1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup *target_mem_cgroup;
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
92 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
96 unsigned int memcg_low_reclaim:1;
97 unsigned int memcg_low_skipped:1;
99 unsigned int hibernation_mode:1;
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready:1;
104 /* Allocation order */
107 /* Scan (total_size >> priority) pages at once */
110 /* The highest zone to isolate pages for reclaim from */
113 /* This context's GFP mask */
116 /* Incremented by the number of inactive pages that were scanned */
117 unsigned long nr_scanned;
119 /* Number of pages freed so far during a call to shrink_zones() */
120 unsigned long nr_reclaimed;
124 unsigned int unqueued_dirty;
125 unsigned int congested;
126 unsigned int writeback;
127 unsigned int immediate;
128 unsigned int file_taken;
132 /* for recording the reclaimed slab by now */
133 struct reclaim_state reclaim_state;
136 #ifdef ARCH_HAS_PREFETCH
137 #define prefetch_prev_lru_page(_page, _base, _field) \
139 if ((_page)->lru.prev != _base) { \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
147 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
150 #ifdef ARCH_HAS_PREFETCHW
151 #define prefetchw_prev_lru_page(_page, _base, _field) \
153 if ((_page)->lru.prev != _base) { \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
161 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
165 * From 0 .. 100. Higher means more swappy.
167 int vm_swappiness = 60;
169 * The total number of pages which are beyond the high watermark within all
172 unsigned long vm_total_pages;
174 static LIST_HEAD(shrinker_list);
175 static DECLARE_RWSEM(shrinker_rwsem);
177 #ifdef CONFIG_MEMCG_KMEM
180 * We allow subsystems to populate their shrinker-related
181 * LRU lists before register_shrinker_prepared() is called
182 * for the shrinker, since we don't want to impose
183 * restrictions on their internal registration order.
184 * In this case shrink_slab_memcg() may find corresponding
185 * bit is set in the shrinkers map.
187 * This value is used by the function to detect registering
188 * shrinkers and to skip do_shrink_slab() calls for them.
190 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
192 static DEFINE_IDR(shrinker_idr);
193 static int shrinker_nr_max;
195 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
197 int id, ret = -ENOMEM;
199 down_write(&shrinker_rwsem);
200 /* This may call shrinker, so it must use down_read_trylock() */
201 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
205 if (id >= shrinker_nr_max) {
206 if (memcg_expand_shrinker_maps(id)) {
207 idr_remove(&shrinker_idr, id);
211 shrinker_nr_max = id + 1;
216 up_write(&shrinker_rwsem);
220 static void unregister_memcg_shrinker(struct shrinker *shrinker)
222 int id = shrinker->id;
226 down_write(&shrinker_rwsem);
227 idr_remove(&shrinker_idr, id);
228 up_write(&shrinker_rwsem);
230 #else /* CONFIG_MEMCG_KMEM */
231 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
236 static void unregister_memcg_shrinker(struct shrinker *shrinker)
239 #endif /* CONFIG_MEMCG_KMEM */
241 static void set_task_reclaim_state(struct task_struct *task,
242 struct reclaim_state *rs)
244 /* Check for an overwrite */
245 WARN_ON_ONCE(rs && task->reclaim_state);
247 /* Check for the nulling of an already-nulled member */
248 WARN_ON_ONCE(!rs && !task->reclaim_state);
250 task->reclaim_state = rs;
254 static bool global_reclaim(struct scan_control *sc)
256 return !sc->target_mem_cgroup;
260 * sane_reclaim - is the usual dirty throttling mechanism operational?
261 * @sc: scan_control in question
263 * The normal page dirty throttling mechanism in balance_dirty_pages() is
264 * completely broken with the legacy memcg and direct stalling in
265 * shrink_page_list() is used for throttling instead, which lacks all the
266 * niceties such as fairness, adaptive pausing, bandwidth proportional
267 * allocation and configurability.
269 * This function tests whether the vmscan currently in progress can assume
270 * that the normal dirty throttling mechanism is operational.
272 static bool sane_reclaim(struct scan_control *sc)
274 struct mem_cgroup *memcg = sc->target_mem_cgroup;
278 #ifdef CONFIG_CGROUP_WRITEBACK
279 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
285 static void set_memcg_congestion(pg_data_t *pgdat,
286 struct mem_cgroup *memcg,
289 struct mem_cgroup_per_node *mn;
294 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
295 WRITE_ONCE(mn->congested, congested);
298 static bool memcg_congested(pg_data_t *pgdat,
299 struct mem_cgroup *memcg)
301 struct mem_cgroup_per_node *mn;
303 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
304 return READ_ONCE(mn->congested);
308 static bool global_reclaim(struct scan_control *sc)
313 static bool sane_reclaim(struct scan_control *sc)
318 static inline void set_memcg_congestion(struct pglist_data *pgdat,
319 struct mem_cgroup *memcg, bool congested)
323 static inline bool memcg_congested(struct pglist_data *pgdat,
324 struct mem_cgroup *memcg)
332 * This misses isolated pages which are not accounted for to save counters.
333 * As the data only determines if reclaim or compaction continues, it is
334 * not expected that isolated pages will be a dominating factor.
336 unsigned long zone_reclaimable_pages(struct zone *zone)
340 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
341 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
342 if (get_nr_swap_pages() > 0)
343 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
344 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
350 * lruvec_lru_size - Returns the number of pages on the given LRU list.
351 * @lruvec: lru vector
353 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
355 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
357 unsigned long lru_size;
360 if (!mem_cgroup_disabled())
361 lru_size = lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
363 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
365 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
366 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
369 if (!managed_zone(zone))
372 if (!mem_cgroup_disabled())
373 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
375 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
376 NR_ZONE_LRU_BASE + lru);
377 lru_size -= min(size, lru_size);
385 * Add a shrinker callback to be called from the vm.
387 int prealloc_shrinker(struct shrinker *shrinker)
389 unsigned int size = sizeof(*shrinker->nr_deferred);
391 if (shrinker->flags & SHRINKER_NUMA_AWARE)
394 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
395 if (!shrinker->nr_deferred)
398 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
399 if (prealloc_memcg_shrinker(shrinker))
406 kfree(shrinker->nr_deferred);
407 shrinker->nr_deferred = NULL;
411 void free_prealloced_shrinker(struct shrinker *shrinker)
413 if (!shrinker->nr_deferred)
416 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
417 unregister_memcg_shrinker(shrinker);
419 kfree(shrinker->nr_deferred);
420 shrinker->nr_deferred = NULL;
423 void register_shrinker_prepared(struct shrinker *shrinker)
425 down_write(&shrinker_rwsem);
426 list_add_tail(&shrinker->list, &shrinker_list);
427 #ifdef CONFIG_MEMCG_KMEM
428 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
429 idr_replace(&shrinker_idr, shrinker, shrinker->id);
431 up_write(&shrinker_rwsem);
434 int register_shrinker(struct shrinker *shrinker)
436 int err = prealloc_shrinker(shrinker);
440 register_shrinker_prepared(shrinker);
443 EXPORT_SYMBOL(register_shrinker);
448 void unregister_shrinker(struct shrinker *shrinker)
450 if (!shrinker->nr_deferred)
452 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
453 unregister_memcg_shrinker(shrinker);
454 down_write(&shrinker_rwsem);
455 list_del(&shrinker->list);
456 up_write(&shrinker_rwsem);
457 kfree(shrinker->nr_deferred);
458 shrinker->nr_deferred = NULL;
460 EXPORT_SYMBOL(unregister_shrinker);
462 #define SHRINK_BATCH 128
464 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
465 struct shrinker *shrinker, int priority)
467 unsigned long freed = 0;
468 unsigned long long delta;
473 int nid = shrinkctl->nid;
474 long batch_size = shrinker->batch ? shrinker->batch
476 long scanned = 0, next_deferred;
478 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
481 freeable = shrinker->count_objects(shrinker, shrinkctl);
482 if (freeable == 0 || freeable == SHRINK_EMPTY)
486 * copy the current shrinker scan count into a local variable
487 * and zero it so that other concurrent shrinker invocations
488 * don't also do this scanning work.
490 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
493 if (shrinker->seeks) {
494 delta = freeable >> priority;
496 do_div(delta, shrinker->seeks);
499 * These objects don't require any IO to create. Trim
500 * them aggressively under memory pressure to keep
501 * them from causing refetches in the IO caches.
503 delta = freeable / 2;
507 if (total_scan < 0) {
508 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
509 shrinker->scan_objects, total_scan);
510 total_scan = freeable;
513 next_deferred = total_scan;
516 * We need to avoid excessive windup on filesystem shrinkers
517 * due to large numbers of GFP_NOFS allocations causing the
518 * shrinkers to return -1 all the time. This results in a large
519 * nr being built up so when a shrink that can do some work
520 * comes along it empties the entire cache due to nr >>>
521 * freeable. This is bad for sustaining a working set in
524 * Hence only allow the shrinker to scan the entire cache when
525 * a large delta change is calculated directly.
527 if (delta < freeable / 4)
528 total_scan = min(total_scan, freeable / 2);
531 * Avoid risking looping forever due to too large nr value:
532 * never try to free more than twice the estimate number of
535 if (total_scan > freeable * 2)
536 total_scan = freeable * 2;
538 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
539 freeable, delta, total_scan, priority);
542 * Normally, we should not scan less than batch_size objects in one
543 * pass to avoid too frequent shrinker calls, but if the slab has less
544 * than batch_size objects in total and we are really tight on memory,
545 * we will try to reclaim all available objects, otherwise we can end
546 * up failing allocations although there are plenty of reclaimable
547 * objects spread over several slabs with usage less than the
550 * We detect the "tight on memory" situations by looking at the total
551 * number of objects we want to scan (total_scan). If it is greater
552 * than the total number of objects on slab (freeable), we must be
553 * scanning at high prio and therefore should try to reclaim as much as
556 while (total_scan >= batch_size ||
557 total_scan >= freeable) {
559 unsigned long nr_to_scan = min(batch_size, total_scan);
561 shrinkctl->nr_to_scan = nr_to_scan;
562 shrinkctl->nr_scanned = nr_to_scan;
563 ret = shrinker->scan_objects(shrinker, shrinkctl);
564 if (ret == SHRINK_STOP)
568 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
569 total_scan -= shrinkctl->nr_scanned;
570 scanned += shrinkctl->nr_scanned;
575 if (next_deferred >= scanned)
576 next_deferred -= scanned;
580 * move the unused scan count back into the shrinker in a
581 * manner that handles concurrent updates. If we exhausted the
582 * scan, there is no need to do an update.
584 if (next_deferred > 0)
585 new_nr = atomic_long_add_return(next_deferred,
586 &shrinker->nr_deferred[nid]);
588 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
590 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
594 #ifdef CONFIG_MEMCG_KMEM
595 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
596 struct mem_cgroup *memcg, int priority)
598 struct memcg_shrinker_map *map;
599 unsigned long ret, freed = 0;
602 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
605 if (!down_read_trylock(&shrinker_rwsem))
608 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
613 for_each_set_bit(i, map->map, shrinker_nr_max) {
614 struct shrink_control sc = {
615 .gfp_mask = gfp_mask,
619 struct shrinker *shrinker;
621 shrinker = idr_find(&shrinker_idr, i);
622 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
624 clear_bit(i, map->map);
628 ret = do_shrink_slab(&sc, shrinker, priority);
629 if (ret == SHRINK_EMPTY) {
630 clear_bit(i, map->map);
632 * After the shrinker reported that it had no objects to
633 * free, but before we cleared the corresponding bit in
634 * the memcg shrinker map, a new object might have been
635 * added. To make sure, we have the bit set in this
636 * case, we invoke the shrinker one more time and reset
637 * the bit if it reports that it is not empty anymore.
638 * The memory barrier here pairs with the barrier in
639 * memcg_set_shrinker_bit():
641 * list_lru_add() shrink_slab_memcg()
642 * list_add_tail() clear_bit()
644 * set_bit() do_shrink_slab()
646 smp_mb__after_atomic();
647 ret = do_shrink_slab(&sc, shrinker, priority);
648 if (ret == SHRINK_EMPTY)
651 memcg_set_shrinker_bit(memcg, nid, i);
655 if (rwsem_is_contended(&shrinker_rwsem)) {
661 up_read(&shrinker_rwsem);
664 #else /* CONFIG_MEMCG_KMEM */
665 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
666 struct mem_cgroup *memcg, int priority)
670 #endif /* CONFIG_MEMCG_KMEM */
673 * shrink_slab - shrink slab caches
674 * @gfp_mask: allocation context
675 * @nid: node whose slab caches to target
676 * @memcg: memory cgroup whose slab caches to target
677 * @priority: the reclaim priority
679 * Call the shrink functions to age shrinkable caches.
681 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
682 * unaware shrinkers will receive a node id of 0 instead.
684 * @memcg specifies the memory cgroup to target. Unaware shrinkers
685 * are called only if it is the root cgroup.
687 * @priority is sc->priority, we take the number of objects and >> by priority
688 * in order to get the scan target.
690 * Returns the number of reclaimed slab objects.
692 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
693 struct mem_cgroup *memcg,
696 unsigned long ret, freed = 0;
697 struct shrinker *shrinker;
700 * The root memcg might be allocated even though memcg is disabled
701 * via "cgroup_disable=memory" boot parameter. This could make
702 * mem_cgroup_is_root() return false, then just run memcg slab
703 * shrink, but skip global shrink. This may result in premature
706 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
707 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
709 if (!down_read_trylock(&shrinker_rwsem))
712 list_for_each_entry(shrinker, &shrinker_list, list) {
713 struct shrink_control sc = {
714 .gfp_mask = gfp_mask,
719 ret = do_shrink_slab(&sc, shrinker, priority);
720 if (ret == SHRINK_EMPTY)
724 * Bail out if someone want to register a new shrinker to
725 * prevent the regsitration from being stalled for long periods
726 * by parallel ongoing shrinking.
728 if (rwsem_is_contended(&shrinker_rwsem)) {
734 up_read(&shrinker_rwsem);
740 void drop_slab_node(int nid)
745 struct mem_cgroup *memcg = NULL;
748 memcg = mem_cgroup_iter(NULL, NULL, NULL);
750 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
751 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
752 } while (freed > 10);
759 for_each_online_node(nid)
763 static inline int is_page_cache_freeable(struct page *page)
766 * A freeable page cache page is referenced only by the caller
767 * that isolated the page, the page cache and optional buffer
768 * heads at page->private.
770 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
772 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
775 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
777 if (current->flags & PF_SWAPWRITE)
779 if (!inode_write_congested(inode))
781 if (inode_to_bdi(inode) == current->backing_dev_info)
787 * We detected a synchronous write error writing a page out. Probably
788 * -ENOSPC. We need to propagate that into the address_space for a subsequent
789 * fsync(), msync() or close().
791 * The tricky part is that after writepage we cannot touch the mapping: nothing
792 * prevents it from being freed up. But we have a ref on the page and once
793 * that page is locked, the mapping is pinned.
795 * We're allowed to run sleeping lock_page() here because we know the caller has
798 static void handle_write_error(struct address_space *mapping,
799 struct page *page, int error)
802 if (page_mapping(page) == mapping)
803 mapping_set_error(mapping, error);
807 /* possible outcome of pageout() */
809 /* failed to write page out, page is locked */
811 /* move page to the active list, page is locked */
813 /* page has been sent to the disk successfully, page is unlocked */
815 /* page is clean and locked */
820 * pageout is called by shrink_page_list() for each dirty page.
821 * Calls ->writepage().
823 static pageout_t pageout(struct page *page, struct address_space *mapping,
824 struct scan_control *sc)
827 * If the page is dirty, only perform writeback if that write
828 * will be non-blocking. To prevent this allocation from being
829 * stalled by pagecache activity. But note that there may be
830 * stalls if we need to run get_block(). We could test
831 * PagePrivate for that.
833 * If this process is currently in __generic_file_write_iter() against
834 * this page's queue, we can perform writeback even if that
837 * If the page is swapcache, write it back even if that would
838 * block, for some throttling. This happens by accident, because
839 * swap_backing_dev_info is bust: it doesn't reflect the
840 * congestion state of the swapdevs. Easy to fix, if needed.
842 if (!is_page_cache_freeable(page))
846 * Some data journaling orphaned pages can have
847 * page->mapping == NULL while being dirty with clean buffers.
849 if (page_has_private(page)) {
850 if (try_to_free_buffers(page)) {
851 ClearPageDirty(page);
852 pr_info("%s: orphaned page\n", __func__);
858 if (mapping->a_ops->writepage == NULL)
859 return PAGE_ACTIVATE;
860 if (!may_write_to_inode(mapping->host, sc))
863 if (clear_page_dirty_for_io(page)) {
865 struct writeback_control wbc = {
866 .sync_mode = WB_SYNC_NONE,
867 .nr_to_write = SWAP_CLUSTER_MAX,
869 .range_end = LLONG_MAX,
873 SetPageReclaim(page);
874 res = mapping->a_ops->writepage(page, &wbc);
876 handle_write_error(mapping, page, res);
877 if (res == AOP_WRITEPAGE_ACTIVATE) {
878 ClearPageReclaim(page);
879 return PAGE_ACTIVATE;
882 if (!PageWriteback(page)) {
883 /* synchronous write or broken a_ops? */
884 ClearPageReclaim(page);
886 trace_mm_vmscan_writepage(page);
887 inc_node_page_state(page, NR_VMSCAN_WRITE);
895 * Same as remove_mapping, but if the page is removed from the mapping, it
896 * gets returned with a refcount of 0.
898 static int __remove_mapping(struct address_space *mapping, struct page *page,
904 BUG_ON(!PageLocked(page));
905 BUG_ON(mapping != page_mapping(page));
907 xa_lock_irqsave(&mapping->i_pages, flags);
909 * The non racy check for a busy page.
911 * Must be careful with the order of the tests. When someone has
912 * a ref to the page, it may be possible that they dirty it then
913 * drop the reference. So if PageDirty is tested before page_count
914 * here, then the following race may occur:
916 * get_user_pages(&page);
917 * [user mapping goes away]
919 * !PageDirty(page) [good]
920 * SetPageDirty(page);
922 * !page_count(page) [good, discard it]
924 * [oops, our write_to data is lost]
926 * Reversing the order of the tests ensures such a situation cannot
927 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
928 * load is not satisfied before that of page->_refcount.
930 * Note that if SetPageDirty is always performed via set_page_dirty,
931 * and thus under the i_pages lock, then this ordering is not required.
933 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
934 refcount = 1 + HPAGE_PMD_NR;
937 if (!page_ref_freeze(page, refcount))
939 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
940 if (unlikely(PageDirty(page))) {
941 page_ref_unfreeze(page, refcount);
945 if (PageSwapCache(page)) {
946 swp_entry_t swap = { .val = page_private(page) };
947 mem_cgroup_swapout(page, swap);
948 __delete_from_swap_cache(page, swap);
949 xa_unlock_irqrestore(&mapping->i_pages, flags);
950 put_swap_page(page, swap);
952 void (*freepage)(struct page *);
955 freepage = mapping->a_ops->freepage;
957 * Remember a shadow entry for reclaimed file cache in
958 * order to detect refaults, thus thrashing, later on.
960 * But don't store shadows in an address space that is
961 * already exiting. This is not just an optizimation,
962 * inode reclaim needs to empty out the radix tree or
963 * the nodes are lost. Don't plant shadows behind its
966 * We also don't store shadows for DAX mappings because the
967 * only page cache pages found in these are zero pages
968 * covering holes, and because we don't want to mix DAX
969 * exceptional entries and shadow exceptional entries in the
970 * same address_space.
972 if (reclaimed && page_is_file_cache(page) &&
973 !mapping_exiting(mapping) && !dax_mapping(mapping))
974 shadow = workingset_eviction(page);
975 __delete_from_page_cache(page, shadow);
976 xa_unlock_irqrestore(&mapping->i_pages, flags);
978 if (freepage != NULL)
985 xa_unlock_irqrestore(&mapping->i_pages, flags);
990 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
991 * someone else has a ref on the page, abort and return 0. If it was
992 * successfully detached, return 1. Assumes the caller has a single ref on
995 int remove_mapping(struct address_space *mapping, struct page *page)
997 if (__remove_mapping(mapping, page, false)) {
999 * Unfreezing the refcount with 1 rather than 2 effectively
1000 * drops the pagecache ref for us without requiring another
1003 page_ref_unfreeze(page, 1);
1010 * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 * @page: page to be put back to appropriate lru list
1013 * Add previously isolated @page to appropriate LRU list.
1014 * Page may still be unevictable for other reasons.
1016 * lru_lock must not be held, interrupts must be enabled.
1018 void putback_lru_page(struct page *page)
1020 lru_cache_add(page);
1021 put_page(page); /* drop ref from isolate */
1024 enum page_references {
1026 PAGEREF_RECLAIM_CLEAN,
1031 static enum page_references page_check_references(struct page *page,
1032 struct scan_control *sc)
1034 int referenced_ptes, referenced_page;
1035 unsigned long vm_flags;
1037 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1039 referenced_page = TestClearPageReferenced(page);
1042 * Mlock lost the isolation race with us. Let try_to_unmap()
1043 * move the page to the unevictable list.
1045 if (vm_flags & VM_LOCKED)
1046 return PAGEREF_RECLAIM;
1048 if (referenced_ptes) {
1049 if (PageSwapBacked(page))
1050 return PAGEREF_ACTIVATE;
1052 * All mapped pages start out with page table
1053 * references from the instantiating fault, so we need
1054 * to look twice if a mapped file page is used more
1057 * Mark it and spare it for another trip around the
1058 * inactive list. Another page table reference will
1059 * lead to its activation.
1061 * Note: the mark is set for activated pages as well
1062 * so that recently deactivated but used pages are
1063 * quickly recovered.
1065 SetPageReferenced(page);
1067 if (referenced_page || referenced_ptes > 1)
1068 return PAGEREF_ACTIVATE;
1071 * Activate file-backed executable pages after first usage.
1073 if (vm_flags & VM_EXEC)
1074 return PAGEREF_ACTIVATE;
1076 return PAGEREF_KEEP;
1079 /* Reclaim if clean, defer dirty pages to writeback */
1080 if (referenced_page && !PageSwapBacked(page))
1081 return PAGEREF_RECLAIM_CLEAN;
1083 return PAGEREF_RECLAIM;
1086 /* Check if a page is dirty or under writeback */
1087 static void page_check_dirty_writeback(struct page *page,
1088 bool *dirty, bool *writeback)
1090 struct address_space *mapping;
1093 * Anonymous pages are not handled by flushers and must be written
1094 * from reclaim context. Do not stall reclaim based on them
1096 if (!page_is_file_cache(page) ||
1097 (PageAnon(page) && !PageSwapBacked(page))) {
1103 /* By default assume that the page flags are accurate */
1104 *dirty = PageDirty(page);
1105 *writeback = PageWriteback(page);
1107 /* Verify dirty/writeback state if the filesystem supports it */
1108 if (!page_has_private(page))
1111 mapping = page_mapping(page);
1112 if (mapping && mapping->a_ops->is_dirty_writeback)
1113 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1117 * shrink_page_list() returns the number of reclaimed pages
1119 static unsigned long shrink_page_list(struct list_head *page_list,
1120 struct pglist_data *pgdat,
1121 struct scan_control *sc,
1122 enum ttu_flags ttu_flags,
1123 struct reclaim_stat *stat,
1126 LIST_HEAD(ret_pages);
1127 LIST_HEAD(free_pages);
1128 unsigned nr_reclaimed = 0;
1129 unsigned pgactivate = 0;
1131 memset(stat, 0, sizeof(*stat));
1134 while (!list_empty(page_list)) {
1135 struct address_space *mapping;
1138 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1139 bool dirty, writeback;
1140 unsigned int nr_pages;
1144 page = lru_to_page(page_list);
1145 list_del(&page->lru);
1147 if (!trylock_page(page))
1150 VM_BUG_ON_PAGE(PageActive(page), page);
1152 nr_pages = compound_nr(page);
1154 /* Account the number of base pages even though THP */
1155 sc->nr_scanned += nr_pages;
1157 if (unlikely(!page_evictable(page)))
1158 goto activate_locked;
1160 if (!sc->may_unmap && page_mapped(page))
1163 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1164 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1167 * The number of dirty pages determines if a node is marked
1168 * reclaim_congested which affects wait_iff_congested. kswapd
1169 * will stall and start writing pages if the tail of the LRU
1170 * is all dirty unqueued pages.
1172 page_check_dirty_writeback(page, &dirty, &writeback);
1173 if (dirty || writeback)
1176 if (dirty && !writeback)
1177 stat->nr_unqueued_dirty++;
1180 * Treat this page as congested if the underlying BDI is or if
1181 * pages are cycling through the LRU so quickly that the
1182 * pages marked for immediate reclaim are making it to the
1183 * end of the LRU a second time.
1185 mapping = page_mapping(page);
1186 if (((dirty || writeback) && mapping &&
1187 inode_write_congested(mapping->host)) ||
1188 (writeback && PageReclaim(page)))
1189 stat->nr_congested++;
1192 * If a page at the tail of the LRU is under writeback, there
1193 * are three cases to consider.
1195 * 1) If reclaim is encountering an excessive number of pages
1196 * under writeback and this page is both under writeback and
1197 * PageReclaim then it indicates that pages are being queued
1198 * for IO but are being recycled through the LRU before the
1199 * IO can complete. Waiting on the page itself risks an
1200 * indefinite stall if it is impossible to writeback the
1201 * page due to IO error or disconnected storage so instead
1202 * note that the LRU is being scanned too quickly and the
1203 * caller can stall after page list has been processed.
1205 * 2) Global or new memcg reclaim encounters a page that is
1206 * not marked for immediate reclaim, or the caller does not
1207 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208 * not to fs). In this case mark the page for immediate
1209 * reclaim and continue scanning.
1211 * Require may_enter_fs because we would wait on fs, which
1212 * may not have submitted IO yet. And the loop driver might
1213 * enter reclaim, and deadlock if it waits on a page for
1214 * which it is needed to do the write (loop masks off
1215 * __GFP_IO|__GFP_FS for this reason); but more thought
1216 * would probably show more reasons.
1218 * 3) Legacy memcg encounters a page that is already marked
1219 * PageReclaim. memcg does not have any dirty pages
1220 * throttling so we could easily OOM just because too many
1221 * pages are in writeback and there is nothing else to
1222 * reclaim. Wait for the writeback to complete.
1224 * In cases 1) and 2) we activate the pages to get them out of
1225 * the way while we continue scanning for clean pages on the
1226 * inactive list and refilling from the active list. The
1227 * observation here is that waiting for disk writes is more
1228 * expensive than potentially causing reloads down the line.
1229 * Since they're marked for immediate reclaim, they won't put
1230 * memory pressure on the cache working set any longer than it
1231 * takes to write them to disk.
1233 if (PageWriteback(page)) {
1235 if (current_is_kswapd() &&
1236 PageReclaim(page) &&
1237 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1238 stat->nr_immediate++;
1239 goto activate_locked;
1242 } else if (sane_reclaim(sc) ||
1243 !PageReclaim(page) || !may_enter_fs) {
1245 * This is slightly racy - end_page_writeback()
1246 * might have just cleared PageReclaim, then
1247 * setting PageReclaim here end up interpreted
1248 * as PageReadahead - but that does not matter
1249 * enough to care. What we do want is for this
1250 * page to have PageReclaim set next time memcg
1251 * reclaim reaches the tests above, so it will
1252 * then wait_on_page_writeback() to avoid OOM;
1253 * and it's also appropriate in global reclaim.
1255 SetPageReclaim(page);
1256 stat->nr_writeback++;
1257 goto activate_locked;
1262 wait_on_page_writeback(page);
1263 /* then go back and try same page again */
1264 list_add_tail(&page->lru, page_list);
1270 references = page_check_references(page, sc);
1272 switch (references) {
1273 case PAGEREF_ACTIVATE:
1274 goto activate_locked;
1276 stat->nr_ref_keep += nr_pages;
1278 case PAGEREF_RECLAIM:
1279 case PAGEREF_RECLAIM_CLEAN:
1280 ; /* try to reclaim the page below */
1284 * Anonymous process memory has backing store?
1285 * Try to allocate it some swap space here.
1286 * Lazyfree page could be freed directly
1288 if (PageAnon(page) && PageSwapBacked(page)) {
1289 if (!PageSwapCache(page)) {
1290 if (!(sc->gfp_mask & __GFP_IO))
1292 if (PageTransHuge(page)) {
1293 /* cannot split THP, skip it */
1294 if (!can_split_huge_page(page, NULL))
1295 goto activate_locked;
1297 * Split pages without a PMD map right
1298 * away. Chances are some or all of the
1299 * tail pages can be freed without IO.
1301 if (!compound_mapcount(page) &&
1302 split_huge_page_to_list(page,
1304 goto activate_locked;
1306 if (!add_to_swap(page)) {
1307 if (!PageTransHuge(page))
1308 goto activate_locked_split;
1309 /* Fallback to swap normal pages */
1310 if (split_huge_page_to_list(page,
1312 goto activate_locked;
1313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314 count_vm_event(THP_SWPOUT_FALLBACK);
1316 if (!add_to_swap(page))
1317 goto activate_locked_split;
1322 /* Adding to swap updated mapping */
1323 mapping = page_mapping(page);
1325 } else if (unlikely(PageTransHuge(page))) {
1326 /* Split file THP */
1327 if (split_huge_page_to_list(page, page_list))
1332 * THP may get split above, need minus tail pages and update
1333 * nr_pages to avoid accounting tail pages twice.
1335 * The tail pages that are added into swap cache successfully
1338 if ((nr_pages > 1) && !PageTransHuge(page)) {
1339 sc->nr_scanned -= (nr_pages - 1);
1344 * The page is mapped into the page tables of one or more
1345 * processes. Try to unmap it here.
1347 if (page_mapped(page)) {
1348 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1350 if (unlikely(PageTransHuge(page)))
1351 flags |= TTU_SPLIT_HUGE_PMD;
1352 if (!try_to_unmap(page, flags)) {
1353 stat->nr_unmap_fail += nr_pages;
1354 goto activate_locked;
1358 if (PageDirty(page)) {
1360 * Only kswapd can writeback filesystem pages
1361 * to avoid risk of stack overflow. But avoid
1362 * injecting inefficient single-page IO into
1363 * flusher writeback as much as possible: only
1364 * write pages when we've encountered many
1365 * dirty pages, and when we've already scanned
1366 * the rest of the LRU for clean pages and see
1367 * the same dirty pages again (PageReclaim).
1369 if (page_is_file_cache(page) &&
1370 (!current_is_kswapd() || !PageReclaim(page) ||
1371 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1373 * Immediately reclaim when written back.
1374 * Similar in principal to deactivate_page()
1375 * except we already have the page isolated
1376 * and know it's dirty
1378 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1379 SetPageReclaim(page);
1381 goto activate_locked;
1384 if (references == PAGEREF_RECLAIM_CLEAN)
1388 if (!sc->may_writepage)
1392 * Page is dirty. Flush the TLB if a writable entry
1393 * potentially exists to avoid CPU writes after IO
1394 * starts and then write it out here.
1396 try_to_unmap_flush_dirty();
1397 switch (pageout(page, mapping, sc)) {
1401 goto activate_locked;
1403 if (PageWriteback(page))
1405 if (PageDirty(page))
1409 * A synchronous write - probably a ramdisk. Go
1410 * ahead and try to reclaim the page.
1412 if (!trylock_page(page))
1414 if (PageDirty(page) || PageWriteback(page))
1416 mapping = page_mapping(page);
1418 ; /* try to free the page below */
1423 * If the page has buffers, try to free the buffer mappings
1424 * associated with this page. If we succeed we try to free
1427 * We do this even if the page is PageDirty().
1428 * try_to_release_page() does not perform I/O, but it is
1429 * possible for a page to have PageDirty set, but it is actually
1430 * clean (all its buffers are clean). This happens if the
1431 * buffers were written out directly, with submit_bh(). ext3
1432 * will do this, as well as the blockdev mapping.
1433 * try_to_release_page() will discover that cleanness and will
1434 * drop the buffers and mark the page clean - it can be freed.
1436 * Rarely, pages can have buffers and no ->mapping. These are
1437 * the pages which were not successfully invalidated in
1438 * truncate_complete_page(). We try to drop those buffers here
1439 * and if that worked, and the page is no longer mapped into
1440 * process address space (page_count == 1) it can be freed.
1441 * Otherwise, leave the page on the LRU so it is swappable.
1443 if (page_has_private(page)) {
1444 if (!try_to_release_page(page, sc->gfp_mask))
1445 goto activate_locked;
1446 if (!mapping && page_count(page) == 1) {
1448 if (put_page_testzero(page))
1452 * rare race with speculative reference.
1453 * the speculative reference will free
1454 * this page shortly, so we may
1455 * increment nr_reclaimed here (and
1456 * leave it off the LRU).
1464 if (PageAnon(page) && !PageSwapBacked(page)) {
1465 /* follow __remove_mapping for reference */
1466 if (!page_ref_freeze(page, 1))
1468 if (PageDirty(page)) {
1469 page_ref_unfreeze(page, 1);
1473 count_vm_event(PGLAZYFREED);
1474 count_memcg_page_event(page, PGLAZYFREED);
1475 } else if (!mapping || !__remove_mapping(mapping, page, true))
1481 * THP may get swapped out in a whole, need account
1484 nr_reclaimed += nr_pages;
1487 * Is there need to periodically free_page_list? It would
1488 * appear not as the counts should be low
1490 if (unlikely(PageTransHuge(page)))
1491 (*get_compound_page_dtor(page))(page);
1493 list_add(&page->lru, &free_pages);
1496 activate_locked_split:
1498 * The tail pages that are failed to add into swap cache
1499 * reach here. Fixup nr_scanned and nr_pages.
1502 sc->nr_scanned -= (nr_pages - 1);
1506 /* Not a candidate for swapping, so reclaim swap space. */
1507 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1509 try_to_free_swap(page);
1510 VM_BUG_ON_PAGE(PageActive(page), page);
1511 if (!PageMlocked(page)) {
1512 int type = page_is_file_cache(page);
1513 SetPageActive(page);
1514 stat->nr_activate[type] += nr_pages;
1515 count_memcg_page_event(page, PGACTIVATE);
1520 list_add(&page->lru, &ret_pages);
1521 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1524 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1526 mem_cgroup_uncharge_list(&free_pages);
1527 try_to_unmap_flush();
1528 free_unref_page_list(&free_pages);
1530 list_splice(&ret_pages, page_list);
1531 count_vm_events(PGACTIVATE, pgactivate);
1533 return nr_reclaimed;
1536 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1537 struct list_head *page_list)
1539 struct scan_control sc = {
1540 .gfp_mask = GFP_KERNEL,
1541 .priority = DEF_PRIORITY,
1544 struct reclaim_stat dummy_stat;
1546 struct page *page, *next;
1547 LIST_HEAD(clean_pages);
1549 list_for_each_entry_safe(page, next, page_list, lru) {
1550 if (page_is_file_cache(page) && !PageDirty(page) &&
1551 !__PageMovable(page) && !PageUnevictable(page)) {
1552 ClearPageActive(page);
1553 list_move(&page->lru, &clean_pages);
1557 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1558 TTU_IGNORE_ACCESS, &dummy_stat, true);
1559 list_splice(&clean_pages, page_list);
1560 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1565 * Attempt to remove the specified page from its LRU. Only take this page
1566 * if it is of the appropriate PageActive status. Pages which are being
1567 * freed elsewhere are also ignored.
1569 * page: page to consider
1570 * mode: one of the LRU isolation modes defined above
1572 * returns 0 on success, -ve errno on failure.
1574 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1578 /* Only take pages on the LRU. */
1582 /* Compaction should not handle unevictable pages but CMA can do so */
1583 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1589 * To minimise LRU disruption, the caller can indicate that it only
1590 * wants to isolate pages it will be able to operate on without
1591 * blocking - clean pages for the most part.
1593 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594 * that it is possible to migrate without blocking
1596 if (mode & ISOLATE_ASYNC_MIGRATE) {
1597 /* All the caller can do on PageWriteback is block */
1598 if (PageWriteback(page))
1601 if (PageDirty(page)) {
1602 struct address_space *mapping;
1606 * Only pages without mappings or that have a
1607 * ->migratepage callback are possible to migrate
1608 * without blocking. However, we can be racing with
1609 * truncation so it's necessary to lock the page
1610 * to stabilise the mapping as truncation holds
1611 * the page lock until after the page is removed
1612 * from the page cache.
1614 if (!trylock_page(page))
1617 mapping = page_mapping(page);
1618 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1625 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1628 if (likely(get_page_unless_zero(page))) {
1630 * Be careful not to clear PageLRU until after we're
1631 * sure the page is not being freed elsewhere -- the
1632 * page release code relies on it.
1643 * Update LRU sizes after isolating pages. The LRU size updates must
1644 * be complete before mem_cgroup_update_lru_size due to a santity check.
1646 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1647 enum lru_list lru, unsigned long *nr_zone_taken)
1651 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1652 if (!nr_zone_taken[zid])
1655 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1657 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1664 * pgdat->lru_lock is heavily contended. Some of the functions that
1665 * shrink the lists perform better by taking out a batch of pages
1666 * and working on them outside the LRU lock.
1668 * For pagecache intensive workloads, this function is the hottest
1669 * spot in the kernel (apart from copy_*_user functions).
1671 * Appropriate locks must be held before calling this function.
1673 * @nr_to_scan: The number of eligible pages to look through on the list.
1674 * @lruvec: The LRU vector to pull pages from.
1675 * @dst: The temp list to put pages on to.
1676 * @nr_scanned: The number of pages that were scanned.
1677 * @sc: The scan_control struct for this reclaim session
1678 * @mode: One of the LRU isolation modes
1679 * @lru: LRU list id for isolating
1681 * returns how many pages were moved onto *@dst.
1683 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1684 struct lruvec *lruvec, struct list_head *dst,
1685 unsigned long *nr_scanned, struct scan_control *sc,
1688 struct list_head *src = &lruvec->lists[lru];
1689 unsigned long nr_taken = 0;
1690 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1691 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1692 unsigned long skipped = 0;
1693 unsigned long scan, total_scan, nr_pages;
1694 LIST_HEAD(pages_skipped);
1695 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1699 while (scan < nr_to_scan && !list_empty(src)) {
1702 page = lru_to_page(src);
1703 prefetchw_prev_lru_page(page, src, flags);
1705 VM_BUG_ON_PAGE(!PageLRU(page), page);
1707 nr_pages = compound_nr(page);
1708 total_scan += nr_pages;
1710 if (page_zonenum(page) > sc->reclaim_idx) {
1711 list_move(&page->lru, &pages_skipped);
1712 nr_skipped[page_zonenum(page)] += nr_pages;
1717 * Do not count skipped pages because that makes the function
1718 * return with no isolated pages if the LRU mostly contains
1719 * ineligible pages. This causes the VM to not reclaim any
1720 * pages, triggering a premature OOM.
1722 * Account all tail pages of THP. This would not cause
1723 * premature OOM since __isolate_lru_page() returns -EBUSY
1724 * only when the page is being freed somewhere else.
1727 switch (__isolate_lru_page(page, mode)) {
1729 nr_taken += nr_pages;
1730 nr_zone_taken[page_zonenum(page)] += nr_pages;
1731 list_move(&page->lru, dst);
1735 /* else it is being freed elsewhere */
1736 list_move(&page->lru, src);
1745 * Splice any skipped pages to the start of the LRU list. Note that
1746 * this disrupts the LRU order when reclaiming for lower zones but
1747 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748 * scanning would soon rescan the same pages to skip and put the
1749 * system at risk of premature OOM.
1751 if (!list_empty(&pages_skipped)) {
1754 list_splice(&pages_skipped, src);
1755 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1756 if (!nr_skipped[zid])
1759 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1760 skipped += nr_skipped[zid];
1763 *nr_scanned = total_scan;
1764 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1765 total_scan, skipped, nr_taken, mode, lru);
1766 update_lru_sizes(lruvec, lru, nr_zone_taken);
1771 * isolate_lru_page - tries to isolate a page from its LRU list
1772 * @page: page to isolate from its LRU list
1774 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 * vmstat statistic corresponding to whatever LRU list the page was on.
1777 * Returns 0 if the page was removed from an LRU list.
1778 * Returns -EBUSY if the page was not on an LRU list.
1780 * The returned page will have PageLRU() cleared. If it was found on
1781 * the active list, it will have PageActive set. If it was found on
1782 * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 * may need to be cleared by the caller before letting the page go.
1785 * The vmstat statistic corresponding to the list on which the page was
1786 * found will be decremented.
1790 * (1) Must be called with an elevated refcount on the page. This is a
1791 * fundamentnal difference from isolate_lru_pages (which is called
1792 * without a stable reference).
1793 * (2) the lru_lock must not be held.
1794 * (3) interrupts must be enabled.
1796 int isolate_lru_page(struct page *page)
1800 VM_BUG_ON_PAGE(!page_count(page), page);
1801 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1803 if (PageLRU(page)) {
1804 pg_data_t *pgdat = page_pgdat(page);
1805 struct lruvec *lruvec;
1807 spin_lock_irq(&pgdat->lru_lock);
1808 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1809 if (PageLRU(page)) {
1810 int lru = page_lru(page);
1813 del_page_from_lru_list(page, lruvec, lru);
1816 spin_unlock_irq(&pgdat->lru_lock);
1822 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 * then get resheduled. When there are massive number of tasks doing page
1824 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 * the LRU list will go small and be scanned faster than necessary, leading to
1826 * unnecessary swapping, thrashing and OOM.
1828 static int too_many_isolated(struct pglist_data *pgdat, int file,
1829 struct scan_control *sc)
1831 unsigned long inactive, isolated;
1833 if (current_is_kswapd())
1836 if (!sane_reclaim(sc))
1840 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1841 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1843 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1844 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1848 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849 * won't get blocked by normal direct-reclaimers, forming a circular
1852 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1855 return isolated > inactive;
1859 * This moves pages from @list to corresponding LRU list.
1861 * We move them the other way if the page is referenced by one or more
1862 * processes, from rmap.
1864 * If the pages are mostly unmapped, the processing is fast and it is
1865 * appropriate to hold zone_lru_lock across the whole operation. But if
1866 * the pages are mapped, the processing is slow (page_referenced()) so we
1867 * should drop zone_lru_lock around each page. It's impossible to balance
1868 * this, so instead we remove the pages from the LRU while processing them.
1869 * It is safe to rely on PG_active against the non-LRU pages in here because
1870 * nobody will play with that bit on a non-LRU page.
1872 * The downside is that we have to touch page->_refcount against each page.
1873 * But we had to alter page->flags anyway.
1875 * Returns the number of pages moved to the given lruvec.
1878 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1879 struct list_head *list)
1881 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1882 int nr_pages, nr_moved = 0;
1883 LIST_HEAD(pages_to_free);
1887 while (!list_empty(list)) {
1888 page = lru_to_page(list);
1889 VM_BUG_ON_PAGE(PageLRU(page), page);
1890 if (unlikely(!page_evictable(page))) {
1891 list_del(&page->lru);
1892 spin_unlock_irq(&pgdat->lru_lock);
1893 putback_lru_page(page);
1894 spin_lock_irq(&pgdat->lru_lock);
1897 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1900 lru = page_lru(page);
1902 nr_pages = hpage_nr_pages(page);
1903 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1904 list_move(&page->lru, &lruvec->lists[lru]);
1906 if (put_page_testzero(page)) {
1907 __ClearPageLRU(page);
1908 __ClearPageActive(page);
1909 del_page_from_lru_list(page, lruvec, lru);
1911 if (unlikely(PageCompound(page))) {
1912 spin_unlock_irq(&pgdat->lru_lock);
1913 (*get_compound_page_dtor(page))(page);
1914 spin_lock_irq(&pgdat->lru_lock);
1916 list_add(&page->lru, &pages_to_free);
1918 nr_moved += nr_pages;
1923 * To save our caller's stack, now use input list for pages to free.
1925 list_splice(&pages_to_free, list);
1931 * If a kernel thread (such as nfsd for loop-back mounts) services
1932 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 * In that case we should only throttle if the backing device it is
1934 * writing to is congested. In other cases it is safe to throttle.
1936 static int current_may_throttle(void)
1938 return !(current->flags & PF_LESS_THROTTLE) ||
1939 current->backing_dev_info == NULL ||
1940 bdi_write_congested(current->backing_dev_info);
1944 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1945 * of reclaimed pages
1947 static noinline_for_stack unsigned long
1948 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1949 struct scan_control *sc, enum lru_list lru)
1951 LIST_HEAD(page_list);
1952 unsigned long nr_scanned;
1953 unsigned long nr_reclaimed = 0;
1954 unsigned long nr_taken;
1955 struct reclaim_stat stat;
1956 int file = is_file_lru(lru);
1957 enum vm_event_item item;
1958 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1959 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1960 bool stalled = false;
1962 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1966 /* wait a bit for the reclaimer. */
1970 /* We are about to die and free our memory. Return now. */
1971 if (fatal_signal_pending(current))
1972 return SWAP_CLUSTER_MAX;
1977 spin_lock_irq(&pgdat->lru_lock);
1979 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1980 &nr_scanned, sc, lru);
1982 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983 reclaim_stat->recent_scanned[file] += nr_taken;
1985 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1986 if (global_reclaim(sc))
1987 __count_vm_events(item, nr_scanned);
1988 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1989 spin_unlock_irq(&pgdat->lru_lock);
1994 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1997 spin_lock_irq(&pgdat->lru_lock);
1999 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2000 if (global_reclaim(sc))
2001 __count_vm_events(item, nr_reclaimed);
2002 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2003 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2004 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2006 move_pages_to_lru(lruvec, &page_list);
2008 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2010 spin_unlock_irq(&pgdat->lru_lock);
2012 mem_cgroup_uncharge_list(&page_list);
2013 free_unref_page_list(&page_list);
2016 * If dirty pages are scanned that are not queued for IO, it
2017 * implies that flushers are not doing their job. This can
2018 * happen when memory pressure pushes dirty pages to the end of
2019 * the LRU before the dirty limits are breached and the dirty
2020 * data has expired. It can also happen when the proportion of
2021 * dirty pages grows not through writes but through memory
2022 * pressure reclaiming all the clean cache. And in some cases,
2023 * the flushers simply cannot keep up with the allocation
2024 * rate. Nudge the flusher threads in case they are asleep.
2026 if (stat.nr_unqueued_dirty == nr_taken)
2027 wakeup_flusher_threads(WB_REASON_VMSCAN);
2029 sc->nr.dirty += stat.nr_dirty;
2030 sc->nr.congested += stat.nr_congested;
2031 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2032 sc->nr.writeback += stat.nr_writeback;
2033 sc->nr.immediate += stat.nr_immediate;
2034 sc->nr.taken += nr_taken;
2036 sc->nr.file_taken += nr_taken;
2038 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2039 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2040 return nr_reclaimed;
2043 static void shrink_active_list(unsigned long nr_to_scan,
2044 struct lruvec *lruvec,
2045 struct scan_control *sc,
2048 unsigned long nr_taken;
2049 unsigned long nr_scanned;
2050 unsigned long vm_flags;
2051 LIST_HEAD(l_hold); /* The pages which were snipped off */
2052 LIST_HEAD(l_active);
2053 LIST_HEAD(l_inactive);
2055 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2056 unsigned nr_deactivate, nr_activate;
2057 unsigned nr_rotated = 0;
2058 int file = is_file_lru(lru);
2059 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2063 spin_lock_irq(&pgdat->lru_lock);
2065 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2066 &nr_scanned, sc, lru);
2068 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2069 reclaim_stat->recent_scanned[file] += nr_taken;
2071 __count_vm_events(PGREFILL, nr_scanned);
2072 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2074 spin_unlock_irq(&pgdat->lru_lock);
2076 while (!list_empty(&l_hold)) {
2078 page = lru_to_page(&l_hold);
2079 list_del(&page->lru);
2081 if (unlikely(!page_evictable(page))) {
2082 putback_lru_page(page);
2086 if (unlikely(buffer_heads_over_limit)) {
2087 if (page_has_private(page) && trylock_page(page)) {
2088 if (page_has_private(page))
2089 try_to_release_page(page, 0);
2094 if (page_referenced(page, 0, sc->target_mem_cgroup,
2096 nr_rotated += hpage_nr_pages(page);
2098 * Identify referenced, file-backed active pages and
2099 * give them one more trip around the active list. So
2100 * that executable code get better chances to stay in
2101 * memory under moderate memory pressure. Anon pages
2102 * are not likely to be evicted by use-once streaming
2103 * IO, plus JVM can create lots of anon VM_EXEC pages,
2104 * so we ignore them here.
2106 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2107 list_add(&page->lru, &l_active);
2112 ClearPageActive(page); /* we are de-activating */
2113 SetPageWorkingset(page);
2114 list_add(&page->lru, &l_inactive);
2118 * Move pages back to the lru list.
2120 spin_lock_irq(&pgdat->lru_lock);
2122 * Count referenced pages from currently used mappings as rotated,
2123 * even though only some of them are actually re-activated. This
2124 * helps balance scan pressure between file and anonymous pages in
2127 reclaim_stat->recent_rotated[file] += nr_rotated;
2129 nr_activate = move_pages_to_lru(lruvec, &l_active);
2130 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2131 /* Keep all free pages in l_active list */
2132 list_splice(&l_inactive, &l_active);
2134 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2135 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2137 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2138 spin_unlock_irq(&pgdat->lru_lock);
2140 mem_cgroup_uncharge_list(&l_active);
2141 free_unref_page_list(&l_active);
2142 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2143 nr_deactivate, nr_rotated, sc->priority, file);
2147 * The inactive anon list should be small enough that the VM never has
2148 * to do too much work.
2150 * The inactive file list should be small enough to leave most memory
2151 * to the established workingset on the scan-resistant active list,
2152 * but large enough to avoid thrashing the aggregate readahead window.
2154 * Both inactive lists should also be large enough that each inactive
2155 * page has a chance to be referenced again before it is reclaimed.
2157 * If that fails and refaulting is observed, the inactive list grows.
2159 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2160 * on this LRU, maintained by the pageout code. An inactive_ratio
2161 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2164 * memory ratio inactive
2165 * -------------------------------------
2174 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2175 struct scan_control *sc, bool trace)
2177 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2178 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2179 enum lru_list inactive_lru = file * LRU_FILE;
2180 unsigned long inactive, active;
2181 unsigned long inactive_ratio;
2182 unsigned long refaults;
2186 * If we don't have swap space, anonymous page deactivation
2189 if (!file && !total_swap_pages)
2192 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2193 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2196 * When refaults are being observed, it means a new workingset
2197 * is being established. Disable active list protection to get
2198 * rid of the stale workingset quickly.
2200 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2201 if (file && lruvec->refaults != refaults) {
2204 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2206 inactive_ratio = int_sqrt(10 * gb);
2212 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2213 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2214 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2215 inactive_ratio, file);
2217 return inactive * inactive_ratio < active;
2220 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2221 struct lruvec *lruvec, struct scan_control *sc)
2223 if (is_active_lru(lru)) {
2224 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2225 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2229 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2240 * Determine how aggressively the anon and file LRU lists should be
2241 * scanned. The relative value of each set of LRU lists is determined
2242 * by looking at the fraction of the pages scanned we did rotate back
2243 * onto the active list instead of evict.
2245 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2246 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2248 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2249 struct scan_control *sc, unsigned long *nr,
2250 unsigned long *lru_pages)
2252 int swappiness = mem_cgroup_swappiness(memcg);
2253 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2255 u64 denominator = 0; /* gcc */
2256 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2257 unsigned long anon_prio, file_prio;
2258 enum scan_balance scan_balance;
2259 unsigned long anon, file;
2260 unsigned long ap, fp;
2263 /* If we have no swap space, do not bother scanning anon pages. */
2264 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2265 scan_balance = SCAN_FILE;
2270 * Global reclaim will swap to prevent OOM even with no
2271 * swappiness, but memcg users want to use this knob to
2272 * disable swapping for individual groups completely when
2273 * using the memory controller's swap limit feature would be
2276 if (!global_reclaim(sc) && !swappiness) {
2277 scan_balance = SCAN_FILE;
2282 * Do not apply any pressure balancing cleverness when the
2283 * system is close to OOM, scan both anon and file equally
2284 * (unless the swappiness setting disagrees with swapping).
2286 if (!sc->priority && swappiness) {
2287 scan_balance = SCAN_EQUAL;
2292 * Prevent the reclaimer from falling into the cache trap: as
2293 * cache pages start out inactive, every cache fault will tip
2294 * the scan balance towards the file LRU. And as the file LRU
2295 * shrinks, so does the window for rotation from references.
2296 * This means we have a runaway feedback loop where a tiny
2297 * thrashing file LRU becomes infinitely more attractive than
2298 * anon pages. Try to detect this based on file LRU size.
2300 if (global_reclaim(sc)) {
2301 unsigned long pgdatfile;
2302 unsigned long pgdatfree;
2304 unsigned long total_high_wmark = 0;
2306 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2307 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2308 node_page_state(pgdat, NR_INACTIVE_FILE);
2310 for (z = 0; z < MAX_NR_ZONES; z++) {
2311 struct zone *zone = &pgdat->node_zones[z];
2312 if (!managed_zone(zone))
2315 total_high_wmark += high_wmark_pages(zone);
2318 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2320 * Force SCAN_ANON if there are enough inactive
2321 * anonymous pages on the LRU in eligible zones.
2322 * Otherwise, the small LRU gets thrashed.
2324 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2325 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2327 scan_balance = SCAN_ANON;
2334 * If there is enough inactive page cache, i.e. if the size of the
2335 * inactive list is greater than that of the active list *and* the
2336 * inactive list actually has some pages to scan on this priority, we
2337 * do not reclaim anything from the anonymous working set right now.
2338 * Without the second condition we could end up never scanning an
2339 * lruvec even if it has plenty of old anonymous pages unless the
2340 * system is under heavy pressure.
2342 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2343 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2344 scan_balance = SCAN_FILE;
2348 scan_balance = SCAN_FRACT;
2351 * With swappiness at 100, anonymous and file have the same priority.
2352 * This scanning priority is essentially the inverse of IO cost.
2354 anon_prio = swappiness;
2355 file_prio = 200 - anon_prio;
2358 * OK, so we have swap space and a fair amount of page cache
2359 * pages. We use the recently rotated / recently scanned
2360 * ratios to determine how valuable each cache is.
2362 * Because workloads change over time (and to avoid overflow)
2363 * we keep these statistics as a floating average, which ends
2364 * up weighing recent references more than old ones.
2366 * anon in [0], file in [1]
2369 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2370 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2371 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2372 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2374 spin_lock_irq(&pgdat->lru_lock);
2375 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2376 reclaim_stat->recent_scanned[0] /= 2;
2377 reclaim_stat->recent_rotated[0] /= 2;
2380 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2381 reclaim_stat->recent_scanned[1] /= 2;
2382 reclaim_stat->recent_rotated[1] /= 2;
2386 * The amount of pressure on anon vs file pages is inversely
2387 * proportional to the fraction of recently scanned pages on
2388 * each list that were recently referenced and in active use.
2390 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2391 ap /= reclaim_stat->recent_rotated[0] + 1;
2393 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2394 fp /= reclaim_stat->recent_rotated[1] + 1;
2395 spin_unlock_irq(&pgdat->lru_lock);
2399 denominator = ap + fp + 1;
2402 for_each_evictable_lru(lru) {
2403 int file = is_file_lru(lru);
2407 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2408 scan = size >> sc->priority;
2410 * If the cgroup's already been deleted, make sure to
2411 * scrape out the remaining cache.
2413 if (!scan && !mem_cgroup_online(memcg))
2414 scan = min(size, SWAP_CLUSTER_MAX);
2416 switch (scan_balance) {
2418 /* Scan lists relative to size */
2422 * Scan types proportional to swappiness and
2423 * their relative recent reclaim efficiency.
2424 * Make sure we don't miss the last page
2425 * because of a round-off error.
2427 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2432 /* Scan one type exclusively */
2433 if ((scan_balance == SCAN_FILE) != file) {
2439 /* Look ma, no brain */
2449 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2451 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2452 struct scan_control *sc, unsigned long *lru_pages)
2454 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2455 unsigned long nr[NR_LRU_LISTS];
2456 unsigned long targets[NR_LRU_LISTS];
2457 unsigned long nr_to_scan;
2459 unsigned long nr_reclaimed = 0;
2460 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2461 struct blk_plug plug;
2464 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2466 /* Record the original scan target for proportional adjustments later */
2467 memcpy(targets, nr, sizeof(nr));
2470 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2471 * event that can occur when there is little memory pressure e.g.
2472 * multiple streaming readers/writers. Hence, we do not abort scanning
2473 * when the requested number of pages are reclaimed when scanning at
2474 * DEF_PRIORITY on the assumption that the fact we are direct
2475 * reclaiming implies that kswapd is not keeping up and it is best to
2476 * do a batch of work at once. For memcg reclaim one check is made to
2477 * abort proportional reclaim if either the file or anon lru has already
2478 * dropped to zero at the first pass.
2480 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2481 sc->priority == DEF_PRIORITY);
2483 blk_start_plug(&plug);
2484 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2485 nr[LRU_INACTIVE_FILE]) {
2486 unsigned long nr_anon, nr_file, percentage;
2487 unsigned long nr_scanned;
2489 for_each_evictable_lru(lru) {
2491 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2492 nr[lru] -= nr_to_scan;
2494 nr_reclaimed += shrink_list(lru, nr_to_scan,
2501 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2505 * For kswapd and memcg, reclaim at least the number of pages
2506 * requested. Ensure that the anon and file LRUs are scanned
2507 * proportionally what was requested by get_scan_count(). We
2508 * stop reclaiming one LRU and reduce the amount scanning
2509 * proportional to the original scan target.
2511 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2512 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2515 * It's just vindictive to attack the larger once the smaller
2516 * has gone to zero. And given the way we stop scanning the
2517 * smaller below, this makes sure that we only make one nudge
2518 * towards proportionality once we've got nr_to_reclaim.
2520 if (!nr_file || !nr_anon)
2523 if (nr_file > nr_anon) {
2524 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2525 targets[LRU_ACTIVE_ANON] + 1;
2527 percentage = nr_anon * 100 / scan_target;
2529 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2530 targets[LRU_ACTIVE_FILE] + 1;
2532 percentage = nr_file * 100 / scan_target;
2535 /* Stop scanning the smaller of the LRU */
2537 nr[lru + LRU_ACTIVE] = 0;
2540 * Recalculate the other LRU scan count based on its original
2541 * scan target and the percentage scanning already complete
2543 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2544 nr_scanned = targets[lru] - nr[lru];
2545 nr[lru] = targets[lru] * (100 - percentage) / 100;
2546 nr[lru] -= min(nr[lru], nr_scanned);
2549 nr_scanned = targets[lru] - nr[lru];
2550 nr[lru] = targets[lru] * (100 - percentage) / 100;
2551 nr[lru] -= min(nr[lru], nr_scanned);
2553 scan_adjusted = true;
2555 blk_finish_plug(&plug);
2556 sc->nr_reclaimed += nr_reclaimed;
2559 * Even if we did not try to evict anon pages at all, we want to
2560 * rebalance the anon lru active/inactive ratio.
2562 if (inactive_list_is_low(lruvec, false, sc, true))
2563 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2564 sc, LRU_ACTIVE_ANON);
2567 /* Use reclaim/compaction for costly allocs or under memory pressure */
2568 static bool in_reclaim_compaction(struct scan_control *sc)
2570 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2571 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2572 sc->priority < DEF_PRIORITY - 2))
2579 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2580 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2581 * true if more pages should be reclaimed such that when the page allocator
2582 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2583 * It will give up earlier than that if there is difficulty reclaiming pages.
2585 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2586 unsigned long nr_reclaimed,
2587 struct scan_control *sc)
2589 unsigned long pages_for_compaction;
2590 unsigned long inactive_lru_pages;
2593 /* If not in reclaim/compaction mode, stop */
2594 if (!in_reclaim_compaction(sc))
2598 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2599 * number of pages that were scanned. This will return to the caller
2600 * with the risk reclaim/compaction and the resulting allocation attempt
2601 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2602 * allocations through requiring that the full LRU list has been scanned
2603 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2604 * scan, but that approximation was wrong, and there were corner cases
2605 * where always a non-zero amount of pages were scanned.
2610 /* If compaction would go ahead or the allocation would succeed, stop */
2611 for (z = 0; z <= sc->reclaim_idx; z++) {
2612 struct zone *zone = &pgdat->node_zones[z];
2613 if (!managed_zone(zone))
2616 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2617 case COMPACT_SUCCESS:
2618 case COMPACT_CONTINUE:
2621 /* check next zone */
2627 * If we have not reclaimed enough pages for compaction and the
2628 * inactive lists are large enough, continue reclaiming
2630 pages_for_compaction = compact_gap(sc->order);
2631 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2632 if (get_nr_swap_pages() > 0)
2633 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2635 return inactive_lru_pages > pages_for_compaction;
2638 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2640 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2641 (memcg && memcg_congested(pgdat, memcg));
2644 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2646 struct reclaim_state *reclaim_state = current->reclaim_state;
2647 unsigned long nr_reclaimed, nr_scanned;
2648 bool reclaimable = false;
2651 struct mem_cgroup *root = sc->target_mem_cgroup;
2652 unsigned long node_lru_pages = 0;
2653 struct mem_cgroup *memcg;
2655 memset(&sc->nr, 0, sizeof(sc->nr));
2657 nr_reclaimed = sc->nr_reclaimed;
2658 nr_scanned = sc->nr_scanned;
2660 memcg = mem_cgroup_iter(root, NULL, NULL);
2662 unsigned long lru_pages;
2663 unsigned long reclaimed;
2664 unsigned long scanned;
2666 switch (mem_cgroup_protected(root, memcg)) {
2667 case MEMCG_PROT_MIN:
2670 * If there is no reclaimable memory, OOM.
2673 case MEMCG_PROT_LOW:
2676 * Respect the protection only as long as
2677 * there is an unprotected supply
2678 * of reclaimable memory from other cgroups.
2680 if (!sc->memcg_low_reclaim) {
2681 sc->memcg_low_skipped = 1;
2684 memcg_memory_event(memcg, MEMCG_LOW);
2686 case MEMCG_PROT_NONE:
2690 reclaimed = sc->nr_reclaimed;
2691 scanned = sc->nr_scanned;
2692 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2693 node_lru_pages += lru_pages;
2695 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2698 /* Record the group's reclaim efficiency */
2699 vmpressure(sc->gfp_mask, memcg, false,
2700 sc->nr_scanned - scanned,
2701 sc->nr_reclaimed - reclaimed);
2703 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2705 if (reclaim_state) {
2706 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2707 reclaim_state->reclaimed_slab = 0;
2710 /* Record the subtree's reclaim efficiency */
2711 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2712 sc->nr_scanned - nr_scanned,
2713 sc->nr_reclaimed - nr_reclaimed);
2715 if (sc->nr_reclaimed - nr_reclaimed)
2718 if (current_is_kswapd()) {
2720 * If reclaim is isolating dirty pages under writeback,
2721 * it implies that the long-lived page allocation rate
2722 * is exceeding the page laundering rate. Either the
2723 * global limits are not being effective at throttling
2724 * processes due to the page distribution throughout
2725 * zones or there is heavy usage of a slow backing
2726 * device. The only option is to throttle from reclaim
2727 * context which is not ideal as there is no guarantee
2728 * the dirtying process is throttled in the same way
2729 * balance_dirty_pages() manages.
2731 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2732 * count the number of pages under pages flagged for
2733 * immediate reclaim and stall if any are encountered
2734 * in the nr_immediate check below.
2736 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2737 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2740 * Tag a node as congested if all the dirty pages
2741 * scanned were backed by a congested BDI and
2742 * wait_iff_congested will stall.
2744 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2745 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2747 /* Allow kswapd to start writing pages during reclaim.*/
2748 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2749 set_bit(PGDAT_DIRTY, &pgdat->flags);
2752 * If kswapd scans pages marked marked for immediate
2753 * reclaim and under writeback (nr_immediate), it
2754 * implies that pages are cycling through the LRU
2755 * faster than they are written so also forcibly stall.
2757 if (sc->nr.immediate)
2758 congestion_wait(BLK_RW_ASYNC, HZ/10);
2762 * Legacy memcg will stall in page writeback so avoid forcibly
2763 * stalling in wait_iff_congested().
2765 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2766 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2767 set_memcg_congestion(pgdat, root, true);
2770 * Stall direct reclaim for IO completions if underlying BDIs
2771 * and node is congested. Allow kswapd to continue until it
2772 * starts encountering unqueued dirty pages or cycling through
2773 * the LRU too quickly.
2775 if (!sc->hibernation_mode && !current_is_kswapd() &&
2776 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2777 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2779 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2783 * Kswapd gives up on balancing particular nodes after too
2784 * many failures to reclaim anything from them and goes to
2785 * sleep. On reclaim progress, reset the failure counter. A
2786 * successful direct reclaim run will revive a dormant kswapd.
2789 pgdat->kswapd_failures = 0;
2795 * Returns true if compaction should go ahead for a costly-order request, or
2796 * the allocation would already succeed without compaction. Return false if we
2797 * should reclaim first.
2799 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2801 unsigned long watermark;
2802 enum compact_result suitable;
2804 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2805 if (suitable == COMPACT_SUCCESS)
2806 /* Allocation should succeed already. Don't reclaim. */
2808 if (suitable == COMPACT_SKIPPED)
2809 /* Compaction cannot yet proceed. Do reclaim. */
2813 * Compaction is already possible, but it takes time to run and there
2814 * are potentially other callers using the pages just freed. So proceed
2815 * with reclaim to make a buffer of free pages available to give
2816 * compaction a reasonable chance of completing and allocating the page.
2817 * Note that we won't actually reclaim the whole buffer in one attempt
2818 * as the target watermark in should_continue_reclaim() is lower. But if
2819 * we are already above the high+gap watermark, don't reclaim at all.
2821 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2823 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2827 * This is the direct reclaim path, for page-allocating processes. We only
2828 * try to reclaim pages from zones which will satisfy the caller's allocation
2831 * If a zone is deemed to be full of pinned pages then just give it a light
2832 * scan then give up on it.
2834 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2838 unsigned long nr_soft_reclaimed;
2839 unsigned long nr_soft_scanned;
2841 pg_data_t *last_pgdat = NULL;
2844 * If the number of buffer_heads in the machine exceeds the maximum
2845 * allowed level, force direct reclaim to scan the highmem zone as
2846 * highmem pages could be pinning lowmem pages storing buffer_heads
2848 orig_mask = sc->gfp_mask;
2849 if (buffer_heads_over_limit) {
2850 sc->gfp_mask |= __GFP_HIGHMEM;
2851 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2854 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2855 sc->reclaim_idx, sc->nodemask) {
2857 * Take care memory controller reclaiming has small influence
2860 if (global_reclaim(sc)) {
2861 if (!cpuset_zone_allowed(zone,
2862 GFP_KERNEL | __GFP_HARDWALL))
2866 * If we already have plenty of memory free for
2867 * compaction in this zone, don't free any more.
2868 * Even though compaction is invoked for any
2869 * non-zero order, only frequent costly order
2870 * reclamation is disruptive enough to become a
2871 * noticeable problem, like transparent huge
2874 if (IS_ENABLED(CONFIG_COMPACTION) &&
2875 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2876 compaction_ready(zone, sc)) {
2877 sc->compaction_ready = true;
2882 * Shrink each node in the zonelist once. If the
2883 * zonelist is ordered by zone (not the default) then a
2884 * node may be shrunk multiple times but in that case
2885 * the user prefers lower zones being preserved.
2887 if (zone->zone_pgdat == last_pgdat)
2891 * This steals pages from memory cgroups over softlimit
2892 * and returns the number of reclaimed pages and
2893 * scanned pages. This works for global memory pressure
2894 * and balancing, not for a memcg's limit.
2896 nr_soft_scanned = 0;
2897 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2898 sc->order, sc->gfp_mask,
2900 sc->nr_reclaimed += nr_soft_reclaimed;
2901 sc->nr_scanned += nr_soft_scanned;
2902 /* need some check for avoid more shrink_zone() */
2905 /* See comment about same check for global reclaim above */
2906 if (zone->zone_pgdat == last_pgdat)
2908 last_pgdat = zone->zone_pgdat;
2909 shrink_node(zone->zone_pgdat, sc);
2913 * Restore to original mask to avoid the impact on the caller if we
2914 * promoted it to __GFP_HIGHMEM.
2916 sc->gfp_mask = orig_mask;
2919 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2921 struct mem_cgroup *memcg;
2923 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2925 unsigned long refaults;
2926 struct lruvec *lruvec;
2928 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2929 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2930 lruvec->refaults = refaults;
2931 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2935 * This is the main entry point to direct page reclaim.
2937 * If a full scan of the inactive list fails to free enough memory then we
2938 * are "out of memory" and something needs to be killed.
2940 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2941 * high - the zone may be full of dirty or under-writeback pages, which this
2942 * caller can't do much about. We kick the writeback threads and take explicit
2943 * naps in the hope that some of these pages can be written. But if the
2944 * allocating task holds filesystem locks which prevent writeout this might not
2945 * work, and the allocation attempt will fail.
2947 * returns: 0, if no pages reclaimed
2948 * else, the number of pages reclaimed
2950 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2951 struct scan_control *sc)
2953 int initial_priority = sc->priority;
2954 pg_data_t *last_pgdat;
2958 delayacct_freepages_start();
2960 if (global_reclaim(sc))
2961 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2964 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2967 shrink_zones(zonelist, sc);
2969 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2972 if (sc->compaction_ready)
2976 * If we're getting trouble reclaiming, start doing
2977 * writepage even in laptop mode.
2979 if (sc->priority < DEF_PRIORITY - 2)
2980 sc->may_writepage = 1;
2981 } while (--sc->priority >= 0);
2984 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2986 if (zone->zone_pgdat == last_pgdat)
2988 last_pgdat = zone->zone_pgdat;
2989 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2990 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
2993 delayacct_freepages_end();
2995 if (sc->nr_reclaimed)
2996 return sc->nr_reclaimed;
2998 /* Aborted reclaim to try compaction? don't OOM, then */
2999 if (sc->compaction_ready)
3002 /* Untapped cgroup reserves? Don't OOM, retry. */
3003 if (sc->memcg_low_skipped) {
3004 sc->priority = initial_priority;
3005 sc->memcg_low_reclaim = 1;
3006 sc->memcg_low_skipped = 0;
3013 static bool allow_direct_reclaim(pg_data_t *pgdat)
3016 unsigned long pfmemalloc_reserve = 0;
3017 unsigned long free_pages = 0;
3021 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3024 for (i = 0; i <= ZONE_NORMAL; i++) {
3025 zone = &pgdat->node_zones[i];
3026 if (!managed_zone(zone))
3029 if (!zone_reclaimable_pages(zone))
3032 pfmemalloc_reserve += min_wmark_pages(zone);
3033 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3036 /* If there are no reserves (unexpected config) then do not throttle */
3037 if (!pfmemalloc_reserve)
3040 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3042 /* kswapd must be awake if processes are being throttled */
3043 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3044 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3045 (enum zone_type)ZONE_NORMAL);
3046 wake_up_interruptible(&pgdat->kswapd_wait);
3053 * Throttle direct reclaimers if backing storage is backed by the network
3054 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3055 * depleted. kswapd will continue to make progress and wake the processes
3056 * when the low watermark is reached.
3058 * Returns true if a fatal signal was delivered during throttling. If this
3059 * happens, the page allocator should not consider triggering the OOM killer.
3061 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3062 nodemask_t *nodemask)
3066 pg_data_t *pgdat = NULL;
3069 * Kernel threads should not be throttled as they may be indirectly
3070 * responsible for cleaning pages necessary for reclaim to make forward
3071 * progress. kjournald for example may enter direct reclaim while
3072 * committing a transaction where throttling it could forcing other
3073 * processes to block on log_wait_commit().
3075 if (current->flags & PF_KTHREAD)
3079 * If a fatal signal is pending, this process should not throttle.
3080 * It should return quickly so it can exit and free its memory
3082 if (fatal_signal_pending(current))
3086 * Check if the pfmemalloc reserves are ok by finding the first node
3087 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3088 * GFP_KERNEL will be required for allocating network buffers when
3089 * swapping over the network so ZONE_HIGHMEM is unusable.
3091 * Throttling is based on the first usable node and throttled processes
3092 * wait on a queue until kswapd makes progress and wakes them. There
3093 * is an affinity then between processes waking up and where reclaim
3094 * progress has been made assuming the process wakes on the same node.
3095 * More importantly, processes running on remote nodes will not compete
3096 * for remote pfmemalloc reserves and processes on different nodes
3097 * should make reasonable progress.
3099 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3100 gfp_zone(gfp_mask), nodemask) {
3101 if (zone_idx(zone) > ZONE_NORMAL)
3104 /* Throttle based on the first usable node */
3105 pgdat = zone->zone_pgdat;
3106 if (allow_direct_reclaim(pgdat))
3111 /* If no zone was usable by the allocation flags then do not throttle */
3115 /* Account for the throttling */
3116 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3119 * If the caller cannot enter the filesystem, it's possible that it
3120 * is due to the caller holding an FS lock or performing a journal
3121 * transaction in the case of a filesystem like ext[3|4]. In this case,
3122 * it is not safe to block on pfmemalloc_wait as kswapd could be
3123 * blocked waiting on the same lock. Instead, throttle for up to a
3124 * second before continuing.
3126 if (!(gfp_mask & __GFP_FS)) {
3127 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3128 allow_direct_reclaim(pgdat), HZ);
3133 /* Throttle until kswapd wakes the process */
3134 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3135 allow_direct_reclaim(pgdat));
3138 if (fatal_signal_pending(current))
3145 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3146 gfp_t gfp_mask, nodemask_t *nodemask)
3148 unsigned long nr_reclaimed;
3149 struct scan_control sc = {
3150 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3151 .gfp_mask = current_gfp_context(gfp_mask),
3152 .reclaim_idx = gfp_zone(gfp_mask),
3154 .nodemask = nodemask,
3155 .priority = DEF_PRIORITY,
3156 .may_writepage = !laptop_mode,
3162 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3163 * Confirm they are large enough for max values.
3165 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3166 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3167 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3170 * Do not enter reclaim if fatal signal was delivered while throttled.
3171 * 1 is returned so that the page allocator does not OOM kill at this
3174 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3177 set_task_reclaim_state(current, &sc.reclaim_state);
3178 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3180 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3182 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3183 set_task_reclaim_state(current, NULL);
3185 return nr_reclaimed;
3190 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3191 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3192 gfp_t gfp_mask, bool noswap,
3194 unsigned long *nr_scanned)
3196 struct scan_control sc = {
3197 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3198 .target_mem_cgroup = memcg,
3199 .may_writepage = !laptop_mode,
3201 .reclaim_idx = MAX_NR_ZONES - 1,
3202 .may_swap = !noswap,
3204 unsigned long lru_pages;
3206 WARN_ON_ONCE(!current->reclaim_state);
3208 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3209 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3211 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3215 * NOTE: Although we can get the priority field, using it
3216 * here is not a good idea, since it limits the pages we can scan.
3217 * if we don't reclaim here, the shrink_node from balance_pgdat
3218 * will pick up pages from other mem cgroup's as well. We hack
3219 * the priority and make it zero.
3221 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3223 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3225 *nr_scanned = sc.nr_scanned;
3227 return sc.nr_reclaimed;
3230 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3231 unsigned long nr_pages,
3235 struct zonelist *zonelist;
3236 unsigned long nr_reclaimed;
3237 unsigned long pflags;
3239 unsigned int noreclaim_flag;
3240 struct scan_control sc = {
3241 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3242 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3243 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3244 .reclaim_idx = MAX_NR_ZONES - 1,
3245 .target_mem_cgroup = memcg,
3246 .priority = DEF_PRIORITY,
3247 .may_writepage = !laptop_mode,
3249 .may_swap = may_swap,
3252 set_task_reclaim_state(current, &sc.reclaim_state);
3254 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3255 * take care of from where we get pages. So the node where we start the
3256 * scan does not need to be the current node.
3258 nid = mem_cgroup_select_victim_node(memcg);
3260 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3262 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3264 psi_memstall_enter(&pflags);
3265 noreclaim_flag = memalloc_noreclaim_save();
3267 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3269 memalloc_noreclaim_restore(noreclaim_flag);
3270 psi_memstall_leave(&pflags);
3272 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3273 set_task_reclaim_state(current, NULL);
3275 return nr_reclaimed;
3279 static void age_active_anon(struct pglist_data *pgdat,
3280 struct scan_control *sc)
3282 struct mem_cgroup *memcg;
3284 if (!total_swap_pages)
3287 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3289 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3291 if (inactive_list_is_low(lruvec, false, sc, true))
3292 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3293 sc, LRU_ACTIVE_ANON);
3295 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3299 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3305 * Check for watermark boosts top-down as the higher zones
3306 * are more likely to be boosted. Both watermarks and boosts
3307 * should not be checked at the time time as reclaim would
3308 * start prematurely when there is no boosting and a lower
3311 for (i = classzone_idx; i >= 0; i--) {
3312 zone = pgdat->node_zones + i;
3313 if (!managed_zone(zone))
3316 if (zone->watermark_boost)
3324 * Returns true if there is an eligible zone balanced for the request order
3327 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3330 unsigned long mark = -1;
3334 * Check watermarks bottom-up as lower zones are more likely to
3337 for (i = 0; i <= classzone_idx; i++) {
3338 zone = pgdat->node_zones + i;
3340 if (!managed_zone(zone))
3343 mark = high_wmark_pages(zone);
3344 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3349 * If a node has no populated zone within classzone_idx, it does not
3350 * need balancing by definition. This can happen if a zone-restricted
3351 * allocation tries to wake a remote kswapd.
3359 /* Clear pgdat state for congested, dirty or under writeback. */
3360 static void clear_pgdat_congested(pg_data_t *pgdat)
3362 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3363 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3364 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3368 * Prepare kswapd for sleeping. This verifies that there are no processes
3369 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3371 * Returns true if kswapd is ready to sleep
3373 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3376 * The throttled processes are normally woken up in balance_pgdat() as
3377 * soon as allow_direct_reclaim() is true. But there is a potential
3378 * race between when kswapd checks the watermarks and a process gets
3379 * throttled. There is also a potential race if processes get
3380 * throttled, kswapd wakes, a large process exits thereby balancing the
3381 * zones, which causes kswapd to exit balance_pgdat() before reaching
3382 * the wake up checks. If kswapd is going to sleep, no process should
3383 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3384 * the wake up is premature, processes will wake kswapd and get
3385 * throttled again. The difference from wake ups in balance_pgdat() is
3386 * that here we are under prepare_to_wait().
3388 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3389 wake_up_all(&pgdat->pfmemalloc_wait);
3391 /* Hopeless node, leave it to direct reclaim */
3392 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3395 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3396 clear_pgdat_congested(pgdat);
3404 * kswapd shrinks a node of pages that are at or below the highest usable
3405 * zone that is currently unbalanced.
3407 * Returns true if kswapd scanned at least the requested number of pages to
3408 * reclaim or if the lack of progress was due to pages under writeback.
3409 * This is used to determine if the scanning priority needs to be raised.
3411 static bool kswapd_shrink_node(pg_data_t *pgdat,
3412 struct scan_control *sc)
3417 /* Reclaim a number of pages proportional to the number of zones */
3418 sc->nr_to_reclaim = 0;
3419 for (z = 0; z <= sc->reclaim_idx; z++) {
3420 zone = pgdat->node_zones + z;
3421 if (!managed_zone(zone))
3424 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3428 * Historically care was taken to put equal pressure on all zones but
3429 * now pressure is applied based on node LRU order.
3431 shrink_node(pgdat, sc);
3434 * Fragmentation may mean that the system cannot be rebalanced for
3435 * high-order allocations. If twice the allocation size has been
3436 * reclaimed then recheck watermarks only at order-0 to prevent
3437 * excessive reclaim. Assume that a process requested a high-order
3438 * can direct reclaim/compact.
3440 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3443 return sc->nr_scanned >= sc->nr_to_reclaim;
3447 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3448 * that are eligible for use by the caller until at least one zone is
3451 * Returns the order kswapd finished reclaiming at.
3453 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3454 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3455 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3456 * or lower is eligible for reclaim until at least one usable zone is
3459 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3462 unsigned long nr_soft_reclaimed;
3463 unsigned long nr_soft_scanned;
3464 unsigned long pflags;
3465 unsigned long nr_boost_reclaim;
3466 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3469 struct scan_control sc = {
3470 .gfp_mask = GFP_KERNEL,
3475 set_task_reclaim_state(current, &sc.reclaim_state);
3476 psi_memstall_enter(&pflags);
3477 __fs_reclaim_acquire();
3479 count_vm_event(PAGEOUTRUN);
3482 * Account for the reclaim boost. Note that the zone boost is left in
3483 * place so that parallel allocations that are near the watermark will
3484 * stall or direct reclaim until kswapd is finished.
3486 nr_boost_reclaim = 0;
3487 for (i = 0; i <= classzone_idx; i++) {
3488 zone = pgdat->node_zones + i;
3489 if (!managed_zone(zone))
3492 nr_boost_reclaim += zone->watermark_boost;
3493 zone_boosts[i] = zone->watermark_boost;
3495 boosted = nr_boost_reclaim;
3498 sc.priority = DEF_PRIORITY;
3500 unsigned long nr_reclaimed = sc.nr_reclaimed;
3501 bool raise_priority = true;
3505 sc.reclaim_idx = classzone_idx;
3508 * If the number of buffer_heads exceeds the maximum allowed
3509 * then consider reclaiming from all zones. This has a dual
3510 * purpose -- on 64-bit systems it is expected that
3511 * buffer_heads are stripped during active rotation. On 32-bit
3512 * systems, highmem pages can pin lowmem memory and shrinking
3513 * buffers can relieve lowmem pressure. Reclaim may still not
3514 * go ahead if all eligible zones for the original allocation
3515 * request are balanced to avoid excessive reclaim from kswapd.
3517 if (buffer_heads_over_limit) {
3518 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3519 zone = pgdat->node_zones + i;
3520 if (!managed_zone(zone))
3529 * If the pgdat is imbalanced then ignore boosting and preserve
3530 * the watermarks for a later time and restart. Note that the
3531 * zone watermarks will be still reset at the end of balancing
3532 * on the grounds that the normal reclaim should be enough to
3533 * re-evaluate if boosting is required when kswapd next wakes.
3535 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3536 if (!balanced && nr_boost_reclaim) {
3537 nr_boost_reclaim = 0;
3542 * If boosting is not active then only reclaim if there are no
3543 * eligible zones. Note that sc.reclaim_idx is not used as
3544 * buffer_heads_over_limit may have adjusted it.
3546 if (!nr_boost_reclaim && balanced)
3549 /* Limit the priority of boosting to avoid reclaim writeback */
3550 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3551 raise_priority = false;
3554 * Do not writeback or swap pages for boosted reclaim. The
3555 * intent is to relieve pressure not issue sub-optimal IO
3556 * from reclaim context. If no pages are reclaimed, the
3557 * reclaim will be aborted.
3559 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3560 sc.may_swap = !nr_boost_reclaim;
3563 * Do some background aging of the anon list, to give
3564 * pages a chance to be referenced before reclaiming. All
3565 * pages are rotated regardless of classzone as this is
3566 * about consistent aging.
3568 age_active_anon(pgdat, &sc);
3571 * If we're getting trouble reclaiming, start doing writepage
3572 * even in laptop mode.
3574 if (sc.priority < DEF_PRIORITY - 2)
3575 sc.may_writepage = 1;
3577 /* Call soft limit reclaim before calling shrink_node. */
3579 nr_soft_scanned = 0;
3580 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3581 sc.gfp_mask, &nr_soft_scanned);
3582 sc.nr_reclaimed += nr_soft_reclaimed;
3585 * There should be no need to raise the scanning priority if
3586 * enough pages are already being scanned that that high
3587 * watermark would be met at 100% efficiency.
3589 if (kswapd_shrink_node(pgdat, &sc))
3590 raise_priority = false;
3593 * If the low watermark is met there is no need for processes
3594 * to be throttled on pfmemalloc_wait as they should not be
3595 * able to safely make forward progress. Wake them
3597 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3598 allow_direct_reclaim(pgdat))
3599 wake_up_all(&pgdat->pfmemalloc_wait);
3601 /* Check if kswapd should be suspending */
3602 __fs_reclaim_release();
3603 ret = try_to_freeze();
3604 __fs_reclaim_acquire();
3605 if (ret || kthread_should_stop())
3609 * Raise priority if scanning rate is too low or there was no
3610 * progress in reclaiming pages
3612 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3613 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3616 * If reclaim made no progress for a boost, stop reclaim as
3617 * IO cannot be queued and it could be an infinite loop in
3618 * extreme circumstances.
3620 if (nr_boost_reclaim && !nr_reclaimed)
3623 if (raise_priority || !nr_reclaimed)
3625 } while (sc.priority >= 1);
3627 if (!sc.nr_reclaimed)
3628 pgdat->kswapd_failures++;
3631 /* If reclaim was boosted, account for the reclaim done in this pass */
3633 unsigned long flags;
3635 for (i = 0; i <= classzone_idx; i++) {
3636 if (!zone_boosts[i])
3639 /* Increments are under the zone lock */
3640 zone = pgdat->node_zones + i;
3641 spin_lock_irqsave(&zone->lock, flags);
3642 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3643 spin_unlock_irqrestore(&zone->lock, flags);
3647 * As there is now likely space, wakeup kcompact to defragment
3650 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3653 snapshot_refaults(NULL, pgdat);
3654 __fs_reclaim_release();
3655 psi_memstall_leave(&pflags);
3656 set_task_reclaim_state(current, NULL);
3659 * Return the order kswapd stopped reclaiming at as
3660 * prepare_kswapd_sleep() takes it into account. If another caller
3661 * entered the allocator slow path while kswapd was awake, order will
3662 * remain at the higher level.
3668 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3669 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3670 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3671 * after previous reclaim attempt (node is still unbalanced). In that case
3672 * return the zone index of the previous kswapd reclaim cycle.
3674 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3675 enum zone_type prev_classzone_idx)
3677 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3678 return prev_classzone_idx;
3679 return pgdat->kswapd_classzone_idx;
3682 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3683 unsigned int classzone_idx)
3688 if (freezing(current) || kthread_should_stop())
3691 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3694 * Try to sleep for a short interval. Note that kcompactd will only be
3695 * woken if it is possible to sleep for a short interval. This is
3696 * deliberate on the assumption that if reclaim cannot keep an
3697 * eligible zone balanced that it's also unlikely that compaction will
3700 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3702 * Compaction records what page blocks it recently failed to
3703 * isolate pages from and skips them in the future scanning.
3704 * When kswapd is going to sleep, it is reasonable to assume
3705 * that pages and compaction may succeed so reset the cache.
3707 reset_isolation_suitable(pgdat);
3710 * We have freed the memory, now we should compact it to make
3711 * allocation of the requested order possible.
3713 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3715 remaining = schedule_timeout(HZ/10);
3718 * If woken prematurely then reset kswapd_classzone_idx and
3719 * order. The values will either be from a wakeup request or
3720 * the previous request that slept prematurely.
3723 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3724 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3727 finish_wait(&pgdat->kswapd_wait, &wait);
3728 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3732 * After a short sleep, check if it was a premature sleep. If not, then
3733 * go fully to sleep until explicitly woken up.
3736 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3737 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3740 * vmstat counters are not perfectly accurate and the estimated
3741 * value for counters such as NR_FREE_PAGES can deviate from the
3742 * true value by nr_online_cpus * threshold. To avoid the zone
3743 * watermarks being breached while under pressure, we reduce the
3744 * per-cpu vmstat threshold while kswapd is awake and restore
3745 * them before going back to sleep.
3747 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3749 if (!kthread_should_stop())
3752 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3755 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3757 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3759 finish_wait(&pgdat->kswapd_wait, &wait);
3763 * The background pageout daemon, started as a kernel thread
3764 * from the init process.
3766 * This basically trickles out pages so that we have _some_
3767 * free memory available even if there is no other activity
3768 * that frees anything up. This is needed for things like routing
3769 * etc, where we otherwise might have all activity going on in
3770 * asynchronous contexts that cannot page things out.
3772 * If there are applications that are active memory-allocators
3773 * (most normal use), this basically shouldn't matter.
3775 static int kswapd(void *p)
3777 unsigned int alloc_order, reclaim_order;
3778 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3779 pg_data_t *pgdat = (pg_data_t*)p;
3780 struct task_struct *tsk = current;
3781 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3783 if (!cpumask_empty(cpumask))
3784 set_cpus_allowed_ptr(tsk, cpumask);
3787 * Tell the memory management that we're a "memory allocator",
3788 * and that if we need more memory we should get access to it
3789 * regardless (see "__alloc_pages()"). "kswapd" should
3790 * never get caught in the normal page freeing logic.
3792 * (Kswapd normally doesn't need memory anyway, but sometimes
3793 * you need a small amount of memory in order to be able to
3794 * page out something else, and this flag essentially protects
3795 * us from recursively trying to free more memory as we're
3796 * trying to free the first piece of memory in the first place).
3798 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3801 pgdat->kswapd_order = 0;
3802 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3806 alloc_order = reclaim_order = pgdat->kswapd_order;
3807 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3810 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3813 /* Read the new order and classzone_idx */
3814 alloc_order = reclaim_order = pgdat->kswapd_order;
3815 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3816 pgdat->kswapd_order = 0;
3817 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3819 ret = try_to_freeze();
3820 if (kthread_should_stop())
3824 * We can speed up thawing tasks if we don't call balance_pgdat
3825 * after returning from the refrigerator
3831 * Reclaim begins at the requested order but if a high-order
3832 * reclaim fails then kswapd falls back to reclaiming for
3833 * order-0. If that happens, kswapd will consider sleeping
3834 * for the order it finished reclaiming at (reclaim_order)
3835 * but kcompactd is woken to compact for the original
3836 * request (alloc_order).
3838 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3840 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3841 if (reclaim_order < alloc_order)
3842 goto kswapd_try_sleep;
3845 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3851 * A zone is low on free memory or too fragmented for high-order memory. If
3852 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3853 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3854 * has failed or is not needed, still wake up kcompactd if only compaction is
3857 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3858 enum zone_type classzone_idx)
3862 if (!managed_zone(zone))
3865 if (!cpuset_zone_allowed(zone, gfp_flags))
3867 pgdat = zone->zone_pgdat;
3869 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3870 pgdat->kswapd_classzone_idx = classzone_idx;
3872 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3874 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3875 if (!waitqueue_active(&pgdat->kswapd_wait))
3878 /* Hopeless node, leave it to direct reclaim if possible */
3879 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3880 (pgdat_balanced(pgdat, order, classzone_idx) &&
3881 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3883 * There may be plenty of free memory available, but it's too
3884 * fragmented for high-order allocations. Wake up kcompactd
3885 * and rely on compaction_suitable() to determine if it's
3886 * needed. If it fails, it will defer subsequent attempts to
3887 * ratelimit its work.
3889 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3890 wakeup_kcompactd(pgdat, order, classzone_idx);
3894 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3896 wake_up_interruptible(&pgdat->kswapd_wait);
3899 #ifdef CONFIG_HIBERNATION
3901 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3904 * Rather than trying to age LRUs the aim is to preserve the overall
3905 * LRU order by reclaiming preferentially
3906 * inactive > active > active referenced > active mapped
3908 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3910 struct scan_control sc = {
3911 .nr_to_reclaim = nr_to_reclaim,
3912 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3913 .reclaim_idx = MAX_NR_ZONES - 1,
3914 .priority = DEF_PRIORITY,
3918 .hibernation_mode = 1,
3920 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3921 unsigned long nr_reclaimed;
3922 unsigned int noreclaim_flag;
3924 fs_reclaim_acquire(sc.gfp_mask);
3925 noreclaim_flag = memalloc_noreclaim_save();
3926 set_task_reclaim_state(current, &sc.reclaim_state);
3928 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3930 set_task_reclaim_state(current, NULL);
3931 memalloc_noreclaim_restore(noreclaim_flag);
3932 fs_reclaim_release(sc.gfp_mask);
3934 return nr_reclaimed;
3936 #endif /* CONFIG_HIBERNATION */
3938 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3939 not required for correctness. So if the last cpu in a node goes
3940 away, we get changed to run anywhere: as the first one comes back,
3941 restore their cpu bindings. */
3942 static int kswapd_cpu_online(unsigned int cpu)
3946 for_each_node_state(nid, N_MEMORY) {
3947 pg_data_t *pgdat = NODE_DATA(nid);
3948 const struct cpumask *mask;
3950 mask = cpumask_of_node(pgdat->node_id);
3952 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3953 /* One of our CPUs online: restore mask */
3954 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3960 * This kswapd start function will be called by init and node-hot-add.
3961 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3963 int kswapd_run(int nid)
3965 pg_data_t *pgdat = NODE_DATA(nid);
3971 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3972 if (IS_ERR(pgdat->kswapd)) {
3973 /* failure at boot is fatal */
3974 BUG_ON(system_state < SYSTEM_RUNNING);
3975 pr_err("Failed to start kswapd on node %d\n", nid);
3976 ret = PTR_ERR(pgdat->kswapd);
3977 pgdat->kswapd = NULL;
3983 * Called by memory hotplug when all memory in a node is offlined. Caller must
3984 * hold mem_hotplug_begin/end().
3986 void kswapd_stop(int nid)
3988 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3991 kthread_stop(kswapd);
3992 NODE_DATA(nid)->kswapd = NULL;
3996 static int __init kswapd_init(void)
4001 for_each_node_state(nid, N_MEMORY)
4003 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4004 "mm/vmscan:online", kswapd_cpu_online,
4010 module_init(kswapd_init)
4016 * If non-zero call node_reclaim when the number of free pages falls below
4019 int node_reclaim_mode __read_mostly;
4021 #define RECLAIM_OFF 0
4022 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4023 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4024 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4027 * Priority for NODE_RECLAIM. This determines the fraction of pages
4028 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4031 #define NODE_RECLAIM_PRIORITY 4
4034 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4037 int sysctl_min_unmapped_ratio = 1;
4040 * If the number of slab pages in a zone grows beyond this percentage then
4041 * slab reclaim needs to occur.
4043 int sysctl_min_slab_ratio = 5;
4045 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4047 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4048 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4049 node_page_state(pgdat, NR_ACTIVE_FILE);
4052 * It's possible for there to be more file mapped pages than
4053 * accounted for by the pages on the file LRU lists because
4054 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4056 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4059 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4060 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4062 unsigned long nr_pagecache_reclaimable;
4063 unsigned long delta = 0;
4066 * If RECLAIM_UNMAP is set, then all file pages are considered
4067 * potentially reclaimable. Otherwise, we have to worry about
4068 * pages like swapcache and node_unmapped_file_pages() provides
4071 if (node_reclaim_mode & RECLAIM_UNMAP)
4072 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4074 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4076 /* If we can't clean pages, remove dirty pages from consideration */
4077 if (!(node_reclaim_mode & RECLAIM_WRITE))
4078 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4080 /* Watch for any possible underflows due to delta */
4081 if (unlikely(delta > nr_pagecache_reclaimable))
4082 delta = nr_pagecache_reclaimable;
4084 return nr_pagecache_reclaimable - delta;
4088 * Try to free up some pages from this node through reclaim.
4090 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4092 /* Minimum pages needed in order to stay on node */
4093 const unsigned long nr_pages = 1 << order;
4094 struct task_struct *p = current;
4095 unsigned int noreclaim_flag;
4096 struct scan_control sc = {
4097 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4098 .gfp_mask = current_gfp_context(gfp_mask),
4100 .priority = NODE_RECLAIM_PRIORITY,
4101 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4102 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4104 .reclaim_idx = gfp_zone(gfp_mask),
4107 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4111 fs_reclaim_acquire(sc.gfp_mask);
4113 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4114 * and we also need to be able to write out pages for RECLAIM_WRITE
4115 * and RECLAIM_UNMAP.
4117 noreclaim_flag = memalloc_noreclaim_save();
4118 p->flags |= PF_SWAPWRITE;
4119 set_task_reclaim_state(p, &sc.reclaim_state);
4121 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4123 * Free memory by calling shrink node with increasing
4124 * priorities until we have enough memory freed.
4127 shrink_node(pgdat, &sc);
4128 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4131 set_task_reclaim_state(p, NULL);
4132 current->flags &= ~PF_SWAPWRITE;
4133 memalloc_noreclaim_restore(noreclaim_flag);
4134 fs_reclaim_release(sc.gfp_mask);
4136 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4138 return sc.nr_reclaimed >= nr_pages;
4141 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4146 * Node reclaim reclaims unmapped file backed pages and
4147 * slab pages if we are over the defined limits.
4149 * A small portion of unmapped file backed pages is needed for
4150 * file I/O otherwise pages read by file I/O will be immediately
4151 * thrown out if the node is overallocated. So we do not reclaim
4152 * if less than a specified percentage of the node is used by
4153 * unmapped file backed pages.
4155 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4156 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4157 return NODE_RECLAIM_FULL;
4160 * Do not scan if the allocation should not be delayed.
4162 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4163 return NODE_RECLAIM_NOSCAN;
4166 * Only run node reclaim on the local node or on nodes that do not
4167 * have associated processors. This will favor the local processor
4168 * over remote processors and spread off node memory allocations
4169 * as wide as possible.
4171 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4172 return NODE_RECLAIM_NOSCAN;
4174 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4175 return NODE_RECLAIM_NOSCAN;
4177 ret = __node_reclaim(pgdat, gfp_mask, order);
4178 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4181 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4188 * page_evictable - test whether a page is evictable
4189 * @page: the page to test
4191 * Test whether page is evictable--i.e., should be placed on active/inactive
4192 * lists vs unevictable list.
4194 * Reasons page might not be evictable:
4195 * (1) page's mapping marked unevictable
4196 * (2) page is part of an mlocked VMA
4199 int page_evictable(struct page *page)
4203 /* Prevent address_space of inode and swap cache from being freed */
4205 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4211 * check_move_unevictable_pages - check pages for evictability and move to
4212 * appropriate zone lru list
4213 * @pvec: pagevec with lru pages to check
4215 * Checks pages for evictability, if an evictable page is in the unevictable
4216 * lru list, moves it to the appropriate evictable lru list. This function
4217 * should be only used for lru pages.
4219 void check_move_unevictable_pages(struct pagevec *pvec)
4221 struct lruvec *lruvec;
4222 struct pglist_data *pgdat = NULL;
4227 for (i = 0; i < pvec->nr; i++) {
4228 struct page *page = pvec->pages[i];
4229 struct pglist_data *pagepgdat = page_pgdat(page);
4232 if (pagepgdat != pgdat) {
4234 spin_unlock_irq(&pgdat->lru_lock);
4236 spin_lock_irq(&pgdat->lru_lock);
4238 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4240 if (!PageLRU(page) || !PageUnevictable(page))
4243 if (page_evictable(page)) {
4244 enum lru_list lru = page_lru_base_type(page);
4246 VM_BUG_ON_PAGE(PageActive(page), page);
4247 ClearPageUnevictable(page);
4248 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4249 add_page_to_lru_list(page, lruvec, lru);
4255 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4256 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4257 spin_unlock_irq(&pgdat->lru_lock);
4260 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);