it is of the appropriate PageActive status. Pages which are being
1540 * freed elsewhere are also ignored.
1542 * page: page to consider
1543 * mode: one of the LRU isolation modes defined above
1545 * returns 0 on success, -ve errno on failure.
1547 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1551 /* Only take pages on the LRU. */
1555 /* Compaction should not handle unevictable pages but CMA can do so */
1556 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1562 * To minimise LRU disruption, the caller can indicate that it only
1563 * wants to isolate pages it will be able to operate on without
1564 * blocking - clean pages for the most part.
1566 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1567 * that it is possible to migrate without blocking
1569 if (mode & ISOLATE_ASYNC_MIGRATE) {
1570 /* All the caller can do on PageWriteback is block */
1571 if (PageWriteback(page))
1574 if (PageDirty(page)) {
1575 struct address_space *mapping;
1579 * Only pages without mappings or that have a
1580 * ->migratepage callback are possible to migrate
1581 * without blocking. However, we can be racing with
1582 * truncation so it's necessary to lock the page
1583 * to stabilise the mapping as truncation holds
1584 * the page lock until after the page is removed
1585 * from the page cache.
1587 if (!trylock_page(page))
1590 mapping = page_mapping(page);
1591 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1598 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1601 if (likely(get_page_unless_zero(page))) {
1603 * Be careful not to clear PageLRU until after we're
1604 * sure the page is not being freed elsewhere -- the
1605 * page release code relies on it.
1616 * Update LRU sizes after isolating pages. The LRU size updates must
1617 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1619 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1620 enum lru_list lru, unsigned long *nr_zone_taken)
1624 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1625 if (!nr_zone_taken[zid])
1628 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1634 * pgdat->lru_lock is heavily contended. Some of the functions that
1635 * shrink the lists perform better by taking out a batch of pages
1636 * and working on them outside the LRU lock.
1638 * For pagecache intensive workloads, this function is the hottest
1639 * spot in the kernel (apart from copy_*_user functions).
1641 * Appropriate locks must be held before calling this function.
1643 * @nr_to_scan: The number of eligible pages to look through on the list.
1644 * @lruvec: The LRU vector to pull pages from.
1645 * @dst: The temp list to put pages on to.
1646 * @nr_scanned: The number of pages that were scanned.
1647 * @sc: The scan_control struct for this reclaim session
1648 * @lru: LRU list id for isolating
1650 * returns how many pages were moved onto *@dst.
1652 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1653 struct lruvec *lruvec, struct list_head *dst,
1654 unsigned long *nr_scanned, struct scan_control *sc,
1657 struct list_head *src = &lruvec->lists[lru];
1658 unsigned long nr_taken = 0;
1659 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1660 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1661 unsigned long skipped = 0;
1662 unsigned long scan, total_scan, nr_pages;
1663 LIST_HEAD(pages_skipped);
1664 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1668 while (scan < nr_to_scan && !list_empty(src)) {
1671 page = lru_to_page(src);
1672 prefetchw_prev_lru_page(page, src, flags);
1674 VM_BUG_ON_PAGE(!PageLRU(page), page);
1676 nr_pages = compound_nr(page);
1677 total_scan += nr_pages;
1679 if (page_zonenum(page) > sc->reclaim_idx) {
1680 list_move(&page->lru, &pages_skipped);
1681 nr_skipped[page_zonenum(page)] += nr_pages;
1686 * Do not count skipped pages because that makes the function
1687 * return with no isolated pages if the LRU mostly contains
1688 * ineligible pages. This causes the VM to not reclaim any
1689 * pages, triggering a premature OOM.
1691 * Account all tail pages of THP. This would not cause
1692 * premature OOM since __isolate_lru_page() returns -EBUSY
1693 * only when the page is being freed somewhere else.
1696 switch (__isolate_lru_page(page, mode)) {
1698 nr_taken += nr_pages;
1699 nr_zone_taken[page_zonenum(page)] += nr_pages;
1700 list_move(&page->lru, dst);
1704 /* else it is being freed elsewhere */
1705 list_move(&page->lru, src);
1714 * Splice any skipped pages to the start of the LRU list. Note that
1715 * this disrupts the LRU order when reclaiming for lower zones but
1716 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1717 * scanning would soon rescan the same pages to skip and put the
1718 * system at risk of premature OOM.
1720 if (!list_empty(&pages_skipped)) {
1723 list_splice(&pages_skipped, src);
1724 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1725 if (!nr_skipped[zid])
1728 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1729 skipped += nr_skipped[zid];
1732 *nr_scanned = total_scan;
1733 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1734 total_scan, skipped, nr_taken, mode, lru);
1735 update_lru_sizes(lruvec, lru, nr_zone_taken);
1740 * isolate_lru_page - tries to isolate a page from its LRU list
1741 * @page: page to isolate from its LRU list
1743 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1744 * vmstat statistic corresponding to whatever LRU list the page was on.
1746 * Returns 0 if the page was removed from an LRU list.
1747 * Returns -EBUSY if the page was not on an LRU list.
1749 * The returned page will have PageLRU() cleared. If it was found on
1750 * the active list, it will have PageActive set. If it was found on
1751 * the unevictable list, it will have the PageUnevictable bit set. That flag
1752 * may need to be cleared by the caller before letting the page go.
1754 * The vmstat statistic corresponding to the list on which the page was
1755 * found will be decremented.
1759 * (1) Must be called with an elevated refcount on the page. This is a
1760 * fundamentnal difference from isolate_lru_pages (which is called
1761 * without a stable reference).
1762 * (2) the lru_lock must not be held.
1763 * (3) interrupts must be enabled.
1765 int isolate_lru_page(struct page *page)
1769 VM_BUG_ON_PAGE(!page_count(page), page);
1770 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1772 if (PageLRU(page)) {
1773 pg_data_t *pgdat = page_pgdat(page);
1774 struct lruvec *lruvec;
1776 spin_lock_irq(&pgdat->lru_lock);
1777 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1778 if (PageLRU(page)) {
1779 int lru = page_lru(page);
1782 del_page_from_lru_list(page, lruvec, lru);
1785 spin_unlock_irq(&pgdat->lru_lock);
1791 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1792 * then get rescheduled. When there are massive number of tasks doing page
1793 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1794 * the LRU list will go small and be scanned faster than necessary, leading to
1795 * unnecessary swapping, thrashing and OOM.
1797 static int too_many_isolated(struct pglist_data *pgdat, int file,
1798 struct scan_control *sc)
1800 unsigned long inactive, isolated;
1802 if (current_is_kswapd())
1805 if (!writeback_throttling_sane(sc))
1809 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1810 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1812 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1813 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1817 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1818 * won't get blocked by normal direct-reclaimers, forming a circular
1821 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1824 return isolated > inactive;
1828 * This moves pages from @list to corresponding LRU list.
1830 * We move them the other way if the page is referenced by one or more
1831 * processes, from rmap.
1833 * If the pages are mostly unmapped, the processing is fast and it is
1834 * appropriate to hold zone_lru_lock across the whole operation. But if
1835 * the pages are mapped, the processing is slow (page_referenced()) so we
1836 * should drop zone_lru_lock around each page. It's impossible to balance
1837 * this, so instead we remove the pages from the LRU while processing them.
1838 * It is safe to rely on PG_active against the non-LRU pages in here because
1839 * nobody will play with that bit on a non-LRU page.
1841 * The downside is that we have to touch page->_refcount against each page.
1842 * But we had to alter page->flags anyway.
1844 * Returns the number of pages moved to the given lruvec.
1847 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1848 struct list_head *list)
1850 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1851 int nr_pages, nr_moved = 0;
1852 LIST_HEAD(pages_to_free);
1856 while (!list_empty(list)) {
1857 page = lru_to_page(list);
1858 VM_BUG_ON_PAGE(PageLRU(page), page);
1859 if (unlikely(!page_evictable(page))) {
1860 list_del(&page->lru);
1861 spin_unlock_irq(&pgdat->lru_lock);
1862 putback_lru_page(page);
1863 spin_lock_irq(&pgdat->lru_lock);
1866 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1869 lru = page_lru(page);
1871 nr_pages = hpage_nr_pages(page);
1872 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1873 list_move(&page->lru, &lruvec->lists[lru]);
1875 if (put_page_testzero(page)) {
1876 __ClearPageLRU(page);
1877 __ClearPageActive(page);
1878 del_page_from_lru_list(page, lruvec, lru);
1880 if (unlikely(PageCompound(page))) {
1881 spin_unlock_irq(&pgdat->lru_lock);
1882 destroy_compound_page(page);
1883 spin_lock_irq(&pgdat->lru_lock);
1885 list_add(&page->lru, &pages_to_free);
1887 nr_moved += nr_pages;
1888 if (PageActive(page))
1889 workingset_age_nonresident(lruvec, nr_pages);
1894 * To save our caller's stack, now use input list for pages to free.
1896 list_splice(&pages_to_free, list);
1902 * If a kernel thread (such as nfsd for loop-back mounts) services
1903 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1904 * In that case we should only throttle if the backing device it is
1905 * writing to is congested. In other cases it is safe to throttle.
1907 static int current_may_throttle(void)
1909 return !(current->flags & PF_LOCAL_THROTTLE) ||
1910 current->backing_dev_info == NULL ||
1911 bdi_write_congested(current->backing_dev_info);
1915 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1916 * of reclaimed pages
1918 static noinline_for_stack unsigned long
1919 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1920 struct scan_control *sc, enum lru_list lru)
1922 LIST_HEAD(page_list);
1923 unsigned long nr_scanned;
1924 unsigned int nr_reclaimed = 0;
1925 unsigned long nr_taken;
1926 struct reclaim_stat stat;
1927 bool file = is_file_lru(lru);
1928 enum vm_event_item item;
1929 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1930 bool stalled = false;
1932 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1936 /* wait a bit for the reclaimer. */
1940 /* We are about to die and free our memory. Return now. */
1941 if (fatal_signal_pending(current))
1942 return SWAP_CLUSTER_MAX;
1947 spin_lock_irq(&pgdat->lru_lock);
1949 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1950 &nr_scanned, sc, lru);
1952 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1953 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1954 if (!cgroup_reclaim(sc))
1955 __count_vm_events(item, nr_scanned);
1956 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1957 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1959 spin_unlock_irq(&pgdat->lru_lock);
1964 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1967 spin_lock_irq(&pgdat->lru_lock);
1969 move_pages_to_lru(lruvec, &page_list);
1971 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1972 lru_note_cost(lruvec, file, stat.nr_pageout);
1973 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1974 if (!cgroup_reclaim(sc))
1975 __count_vm_events(item, nr_reclaimed);
1976 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1977 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1979 spin_unlock_irq(&pgdat->lru_lock);
1981 mem_cgroup_uncharge_list(&page_list);
1982 free_unref_page_list(&page_list);
1985 * If dirty pages are scanned that are not queued for IO, it
1986 * implies that flushers are not doing their job. This can
1987 * happen when memory pressure pushes dirty pages to the end of
1988 * the LRU before the dirty limits are breached and the dirty
1989 * data has expired. It can also happen when the proportion of
1990 * dirty pages grows not through writes but through memory
1991 * pressure reclaiming all the clean cache. And in some cases,
1992 * the flushers simply cannot keep up with the allocation
1993 * rate. Nudge the flusher threads in case they are asleep.
1995 if (stat.nr_unqueued_dirty == nr_taken)
1996 wakeup_flusher_threads(WB_REASON_VMSCAN);
1998 sc->nr.dirty += stat.nr_dirty;
1999 sc->nr.congested += stat.nr_congested;
2000 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2001 sc->nr.writeback += stat.nr_writeback;
2002 sc->nr.immediate += stat.nr_immediate;
2003 sc->nr.taken += nr_taken;
2005 sc->nr.file_taken += nr_taken;
2007 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2008 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2009 return nr_reclaimed;
2012 static void shrink_active_list(unsigned long nr_to_scan,
2013 struct lruvec *lruvec,
2014 struct scan_control *sc,
2017 unsigned long nr_taken;
2018 unsigned long nr_scanned;
2019 unsigned long vm_flags;
2020 LIST_HEAD(l_hold); /* The pages which were snipped off */
2021 LIST_HEAD(l_active);
2022 LIST_HEAD(l_inactive);
2024 unsigned nr_deactivate, nr_activate;
2025 unsigned nr_rotated = 0;
2026 int file = is_file_lru(lru);
2027 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2031 spin_lock_irq(&pgdat->lru_lock);
2033 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2034 &nr_scanned, sc, lru);
2036 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2038 __count_vm_events(PGREFILL, nr_scanned);
2039 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2041 spin_unlock_irq(&pgdat->lru_lock);
2043 while (!list_empty(&l_hold)) {
2045 page = lru_to_page(&l_hold);
2046 list_del(&page->lru);
2048 if (unlikely(!page_evictable(page))) {
2049 putback_lru_page(page);
2053 if (unlikely(buffer_heads_over_limit)) {
2054 if (page_has_private(page) && trylock_page(page)) {
2055 if (page_has_private(page))
2056 try_to_release_page(page, 0);
2061 if (page_referenced(page, 0, sc->target_mem_cgroup,
2064 * Identify referenced, file-backed active pages and
2065 * give them one more trip around the active list. So
2066 * that executable code get better chances to stay in
2067 * memory under moderate memory pressure. Anon pages
2068 * are not likely to be evicted by use-once streaming
2069 * IO, plus JVM can create lots of anon VM_EXEC pages,
2070 * so we ignore them here.
2072 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2073 nr_rotated += hpage_nr_pages(page);
2074 list_add(&page->lru, &l_active);
2079 ClearPageActive(page); /* we are de-activating */
2080 SetPageWorkingset(page);
2081 list_add(&page->lru, &l_inactive);
2085 * Move pages back to the lru list.
2087 spin_lock_irq(&pgdat->lru_lock);
2089 nr_activate = move_pages_to_lru(lruvec, &l_active);
2090 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2091 /* Keep all free pages in l_active list */
2092 list_splice(&l_inactive, &l_active);
2094 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2095 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2097 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2098 spin_unlock_irq(&pgdat->lru_lock);
2100 mem_cgroup_uncharge_list(&l_active);
2101 free_unref_page_list(&l_active);
2102 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2103 nr_deactivate, nr_rotated, sc->priority, file);
2106 unsigned long reclaim_pages(struct list_head *page_list)
2108 int nid = NUMA_NO_NODE;
2109 unsigned int nr_reclaimed = 0;
2110 LIST_HEAD(node_page_list);
2111 struct reclaim_stat dummy_stat;
2113 struct scan_control sc = {
2114 .gfp_mask = GFP_KERNEL,
2115 .priority = DEF_PRIORITY,
2121 while (!list_empty(page_list)) {
2122 page = lru_to_page(page_list);
2123 if (nid == NUMA_NO_NODE) {
2124 nid = page_to_nid(page);
2125 INIT_LIST_HEAD(&node_page_list);
2128 if (nid == page_to_nid(page)) {
2129 ClearPageActive(page);
2130 list_move(&page->lru, &node_page_list);
2134 nr_reclaimed += shrink_page_list(&node_page_list,
2137 &dummy_stat, false);
2138 while (!list_empty(&node_page_list)) {
2139 page = lru_to_page(&node_page_list);
2140 list_del(&page->lru);
2141 putback_lru_page(page);
2147 if (!list_empty(&node_page_list)) {
2148 nr_reclaimed += shrink_page_list(&node_page_list,
2151 &dummy_stat, false);
2152 while (!list_empty(&node_page_list)) {
2153 page = lru_to_page(&node_page_list);
2154 list_del(&page->lru);
2155 putback_lru_page(page);
2159 return nr_reclaimed;
2162 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2163 struct lruvec *lruvec, struct scan_control *sc)
2165 if (is_active_lru(lru)) {
2166 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2167 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2169 sc->skipped_deactivate = 1;
2173 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2177 * The inactive anon list should be small enough that the VM never has
2178 * to do too much work.
2180 * The inactive file list should be small enough to leave most memory
2181 * to the established workingset on the scan-resistant active list,
2182 * but large enough to avoid thrashing the aggregate readahead window.
2184 * Both inactive lists should also be large enough that each inactive
2185 * page has a chance to be referenced again before it is reclaimed.
2187 * If that fails and refaulting is observed, the inactive list grows.
2189 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2190 * on this LRU, maintained by the pageout code. An inactive_ratio
2191 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2194 * memory ratio inactive
2195 * -------------------------------------
2204 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2206 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2207 unsigned long inactive, active;
2208 unsigned long inactive_ratio;
2211 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2212 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2214 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2216 inactive_ratio = int_sqrt(10 * gb);
2220 return inactive * inactive_ratio < active;
2231 * Determine how aggressively the anon and file LRU lists should be
2232 * scanned. The relative value of each set of LRU lists is determined
2233 * by looking at the fraction of the pages scanned we did rotate back
2234 * onto the active list instead of evict.
2236 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2237 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2239 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2242 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2243 unsigned long anon_cost, file_cost, total_cost;
2244 int swappiness = mem_cgroup_swappiness(memcg);
2246 u64 denominator = 0; /* gcc */
2247 enum scan_balance scan_balance;
2248 unsigned long ap, fp;
2251 /* If we have no swap space, do not bother scanning anon pages. */
2252 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2253 scan_balance = SCAN_FILE;
2258 * Global reclaim will swap to prevent OOM even with no
2259 * swappiness, but memcg users want to use this knob to
2260 * disable swapping for individual groups completely when
2261 * using the memory controller's swap limit feature would be
2264 if (cgroup_reclaim(sc) && !swappiness) {
2265 scan_balance = SCAN_FILE;
2270 * Do not apply any pressure balancing cleverness when the
2271 * system is close to OOM, scan both anon and file equally
2272 * (unless the swappiness setting disagrees with swapping).
2274 if (!sc->priority && swappiness) {
2275 scan_balance = SCAN_EQUAL;
2280 * If the system is almost out of file pages, force-scan anon.
2282 if (sc->file_is_tiny) {
2283 scan_balance = SCAN_ANON;
2288 * If there is enough inactive page cache, we do not reclaim
2289 * anything from the anonymous working right now.
2291 if (sc->cache_trim_mode) {
2292 scan_balance = SCAN_FILE;
2296 scan_balance = SCAN_FRACT;
2298 * Calculate the pressure balance between anon and file pages.
2300 * The amount of pressure we put on each LRU is inversely
2301 * proportional to the cost of reclaiming each list, as
2302 * determined by the share of pages that are refaulting, times
2303 * the relative IO cost of bringing back a swapped out
2304 * anonymous page vs reloading a filesystem page (swappiness).
2306 * Although we limit that influence to ensure no list gets
2307 * left behind completely: at least a third of the pressure is
2308 * applied, before swappiness.
2310 * With swappiness at 100, anon and file have equal IO cost.
2312 total_cost = sc->anon_cost + sc->file_cost;
2313 anon_cost = total_cost + sc->anon_cost;
2314 file_cost = total_cost + sc->file_cost;
2315 total_cost = anon_cost + file_cost;
2317 ap = swappiness * (total_cost + 1);
2318 ap /= anon_cost + 1;
2320 fp = (200 - swappiness) * (total_cost + 1);
2321 fp /= file_cost + 1;
2325 denominator = ap + fp;
2327 for_each_evictable_lru(lru) {
2328 int file = is_file_lru(lru);
2329 unsigned long lruvec_size;
2331 unsigned long protection;
2333 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2334 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2336 sc->memcg_low_reclaim);
2340 * Scale a cgroup's reclaim pressure by proportioning
2341 * its current usage to its memory.low or memory.min
2344 * This is important, as otherwise scanning aggression
2345 * becomes extremely binary -- from nothing as we
2346 * approach the memory protection threshold, to totally
2347 * nominal as we exceed it. This results in requiring
2348 * setting extremely liberal protection thresholds. It
2349 * also means we simply get no protection at all if we
2350 * set it too low, which is not ideal.
2352 * If there is any protection in place, we reduce scan
2353 * pressure by how much of the total memory used is
2354 * within protection thresholds.
2356 * There is one special case: in the first reclaim pass,
2357 * we skip over all groups that are within their low
2358 * protection. If that fails to reclaim enough pages to
2359 * satisfy the reclaim goal, we come back and override
2360 * the best-effort low protection. However, we still
2361 * ideally want to honor how well-behaved groups are in
2362 * that case instead of simply punishing them all
2363 * equally. As such, we reclaim them based on how much
2364 * memory they are using, reducing the scan pressure
2365 * again by how much of the total memory used is under
2368 unsigned long cgroup_size = mem_cgroup_size(memcg);
2370 /* Avoid TOCTOU with earlier protection check */
2371 cgroup_size = max(cgroup_size, protection);
2373 scan = lruvec_size - lruvec_size * protection /
2377 * Minimally target SWAP_CLUSTER_MAX pages to keep
2378 * reclaim moving forwards, avoiding decrementing
2379 * sc->priority further than desirable.
2381 scan = max(scan, SWAP_CLUSTER_MAX);
2386 scan >>= sc->priority;
2389 * If the cgroup's already been deleted, make sure to
2390 * scrape out the remaining cache.
2392 if (!scan && !mem_cgroup_online(memcg))
2393 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2395 switch (scan_balance) {
2397 /* Scan lists relative to size */
2401 * Scan types proportional to swappiness and
2402 * their relative recent reclaim efficiency.
2403 * Make sure we don't miss the last page on
2404 * the offlined memory cgroups because of a
2407 scan = mem_cgroup_online(memcg) ?
2408 div64_u64(scan * fraction[file], denominator) :
2409 DIV64_U64_ROUND_UP(scan * fraction[file],
2414 /* Scan one type exclusively */
2415 if ((scan_balance == SCAN_FILE) != file)
2419 /* Look ma, no brain */
2427 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2429 unsigned long nr[NR_LRU_LISTS];
2430 unsigned long targets[NR_LRU_LISTS];
2431 unsigned long nr_to_scan;
2433 unsigned long nr_reclaimed = 0;
2434 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2435 struct blk_plug plug;
2438 get_scan_count(lruvec, sc, nr);
2440 /* Record the original scan target for proportional adjustments later */
2441 memcpy(targets, nr, sizeof(nr));
2444 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2445 * event that can occur when there is little memory pressure e.g.
2446 * multiple streaming readers/writers. Hence, we do not abort scanning
2447 * when the requested number of pages are reclaimed when scanning at
2448 * DEF_PRIORITY on the assumption that the fact we are direct
2449 * reclaiming implies that kswapd is not keeping up and it is best to
2450 * do a batch of work at once. For memcg reclaim one check is made to
2451 * abort proportional reclaim if either the file or anon lru has already
2452 * dropped to zero at the first pass.
2454 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2455 sc->priority == DEF_PRIORITY);
2457 blk_start_plug(&plug);
2458 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2459 nr[LRU_INACTIVE_FILE]) {
2460 unsigned long nr_anon, nr_file, percentage;
2461 unsigned long nr_scanned;
2463 for_each_evictable_lru(lru) {
2465 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2466 nr[lru] -= nr_to_scan;
2468 nr_reclaimed += shrink_list(lru, nr_to_scan,
2475 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2479 * For kswapd and memcg, reclaim at least the number of pages
2480 * requested. Ensure that the anon and file LRUs are scanned
2481 * proportionally what was requested by get_scan_count(). We
2482 * stop reclaiming one LRU and reduce the amount scanning
2483 * proportional to the original scan target.
2485 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2486 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2489 * It's just vindictive to attack the larger once the smaller
2490 * has gone to zero. And given the way we stop scanning the
2491 * smaller below, this makes sure that we only make one nudge
2492 * towards proportionality once we've got nr_to_reclaim.
2494 if (!nr_file || !nr_anon)
2497 if (nr_file > nr_anon) {
2498 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2499 targets[LRU_ACTIVE_ANON] + 1;
2501 percentage = nr_anon * 100 / scan_target;
2503 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2504 targets[LRU_ACTIVE_FILE] + 1;
2506 percentage = nr_file * 100 / scan_target;
2509 /* Stop scanning the smaller of the LRU */
2511 nr[lru + LRU_ACTIVE] = 0;
2514 * Recalculate the other LRU scan count based on its original
2515 * scan target and the percentage scanning already complete
2517 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2518 nr_scanned = targets[lru] - nr[lru];
2519 nr[lru] = targets[lru] * (100 - percentage) / 100;
2520 nr[lru] -= min(nr[lru], nr_scanned);
2523 nr_scanned = targets[lru] - nr[lru];
2524 nr[lru] = targets[lru] * (100 - percentage) / 100;
2525 nr[lru] -= min(nr[lru], nr_scanned);
2527 scan_adjusted = true;
2529 blk_finish_plug(&plug);
2530 sc->nr_reclaimed += nr_reclaimed;
2533 * Even if we did not try to evict anon pages at all, we want to
2534 * rebalance the anon lru active/inactive ratio.
2536 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2537 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2538 sc, LRU_ACTIVE_ANON);
2541 /* Use reclaim/compaction for costly allocs or under memory pressure */
2542 static bool in_reclaim_compaction(struct scan_control *sc)
2544 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2545 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2546 sc->priority < DEF_PRIORITY - 2))
2553 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2554 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2555 * true if more pages should be reclaimed such that when the page allocator
2556 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2557 * It will give up earlier than that if there is difficulty reclaiming pages.
2559 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2560 unsigned long nr_reclaimed,
2561 struct scan_control *sc)
2563 unsigned long pages_for_compaction;
2564 unsigned long inactive_lru_pages;
2567 /* If not in reclaim/compaction mode, stop */
2568 if (!in_reclaim_compaction(sc))
2572 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2573 * number of pages that were scanned. This will return to the caller
2574 * with the risk reclaim/compaction and the resulting allocation attempt
2575 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2576 * allocations through requiring that the full LRU list has been scanned
2577 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2578 * scan, but that approximation was wrong, and there were corner cases
2579 * where always a non-zero amount of pages were scanned.
2584 /* If compaction would go ahead or the allocation would succeed, stop */
2585 for (z = 0; z <= sc->reclaim_idx; z++) {
2586 struct zone *zone = &pgdat->node_zones[z];
2587 if (!managed_zone(zone))
2590 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2591 case COMPACT_SUCCESS:
2592 case COMPACT_CONTINUE:
2595 /* check next zone */
2601 * If we have not reclaimed enough pages for compaction and the
2602 * inactive lists are large enough, continue reclaiming
2604 pages_for_compaction = compact_gap(sc->order);
2605 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2606 if (get_nr_swap_pages() > 0)
2607 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2609 return inactive_lru_pages > pages_for_compaction;
2612 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2614 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2615 struct mem_cgroup *memcg;
2617 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2619 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2620 unsigned long reclaimed;
2621 unsigned long scanned;
2623 mem_cgroup_calculate_protection(target_memcg, memcg);
2625 if (mem_cgroup_below_min(memcg)) {
2628 * If there is no reclaimable memory, OOM.
2631 } else if (mem_cgroup_below_low(memcg)) {
2634 * Respect the protection only as long as
2635 * there is an unprotected supply
2636 * of reclaimable memory from other cgroups.
2638 if (!sc->memcg_low_reclaim) {
2639 sc->memcg_low_skipped = 1;
2642 memcg_memory_event(memcg, MEMCG_LOW);
2645 reclaimed = sc->nr_reclaimed;
2646 scanned = sc->nr_scanned;
2648 shrink_lruvec(lruvec, sc);
2650 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2653 /* Record the group's reclaim efficiency */
2654 vmpressure(sc->gfp_mask, memcg, false,
2655 sc->nr_scanned - scanned,
2656 sc->nr_reclaimed - reclaimed);
2658 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2661 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2663 struct reclaim_state *reclaim_state = current->reclaim_state;
2664 unsigned long nr_reclaimed, nr_scanned;
2665 struct lruvec *target_lruvec;
2666 bool reclaimable = false;
2669 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2672 memset(&sc->nr, 0, sizeof(sc->nr));
2674 nr_reclaimed = sc->nr_reclaimed;
2675 nr_scanned = sc->nr_scanned;
2678 * Determine the scan balance between anon and file LRUs.
2680 spin_lock_irq(&pgdat->lru_lock);
2681 sc->anon_cost = target_lruvec->anon_cost;
2682 sc->file_cost = target_lruvec->file_cost;
2683 spin_unlock_irq(&pgdat->lru_lock);
2686 * Target desirable inactive:active list ratios for the anon
2687 * and file LRU lists.
2689 if (!sc->force_deactivate) {
2690 unsigned long refaults;
2692 if (inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2693 sc->may_deactivate |= DEACTIVATE_ANON;
2695 sc->may_deactivate &= ~DEACTIVATE_ANON;
2698 * When refaults are being observed, it means a new
2699 * workingset is being established. Deactivate to get
2700 * rid of any stale active pages quickly.
2702 refaults = lruvec_page_state(target_lruvec,
2703 WORKINGSET_ACTIVATE);
2704 if (refaults != target_lruvec->refaults ||
2705 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2706 sc->may_deactivate |= DEACTIVATE_FILE;
2708 sc->may_deactivate &= ~DEACTIVATE_FILE;
2710 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2713 * If we have plenty of inactive file pages that aren't
2714 * thrashing, try to reclaim those first before touching
2717 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2718 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2719 sc->cache_trim_mode = 1;
2721 sc->cache_trim_mode = 0;
2724 * Prevent the reclaimer from falling into the cache trap: as
2725 * cache pages start out inactive, every cache fault will tip
2726 * the scan balance towards the file LRU. And as the file LRU
2727 * shrinks, so does the window for rotation from references.
2728 * This means we have a runaway feedback loop where a tiny
2729 * thrashing file LRU becomes infinitely more attractive than
2730 * anon pages. Try to detect this based on file LRU size.
2732 if (!cgroup_reclaim(sc)) {
2733 unsigned long total_high_wmark = 0;
2734 unsigned long free, anon;
2737 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2738 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2739 node_page_state(pgdat, NR_INACTIVE_FILE);
2741 for (z = 0; z < MAX_NR_ZONES; z++) {
2742 struct zone *zone = &pgdat->node_zones[z];
2743 if (!managed_zone(zone))
2746 total_high_wmark += high_wmark_pages(zone);
2750 * Consider anon: if that's low too, this isn't a
2751 * runaway file reclaim problem, but rather just
2752 * extreme pressure. Reclaim as per usual then.
2754 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2757 file + free <= total_high_wmark &&
2758 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2759 anon >> sc->priority;
2762 shrink_node_memcgs(pgdat, sc);
2764 if (reclaim_state) {
2765 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2766 reclaim_state->reclaimed_slab = 0;
2769 /* Record the subtree's reclaim efficiency */
2770 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2771 sc->nr_scanned - nr_scanned,
2772 sc->nr_reclaimed - nr_reclaimed);
2774 if (sc->nr_reclaimed - nr_reclaimed)
2777 if (current_is_kswapd()) {
2779 * If reclaim is isolating dirty pages under writeback,
2780 * it implies that the long-lived page allocation rate
2781 * is exceeding the page laundering rate. Either the
2782 * global limits are not being effective at throttling
2783 * processes due to the page distribution throughout
2784 * zones or there is heavy usage of a slow backing
2785 * device. The only option is to throttle from reclaim
2786 * context which is not ideal as there is no guarantee
2787 * the dirtying process is throttled in the same way
2788 * balance_dirty_pages() manages.
2790 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2791 * count the number of pages under pages flagged for
2792 * immediate reclaim and stall if any are encountered
2793 * in the nr_immediate check below.
2795 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2796 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2798 /* Allow kswapd to start writing pages during reclaim.*/
2799 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2800 set_bit(PGDAT_DIRTY, &pgdat->flags);
2803 * If kswapd scans pages marked marked for immediate
2804 * reclaim and under writeback (nr_immediate), it
2805 * implies that pages are cycling through the LRU
2806 * faster than they are written so also forcibly stall.
2808 if (sc->nr.immediate)
2809 congestion_wait(BLK_RW_ASYNC, HZ/10);
2813 * Tag a node/memcg as congested if all the dirty pages
2814 * scanned were backed by a congested BDI and
2815 * wait_iff_congested will stall.
2817 * Legacy memcg will stall in page writeback so avoid forcibly
2818 * stalling in wait_iff_congested().
2820 if ((current_is_kswapd() ||
2821 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2822 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2823 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2826 * Stall direct reclaim for IO completions if underlying BDIs
2827 * and node is congested. Allow kswapd to continue until it
2828 * starts encountering unqueued dirty pages or cycling through
2829 * the LRU too quickly.
2831 if (!current_is_kswapd() && current_may_throttle() &&
2832 !sc->hibernation_mode &&
2833 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2834 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2836 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2841 * Kswapd gives up on balancing particular nodes after too
2842 * many failures to reclaim anything from them and goes to
2843 * sleep. On reclaim progress, reset the failure counter. A
2844 * successful direct reclaim run will revive a dormant kswapd.
2847 pgdat->kswapd_failures = 0;
2851 * Returns true if compaction should go ahead for a costly-order request, or
2852 * the allocation would already succeed without compaction. Return false if we
2853 * should reclaim first.
2855 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2857 unsigned long watermark;
2858 enum compact_result suitable;
2860 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2861 if (suitable == COMPACT_SUCCESS)
2862 /* Allocation should succeed already. Don't reclaim. */
2864 if (suitable == COMPACT_SKIPPED)
2865 /* Compaction cannot yet proceed. Do reclaim. */
2869 * Compaction is already possible, but it takes time to run and there
2870 * are potentially other callers using the pages just freed. So proceed
2871 * with reclaim to make a buffer of free pages available to give
2872 * compaction a reasonable chance of completing and allocating the page.
2873 * Note that we won't actually reclaim the whole buffer in one attempt
2874 * as the target watermark in should_continue_reclaim() is lower. But if
2875 * we are already above the high+gap watermark, don't reclaim at all.
2877 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2879 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2883 * This is the direct reclaim path, for page-allocating processes. We only
2884 * try to reclaim pages from zones which will satisfy the caller's allocation
2887 * If a zone is deemed to be full of pinned pages then just give it a light
2888 * scan then give up on it.
2890 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2894 unsigned long nr_soft_reclaimed;
2895 unsigned long nr_soft_scanned;
2897 pg_data_t *last_pgdat = NULL;
2900 * If the number of buffer_heads in the machine exceeds the maximum
2901 * allowed level, force direct reclaim to scan the highmem zone as
2902 * highmem pages could be pinning lowmem pages storing buffer_heads
2904 orig_mask = sc->gfp_mask;
2905 if (buffer_heads_over_limit) {
2906 sc->gfp_mask |= __GFP_HIGHMEM;
2907 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2910 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2911 sc->reclaim_idx, sc->nodemask) {
2913 * Take care memory controller reclaiming has small influence
2916 if (!cgroup_reclaim(sc)) {
2917 if (!cpuset_zone_allowed(zone,
2918 GFP_KERNEL | __GFP_HARDWALL))
2922 * If we already have plenty of memory free for
2923 * compaction in this zone, don't free any more.
2924 * Even though compaction is invoked for any
2925 * non-zero order, only frequent costly order
2926 * reclamation is disruptive enough to become a
2927 * noticeable problem, like transparent huge
2930 if (IS_ENABLED(CONFIG_COMPACTION) &&
2931 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2932 compaction_ready(zone, sc)) {
2933 sc->compaction_ready = true;
2938 * Shrink each node in the zonelist once. If the
2939 * zonelist is ordered by zone (not the default) then a
2940 * node may be shrunk multiple times but in that case
2941 * the user prefers lower zones being preserved.
2943 if (zone->zone_pgdat == last_pgdat)
2947 * This steals pages from memory cgroups over softlimit
2948 * and returns the number of reclaimed pages and
2949 * scanned pages. This works for global memory pressure
2950 * and balancing, not for a memcg's limit.
2952 nr_soft_scanned = 0;
2953 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2954 sc->order, sc->gfp_mask,
2956 sc->nr_reclaimed += nr_soft_reclaimed;
2957 sc->nr_scanned += nr_soft_scanned;
2958 /* need some check for avoid more shrink_zone() */
2961 /* See comment about same check for global reclaim above */
2962 if (zone->zone_pgdat == last_pgdat)
2964 last_pgdat = zone->zone_pgdat;
2965 shrink_node(zone->zone_pgdat, sc);
2969 * Restore to original mask to avoid the impact on the caller if we
2970 * promoted it to __GFP_HIGHMEM.
2972 sc->gfp_mask = orig_mask;
2975 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2977 struct lruvec *target_lruvec;
2978 unsigned long refaults;
2980 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2981 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE);
2982 target_lruvec->refaults = refaults;
2986 * This is the main entry point to direct page reclaim.
2988 * If a full scan of the inactive list fails to free enough memory then we
2989 * are "out of memory" and something needs to be killed.
2991 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2992 * high - the zone may be full of dirty or under-writeback pages, which this
2993 * caller can't do much about. We kick the writeback threads and take explicit
2994 * naps in the hope that some of these pages can be written. But if the
2995 * allocating task holds filesystem locks which prevent writeout this might not
2996 * work, and the allocation attempt will fail.
2998 * returns: 0, if no pages reclaimed
2999 * else, the number of pages reclaimed
3001 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3002 struct scan_control *sc)
3004 int initial_priority = sc->priority;
3005 pg_data_t *last_pgdat;
3009 delayacct_freepages_start();
3011 if (!cgroup_reclaim(sc))
3012 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3015 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3018 shrink_zones(zonelist, sc);
3020 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3023 if (sc->compaction_ready)
3027 * If we're getting trouble reclaiming, start doing
3028 * writepage even in laptop mode.
3030 if (sc->priority < DEF_PRIORITY - 2)
3031 sc->may_writepage = 1;
3032 } while (--sc->priority >= 0);
3035 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3037 if (zone->zone_pgdat == last_pgdat)
3039 last_pgdat = zone->zone_pgdat;
3041 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3043 if (cgroup_reclaim(sc)) {
3044 struct lruvec *lruvec;
3046 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3048 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3052 delayacct_freepages_end();
3054 if (sc->nr_reclaimed)
3055 return sc->nr_reclaimed;
3057 /* Aborted reclaim to try compaction? don't OOM, then */
3058 if (sc->compaction_ready)
3062 * We make inactive:active ratio decisions based on the node's
3063 * composition of memory, but a restrictive reclaim_idx or a
3064 * memory.low cgroup setting can exempt large amounts of
3065 * memory from reclaim. Neither of which are very common, so
3066 * instead of doing costly eligibility calculations of the
3067 * entire cgroup subtree up front, we assume the estimates are
3068 * good, and retry with forcible deactivation if that fails.
3070 if (sc->skipped_deactivate) {
3071 sc->priority = initial_priority;
3072 sc->force_deactivate = 1;
3073 sc->skipped_deactivate = 0;
3077 /* Untapped cgroup reserves? Don't OOM, retry. */
3078 if (sc->memcg_low_skipped) {
3079 sc->priority = initial_priority;
3080 sc->force_deactivate = 0;
3081 sc->memcg_low_reclaim = 1;
3082 sc->memcg_low_skipped = 0;
3089 static bool allow_direct_reclaim(pg_data_t *pgdat)
3092 unsigned long pfmemalloc_reserve = 0;
3093 unsigned long free_pages = 0;
3097 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3100 for (i = 0; i <= ZONE_NORMAL; i++) {
3101 zone = &pgdat->node_zones[i];
3102 if (!managed_zone(zone))
3105 if (!zone_reclaimable_pages(zone))
3108 pfmemalloc_reserve += min_wmark_pages(zone);
3109 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3112 /* If there are no reserves (unexpected config) then do not throttle */
3113 if (!pfmemalloc_reserve)
3116 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3118 /* kswapd must be awake if processes are being throttled */
3119 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3120 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3121 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3123 wake_up_interruptible(&pgdat->kswapd_wait);
3130 * Throttle direct reclaimers if backing storage is backed by the network
3131 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3132 * depleted. kswapd will continue to make progress and wake the processes
3133 * when the low watermark is reached.
3135 * Returns true if a fatal signal was delivered during throttling. If this
3136 * happens, the page allocator should not consider triggering the OOM killer.
3138 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3139 nodemask_t *nodemask)
3143 pg_data_t *pgdat = NULL;
3146 * Kernel threads should not be throttled as they may be indirectly
3147 * responsible for cleaning pages necessary for reclaim to make forward
3148 * progress. kjournald for example may enter direct reclaim while
3149 * committing a transaction where throttling it could forcing other
3150 * processes to block on log_wait_commit().
3152 if (current->flags & PF_KTHREAD)
3156 * If a fatal signal is pending, this process should not throttle.
3157 * It should return quickly so it can exit and free its memory
3159 if (fatal_signal_pending(current))
3163 * Check if the pfmemalloc reserves are ok by finding the first node
3164 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3165 * GFP_KERNEL will be required for allocating network buffers when
3166 * swapping over the network so ZONE_HIGHMEM is unusable.
3168 * Throttling is based on the first usable node and throttled processes
3169 * wait on a queue until kswapd makes progress and wakes them. There
3170 * is an affinity then between processes waking up and where reclaim
3171 * progress has been made assuming the process wakes on the same node.
3172 * More importantly, processes running on remote nodes will not compete
3173 * for remote pfmemalloc reserves and processes on different nodes
3174 * should make reasonable progress.
3176 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3177 gfp_zone(gfp_mask), nodemask) {
3178 if (zone_idx(zone) > ZONE_NORMAL)
3181 /* Throttle based on the first usable node */
3182 pgdat = zone->zone_pgdat;
3183 if (allow_direct_reclaim(pgdat))
3188 /* If no zone was usable by the allocation flags then do not throttle */
3192 /* Account for the throttling */
3193 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3196 * If the caller cannot enter the filesystem, it's possible that it
3197 * is due to the caller holding an FS lock or performing a journal
3198 * transaction in the case of a filesystem like ext[3|4]. In this case,
3199 * it is not safe to block on pfmemalloc_wait as kswapd could be
3200 * blocked waiting on the same lock. Instead, throttle for up to a
3201 * second before continuing.
3203 if (!(gfp_mask & __GFP_FS)) {
3204 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3205 allow_direct_reclaim(pgdat), HZ);
3210 /* Throttle until kswapd wakes the process */
3211 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3212 allow_direct_reclaim(pgdat));
3215 if (fatal_signal_pending(current))
3222 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3223 gfp_t gfp_mask, nodemask_t *nodemask)
3225 unsigned long nr_reclaimed;
3226 struct scan_control sc = {
3227 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3228 .gfp_mask = current_gfp_context(gfp_mask),
3229 .reclaim_idx = gfp_zone(gfp_mask),
3231 .nodemask = nodemask,
3232 .priority = DEF_PRIORITY,
3233 .may_writepage = !laptop_mode,
3239 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3240 * Confirm they are large enough for max values.
3242 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3243 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3244 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3247 * Do not enter reclaim if fatal signal was delivered while throttled.
3248 * 1 is returned so that the page allocator does not OOM kill at this
3251 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3254 set_task_reclaim_state(current, &sc.reclaim_state);
3255 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3257 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3259 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3260 set_task_reclaim_state(current, NULL);
3262 return nr_reclaimed;
3267 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3268 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3269 gfp_t gfp_mask, bool noswap,
3271 unsigned long *nr_scanned)
3273 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3274 struct scan_control sc = {
3275 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3276 .target_mem_cgroup = memcg,
3277 .may_writepage = !laptop_mode,
3279 .reclaim_idx = MAX_NR_ZONES - 1,
3280 .may_swap = !noswap,
3283 WARN_ON_ONCE(!current->reclaim_state);
3285 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3286 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3288 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3292 * NOTE: Although we can get the priority field, using it
3293 * here is not a good idea, since it limits the pages we can scan.
3294 * if we don't reclaim here, the shrink_node from balance_pgdat
3295 * will pick up pages from other mem cgroup's as well. We hack
3296 * the priority and make it zero.
3298 shrink_lruvec(lruvec, &sc);
3300 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3302 *nr_scanned = sc.nr_scanned;
3304 return sc.nr_reclaimed;
3307 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3308 unsigned long nr_pages,
3312 unsigned long nr_reclaimed;
3313 unsigned int noreclaim_flag;
3314 struct scan_control sc = {
3315 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3316 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3317 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3318 .reclaim_idx = MAX_NR_ZONES - 1,
3319 .target_mem_cgroup = memcg,
3320 .priority = DEF_PRIORITY,
3321 .may_writepage = !laptop_mode,
3323 .may_swap = may_swap,
3326 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3327 * equal pressure on all the nodes. This is based on the assumption that
3328 * the reclaim does not bail out early.
3330 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3332 set_task_reclaim_state(current, &sc.reclaim_state);
3333 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3334 noreclaim_flag = memalloc_noreclaim_save();
3336 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3338 memalloc_noreclaim_restore(noreclaim_flag);
3339 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3340 set_task_reclaim_state(current, NULL);
3342 return nr_reclaimed;
3346 static void age_active_anon(struct pglist_data *pgdat,
3347 struct scan_control *sc)
3349 struct mem_cgroup *memcg;
3350 struct lruvec *lruvec;
3352 if (!total_swap_pages)
3355 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3356 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3359 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3361 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3362 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3363 sc, LRU_ACTIVE_ANON);
3364 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3368 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3374 * Check for watermark boosts top-down as the higher zones
3375 * are more likely to be boosted. Both watermarks and boosts
3376 * should not be checked at the time time as reclaim would
3377 * start prematurely when there is no boosting and a lower
3380 for (i = highest_zoneidx; i >= 0; i--) {
3381 zone = pgdat->node_zones + i;
3382 if (!managed_zone(zone))
3385 if (zone->watermark_boost)
3393 * Returns true if there is an eligible zone balanced for the request order
3394 * and highest_zoneidx
3396 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3399 unsigned long mark = -1;
3403 * Check watermarks bottom-up as lower zones are more likely to
3406 for (i = 0; i <= highest_zoneidx; i++) {
3407 zone = pgdat->node_zones + i;
3409 if (!managed_zone(zone))
3412 mark = high_wmark_pages(zone);
3413 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3418 * If a node has no populated zone within highest_zoneidx, it does not
3419 * need balancing by definition. This can happen if a zone-restricted
3420 * allocation tries to wake a remote kswapd.
3428 /* Clear pgdat state for congested, dirty or under writeback. */
3429 static void clear_pgdat_congested(pg_data_t *pgdat)
3431 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3433 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3434 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3435 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3439 * Prepare kswapd for sleeping. This verifies that there are no processes
3440 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3442 * Returns true if kswapd is ready to sleep
3444 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3445 int highest_zoneidx)
3448 * The throttled processes are normally woken up in balance_pgdat() as
3449 * soon as allow_direct_reclaim() is true. But there is a potential
3450 * race between when kswapd checks the watermarks and a process gets
3451 * throttled. There is also a potential race if processes get
3452 * throttled, kswapd wakes, a large process exits thereby balancing the
3453 * zones, which causes kswapd to exit balance_pgdat() before reaching
3454 * the wake up checks. If kswapd is going to sleep, no process should
3455 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3456 * the wake up is premature, processes will wake kswapd and get
3457 * throttled again. The difference from wake ups in balance_pgdat() is
3458 * that here we are under prepare_to_wait().
3460 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3461 wake_up_all(&pgdat->pfmemalloc_wait);
3463 /* Hopeless node, leave it to direct reclaim */
3464 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3467 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3468 clear_pgdat_congested(pgdat);
3476 * kswapd shrinks a node of pages that are at or below the highest usable
3477 * zone that is currently unbalanced.
3479 * Returns true if kswapd scanned at least the requested number of pages to
3480 * reclaim or if the lack of progress was due to pages under writeback.
3481 * This is used to determine if the scanning priority needs to be raised.
3483 static bool kswapd_shrink_node(pg_data_t *pgdat,
3484 struct scan_control *sc)
3489 /* Reclaim a number of pages proportional to the number of zones */
3490 sc->nr_to_reclaim = 0;
3491 for (z = 0; z <= sc->reclaim_idx; z++) {
3492 zone = pgdat->node_zones + z;
3493 if (!managed_zone(zone))
3496 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3500 * Historically care was taken to put equal pressure on all zones but
3501 * now pressure is applied based on node LRU order.
3503 shrink_node(pgdat, sc);
3506 * Fragmentation may mean that the system cannot be rebalanced for
3507 * high-order allocations. If twice the allocation size has been
3508 * reclaimed then recheck watermarks only at order-0 to prevent
3509 * excessive reclaim. Assume that a process requested a high-order
3510 * can direct reclaim/compact.
3512 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3515 return sc->nr_scanned >= sc->nr_to_reclaim;
3519 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3520 * that are eligible for use by the caller until at least one zone is
3523 * Returns the order kswapd finished reclaiming at.
3525 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3526 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3527 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3528 * or lower is eligible for reclaim until at least one usable zone is
3531 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3534 unsigned long nr_soft_reclaimed;
3535 unsigned long nr_soft_scanned;
3536 unsigned long pflags;
3537 unsigned long nr_boost_reclaim;
3538 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3541 struct scan_control sc = {
3542 .gfp_mask = GFP_KERNEL,
3547 set_task_reclaim_state(current, &sc.reclaim_state);
3548 psi_memstall_enter(&pflags);
3549 __fs_reclaim_acquire();
3551 count_vm_event(PAGEOUTRUN);
3554 * Account for the reclaim boost. Note that the zone boost is left in
3555 * place so that parallel allocations that are near the watermark will
3556 * stall or direct reclaim until kswapd is finished.
3558 nr_boost_reclaim = 0;
3559 for (i = 0; i <= highest_zoneidx; i++) {
3560 zone = pgdat->node_zones + i;
3561 if (!managed_zone(zone))
3564 nr_boost_reclaim += zone->watermark_boost;
3565 zone_boosts[i] = zone->watermark_boost;
3567 boosted = nr_boost_reclaim;
3570 sc.priority = DEF_PRIORITY;
3572 unsigned long nr_reclaimed = sc.nr_reclaimed;
3573 bool raise_priority = true;
3577 sc.reclaim_idx = highest_zoneidx;
3580 * If the number of buffer_heads exceeds the maximum allowed
3581 * then consider reclaiming from all zones. This has a dual
3582 * purpose -- on 64-bit systems it is expected that
3583 * buffer_heads are stripped during active rotation. On 32-bit
3584 * systems, highmem pages can pin lowmem memory and shrinking
3585 * buffers can relieve lowmem pressure. Reclaim may still not
3586 * go ahead if all eligible zones for the original allocation
3587 * request are balanced to avoid excessive reclaim from kswapd.
3589 if (buffer_heads_over_limit) {
3590 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3591 zone = pgdat->node_zones + i;
3592 if (!managed_zone(zone))
3601 * If the pgdat is imbalanced then ignore boosting and preserve
3602 * the watermarks for a later time and restart. Note that the
3603 * zone watermarks will be still reset at the end of balancing
3604 * on the grounds that the normal reclaim should be enough to
3605 * re-evaluate if boosting is required when kswapd next wakes.
3607 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3608 if (!balanced && nr_boost_reclaim) {
3609 nr_boost_reclaim = 0;
3614 * If boosting is not active then only reclaim if there are no
3615 * eligible zones. Note that sc.reclaim_idx is not used as
3616 * buffer_heads_over_limit may have adjusted it.
3618 if (!nr_boost_reclaim && balanced)
3621 /* Limit the priority of boosting to avoid reclaim writeback */
3622 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3623 raise_priority = false;
3626 * Do not writeback or swap pages for boosted reclaim. The
3627 * intent is to relieve pressure not issue sub-optimal IO
3628 * from reclaim context. If no pages are reclaimed, the
3629 * reclaim will be aborted.
3631 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3632 sc.may_swap = !nr_boost_reclaim;
3635 * Do some background aging of the anon list, to give
3636 * pages a chance to be referenced before reclaiming. All
3637 * pages are rotated regardless of classzone as this is
3638 * about consistent aging.
3640 age_active_anon(pgdat, &sc);
3643 * If we're getting trouble reclaiming, start doing writepage
3644 * even in laptop mode.
3646 if (sc.priority < DEF_PRIORITY - 2)
3647 sc.may_writepage = 1;
3649 /* Call soft limit reclaim before calling shrink_node. */
3651 nr_soft_scanned = 0;
3652 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3653 sc.gfp_mask, &nr_soft_scanned);
3654 sc.nr_reclaimed += nr_soft_reclaimed;
3657 * There should be no need to raise the scanning priority if
3658 * enough pages are already being scanned that that high
3659 * watermark would be met at 100% efficiency.
3661 if (kswapd_shrink_node(pgdat, &sc))
3662 raise_priority = false;
3665 * If the low watermark is met there is no need for processes
3666 * to be throttled on pfmemalloc_wait as they should not be
3667 * able to safely make forward progress. Wake them
3669 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3670 allow_direct_reclaim(pgdat))
3671 wake_up_all(&pgdat->pfmemalloc_wait);
3673 /* Check if kswapd should be suspending */
3674 __fs_reclaim_release();
3675 ret = try_to_freeze();
3676 __fs_reclaim_acquire();
3677 if (ret || kthread_should_stop())
3681 * Raise priority if scanning rate is too low or there was no
3682 * progress in reclaiming pages
3684 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3685 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3688 * If reclaim made no progress for a boost, stop reclaim as
3689 * IO cannot be queued and it could be an infinite loop in
3690 * extreme circumstances.
3692 if (nr_boost_reclaim && !nr_reclaimed)
3695 if (raise_priority || !nr_reclaimed)
3697 } while (sc.priority >= 1);
3699 if (!sc.nr_reclaimed)
3700 pgdat->kswapd_failures++;
3703 /* If reclaim was boosted, account for the reclaim done in this pass */
3705 unsigned long flags;
3707 for (i = 0; i <= highest_zoneidx; i++) {
3708 if (!zone_boosts[i])
3711 /* Increments are under the zone lock */
3712 zone = pgdat->node_zones + i;
3713 spin_lock_irqsave(&zone->lock, flags);
3714 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3715 spin_unlock_irqrestore(&zone->lock, flags);
3719 * As there is now likely space, wakeup kcompact to defragment
3722 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3725 snapshot_refaults(NULL, pgdat);
3726 __fs_reclaim_release();
3727 psi_memstall_leave(&pflags);
3728 set_task_reclaim_state(current, NULL);
3731 * Return the order kswapd stopped reclaiming at as
3732 * prepare_kswapd_sleep() takes it into account. If another caller
3733 * entered the allocator slow path while kswapd was awake, order will
3734 * remain at the higher level.
3740 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3741 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3742 * not a valid index then either kswapd runs for first time or kswapd couldn't
3743 * sleep after previous reclaim attempt (node is still unbalanced). In that
3744 * case return the zone index of the previous kswapd reclaim cycle.
3746 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3747 enum zone_type prev_highest_zoneidx)
3749 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3751 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3754 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3755 unsigned int highest_zoneidx)
3760 if (freezing(current) || kthread_should_stop())
3763 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3766 * Try to sleep for a short interval. Note that kcompactd will only be
3767 * woken if it is possible to sleep for a short interval. This is
3768 * deliberate on the assumption that if reclaim cannot keep an
3769 * eligible zone balanced that it's also unlikely that compaction will
3772 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3774 * Compaction records what page blocks it recently failed to
3775 * isolate pages from and skips them in the future scanning.
3776 * When kswapd is going to sleep, it is reasonable to assume
3777 * that pages and compaction may succeed so reset the cache.
3779 reset_isolation_suitable(pgdat);
3782 * We have freed the memory, now we should compact it to make
3783 * allocation of the requested order possible.
3785 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3787 remaining = schedule_timeout(HZ/10);
3790 * If woken prematurely then reset kswapd_highest_zoneidx and
3791 * order. The values will either be from a wakeup request or
3792 * the previous request that slept prematurely.
3795 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3796 kswapd_highest_zoneidx(pgdat,
3799 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3800 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3803 finish_wait(&pgdat->kswapd_wait, &wait);
3804 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3808 * After a short sleep, check if it was a premature sleep. If not, then
3809 * go fully to sleep until explicitly woken up.
3812 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3813 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3816 * vmstat counters are not perfectly accurate and the estimated
3817 * value for counters such as NR_FREE_PAGES can deviate from the
3818 * true value by nr_online_cpus * threshold. To avoid the zone
3819 * watermarks being breached while under pressure, we reduce the
3820 * per-cpu vmstat threshold while kswapd is awake and restore
3821 * them before going back to sleep.
3823 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3825 if (!kthread_should_stop())
3828 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3831 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3833 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3835 finish_wait(&pgdat->kswapd_wait, &wait);
3839 * The background pageout daemon, started as a kernel thread
3840 * from the init process.
3842 * This basically trickles out pages so that we have _some_
3843 * free memory available even if there is no other activity
3844 * that frees anything up. This is needed for things like routing
3845 * etc, where we otherwise might have all activity going on in
3846 * asynchronous contexts that cannot page things out.
3848 * If there are applications that are active memory-allocators
3849 * (most normal use), this basically shouldn't matter.
3851 static int kswapd(void *p)
3853 unsigned int alloc_order, reclaim_order;
3854 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3855 pg_data_t *pgdat = (pg_data_t*)p;
3856 struct task_struct *tsk = current;
3857 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3859 if (!cpumask_empty(cpumask))
3860 set_cpus_allowed_ptr(tsk, cpumask);
3863 * Tell the memory management that we're a "memory allocator",
3864 * and that if we need more memory we should get access to it
3865 * regardless (see "__alloc_pages()"). "kswapd" should
3866 * never get caught in the normal page freeing logic.
3868 * (Kswapd normally doesn't need memory anyway, but sometimes
3869 * you need a small amount of memory in order to be able to
3870 * page out something else, and this flag essentially protects
3871 * us from recursively trying to free more memory as we're
3872 * trying to free the first piece of memory in the first place).
3874 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3877 WRITE_ONCE(pgdat->kswapd_order, 0);
3878 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3882 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3883 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3887 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3890 /* Read the new order and highest_zoneidx */
3891 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3892 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3894 WRITE_ONCE(pgdat->kswapd_order, 0);
3895 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3897 ret = try_to_freeze();
3898 if (kthread_should_stop())
3902 * We can speed up thawing tasks if we don't call balance_pgdat
3903 * after returning from the refrigerator
3909 * Reclaim begins at the requested order but if a high-order
3910 * reclaim fails then kswapd falls back to reclaiming for
3911 * order-0. If that happens, kswapd will consider sleeping
3912 * for the order it finished reclaiming at (reclaim_order)
3913 * but kcompactd is woken to compact for the original
3914 * request (alloc_order).
3916 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3918 reclaim_order = balance_pgdat(pgdat, alloc_order,
3920 if (reclaim_order < alloc_order)
3921 goto kswapd_try_sleep;
3924 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3930 * A zone is low on free memory or too fragmented for high-order memory. If
3931 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3932 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3933 * has failed or is not needed, still wake up kcompactd if only compaction is
3936 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3937 enum zone_type highest_zoneidx)
3940 enum zone_type curr_idx;
3942 if (!managed_zone(zone))
3945 if (!cpuset_zone_allowed(zone, gfp_flags))
3948 pgdat = zone->zone_pgdat;
3949 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3951 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3952 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3954 if (READ_ONCE(pgdat->kswapd_order) < order)
3955 WRITE_ONCE(pgdat->kswapd_order, order);
3957 if (!waitqueue_active(&pgdat->kswapd_wait))
3960 /* Hopeless node, leave it to direct reclaim if possible */
3961 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3962 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3963 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3965 * There may be plenty of free memory available, but it's too
3966 * fragmented for high-order allocations. Wake up kcompactd
3967 * and rely on compaction_suitable() to determine if it's
3968 * needed. If it fails, it will defer subsequent attempts to
3969 * ratelimit its work.
3971 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3972 wakeup_kcompactd(pgdat, order, highest_zoneidx);
3976 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3978 wake_up_interruptible(&pgdat->kswapd_wait);
3981 #ifdef CONFIG_HIBERNATION
3983 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3986 * Rather than trying to age LRUs the aim is to preserve the overall
3987 * LRU order by reclaiming preferentially
3988 * inactive > active > active referenced > active mapped
3990 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3992 struct scan_control sc = {
3993 .nr_to_reclaim = nr_to_reclaim,
3994 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3995 .reclaim_idx = MAX_NR_ZONES - 1,
3996 .priority = DEF_PRIORITY,
4000 .hibernation_mode = 1,
4002 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4003 unsigned long nr_reclaimed;
4004 unsigned int noreclaim_flag;
4006 fs_reclaim_acquire(sc.gfp_mask);
4007 noreclaim_flag = memalloc_noreclaim_save();
4008 set_task_reclaim_state(current, &sc.reclaim_state);
4010 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4012 set_task_reclaim_state(current, NULL);
4013 memalloc_noreclaim_restore(noreclaim_flag);
4014 fs_reclaim_release(sc.gfp_mask);
4016 return nr_reclaimed;
4018 #endif /* CONFIG_HIBERNATION */
4021 * This kswapd start function will be called by init and node-hot-add.
4022 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4024 int kswapd_run(int nid)
4026 pg_data_t *pgdat = NODE_DATA(nid);
4032 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4033 if (IS_ERR(pgdat->kswapd)) {
4034 /* failure at boot is fatal */
4035 BUG_ON(system_state < SYSTEM_RUNNING);
4036 pr_err("Failed to start kswapd on node %d\n", nid);
4037 ret = PTR_ERR(pgdat->kswapd);
4038 pgdat->kswapd = NULL;
4044 * Called by memory hotplug when all memory in a node is offlined. Caller must
4045 * hold mem_hotplug_begin/end().
4047 void kswapd_stop(int nid)
4049 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4052 kthread_stop(kswapd);
4053 NODE_DATA(nid)->kswapd = NULL;
4057 static int __init kswapd_init(void)
4062 for_each_node_state(nid, N_MEMORY)
4067 module_init(kswapd_init)
4073 * If non-zero call node_reclaim when the number of free pages falls below
4076 int node_reclaim_mode __read_mostly;
4078 #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4079 #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4082 * Priority for NODE_RECLAIM. This determines the fraction of pages
4083 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4086 #define NODE_RECLAIM_PRIORITY 4
4089 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4092 int sysctl_min_unmapped_ratio = 1;
4095 * If the number of slab pages in a zone grows beyond this percentage then
4096 * slab reclaim needs to occur.
4098 int sysctl_min_slab_ratio = 5;
4100 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4102 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4103 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4104 node_page_state(pgdat, NR_ACTIVE_FILE);
4107 * It's possible for there to be more file mapped pages than
4108 * accounted for by the pages on the file LRU lists because
4109 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4111 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4114 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4115 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4117 unsigned long nr_pagecache_reclaimable;
4118 unsigned long delta = 0;
4121 * If RECLAIM_UNMAP is set, then all file pages are considered
4122 * potentially reclaimable. Otherwise, we have to worry about
4123 * pages like swapcache and node_unmapped_file_pages() provides
4126 if (node_reclaim_mode & RECLAIM_UNMAP)
4127 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4129 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4131 /* If we can't clean pages, remove dirty pages from consideration */
4132 if (!(node_reclaim_mode & RECLAIM_WRITE))
4133 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4135 /* Watch for any possible underflows due to delta */
4136 if (unlikely(delta > nr_pagecache_reclaimable))
4137 delta = nr_pagecache_reclaimable;
4139 return nr_pagecache_reclaimable - delta;
4143 * Try to free up some pages from this node through reclaim.
4145 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4147 /* Minimum pages needed in order to stay on node */
4148 const unsigned long nr_pages = 1 << order;
4149 struct task_struct *p = current;
4150 unsigned int noreclaim_flag;
4151 struct scan_control sc = {
4152 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4153 .gfp_mask = current_gfp_context(gfp_mask),
4155 .priority = NODE_RECLAIM_PRIORITY,
4156 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4157 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4159 .reclaim_idx = gfp_zone(gfp_mask),
4162 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4166 fs_reclaim_acquire(sc.gfp_mask);
4168 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4169 * and we also need to be able to write out pages for RECLAIM_WRITE
4170 * and RECLAIM_UNMAP.
4172 noreclaim_flag = memalloc_noreclaim_save();
4173 p->flags |= PF_SWAPWRITE;
4174 set_task_reclaim_state(p, &sc.reclaim_state);
4176 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4178 * Free memory by calling shrink node with increasing
4179 * priorities until we have enough memory freed.
4182 shrink_node(pgdat, &sc);
4183 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4186 set_task_reclaim_state(p, NULL);
4187 current->flags &= ~PF_SWAPWRITE;
4188 memalloc_noreclaim_restore(noreclaim_flag);
4189 fs_reclaim_release(sc.gfp_mask);
4191 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4193 return sc.nr_reclaimed >= nr_pages;
4196 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4201 * Node reclaim reclaims unmapped file backed pages and
4202 * slab pages if we are over the defined limits.
4204 * A small portion of unmapped file backed pages is needed for
4205 * file I/O otherwise pages read by file I/O will be immediately
4206 * thrown out if the node is overallocated. So we do not reclaim
4207 * if less than a specified percentage of the node is used by
4208 * unmapped file backed pages.
4210 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4211 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4212 pgdat->min_slab_pages)
4213 return NODE_RECLAIM_FULL;
4216 * Do not scan if the allocation should not be delayed.
4218 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4219 return NODE_RECLAIM_NOSCAN;
4222 * Only run node reclaim on the local node or on nodes that do not
4223 * have associated processors. This will favor the local processor
4224 * over remote processors and spread off node memory allocations
4225 * as wide as possible.
4227 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4228 return NODE_RECLAIM_NOSCAN;
4230 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4231 return NODE_RECLAIM_NOSCAN;
4233 ret = __node_reclaim(pgdat, gfp_mask, order);
4234 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4237 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4244 * check_move_unevictable_pages - check pages for evictability and move to
4245 * appropriate zone lru list
4246 * @pvec: pagevec with lru pages to check
4248 * Checks pages for evictability, if an evictable page is in the unevictable
4249 * lru list, moves it to the appropriate evictable lru list. This function
4250 * should be only used for lru pages.
4252 void check_move_unevictable_pages(struct pagevec *pvec)
4254 struct lruvec *lruvec;
4255 struct pglist_data *pgdat = NULL;
4260 for (i = 0; i < pvec->nr; i++) {
4261 struct page *page = pvec->pages[i];
4262 struct pglist_data *pagepgdat = page_pgdat(page);
4265 if (pagepgdat != pgdat) {
4267 spin_unlock_irq(&pgdat->lru_lock);
4269 spin_lock_irq(&pgdat->lru_lock);
4271 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4273 if (!PageLRU(page) || !PageUnevictable(page))
4276 if (page_evictable(page)) {
4277 enum lru_list lru = page_lru_base_type(page);
4279 VM_BUG_ON_PAGE(PageActive(page), page);
4280 ClearPageUnevictable(page);
4281 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4282 add_page_to_lru_list(page, lruvec, lru);
4288 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4289 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4290 spin_unlock_irq(&pgdat->lru_lock);
4293 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);