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;
91 /* e.g. boosted watermark reclaim leaves slabs alone */
92 unsigned int may_shrinkslab:1;
95 * Cgroups are not reclaimed below their configured memory.low,
96 * unless we threaten to OOM. If any cgroups are skipped due to
97 * memory.low and nothing was reclaimed, go back for memory.low.
99 unsigned int memcg_low_reclaim:1;
100 unsigned int memcg_low_skipped:1;
102 unsigned int hibernation_mode:1;
104 /* One of the zones is ready for compaction */
105 unsigned int compaction_ready:1;
107 /* Allocation order */
110 /* Scan (total_size >> priority) pages at once */
113 /* The highest zone to isolate pages for reclaim from */
116 /* This context's GFP mask */
119 /* Incremented by the number of inactive pages that were scanned */
120 unsigned long nr_scanned;
122 /* Number of pages freed so far during a call to shrink_zones() */
123 unsigned long nr_reclaimed;
127 unsigned int unqueued_dirty;
128 unsigned int congested;
129 unsigned int writeback;
130 unsigned int immediate;
131 unsigned int file_taken;
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 */
242 static bool global_reclaim(struct scan_control *sc)
244 return !sc->target_mem_cgroup;
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
260 static bool sane_reclaim(struct scan_control *sc)
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
266 #ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
273 static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
277 struct mem_cgroup_per_node *mn;
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
286 static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
289 struct mem_cgroup_per_node *mn;
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
296 static bool global_reclaim(struct scan_control *sc)
301 static bool sane_reclaim(struct scan_control *sc)
306 static inline void set_memcg_congestion(struct pglist_data *pgdat,
307 struct mem_cgroup *memcg, bool congested)
311 static inline bool memcg_congested(struct pglist_data *pgdat,
312 struct mem_cgroup *memcg)
320 * This misses isolated pages which are not accounted for to save counters.
321 * As the data only determines if reclaim or compaction continues, it is
322 * not expected that isolated pages will be a dominating factor.
324 unsigned long zone_reclaimable_pages(struct zone *zone)
328 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
329 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
330 if (get_nr_swap_pages() > 0)
331 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
332 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
338 * lruvec_lru_size - Returns the number of pages on the given LRU list.
339 * @lruvec: lru vector
341 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
343 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
345 unsigned long lru_size;
348 if (!mem_cgroup_disabled())
349 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
351 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
353 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
354 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
357 if (!managed_zone(zone))
360 if (!mem_cgroup_disabled())
361 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
363 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
364 NR_ZONE_LRU_BASE + lru);
365 lru_size -= min(size, lru_size);
373 * Add a shrinker callback to be called from the vm.
375 int prealloc_shrinker(struct shrinker *shrinker)
377 size_t size = sizeof(*shrinker->nr_deferred);
379 if (shrinker->flags & SHRINKER_NUMA_AWARE)
382 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
383 if (!shrinker->nr_deferred)
386 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
387 if (prealloc_memcg_shrinker(shrinker))
394 kfree(shrinker->nr_deferred);
395 shrinker->nr_deferred = NULL;
399 void free_prealloced_shrinker(struct shrinker *shrinker)
401 if (!shrinker->nr_deferred)
404 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
405 unregister_memcg_shrinker(shrinker);
407 kfree(shrinker->nr_deferred);
408 shrinker->nr_deferred = NULL;
411 void register_shrinker_prepared(struct shrinker *shrinker)
413 down_write(&shrinker_rwsem);
414 list_add_tail(&shrinker->list, &shrinker_list);
415 #ifdef CONFIG_MEMCG_KMEM
416 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
417 idr_replace(&shrinker_idr, shrinker, shrinker->id);
419 up_write(&shrinker_rwsem);
422 int register_shrinker(struct shrinker *shrinker)
424 int err = prealloc_shrinker(shrinker);
428 register_shrinker_prepared(shrinker);
431 EXPORT_SYMBOL(register_shrinker);
436 void unregister_shrinker(struct shrinker *shrinker)
438 if (!shrinker->nr_deferred)
440 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
441 unregister_memcg_shrinker(shrinker);
442 down_write(&shrinker_rwsem);
443 list_del(&shrinker->list);
444 up_write(&shrinker_rwsem);
445 kfree(shrinker->nr_deferred);
446 shrinker->nr_deferred = NULL;
448 EXPORT_SYMBOL(unregister_shrinker);
450 #define SHRINK_BATCH 128
452 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
453 struct shrinker *shrinker, int priority)
455 unsigned long freed = 0;
456 unsigned long long delta;
461 int nid = shrinkctl->nid;
462 long batch_size = shrinker->batch ? shrinker->batch
464 long scanned = 0, next_deferred;
466 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
469 freeable = shrinker->count_objects(shrinker, shrinkctl);
470 if (freeable == 0 || freeable == SHRINK_EMPTY)
474 * copy the current shrinker scan count into a local variable
475 * and zero it so that other concurrent shrinker invocations
476 * don't also do this scanning work.
478 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
481 if (shrinker->seeks) {
482 delta = freeable >> priority;
484 do_div(delta, shrinker->seeks);
487 * These objects don't require any IO to create. Trim
488 * them aggressively under memory pressure to keep
489 * them from causing refetches in the IO caches.
491 delta = freeable / 2;
495 if (total_scan < 0) {
496 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
497 shrinker->scan_objects, total_scan);
498 total_scan = freeable;
501 next_deferred = total_scan;
504 * We need to avoid excessive windup on filesystem shrinkers
505 * due to large numbers of GFP_NOFS allocations causing the
506 * shrinkers to return -1 all the time. This results in a large
507 * nr being built up so when a shrink that can do some work
508 * comes along it empties the entire cache due to nr >>>
509 * freeable. This is bad for sustaining a working set in
512 * Hence only allow the shrinker to scan the entire cache when
513 * a large delta change is calculated directly.
515 if (delta < freeable / 4)
516 total_scan = min(total_scan, freeable / 2);
519 * Avoid risking looping forever due to too large nr value:
520 * never try to free more than twice the estimate number of
523 if (total_scan > freeable * 2)
524 total_scan = freeable * 2;
526 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
527 freeable, delta, total_scan, priority);
530 * Normally, we should not scan less than batch_size objects in one
531 * pass to avoid too frequent shrinker calls, but if the slab has less
532 * than batch_size objects in total and we are really tight on memory,
533 * we will try to reclaim all available objects, otherwise we can end
534 * up failing allocations although there are plenty of reclaimable
535 * objects spread over several slabs with usage less than the
538 * We detect the "tight on memory" situations by looking at the total
539 * number of objects we want to scan (total_scan). If it is greater
540 * than the total number of objects on slab (freeable), we must be
541 * scanning at high prio and therefore should try to reclaim as much as
544 while (total_scan >= batch_size ||
545 total_scan >= freeable) {
547 unsigned long nr_to_scan = min(batch_size, total_scan);
549 shrinkctl->nr_to_scan = nr_to_scan;
550 shrinkctl->nr_scanned = nr_to_scan;
551 ret = shrinker->scan_objects(shrinker, shrinkctl);
552 if (ret == SHRINK_STOP)
556 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
557 total_scan -= shrinkctl->nr_scanned;
558 scanned += shrinkctl->nr_scanned;
563 if (next_deferred >= scanned)
564 next_deferred -= scanned;
568 * move the unused scan count back into the shrinker in a
569 * manner that handles concurrent updates. If we exhausted the
570 * scan, there is no need to do an update.
572 if (next_deferred > 0)
573 new_nr = atomic_long_add_return(next_deferred,
574 &shrinker->nr_deferred[nid]);
576 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
578 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
582 #ifdef CONFIG_MEMCG_KMEM
583 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
584 struct mem_cgroup *memcg, int priority)
586 struct memcg_shrinker_map *map;
587 unsigned long ret, freed = 0;
590 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))
593 if (!down_read_trylock(&shrinker_rwsem))
596 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
601 for_each_set_bit(i, map->map, shrinker_nr_max) {
602 struct shrink_control sc = {
603 .gfp_mask = gfp_mask,
607 struct shrinker *shrinker;
609 shrinker = idr_find(&shrinker_idr, i);
610 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
612 clear_bit(i, map->map);
616 ret = do_shrink_slab(&sc, shrinker, priority);
617 if (ret == SHRINK_EMPTY) {
618 clear_bit(i, map->map);
620 * After the shrinker reported that it had no objects to
621 * free, but before we cleared the corresponding bit in
622 * the memcg shrinker map, a new object might have been
623 * added. To make sure, we have the bit set in this
624 * case, we invoke the shrinker one more time and reset
625 * the bit if it reports that it is not empty anymore.
626 * The memory barrier here pairs with the barrier in
627 * memcg_set_shrinker_bit():
629 * list_lru_add() shrink_slab_memcg()
630 * list_add_tail() clear_bit()
632 * set_bit() do_shrink_slab()
634 smp_mb__after_atomic();
635 ret = do_shrink_slab(&sc, shrinker, priority);
636 if (ret == SHRINK_EMPTY)
639 memcg_set_shrinker_bit(memcg, nid, i);
643 if (rwsem_is_contended(&shrinker_rwsem)) {
649 up_read(&shrinker_rwsem);
652 #else /* CONFIG_MEMCG_KMEM */
653 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
654 struct mem_cgroup *memcg, int priority)
658 #endif /* CONFIG_MEMCG_KMEM */
661 * shrink_slab - shrink slab caches
662 * @gfp_mask: allocation context
663 * @nid: node whose slab caches to target
664 * @memcg: memory cgroup whose slab caches to target
665 * @priority: the reclaim priority
667 * Call the shrink functions to age shrinkable caches.
669 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
670 * unaware shrinkers will receive a node id of 0 instead.
672 * @memcg specifies the memory cgroup to target. Unaware shrinkers
673 * are called only if it is the root cgroup.
675 * @priority is sc->priority, we take the number of objects and >> by priority
676 * in order to get the scan target.
678 * Returns the number of reclaimed slab objects.
680 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
681 struct mem_cgroup *memcg,
684 unsigned long ret, freed = 0;
685 struct shrinker *shrinker;
687 if (!mem_cgroup_is_root(memcg))
688 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
690 if (!down_read_trylock(&shrinker_rwsem))
693 list_for_each_entry(shrinker, &shrinker_list, list) {
694 struct shrink_control sc = {
695 .gfp_mask = gfp_mask,
700 ret = do_shrink_slab(&sc, shrinker, priority);
701 if (ret == SHRINK_EMPTY)
705 * Bail out if someone want to register a new shrinker to
706 * prevent the regsitration from being stalled for long periods
707 * by parallel ongoing shrinking.
709 if (rwsem_is_contended(&shrinker_rwsem)) {
715 up_read(&shrinker_rwsem);
721 void drop_slab_node(int nid)
726 struct mem_cgroup *memcg = NULL;
729 memcg = mem_cgroup_iter(NULL, NULL, NULL);
731 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
732 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
733 } while (freed > 10);
740 for_each_online_node(nid)
744 static inline int is_page_cache_freeable(struct page *page)
747 * A freeable page cache page is referenced only by the caller
748 * that isolated the page, the page cache and optional buffer
749 * heads at page->private.
751 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
753 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
756 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
758 if (current->flags & PF_SWAPWRITE)
760 if (!inode_write_congested(inode))
762 if (inode_to_bdi(inode) == current->backing_dev_info)
768 * We detected a synchronous write error writing a page out. Probably
769 * -ENOSPC. We need to propagate that into the address_space for a subsequent
770 * fsync(), msync() or close().
772 * The tricky part is that after writepage we cannot touch the mapping: nothing
773 * prevents it from being freed up. But we have a ref on the page and once
774 * that page is locked, the mapping is pinned.
776 * We're allowed to run sleeping lock_page() here because we know the caller has
779 static void handle_write_error(struct address_space *mapping,
780 struct page *page, int error)
783 if (page_mapping(page) == mapping)
784 mapping_set_error(mapping, error);
788 /* possible outcome of pageout() */
790 /* failed to write page out, page is locked */
792 /* move page to the active list, page is locked */
794 /* page has been sent to the disk successfully, page is unlocked */
796 /* page is clean and locked */
801 * pageout is called by shrink_page_list() for each dirty page.
802 * Calls ->writepage().
804 static pageout_t pageout(struct page *page, struct address_space *mapping,
805 struct scan_control *sc)
808 * If the page is dirty, only perform writeback if that write
809 * will be non-blocking. To prevent this allocation from being
810 * stalled by pagecache activity. But note that there may be
811 * stalls if we need to run get_block(). We could test
812 * PagePrivate for that.
814 * If this process is currently in __generic_file_write_iter() against
815 * this page's queue, we can perform writeback even if that
818 * If the page is swapcache, write it back even if that would
819 * block, for some throttling. This happens by accident, because
820 * swap_backing_dev_info is bust: it doesn't reflect the
821 * congestion state of the swapdevs. Easy to fix, if needed.
823 if (!is_page_cache_freeable(page))
827 * Some data journaling orphaned pages can have
828 * page->mapping == NULL while being dirty with clean buffers.
830 if (page_has_private(page)) {
831 if (try_to_free_buffers(page)) {
832 ClearPageDirty(page);
833 pr_info("%s: orphaned page\n", __func__);
839 if (mapping->a_ops->writepage == NULL)
840 return PAGE_ACTIVATE;
841 if (!may_write_to_inode(mapping->host, sc))
844 if (clear_page_dirty_for_io(page)) {
846 struct writeback_control wbc = {
847 .sync_mode = WB_SYNC_NONE,
848 .nr_to_write = SWAP_CLUSTER_MAX,
850 .range_end = LLONG_MAX,
854 SetPageReclaim(page);
855 res = mapping->a_ops->writepage(page, &wbc);
857 handle_write_error(mapping, page, res);
858 if (res == AOP_WRITEPAGE_ACTIVATE) {
859 ClearPageReclaim(page);
860 return PAGE_ACTIVATE;
863 if (!PageWriteback(page)) {
864 /* synchronous write or broken a_ops? */
865 ClearPageReclaim(page);
867 trace_mm_vmscan_writepage(page);
868 inc_node_page_state(page, NR_VMSCAN_WRITE);
876 * Same as remove_mapping, but if the page is removed from the mapping, it
877 * gets returned with a refcount of 0.
879 static int __remove_mapping(struct address_space *mapping, struct page *page,
885 BUG_ON(!PageLocked(page));
886 BUG_ON(mapping != page_mapping(page));
888 xa_lock_irqsave(&mapping->i_pages, flags);
890 * The non racy check for a busy page.
892 * Must be careful with the order of the tests. When someone has
893 * a ref to the page, it may be possible that they dirty it then
894 * drop the reference. So if PageDirty is tested before page_count
895 * here, then the following race may occur:
897 * get_user_pages(&page);
898 * [user mapping goes away]
900 * !PageDirty(page) [good]
901 * SetPageDirty(page);
903 * !page_count(page) [good, discard it]
905 * [oops, our write_to data is lost]
907 * Reversing the order of the tests ensures such a situation cannot
908 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
909 * load is not satisfied before that of page->_refcount.
911 * Note that if SetPageDirty is always performed via set_page_dirty,
912 * and thus under the i_pages lock, then this ordering is not required.
914 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
915 refcount = 1 + HPAGE_PMD_NR;
918 if (!page_ref_freeze(page, refcount))
920 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
921 if (unlikely(PageDirty(page))) {
922 page_ref_unfreeze(page, refcount);
926 if (PageSwapCache(page)) {
927 swp_entry_t swap = { .val = page_private(page) };
928 mem_cgroup_swapout(page, swap);
929 __delete_from_swap_cache(page, swap);
930 xa_unlock_irqrestore(&mapping->i_pages, flags);
931 put_swap_page(page, swap);
933 void (*freepage)(struct page *);
936 freepage = mapping->a_ops->freepage;
938 * Remember a shadow entry for reclaimed file cache in
939 * order to detect refaults, thus thrashing, later on.
941 * But don't store shadows in an address space that is
942 * already exiting. This is not just an optizimation,
943 * inode reclaim needs to empty out the radix tree or
944 * the nodes are lost. Don't plant shadows behind its
947 * We also don't store shadows for DAX mappings because the
948 * only page cache pages found in these are zero pages
949 * covering holes, and because we don't want to mix DAX
950 * exceptional entries and shadow exceptional entries in the
951 * same address_space.
953 if (reclaimed && page_is_file_cache(page) &&
954 !mapping_exiting(mapping) && !dax_mapping(mapping))
955 shadow = workingset_eviction(mapping, page);
956 __delete_from_page_cache(page, shadow);
957 xa_unlock_irqrestore(&mapping->i_pages, flags);
959 if (freepage != NULL)
966 xa_unlock_irqrestore(&mapping->i_pages, flags);
971 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
972 * someone else has a ref on the page, abort and return 0. If it was
973 * successfully detached, return 1. Assumes the caller has a single ref on
976 int remove_mapping(struct address_space *mapping, struct page *page)
978 if (__remove_mapping(mapping, page, false)) {
980 * Unfreezing the refcount with 1 rather than 2 effectively
981 * drops the pagecache ref for us without requiring another
984 page_ref_unfreeze(page, 1);
991 * putback_lru_page - put previously isolated page onto appropriate LRU list
992 * @page: page to be put back to appropriate lru list
994 * Add previously isolated @page to appropriate LRU list.
995 * Page may still be unevictable for other reasons.
997 * lru_lock must not be held, interrupts must be enabled.
999 void putback_lru_page(struct page *page)
1001 lru_cache_add(page);
1002 put_page(page); /* drop ref from isolate */
1005 enum page_references {
1007 PAGEREF_RECLAIM_CLEAN,
1012 static enum page_references page_check_references(struct page *page,
1013 struct scan_control *sc)
1015 int referenced_ptes, referenced_page;
1016 unsigned long vm_flags;
1018 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1020 referenced_page = TestClearPageReferenced(page);
1023 * Mlock lost the isolation race with us. Let try_to_unmap()
1024 * move the page to the unevictable list.
1026 if (vm_flags & VM_LOCKED)
1027 return PAGEREF_RECLAIM;
1029 if (referenced_ptes) {
1030 if (PageSwapBacked(page))
1031 return PAGEREF_ACTIVATE;
1033 * All mapped pages start out with page table
1034 * references from the instantiating fault, so we need
1035 * to look twice if a mapped file page is used more
1038 * Mark it and spare it for another trip around the
1039 * inactive list. Another page table reference will
1040 * lead to its activation.
1042 * Note: the mark is set for activated pages as well
1043 * so that recently deactivated but used pages are
1044 * quickly recovered.
1046 SetPageReferenced(page);
1048 if (referenced_page || referenced_ptes > 1)
1049 return PAGEREF_ACTIVATE;
1052 * Activate file-backed executable pages after first usage.
1054 if (vm_flags & VM_EXEC)
1055 return PAGEREF_ACTIVATE;
1057 return PAGEREF_KEEP;
1060 /* Reclaim if clean, defer dirty pages to writeback */
1061 if (referenced_page && !PageSwapBacked(page))
1062 return PAGEREF_RECLAIM_CLEAN;
1064 return PAGEREF_RECLAIM;
1067 /* Check if a page is dirty or under writeback */
1068 static void page_check_dirty_writeback(struct page *page,
1069 bool *dirty, bool *writeback)
1071 struct address_space *mapping;
1074 * Anonymous pages are not handled by flushers and must be written
1075 * from reclaim context. Do not stall reclaim based on them
1077 if (!page_is_file_cache(page) ||
1078 (PageAnon(page) && !PageSwapBacked(page))) {
1084 /* By default assume that the page flags are accurate */
1085 *dirty = PageDirty(page);
1086 *writeback = PageWriteback(page);
1088 /* Verify dirty/writeback state if the filesystem supports it */
1089 if (!page_has_private(page))
1092 mapping = page_mapping(page);
1093 if (mapping && mapping->a_ops->is_dirty_writeback)
1094 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1098 * shrink_page_list() returns the number of reclaimed pages
1100 static unsigned long shrink_page_list(struct list_head *page_list,
1101 struct pglist_data *pgdat,
1102 struct scan_control *sc,
1103 enum ttu_flags ttu_flags,
1104 struct reclaim_stat *stat,
1107 LIST_HEAD(ret_pages);
1108 LIST_HEAD(free_pages);
1110 unsigned nr_unqueued_dirty = 0;
1111 unsigned nr_dirty = 0;
1112 unsigned nr_congested = 0;
1113 unsigned nr_reclaimed = 0;
1114 unsigned nr_writeback = 0;
1115 unsigned nr_immediate = 0;
1116 unsigned nr_ref_keep = 0;
1117 unsigned nr_unmap_fail = 0;
1121 while (!list_empty(page_list)) {
1122 struct address_space *mapping;
1125 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1126 bool dirty, writeback;
1130 page = lru_to_page(page_list);
1131 list_del(&page->lru);
1133 if (!trylock_page(page))
1136 VM_BUG_ON_PAGE(PageActive(page), page);
1140 if (unlikely(!page_evictable(page)))
1141 goto activate_locked;
1143 if (!sc->may_unmap && page_mapped(page))
1146 /* Double the slab pressure for mapped and swapcache pages */
1147 if ((page_mapped(page) || PageSwapCache(page)) &&
1148 !(PageAnon(page) && !PageSwapBacked(page)))
1151 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1152 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1155 * The number of dirty pages determines if a node is marked
1156 * reclaim_congested which affects wait_iff_congested. kswapd
1157 * will stall and start writing pages if the tail of the LRU
1158 * is all dirty unqueued pages.
1160 page_check_dirty_writeback(page, &dirty, &writeback);
1161 if (dirty || writeback)
1164 if (dirty && !writeback)
1165 nr_unqueued_dirty++;
1168 * Treat this page as congested if the underlying BDI is or if
1169 * pages are cycling through the LRU so quickly that the
1170 * pages marked for immediate reclaim are making it to the
1171 * end of the LRU a second time.
1173 mapping = page_mapping(page);
1174 if (((dirty || writeback) && mapping &&
1175 inode_write_congested(mapping->host)) ||
1176 (writeback && PageReclaim(page)))
1180 * If a page at the tail of the LRU is under writeback, there
1181 * are three cases to consider.
1183 * 1) If reclaim is encountering an excessive number of pages
1184 * under writeback and this page is both under writeback and
1185 * PageReclaim then it indicates that pages are being queued
1186 * for IO but are being recycled through the LRU before the
1187 * IO can complete. Waiting on the page itself risks an
1188 * indefinite stall if it is impossible to writeback the
1189 * page due to IO error or disconnected storage so instead
1190 * note that the LRU is being scanned too quickly and the
1191 * caller can stall after page list has been processed.
1193 * 2) Global or new memcg reclaim encounters a page that is
1194 * not marked for immediate reclaim, or the caller does not
1195 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1196 * not to fs). In this case mark the page for immediate
1197 * reclaim and continue scanning.
1199 * Require may_enter_fs because we would wait on fs, which
1200 * may not have submitted IO yet. And the loop driver might
1201 * enter reclaim, and deadlock if it waits on a page for
1202 * which it is needed to do the write (loop masks off
1203 * __GFP_IO|__GFP_FS for this reason); but more thought
1204 * would probably show more reasons.
1206 * 3) Legacy memcg encounters a page that is already marked
1207 * PageReclaim. memcg does not have any dirty pages
1208 * throttling so we could easily OOM just because too many
1209 * pages are in writeback and there is nothing else to
1210 * reclaim. Wait for the writeback to complete.
1212 * In cases 1) and 2) we activate the pages to get them out of
1213 * the way while we continue scanning for clean pages on the
1214 * inactive list and refilling from the active list. The
1215 * observation here is that waiting for disk writes is more
1216 * expensive than potentially causing reloads down the line.
1217 * Since they're marked for immediate reclaim, they won't put
1218 * memory pressure on the cache working set any longer than it
1219 * takes to write them to disk.
1221 if (PageWriteback(page)) {
1223 if (current_is_kswapd() &&
1224 PageReclaim(page) &&
1225 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1227 goto activate_locked;
1230 } else if (sane_reclaim(sc) ||
1231 !PageReclaim(page) || !may_enter_fs) {
1233 * This is slightly racy - end_page_writeback()
1234 * might have just cleared PageReclaim, then
1235 * setting PageReclaim here end up interpreted
1236 * as PageReadahead - but that does not matter
1237 * enough to care. What we do want is for this
1238 * page to have PageReclaim set next time memcg
1239 * reclaim reaches the tests above, so it will
1240 * then wait_on_page_writeback() to avoid OOM;
1241 * and it's also appropriate in global reclaim.
1243 SetPageReclaim(page);
1245 goto activate_locked;
1250 wait_on_page_writeback(page);
1251 /* then go back and try same page again */
1252 list_add_tail(&page->lru, page_list);
1258 references = page_check_references(page, sc);
1260 switch (references) {
1261 case PAGEREF_ACTIVATE:
1262 goto activate_locked;
1266 case PAGEREF_RECLAIM:
1267 case PAGEREF_RECLAIM_CLEAN:
1268 ; /* try to reclaim the page below */
1272 * Anonymous process memory has backing store?
1273 * Try to allocate it some swap space here.
1274 * Lazyfree page could be freed directly
1276 if (PageAnon(page) && PageSwapBacked(page)) {
1277 if (!PageSwapCache(page)) {
1278 if (!(sc->gfp_mask & __GFP_IO))
1280 if (PageTransHuge(page)) {
1281 /* cannot split THP, skip it */
1282 if (!can_split_huge_page(page, NULL))
1283 goto activate_locked;
1285 * Split pages without a PMD map right
1286 * away. Chances are some or all of the
1287 * tail pages can be freed without IO.
1289 if (!compound_mapcount(page) &&
1290 split_huge_page_to_list(page,
1292 goto activate_locked;
1294 if (!add_to_swap(page)) {
1295 if (!PageTransHuge(page))
1296 goto activate_locked;
1297 /* Fallback to swap normal pages */
1298 if (split_huge_page_to_list(page,
1300 goto activate_locked;
1301 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1302 count_vm_event(THP_SWPOUT_FALLBACK);
1304 if (!add_to_swap(page))
1305 goto activate_locked;
1310 /* Adding to swap updated mapping */
1311 mapping = page_mapping(page);
1313 } else if (unlikely(PageTransHuge(page))) {
1314 /* Split file THP */
1315 if (split_huge_page_to_list(page, page_list))
1320 * The page is mapped into the page tables of one or more
1321 * processes. Try to unmap it here.
1323 if (page_mapped(page)) {
1324 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1326 if (unlikely(PageTransHuge(page)))
1327 flags |= TTU_SPLIT_HUGE_PMD;
1328 if (!try_to_unmap(page, flags)) {
1330 goto activate_locked;
1334 if (PageDirty(page)) {
1336 * Only kswapd can writeback filesystem pages
1337 * to avoid risk of stack overflow. But avoid
1338 * injecting inefficient single-page IO into
1339 * flusher writeback as much as possible: only
1340 * write pages when we've encountered many
1341 * dirty pages, and when we've already scanned
1342 * the rest of the LRU for clean pages and see
1343 * the same dirty pages again (PageReclaim).
1345 if (page_is_file_cache(page) &&
1346 (!current_is_kswapd() || !PageReclaim(page) ||
1347 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1349 * Immediately reclaim when written back.
1350 * Similar in principal to deactivate_page()
1351 * except we already have the page isolated
1352 * and know it's dirty
1354 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1355 SetPageReclaim(page);
1357 goto activate_locked;
1360 if (references == PAGEREF_RECLAIM_CLEAN)
1364 if (!sc->may_writepage)
1368 * Page is dirty. Flush the TLB if a writable entry
1369 * potentially exists to avoid CPU writes after IO
1370 * starts and then write it out here.
1372 try_to_unmap_flush_dirty();
1373 switch (pageout(page, mapping, sc)) {
1377 goto activate_locked;
1379 if (PageWriteback(page))
1381 if (PageDirty(page))
1385 * A synchronous write - probably a ramdisk. Go
1386 * ahead and try to reclaim the page.
1388 if (!trylock_page(page))
1390 if (PageDirty(page) || PageWriteback(page))
1392 mapping = page_mapping(page);
1394 ; /* try to free the page below */
1399 * If the page has buffers, try to free the buffer mappings
1400 * associated with this page. If we succeed we try to free
1403 * We do this even if the page is PageDirty().
1404 * try_to_release_page() does not perform I/O, but it is
1405 * possible for a page to have PageDirty set, but it is actually
1406 * clean (all its buffers are clean). This happens if the
1407 * buffers were written out directly, with submit_bh(). ext3
1408 * will do this, as well as the blockdev mapping.
1409 * try_to_release_page() will discover that cleanness and will
1410 * drop the buffers and mark the page clean - it can be freed.
1412 * Rarely, pages can have buffers and no ->mapping. These are
1413 * the pages which were not successfully invalidated in
1414 * truncate_complete_page(). We try to drop those buffers here
1415 * and if that worked, and the page is no longer mapped into
1416 * process address space (page_count == 1) it can be freed.
1417 * Otherwise, leave the page on the LRU so it is swappable.
1419 if (page_has_private(page)) {
1420 if (!try_to_release_page(page, sc->gfp_mask))
1421 goto activate_locked;
1422 if (!mapping && page_count(page) == 1) {
1424 if (put_page_testzero(page))
1428 * rare race with speculative reference.
1429 * the speculative reference will free
1430 * this page shortly, so we may
1431 * increment nr_reclaimed here (and
1432 * leave it off the LRU).
1440 if (PageAnon(page) && !PageSwapBacked(page)) {
1441 /* follow __remove_mapping for reference */
1442 if (!page_ref_freeze(page, 1))
1444 if (PageDirty(page)) {
1445 page_ref_unfreeze(page, 1);
1449 count_vm_event(PGLAZYFREED);
1450 count_memcg_page_event(page, PGLAZYFREED);
1451 } else if (!mapping || !__remove_mapping(mapping, page, true))
1459 * Is there need to periodically free_page_list? It would
1460 * appear not as the counts should be low
1462 if (unlikely(PageTransHuge(page))) {
1463 mem_cgroup_uncharge(page);
1464 (*get_compound_page_dtor(page))(page);
1466 list_add(&page->lru, &free_pages);
1470 /* Not a candidate for swapping, so reclaim swap space. */
1471 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1473 try_to_free_swap(page);
1474 VM_BUG_ON_PAGE(PageActive(page), page);
1475 if (!PageMlocked(page)) {
1476 SetPageActive(page);
1478 count_memcg_page_event(page, PGACTIVATE);
1483 list_add(&page->lru, &ret_pages);
1484 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1487 mem_cgroup_uncharge_list(&free_pages);
1488 try_to_unmap_flush();
1489 free_unref_page_list(&free_pages);
1491 list_splice(&ret_pages, page_list);
1492 count_vm_events(PGACTIVATE, pgactivate);
1495 stat->nr_dirty = nr_dirty;
1496 stat->nr_congested = nr_congested;
1497 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1498 stat->nr_writeback = nr_writeback;
1499 stat->nr_immediate = nr_immediate;
1500 stat->nr_activate = pgactivate;
1501 stat->nr_ref_keep = nr_ref_keep;
1502 stat->nr_unmap_fail = nr_unmap_fail;
1504 return nr_reclaimed;
1507 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1508 struct list_head *page_list)
1510 struct scan_control sc = {
1511 .gfp_mask = GFP_KERNEL,
1512 .priority = DEF_PRIORITY,
1516 struct page *page, *next;
1517 LIST_HEAD(clean_pages);
1519 list_for_each_entry_safe(page, next, page_list, lru) {
1520 if (page_is_file_cache(page) && !PageDirty(page) &&
1521 !__PageMovable(page)) {
1522 ClearPageActive(page);
1523 list_move(&page->lru, &clean_pages);
1527 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1528 TTU_IGNORE_ACCESS, NULL, true);
1529 list_splice(&clean_pages, page_list);
1530 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1535 * Attempt to remove the specified page from its LRU. Only take this page
1536 * if it is of the appropriate PageActive status. Pages which are being
1537 * freed elsewhere are also ignored.
1539 * page: page to consider
1540 * mode: one of the LRU isolation modes defined above
1542 * returns 0 on success, -ve errno on failure.
1544 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1548 /* Only take pages on the LRU. */
1552 /* Compaction should not handle unevictable pages but CMA can do so */
1553 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1559 * To minimise LRU disruption, the caller can indicate that it only
1560 * wants to isolate pages it will be able to operate on without
1561 * blocking - clean pages for the most part.
1563 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1564 * that it is possible to migrate without blocking
1566 if (mode & ISOLATE_ASYNC_MIGRATE) {
1567 /* All the caller can do on PageWriteback is block */
1568 if (PageWriteback(page))
1571 if (PageDirty(page)) {
1572 struct address_space *mapping;
1576 * Only pages without mappings or that have a
1577 * ->migratepage callback are possible to migrate
1578 * without blocking. However, we can be racing with
1579 * truncation so it's necessary to lock the page
1580 * to stabilise the mapping as truncation holds
1581 * the page lock until after the page is removed
1582 * from the page cache.
1584 if (!trylock_page(page))
1587 mapping = page_mapping(page);
1588 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1595 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1598 if (likely(get_page_unless_zero(page))) {
1600 * Be careful not to clear PageLRU until after we're
1601 * sure the page is not being freed elsewhere -- the
1602 * page release code relies on it.
1613 * Update LRU sizes after isolating pages. The LRU size updates must
1614 * be complete before mem_cgroup_update_lru_size due to a santity check.
1616 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1617 enum lru_list lru, unsigned long *nr_zone_taken)
1621 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1622 if (!nr_zone_taken[zid])
1625 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1627 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1634 * zone_lru_lock is heavily contended. Some of the functions that
1635 * shrink the lists perform better by taking out a batch of pages
1636 * and working on them outside the LRU lock.
1638 * For pagecache intensive workloads, this function is the hottest
1639 * spot in the kernel (apart from copy_*_user functions).
1641 * Appropriate locks must be held before calling this function.
1643 * @nr_to_scan: The number of eligible pages to look through on the list.
1644 * @lruvec: The LRU vector to pull pages from.
1645 * @dst: The temp list to put pages on to.
1646 * @nr_scanned: The number of pages that were scanned.
1647 * @sc: The scan_control struct for this reclaim session
1648 * @mode: One of the LRU isolation modes
1649 * @lru: LRU list id for isolating
1651 * returns how many pages were moved onto *@dst.
1653 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1654 struct lruvec *lruvec, struct list_head *dst,
1655 unsigned long *nr_scanned, struct scan_control *sc,
1656 isolate_mode_t mode, enum lru_list lru)
1658 struct list_head *src = &lruvec->lists[lru];
1659 unsigned long nr_taken = 0;
1660 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1661 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1662 unsigned long skipped = 0;
1663 unsigned long scan, total_scan, nr_pages;
1664 LIST_HEAD(pages_skipped);
1667 for (total_scan = 0;
1668 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1672 page = lru_to_page(src);
1673 prefetchw_prev_lru_page(page, src, flags);
1675 VM_BUG_ON_PAGE(!PageLRU(page), page);
1677 if (page_zonenum(page) > sc->reclaim_idx) {
1678 list_move(&page->lru, &pages_skipped);
1679 nr_skipped[page_zonenum(page)]++;
1684 * Do not count skipped pages because that makes the function
1685 * return with no isolated pages if the LRU mostly contains
1686 * ineligible pages. This causes the VM to not reclaim any
1687 * pages, triggering a premature OOM.
1690 switch (__isolate_lru_page(page, mode)) {
1692 nr_pages = hpage_nr_pages(page);
1693 nr_taken += nr_pages;
1694 nr_zone_taken[page_zonenum(page)] += nr_pages;
1695 list_move(&page->lru, dst);
1699 /* else it is being freed elsewhere */
1700 list_move(&page->lru, src);
1709 * Splice any skipped pages to the start of the LRU list. Note that
1710 * this disrupts the LRU order when reclaiming for lower zones but
1711 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1712 * scanning would soon rescan the same pages to skip and put the
1713 * system at risk of premature OOM.
1715 if (!list_empty(&pages_skipped)) {
1718 list_splice(&pages_skipped, src);
1719 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1720 if (!nr_skipped[zid])
1723 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1724 skipped += nr_skipped[zid];
1727 *nr_scanned = total_scan;
1728 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1729 total_scan, skipped, nr_taken, mode, lru);
1730 update_lru_sizes(lruvec, lru, nr_zone_taken);
1735 * isolate_lru_page - tries to isolate a page from its LRU list
1736 * @page: page to isolate from its LRU list
1738 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1739 * vmstat statistic corresponding to whatever LRU list the page was on.
1741 * Returns 0 if the page was removed from an LRU list.
1742 * Returns -EBUSY if the page was not on an LRU list.
1744 * The returned page will have PageLRU() cleared. If it was found on
1745 * the active list, it will have PageActive set. If it was found on
1746 * the unevictable list, it will have the PageUnevictable bit set. That flag
1747 * may need to be cleared by the caller before letting the page go.
1749 * The vmstat statistic corresponding to the list on which the page was
1750 * found will be decremented.
1754 * (1) Must be called with an elevated refcount on the page. This is a
1755 * fundamentnal difference from isolate_lru_pages (which is called
1756 * without a stable reference).
1757 * (2) the lru_lock must not be held.
1758 * (3) interrupts must be enabled.
1760 int isolate_lru_page(struct page *page)
1764 VM_BUG_ON_PAGE(!page_count(page), page);
1765 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1767 if (PageLRU(page)) {
1768 struct zone *zone = page_zone(page);
1769 struct lruvec *lruvec;
1771 spin_lock_irq(zone_lru_lock(zone));
1772 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1773 if (PageLRU(page)) {
1774 int lru = page_lru(page);
1777 del_page_from_lru_list(page, lruvec, lru);
1780 spin_unlock_irq(zone_lru_lock(zone));
1786 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1787 * then get resheduled. When there are massive number of tasks doing page
1788 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1789 * the LRU list will go small and be scanned faster than necessary, leading to
1790 * unnecessary swapping, thrashing and OOM.
1792 static int too_many_isolated(struct pglist_data *pgdat, int file,
1793 struct scan_control *sc)
1795 unsigned long inactive, isolated;
1797 if (current_is_kswapd())
1800 if (!sane_reclaim(sc))
1804 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1805 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1807 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1808 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1812 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1813 * won't get blocked by normal direct-reclaimers, forming a circular
1816 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1819 return isolated > inactive;
1822 static noinline_for_stack void
1823 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1825 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1826 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1827 LIST_HEAD(pages_to_free);
1830 * Put back any unfreeable pages.
1832 while (!list_empty(page_list)) {
1833 struct page *page = lru_to_page(page_list);
1836 VM_BUG_ON_PAGE(PageLRU(page), page);
1837 list_del(&page->lru);
1838 if (unlikely(!page_evictable(page))) {
1839 spin_unlock_irq(&pgdat->lru_lock);
1840 putback_lru_page(page);
1841 spin_lock_irq(&pgdat->lru_lock);
1845 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1848 lru = page_lru(page);
1849 add_page_to_lru_list(page, lruvec, lru);
1851 if (is_active_lru(lru)) {
1852 int file = is_file_lru(lru);
1853 int numpages = hpage_nr_pages(page);
1854 reclaim_stat->recent_rotated[file] += numpages;
1856 if (put_page_testzero(page)) {
1857 __ClearPageLRU(page);
1858 __ClearPageActive(page);
1859 del_page_from_lru_list(page, lruvec, lru);
1861 if (unlikely(PageCompound(page))) {
1862 spin_unlock_irq(&pgdat->lru_lock);
1863 mem_cgroup_uncharge(page);
1864 (*get_compound_page_dtor(page))(page);
1865 spin_lock_irq(&pgdat->lru_lock);
1867 list_add(&page->lru, &pages_to_free);
1872 * To save our caller's stack, now use input list for pages to free.
1874 list_splice(&pages_to_free, page_list);
1878 * If a kernel thread (such as nfsd for loop-back mounts) services
1879 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1880 * In that case we should only throttle if the backing device it is
1881 * writing to is congested. In other cases it is safe to throttle.
1883 static int current_may_throttle(void)
1885 return !(current->flags & PF_LESS_THROTTLE) ||
1886 current->backing_dev_info == NULL ||
1887 bdi_write_congested(current->backing_dev_info);
1891 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1892 * of reclaimed pages
1894 static noinline_for_stack unsigned long
1895 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1896 struct scan_control *sc, enum lru_list lru)
1898 LIST_HEAD(page_list);
1899 unsigned long nr_scanned;
1900 unsigned long nr_reclaimed = 0;
1901 unsigned long nr_taken;
1902 struct reclaim_stat stat = {};
1903 isolate_mode_t isolate_mode = 0;
1904 int file = is_file_lru(lru);
1905 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1906 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1907 bool stalled = false;
1909 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1913 /* wait a bit for the reclaimer. */
1917 /* We are about to die and free our memory. Return now. */
1918 if (fatal_signal_pending(current))
1919 return SWAP_CLUSTER_MAX;
1925 isolate_mode |= ISOLATE_UNMAPPED;
1927 spin_lock_irq(&pgdat->lru_lock);
1929 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1930 &nr_scanned, sc, isolate_mode, lru);
1932 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1933 reclaim_stat->recent_scanned[file] += nr_taken;
1935 if (current_is_kswapd()) {
1936 if (global_reclaim(sc))
1937 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1938 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1941 if (global_reclaim(sc))
1942 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1943 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1946 spin_unlock_irq(&pgdat->lru_lock);
1951 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1954 spin_lock_irq(&pgdat->lru_lock);
1956 if (current_is_kswapd()) {
1957 if (global_reclaim(sc))
1958 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1959 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1962 if (global_reclaim(sc))
1963 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1964 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1968 putback_inactive_pages(lruvec, &page_list);
1970 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1972 spin_unlock_irq(&pgdat->lru_lock);
1974 mem_cgroup_uncharge_list(&page_list);
1975 free_unref_page_list(&page_list);
1978 * If dirty pages are scanned that are not queued for IO, it
1979 * implies that flushers are not doing their job. This can
1980 * happen when memory pressure pushes dirty pages to the end of
1981 * the LRU before the dirty limits are breached and the dirty
1982 * data has expired. It can also happen when the proportion of
1983 * dirty pages grows not through writes but through memory
1984 * pressure reclaiming all the clean cache. And in some cases,
1985 * the flushers simply cannot keep up with the allocation
1986 * rate. Nudge the flusher threads in case they are asleep.
1988 if (stat.nr_unqueued_dirty == nr_taken)
1989 wakeup_flusher_threads(WB_REASON_VMSCAN);
1991 sc->nr.dirty += stat.nr_dirty;
1992 sc->nr.congested += stat.nr_congested;
1993 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1994 sc->nr.writeback += stat.nr_writeback;
1995 sc->nr.immediate += stat.nr_immediate;
1996 sc->nr.taken += nr_taken;
1998 sc->nr.file_taken += nr_taken;
2000 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2001 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2002 return nr_reclaimed;
2006 * This moves pages from the active list to the inactive list.
2008 * We move them the other way if the page is referenced by one or more
2009 * processes, from rmap.
2011 * If the pages are mostly unmapped, the processing is fast and it is
2012 * appropriate to hold zone_lru_lock across the whole operation. But if
2013 * the pages are mapped, the processing is slow (page_referenced()) so we
2014 * should drop zone_lru_lock around each page. It's impossible to balance
2015 * this, so instead we remove the pages from the LRU while processing them.
2016 * It is safe to rely on PG_active against the non-LRU pages in here because
2017 * nobody will play with that bit on a non-LRU page.
2019 * The downside is that we have to touch page->_refcount against each page.
2020 * But we had to alter page->flags anyway.
2022 * Returns the number of pages moved to the given lru.
2025 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
2026 struct list_head *list,
2027 struct list_head *pages_to_free,
2030 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2035 while (!list_empty(list)) {
2036 page = lru_to_page(list);
2037 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2039 VM_BUG_ON_PAGE(PageLRU(page), page);
2042 nr_pages = hpage_nr_pages(page);
2043 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
2044 list_move(&page->lru, &lruvec->lists[lru]);
2046 if (put_page_testzero(page)) {
2047 __ClearPageLRU(page);
2048 __ClearPageActive(page);
2049 del_page_from_lru_list(page, lruvec, lru);
2051 if (unlikely(PageCompound(page))) {
2052 spin_unlock_irq(&pgdat->lru_lock);
2053 mem_cgroup_uncharge(page);
2054 (*get_compound_page_dtor(page))(page);
2055 spin_lock_irq(&pgdat->lru_lock);
2057 list_add(&page->lru, pages_to_free);
2059 nr_moved += nr_pages;
2063 if (!is_active_lru(lru)) {
2064 __count_vm_events(PGDEACTIVATE, nr_moved);
2065 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
2072 static void shrink_active_list(unsigned long nr_to_scan,
2073 struct lruvec *lruvec,
2074 struct scan_control *sc,
2077 unsigned long nr_taken;
2078 unsigned long nr_scanned;
2079 unsigned long vm_flags;
2080 LIST_HEAD(l_hold); /* The pages which were snipped off */
2081 LIST_HEAD(l_active);
2082 LIST_HEAD(l_inactive);
2084 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2085 unsigned nr_deactivate, nr_activate;
2086 unsigned nr_rotated = 0;
2087 isolate_mode_t isolate_mode = 0;
2088 int file = is_file_lru(lru);
2089 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2094 isolate_mode |= ISOLATE_UNMAPPED;
2096 spin_lock_irq(&pgdat->lru_lock);
2098 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2099 &nr_scanned, sc, isolate_mode, lru);
2101 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2102 reclaim_stat->recent_scanned[file] += nr_taken;
2104 __count_vm_events(PGREFILL, nr_scanned);
2105 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2107 spin_unlock_irq(&pgdat->lru_lock);
2109 while (!list_empty(&l_hold)) {
2111 page = lru_to_page(&l_hold);
2112 list_del(&page->lru);
2114 if (unlikely(!page_evictable(page))) {
2115 putback_lru_page(page);
2119 if (unlikely(buffer_heads_over_limit)) {
2120 if (page_has_private(page) && trylock_page(page)) {
2121 if (page_has_private(page))
2122 try_to_release_page(page, 0);
2127 if (page_referenced(page, 0, sc->target_mem_cgroup,
2129 nr_rotated += hpage_nr_pages(page);
2131 * Identify referenced, file-backed active pages and
2132 * give them one more trip around the active list. So
2133 * that executable code get better chances to stay in
2134 * memory under moderate memory pressure. Anon pages
2135 * are not likely to be evicted by use-once streaming
2136 * IO, plus JVM can create lots of anon VM_EXEC pages,
2137 * so we ignore them here.
2139 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2140 list_add(&page->lru, &l_active);
2145 ClearPageActive(page); /* we are de-activating */
2146 SetPageWorkingset(page);
2147 list_add(&page->lru, &l_inactive);
2151 * Move pages back to the lru list.
2153 spin_lock_irq(&pgdat->lru_lock);
2155 * Count referenced pages from currently used mappings as rotated,
2156 * even though only some of them are actually re-activated. This
2157 * helps balance scan pressure between file and anonymous pages in
2160 reclaim_stat->recent_rotated[file] += nr_rotated;
2162 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2163 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2164 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2165 spin_unlock_irq(&pgdat->lru_lock);
2167 mem_cgroup_uncharge_list(&l_hold);
2168 free_unref_page_list(&l_hold);
2169 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2170 nr_deactivate, nr_rotated, sc->priority, file);
2174 * The inactive anon list should be small enough that the VM never has
2175 * to do too much work.
2177 * The inactive file list should be small enough to leave most memory
2178 * to the established workingset on the scan-resistant active list,
2179 * but large enough to avoid thrashing the aggregate readahead window.
2181 * Both inactive lists should also be large enough that each inactive
2182 * page has a chance to be referenced again before it is reclaimed.
2184 * If that fails and refaulting is observed, the inactive list grows.
2186 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2187 * on this LRU, maintained by the pageout code. An inactive_ratio
2188 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2191 * memory ratio inactive
2192 * -------------------------------------
2201 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2202 struct mem_cgroup *memcg,
2203 struct scan_control *sc, bool actual_reclaim)
2205 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2206 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2207 enum lru_list inactive_lru = file * LRU_FILE;
2208 unsigned long inactive, active;
2209 unsigned long inactive_ratio;
2210 unsigned long refaults;
2214 * If we don't have swap space, anonymous page deactivation
2217 if (!file && !total_swap_pages)
2220 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2221 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2224 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2226 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2229 * When refaults are being observed, it means a new workingset
2230 * is being established. Disable active list protection to get
2231 * rid of the stale workingset quickly.
2233 if (file && actual_reclaim && lruvec->refaults != refaults) {
2236 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2238 inactive_ratio = int_sqrt(10 * gb);
2244 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2245 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2246 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2247 inactive_ratio, file);
2249 return inactive * inactive_ratio < active;
2252 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2253 struct lruvec *lruvec, struct mem_cgroup *memcg,
2254 struct scan_control *sc)
2256 if (is_active_lru(lru)) {
2257 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2259 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2263 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2274 * Determine how aggressively the anon and file LRU lists should be
2275 * scanned. The relative value of each set of LRU lists is determined
2276 * by looking at the fraction of the pages scanned we did rotate back
2277 * onto the active list instead of evict.
2279 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2280 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2282 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2283 struct scan_control *sc, unsigned long *nr,
2284 unsigned long *lru_pages)
2286 int swappiness = mem_cgroup_swappiness(memcg);
2287 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2289 u64 denominator = 0; /* gcc */
2290 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2291 unsigned long anon_prio, file_prio;
2292 enum scan_balance scan_balance;
2293 unsigned long anon, file;
2294 unsigned long ap, fp;
2297 /* If we have no swap space, do not bother scanning anon pages. */
2298 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2299 scan_balance = SCAN_FILE;
2304 * Global reclaim will swap to prevent OOM even with no
2305 * swappiness, but memcg users want to use this knob to
2306 * disable swapping for individual groups completely when
2307 * using the memory controller's swap limit feature would be
2310 if (!global_reclaim(sc) && !swappiness) {
2311 scan_balance = SCAN_FILE;
2316 * Do not apply any pressure balancing cleverness when the
2317 * system is close to OOM, scan both anon and file equally
2318 * (unless the swappiness setting disagrees with swapping).
2320 if (!sc->priority && swappiness) {
2321 scan_balance = SCAN_EQUAL;
2326 * Prevent the reclaimer from falling into the cache trap: as
2327 * cache pages start out inactive, every cache fault will tip
2328 * the scan balance towards the file LRU. And as the file LRU
2329 * shrinks, so does the window for rotation from references.
2330 * This means we have a runaway feedback loop where a tiny
2331 * thrashing file LRU becomes infinitely more attractive than
2332 * anon pages. Try to detect this based on file LRU size.
2334 if (global_reclaim(sc)) {
2335 unsigned long pgdatfile;
2336 unsigned long pgdatfree;
2338 unsigned long total_high_wmark = 0;
2340 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2341 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2342 node_page_state(pgdat, NR_INACTIVE_FILE);
2344 for (z = 0; z < MAX_NR_ZONES; z++) {
2345 struct zone *zone = &pgdat->node_zones[z];
2346 if (!managed_zone(zone))
2349 total_high_wmark += high_wmark_pages(zone);
2352 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2354 * Force SCAN_ANON if there are enough inactive
2355 * anonymous pages on the LRU in eligible zones.
2356 * Otherwise, the small LRU gets thrashed.
2358 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2359 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2361 scan_balance = SCAN_ANON;
2368 * If there is enough inactive page cache, i.e. if the size of the
2369 * inactive list is greater than that of the active list *and* the
2370 * inactive list actually has some pages to scan on this priority, we
2371 * do not reclaim anything from the anonymous working set right now.
2372 * Without the second condition we could end up never scanning an
2373 * lruvec even if it has plenty of old anonymous pages unless the
2374 * system is under heavy pressure.
2376 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2377 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2378 scan_balance = SCAN_FILE;
2382 scan_balance = SCAN_FRACT;
2385 * With swappiness at 100, anonymous and file have the same priority.
2386 * This scanning priority is essentially the inverse of IO cost.
2388 anon_prio = swappiness;
2389 file_prio = 200 - anon_prio;
2392 * OK, so we have swap space and a fair amount of page cache
2393 * pages. We use the recently rotated / recently scanned
2394 * ratios to determine how valuable each cache is.
2396 * Because workloads change over time (and to avoid overflow)
2397 * we keep these statistics as a floating average, which ends
2398 * up weighing recent references more than old ones.
2400 * anon in [0], file in [1]
2403 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2404 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2405 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2406 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2408 spin_lock_irq(&pgdat->lru_lock);
2409 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2410 reclaim_stat->recent_scanned[0] /= 2;
2411 reclaim_stat->recent_rotated[0] /= 2;
2414 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2415 reclaim_stat->recent_scanned[1] /= 2;
2416 reclaim_stat->recent_rotated[1] /= 2;
2420 * The amount of pressure on anon vs file pages is inversely
2421 * proportional to the fraction of recently scanned pages on
2422 * each list that were recently referenced and in active use.
2424 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2425 ap /= reclaim_stat->recent_rotated[0] + 1;
2427 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2428 fp /= reclaim_stat->recent_rotated[1] + 1;
2429 spin_unlock_irq(&pgdat->lru_lock);
2433 denominator = ap + fp + 1;
2436 for_each_evictable_lru(lru) {
2437 int file = is_file_lru(lru);
2441 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2442 scan = size >> sc->priority;
2444 * If the cgroup's already been deleted, make sure to
2445 * scrape out the remaining cache.
2447 if (!scan && !mem_cgroup_online(memcg))
2448 scan = min(size, SWAP_CLUSTER_MAX);
2450 switch (scan_balance) {
2452 /* Scan lists relative to size */
2456 * Scan types proportional to swappiness and
2457 * their relative recent reclaim efficiency.
2458 * Make sure we don't miss the last page
2459 * because of a round-off error.
2461 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2466 /* Scan one type exclusively */
2467 if ((scan_balance == SCAN_FILE) != file) {
2473 /* Look ma, no brain */
2483 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2485 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2486 struct scan_control *sc, unsigned long *lru_pages)
2488 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2489 unsigned long nr[NR_LRU_LISTS];
2490 unsigned long targets[NR_LRU_LISTS];
2491 unsigned long nr_to_scan;
2493 unsigned long nr_reclaimed = 0;
2494 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2495 struct blk_plug plug;
2498 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2500 /* Record the original scan target for proportional adjustments later */
2501 memcpy(targets, nr, sizeof(nr));
2504 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2505 * event that can occur when there is little memory pressure e.g.
2506 * multiple streaming readers/writers. Hence, we do not abort scanning
2507 * when the requested number of pages are reclaimed when scanning at
2508 * DEF_PRIORITY on the assumption that the fact we are direct
2509 * reclaiming implies that kswapd is not keeping up and it is best to
2510 * do a batch of work at once. For memcg reclaim one check is made to
2511 * abort proportional reclaim if either the file or anon lru has already
2512 * dropped to zero at the first pass.
2514 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2515 sc->priority == DEF_PRIORITY);
2517 blk_start_plug(&plug);
2518 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2519 nr[LRU_INACTIVE_FILE]) {
2520 unsigned long nr_anon, nr_file, percentage;
2521 unsigned long nr_scanned;
2523 for_each_evictable_lru(lru) {
2525 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2526 nr[lru] -= nr_to_scan;
2528 nr_reclaimed += shrink_list(lru, nr_to_scan,
2535 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2539 * For kswapd and memcg, reclaim at least the number of pages
2540 * requested. Ensure that the anon and file LRUs are scanned
2541 * proportionally what was requested by get_scan_count(). We
2542 * stop reclaiming one LRU and reduce the amount scanning
2543 * proportional to the original scan target.
2545 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2546 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2549 * It's just vindictive to attack the larger once the smaller
2550 * has gone to zero. And given the way we stop scanning the
2551 * smaller below, this makes sure that we only make one nudge
2552 * towards proportionality once we've got nr_to_reclaim.
2554 if (!nr_file || !nr_anon)
2557 if (nr_file > nr_anon) {
2558 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2559 targets[LRU_ACTIVE_ANON] + 1;
2561 percentage = nr_anon * 100 / scan_target;
2563 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2564 targets[LRU_ACTIVE_FILE] + 1;
2566 percentage = nr_file * 100 / scan_target;
2569 /* Stop scanning the smaller of the LRU */
2571 nr[lru + LRU_ACTIVE] = 0;
2574 * Recalculate the other LRU scan count based on its original
2575 * scan target and the percentage scanning already complete
2577 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2578 nr_scanned = targets[lru] - nr[lru];
2579 nr[lru] = targets[lru] * (100 - percentage) / 100;
2580 nr[lru] -= min(nr[lru], nr_scanned);
2583 nr_scanned = targets[lru] - nr[lru];
2584 nr[lru] = targets[lru] * (100 - percentage) / 100;
2585 nr[lru] -= min(nr[lru], nr_scanned);
2587 scan_adjusted = true;
2589 blk_finish_plug(&plug);
2590 sc->nr_reclaimed += nr_reclaimed;
2593 * Even if we did not try to evict anon pages at all, we want to
2594 * rebalance the anon lru active/inactive ratio.
2596 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2597 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2598 sc, LRU_ACTIVE_ANON);
2601 /* Use reclaim/compaction for costly allocs or under memory pressure */
2602 static bool in_reclaim_compaction(struct scan_control *sc)
2604 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2605 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2606 sc->priority < DEF_PRIORITY - 2))
2613 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2614 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2615 * true if more pages should be reclaimed such that when the page allocator
2616 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2617 * It will give up earlier than that if there is difficulty reclaiming pages.
2619 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2620 unsigned long nr_reclaimed,
2621 unsigned long nr_scanned,
2622 struct scan_control *sc)
2624 unsigned long pages_for_compaction;
2625 unsigned long inactive_lru_pages;
2628 /* If not in reclaim/compaction mode, stop */
2629 if (!in_reclaim_compaction(sc))
2632 /* Consider stopping depending on scan and reclaim activity */
2633 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2635 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2636 * full LRU list has been scanned and we are still failing
2637 * to reclaim pages. This full LRU scan is potentially
2638 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2640 if (!nr_reclaimed && !nr_scanned)
2644 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2645 * fail without consequence, stop if we failed to reclaim
2646 * any pages from the last SWAP_CLUSTER_MAX number of
2647 * pages that were scanned. This will return to the
2648 * caller faster at the risk reclaim/compaction and
2649 * the resulting allocation attempt fails
2656 * If we have not reclaimed enough pages for compaction and the
2657 * inactive lists are large enough, continue reclaiming
2659 pages_for_compaction = compact_gap(sc->order);
2660 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2661 if (get_nr_swap_pages() > 0)
2662 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2663 if (sc->nr_reclaimed < pages_for_compaction &&
2664 inactive_lru_pages > pages_for_compaction)
2667 /* If compaction would go ahead or the allocation would succeed, stop */
2668 for (z = 0; z <= sc->reclaim_idx; z++) {
2669 struct zone *zone = &pgdat->node_zones[z];
2670 if (!managed_zone(zone))
2673 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2674 case COMPACT_SUCCESS:
2675 case COMPACT_CONTINUE:
2678 /* check next zone */
2685 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2687 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2688 (memcg && memcg_congested(pgdat, memcg));
2691 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2693 struct reclaim_state *reclaim_state = current->reclaim_state;
2694 unsigned long nr_reclaimed, nr_scanned;
2695 bool reclaimable = false;
2698 struct mem_cgroup *root = sc->target_mem_cgroup;
2699 struct mem_cgroup_reclaim_cookie reclaim = {
2701 .priority = sc->priority,
2703 unsigned long node_lru_pages = 0;
2704 struct mem_cgroup *memcg;
2706 memset(&sc->nr, 0, sizeof(sc->nr));
2708 nr_reclaimed = sc->nr_reclaimed;
2709 nr_scanned = sc->nr_scanned;
2711 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2713 unsigned long lru_pages;
2714 unsigned long reclaimed;
2715 unsigned long scanned;
2717 switch (mem_cgroup_protected(root, memcg)) {
2718 case MEMCG_PROT_MIN:
2721 * If there is no reclaimable memory, OOM.
2724 case MEMCG_PROT_LOW:
2727 * Respect the protection only as long as
2728 * there is an unprotected supply
2729 * of reclaimable memory from other cgroups.
2731 if (!sc->memcg_low_reclaim) {
2732 sc->memcg_low_skipped = 1;
2735 memcg_memory_event(memcg, MEMCG_LOW);
2737 case MEMCG_PROT_NONE:
2741 reclaimed = sc->nr_reclaimed;
2742 scanned = sc->nr_scanned;
2743 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2744 node_lru_pages += lru_pages;
2746 if (sc->may_shrinkslab) {
2747 shrink_slab(sc->gfp_mask, pgdat->node_id,
2748 memcg, sc->priority);
2751 /* Record the group's reclaim efficiency */
2752 vmpressure(sc->gfp_mask, memcg, false,
2753 sc->nr_scanned - scanned,
2754 sc->nr_reclaimed - reclaimed);
2757 * Direct reclaim and kswapd have to scan all memory
2758 * cgroups to fulfill the overall scan target for the
2761 * Limit reclaim, on the other hand, only cares about
2762 * nr_to_reclaim pages to be reclaimed and it will
2763 * retry with decreasing priority if one round over the
2764 * whole hierarchy is not sufficient.
2766 if (!global_reclaim(sc) &&
2767 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2768 mem_cgroup_iter_break(root, memcg);
2771 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2773 if (reclaim_state) {
2774 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2775 reclaim_state->reclaimed_slab = 0;
2778 /* Record the subtree's reclaim efficiency */
2779 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2780 sc->nr_scanned - nr_scanned,
2781 sc->nr_reclaimed - nr_reclaimed);
2783 if (sc->nr_reclaimed - nr_reclaimed)
2786 if (current_is_kswapd()) {
2788 * If reclaim is isolating dirty pages under writeback,
2789 * it implies that the long-lived page allocation rate
2790 * is exceeding the page laundering rate. Either the
2791 * global limits are not being effective at throttling
2792 * processes due to the page distribution throughout
2793 * zones or there is heavy usage of a slow backing
2794 * device. The only option is to throttle from reclaim
2795 * context which is not ideal as there is no guarantee
2796 * the dirtying process is throttled in the same way
2797 * balance_dirty_pages() manages.
2799 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2800 * count the number of pages under pages flagged for
2801 * immediate reclaim and stall if any are encountered
2802 * in the nr_immediate check below.
2804 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2805 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2808 * Tag a node as congested if all the dirty pages
2809 * scanned were backed by a congested BDI and
2810 * wait_iff_congested will stall.
2812 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2813 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2815 /* Allow kswapd to start writing pages during reclaim.*/
2816 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2817 set_bit(PGDAT_DIRTY, &pgdat->flags);
2820 * If kswapd scans pages marked marked for immediate
2821 * reclaim and under writeback (nr_immediate), it
2822 * implies that pages are cycling through the LRU
2823 * faster than they are written so also forcibly stall.
2825 if (sc->nr.immediate)
2826 congestion_wait(BLK_RW_ASYNC, HZ/10);
2830 * Legacy memcg will stall in page writeback so avoid forcibly
2831 * stalling in wait_iff_congested().
2833 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2834 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2835 set_memcg_congestion(pgdat, root, true);
2838 * Stall direct reclaim for IO completions if underlying BDIs
2839 * and node is congested. Allow kswapd to continue until it
2840 * starts encountering unqueued dirty pages or cycling through
2841 * the LRU too quickly.
2843 if (!sc->hibernation_mode && !current_is_kswapd() &&
2844 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2845 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2847 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2848 sc->nr_scanned - nr_scanned, sc));
2851 * Kswapd gives up on balancing particular nodes after too
2852 * many failures to reclaim anything from them and goes to
2853 * sleep. On reclaim progress, reset the failure counter. A
2854 * successful direct reclaim run will revive a dormant kswapd.
2857 pgdat->kswapd_failures = 0;
2863 * Returns true if compaction should go ahead for a costly-order request, or
2864 * the allocation would already succeed without compaction. Return false if we
2865 * should reclaim first.
2867 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2869 unsigned long watermark;
2870 enum compact_result suitable;
2872 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2873 if (suitable == COMPACT_SUCCESS)
2874 /* Allocation should succeed already. Don't reclaim. */
2876 if (suitable == COMPACT_SKIPPED)
2877 /* Compaction cannot yet proceed. Do reclaim. */
2881 * Compaction is already possible, but it takes time to run and there
2882 * are potentially other callers using the pages just freed. So proceed
2883 * with reclaim to make a buffer of free pages available to give
2884 * compaction a reasonable chance of completing and allocating the page.
2885 * Note that we won't actually reclaim the whole buffer in one attempt
2886 * as the target watermark in should_continue_reclaim() is lower. But if
2887 * we are already above the high+gap watermark, don't reclaim at all.
2889 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2891 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2895 * This is the direct reclaim path, for page-allocating processes. We only
2896 * try to reclaim pages from zones which will satisfy the caller's allocation
2899 * If a zone is deemed to be full of pinned pages then just give it a light
2900 * scan then give up on it.
2902 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2906 unsigned long nr_soft_reclaimed;
2907 unsigned long nr_soft_scanned;
2909 pg_data_t *last_pgdat = NULL;
2912 * If the number of buffer_heads in the machine exceeds the maximum
2913 * allowed level, force direct reclaim to scan the highmem zone as
2914 * highmem pages could be pinning lowmem pages storing buffer_heads
2916 orig_mask = sc->gfp_mask;
2917 if (buffer_heads_over_limit) {
2918 sc->gfp_mask |= __GFP_HIGHMEM;
2919 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2922 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2923 sc->reclaim_idx, sc->nodemask) {
2925 * Take care memory controller reclaiming has small influence
2928 if (global_reclaim(sc)) {
2929 if (!cpuset_zone_allowed(zone,
2930 GFP_KERNEL | __GFP_HARDWALL))
2934 * If we already have plenty of memory free for
2935 * compaction in this zone, don't free any more.
2936 * Even though compaction is invoked for any
2937 * non-zero order, only frequent costly order
2938 * reclamation is disruptive enough to become a
2939 * noticeable problem, like transparent huge
2942 if (IS_ENABLED(CONFIG_COMPACTION) &&
2943 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2944 compaction_ready(zone, sc)) {
2945 sc->compaction_ready = true;
2950 * Shrink each node in the zonelist once. If the
2951 * zonelist is ordered by zone (not the default) then a
2952 * node may be shrunk multiple times but in that case
2953 * the user prefers lower zones being preserved.
2955 if (zone->zone_pgdat == last_pgdat)
2959 * This steals pages from memory cgroups over softlimit
2960 * and returns the number of reclaimed pages and
2961 * scanned pages. This works for global memory pressure
2962 * and balancing, not for a memcg's limit.
2964 nr_soft_scanned = 0;
2965 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2966 sc->order, sc->gfp_mask,
2968 sc->nr_reclaimed += nr_soft_reclaimed;
2969 sc->nr_scanned += nr_soft_scanned;
2970 /* need some check for avoid more shrink_zone() */
2973 /* See comment about same check for global reclaim above */
2974 if (zone->zone_pgdat == last_pgdat)
2976 last_pgdat = zone->zone_pgdat;
2977 shrink_node(zone->zone_pgdat, sc);
2981 * Restore to original mask to avoid the impact on the caller if we
2982 * promoted it to __GFP_HIGHMEM.
2984 sc->gfp_mask = orig_mask;
2987 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2989 struct mem_cgroup *memcg;
2991 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2993 unsigned long refaults;
2994 struct lruvec *lruvec;
2997 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2999 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
3001 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3002 lruvec->refaults = refaults;
3003 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3007 * This is the main entry point to direct page reclaim.
3009 * If a full scan of the inactive list fails to free enough memory then we
3010 * are "out of memory" and something needs to be killed.
3012 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3013 * high - the zone may be full of dirty or under-writeback pages, which this
3014 * caller can't do much about. We kick the writeback threads and take explicit
3015 * naps in the hope that some of these pages can be written. But if the
3016 * allocating task holds filesystem locks which prevent writeout this might not
3017 * work, and the allocation attempt will fail.
3019 * returns: 0, if no pages reclaimed
3020 * else, the number of pages reclaimed
3022 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3023 struct scan_control *sc)
3025 int initial_priority = sc->priority;
3026 pg_data_t *last_pgdat;
3030 delayacct_freepages_start();
3032 if (global_reclaim(sc))
3033 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3036 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3039 shrink_zones(zonelist, sc);
3041 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3044 if (sc->compaction_ready)
3048 * If we're getting trouble reclaiming, start doing
3049 * writepage even in laptop mode.
3051 if (sc->priority < DEF_PRIORITY - 2)
3052 sc->may_writepage = 1;
3053 } while (--sc->priority >= 0);
3056 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3058 if (zone->zone_pgdat == last_pgdat)
3060 last_pgdat = zone->zone_pgdat;
3061 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3062 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3065 delayacct_freepages_end();
3067 if (sc->nr_reclaimed)
3068 return sc->nr_reclaimed;
3070 /* Aborted reclaim to try compaction? don't OOM, then */
3071 if (sc->compaction_ready)
3074 /* Untapped cgroup reserves? Don't OOM, retry. */
3075 if (sc->memcg_low_skipped) {
3076 sc->priority = initial_priority;
3077 sc->memcg_low_reclaim = 1;
3078 sc->memcg_low_skipped = 0;
3085 static bool allow_direct_reclaim(pg_data_t *pgdat)
3088 unsigned long pfmemalloc_reserve = 0;
3089 unsigned long free_pages = 0;
3093 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3096 for (i = 0; i <= ZONE_NORMAL; i++) {
3097 zone = &pgdat->node_zones[i];
3098 if (!managed_zone(zone))
3101 if (!zone_reclaimable_pages(zone))
3104 pfmemalloc_reserve += min_wmark_pages(zone);
3105 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3108 /* If there are no reserves (unexpected config) then do not throttle */
3109 if (!pfmemalloc_reserve)
3112 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3114 /* kswapd must be awake if processes are being throttled */
3115 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3116 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3117 (enum zone_type)ZONE_NORMAL);
3118 wake_up_interruptible(&pgdat->kswapd_wait);
3125 * Throttle direct reclaimers if backing storage is backed by the network
3126 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3127 * depleted. kswapd will continue to make progress and wake the processes
3128 * when the low watermark is reached.
3130 * Returns true if a fatal signal was delivered during throttling. If this
3131 * happens, the page allocator should not consider triggering the OOM killer.
3133 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3134 nodemask_t *nodemask)
3138 pg_data_t *pgdat = NULL;
3141 * Kernel threads should not be throttled as they may be indirectly
3142 * responsible for cleaning pages necessary for reclaim to make forward
3143 * progress. kjournald for example may enter direct reclaim while
3144 * committing a transaction where throttling it could forcing other
3145 * processes to block on log_wait_commit().
3147 if (current->flags & PF_KTHREAD)
3151 * If a fatal signal is pending, this process should not throttle.
3152 * It should return quickly so it can exit and free its memory
3154 if (fatal_signal_pending(current))
3158 * Check if the pfmemalloc reserves are ok by finding the first node
3159 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3160 * GFP_KERNEL will be required for allocating network buffers when
3161 * swapping over the network so ZONE_HIGHMEM is unusable.
3163 * Throttling is based on the first usable node and throttled processes
3164 * wait on a queue until kswapd makes progress and wakes them. There
3165 * is an affinity then between processes waking up and where reclaim
3166 * progress has been made assuming the process wakes on the same node.
3167 * More importantly, processes running on remote nodes will not compete
3168 * for remote pfmemalloc reserves and processes on different nodes
3169 * should make reasonable progress.
3171 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3172 gfp_zone(gfp_mask), nodemask) {
3173 if (zone_idx(zone) > ZONE_NORMAL)
3176 /* Throttle based on the first usable node */
3177 pgdat = zone->zone_pgdat;
3178 if (allow_direct_reclaim(pgdat))
3183 /* If no zone was usable by the allocation flags then do not throttle */
3187 /* Account for the throttling */
3188 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3191 * If the caller cannot enter the filesystem, it's possible that it
3192 * is due to the caller holding an FS lock or performing a journal
3193 * transaction in the case of a filesystem like ext[3|4]. In this case,
3194 * it is not safe to block on pfmemalloc_wait as kswapd could be
3195 * blocked waiting on the same lock. Instead, throttle for up to a
3196 * second before continuing.
3198 if (!(gfp_mask & __GFP_FS)) {
3199 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3200 allow_direct_reclaim(pgdat), HZ);
3205 /* Throttle until kswapd wakes the process */
3206 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3207 allow_direct_reclaim(pgdat));
3210 if (fatal_signal_pending(current))
3217 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3218 gfp_t gfp_mask, nodemask_t *nodemask)
3220 unsigned long nr_reclaimed;
3221 struct scan_control sc = {
3222 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3223 .gfp_mask = current_gfp_context(gfp_mask),
3224 .reclaim_idx = gfp_zone(gfp_mask),
3226 .nodemask = nodemask,
3227 .priority = DEF_PRIORITY,
3228 .may_writepage = !laptop_mode,
3231 .may_shrinkslab = 1,
3235 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3236 * Confirm they are large enough for max values.
3238 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3239 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3240 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3243 * Do not enter reclaim if fatal signal was delivered while throttled.
3244 * 1 is returned so that the page allocator does not OOM kill at this
3247 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3250 trace_mm_vmscan_direct_reclaim_begin(order,
3255 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3257 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3259 return nr_reclaimed;
3264 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3265 gfp_t gfp_mask, bool noswap,
3267 unsigned long *nr_scanned)
3269 struct scan_control sc = {
3270 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3271 .target_mem_cgroup = memcg,
3272 .may_writepage = !laptop_mode,
3274 .reclaim_idx = MAX_NR_ZONES - 1,
3275 .may_swap = !noswap,
3276 .may_shrinkslab = 1,
3278 unsigned long lru_pages;
3280 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3281 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3283 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3289 * NOTE: Although we can get the priority field, using it
3290 * here is not a good idea, since it limits the pages we can scan.
3291 * if we don't reclaim here, the shrink_node from balance_pgdat
3292 * will pick up pages from other mem cgroup's as well. We hack
3293 * the priority and make it zero.
3295 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3297 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3299 *nr_scanned = sc.nr_scanned;
3300 return sc.nr_reclaimed;
3303 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3304 unsigned long nr_pages,
3308 struct zonelist *zonelist;
3309 unsigned long nr_reclaimed;
3310 unsigned long pflags;
3312 unsigned int noreclaim_flag;
3313 struct scan_control sc = {
3314 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3315 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3316 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3317 .reclaim_idx = MAX_NR_ZONES - 1,
3318 .target_mem_cgroup = memcg,
3319 .priority = DEF_PRIORITY,
3320 .may_writepage = !laptop_mode,
3322 .may_swap = may_swap,
3323 .may_shrinkslab = 1,
3327 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3328 * take care of from where we get pages. So the node where we start the
3329 * scan does not need to be the current node.
3331 nid = mem_cgroup_select_victim_node(memcg);
3333 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3335 trace_mm_vmscan_memcg_reclaim_begin(0,
3340 psi_memstall_enter(&pflags);
3341 noreclaim_flag = memalloc_noreclaim_save();
3343 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3345 memalloc_noreclaim_restore(noreclaim_flag);
3346 psi_memstall_leave(&pflags);
3348 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3350 return nr_reclaimed;
3354 static void age_active_anon(struct pglist_data *pgdat,
3355 struct scan_control *sc)
3357 struct mem_cgroup *memcg;
3359 if (!total_swap_pages)
3362 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3364 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3366 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3367 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3368 sc, LRU_ACTIVE_ANON);
3370 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3374 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3380 * Check for watermark boosts top-down as the higher zones
3381 * are more likely to be boosted. Both watermarks and boosts
3382 * should not be checked at the time time as reclaim would
3383 * start prematurely when there is no boosting and a lower
3386 for (i = classzone_idx; i >= 0; i--) {
3387 zone = pgdat->node_zones + i;
3388 if (!managed_zone(zone))
3391 if (zone->watermark_boost)
3399 * Returns true if there is an eligible zone balanced for the request order
3402 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3405 unsigned long mark = -1;
3409 * Check watermarks bottom-up as lower zones are more likely to
3412 for (i = 0; i <= classzone_idx; i++) {
3413 zone = pgdat->node_zones + i;
3415 if (!managed_zone(zone))
3418 mark = high_wmark_pages(zone);
3419 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3424 * If a node has no populated zone within classzone_idx, it does not
3425 * need balancing by definition. This can happen if a zone-restricted
3426 * allocation tries to wake a remote kswapd.
3434 /* Clear pgdat state for congested, dirty or under writeback. */
3435 static void clear_pgdat_congested(pg_data_t *pgdat)
3437 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3438 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3439 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3443 * Prepare kswapd for sleeping. This verifies that there are no processes
3444 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3446 * Returns true if kswapd is ready to sleep
3448 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3451 * The throttled processes are normally woken up in balance_pgdat() as
3452 * soon as allow_direct_reclaim() is true. But there is a potential
3453 * race between when kswapd checks the watermarks and a process gets
3454 * throttled. There is also a potential race if processes get
3455 * throttled, kswapd wakes, a large process exits thereby balancing the
3456 * zones, which causes kswapd to exit balance_pgdat() before reaching
3457 * the wake up checks. If kswapd is going to sleep, no process should
3458 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3459 * the wake up is premature, processes will wake kswapd and get
3460 * throttled again. The difference from wake ups in balance_pgdat() is
3461 * that here we are under prepare_to_wait().
3463 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3464 wake_up_all(&pgdat->pfmemalloc_wait);
3466 /* Hopeless node, leave it to direct reclaim */
3467 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3470 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3471 clear_pgdat_congested(pgdat);
3479 * kswapd shrinks a node of pages that are at or below the highest usable
3480 * zone that is currently unbalanced.
3482 * Returns true if kswapd scanned at least the requested number of pages to
3483 * reclaim or if the lack of progress was due to pages under writeback.
3484 * This is used to determine if the scanning priority needs to be raised.
3486 static bool kswapd_shrink_node(pg_data_t *pgdat,
3487 struct scan_control *sc)
3492 /* Reclaim a number of pages proportional to the number of zones */
3493 sc->nr_to_reclaim = 0;
3494 for (z = 0; z <= sc->reclaim_idx; z++) {
3495 zone = pgdat->node_zones + z;
3496 if (!managed_zone(zone))
3499 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3503 * Historically care was taken to put equal pressure on all zones but
3504 * now pressure is applied based on node LRU order.
3506 shrink_node(pgdat, sc);
3509 * Fragmentation may mean that the system cannot be rebalanced for
3510 * high-order allocations. If twice the allocation size has been
3511 * reclaimed then recheck watermarks only at order-0 to prevent
3512 * excessive reclaim. Assume that a process requested a high-order
3513 * can direct reclaim/compact.
3515 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3518 return sc->nr_scanned >= sc->nr_to_reclaim;
3522 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3523 * that are eligible for use by the caller until at least one zone is
3526 * Returns the order kswapd finished reclaiming at.
3528 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3529 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3530 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3531 * or lower is eligible for reclaim until at least one usable zone is
3534 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3537 unsigned long nr_soft_reclaimed;
3538 unsigned long nr_soft_scanned;
3539 unsigned long pflags;
3540 unsigned long nr_boost_reclaim;
3541 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3544 struct scan_control sc = {
3545 .gfp_mask = GFP_KERNEL,
3550 psi_memstall_enter(&pflags);
3551 __fs_reclaim_acquire();
3553 count_vm_event(PAGEOUTRUN);
3556 * Account for the reclaim boost. Note that the zone boost is left in
3557 * place so that parallel allocations that are near the watermark will
3558 * stall or direct reclaim until kswapd is finished.
3560 nr_boost_reclaim = 0;
3561 for (i = 0; i <= classzone_idx; i++) {
3562 zone = pgdat->node_zones + i;
3563 if (!managed_zone(zone))
3566 nr_boost_reclaim += zone->watermark_boost;
3567 zone_boosts[i] = zone->watermark_boost;
3569 boosted = nr_boost_reclaim;
3572 sc.priority = DEF_PRIORITY;
3574 unsigned long nr_reclaimed = sc.nr_reclaimed;
3575 bool raise_priority = true;
3579 sc.reclaim_idx = classzone_idx;
3582 * If the number of buffer_heads exceeds the maximum allowed
3583 * then consider reclaiming from all zones. This has a dual
3584 * purpose -- on 64-bit systems it is expected that
3585 * buffer_heads are stripped during active rotation. On 32-bit
3586 * systems, highmem pages can pin lowmem memory and shrinking
3587 * buffers can relieve lowmem pressure. Reclaim may still not
3588 * go ahead if all eligible zones for the original allocation
3589 * request are balanced to avoid excessive reclaim from kswapd.
3591 if (buffer_heads_over_limit) {
3592 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3593 zone = pgdat->node_zones + i;
3594 if (!managed_zone(zone))
3603 * If the pgdat is imbalanced then ignore boosting and preserve
3604 * the watermarks for a later time and restart. Note that the
3605 * zone watermarks will be still reset at the end of balancing
3606 * on the grounds that the normal reclaim should be enough to
3607 * re-evaluate if boosting is required when kswapd next wakes.
3609 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3610 if (!balanced && nr_boost_reclaim) {
3611 nr_boost_reclaim = 0;
3616 * If boosting is not active then only reclaim if there are no
3617 * eligible zones. Note that sc.reclaim_idx is not used as
3618 * buffer_heads_over_limit may have adjusted it.
3620 if (!nr_boost_reclaim && balanced)
3623 /* Limit the priority of boosting to avoid reclaim writeback */
3624 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3625 raise_priority = false;
3628 * Do not writeback or swap pages for boosted reclaim. The
3629 * intent is to relieve pressure not issue sub-optimal IO
3630 * from reclaim context. If no pages are reclaimed, the
3631 * reclaim will be aborted.
3633 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3634 sc.may_swap = !nr_boost_reclaim;
3635 sc.may_shrinkslab = !nr_boost_reclaim;
3638 * Do some background aging of the anon list, to give
3639 * pages a chance to be referenced before reclaiming. All
3640 * pages are rotated regardless of classzone as this is
3641 * about consistent aging.
3643 age_active_anon(pgdat, &sc);
3646 * If we're getting trouble reclaiming, start doing writepage
3647 * even in laptop mode.
3649 if (sc.priority < DEF_PRIORITY - 2)
3650 sc.may_writepage = 1;
3652 /* Call soft limit reclaim before calling shrink_node. */
3654 nr_soft_scanned = 0;
3655 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3656 sc.gfp_mask, &nr_soft_scanned);
3657 sc.nr_reclaimed += nr_soft_reclaimed;
3660 * There should be no need to raise the scanning priority if
3661 * enough pages are already being scanned that that high
3662 * watermark would be met at 100% efficiency.
3664 if (kswapd_shrink_node(pgdat, &sc))
3665 raise_priority = false;
3668 * If the low watermark is met there is no need for processes
3669 * to be throttled on pfmemalloc_wait as they should not be
3670 * able to safely make forward progress. Wake them
3672 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3673 allow_direct_reclaim(pgdat))
3674 wake_up_all(&pgdat->pfmemalloc_wait);
3676 /* Check if kswapd should be suspending */
3677 __fs_reclaim_release();
3678 ret = try_to_freeze();
3679 __fs_reclaim_acquire();
3680 if (ret || kthread_should_stop())
3684 * Raise priority if scanning rate is too low or there was no
3685 * progress in reclaiming pages
3687 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3688 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3691 * If reclaim made no progress for a boost, stop reclaim as
3692 * IO cannot be queued and it could be an infinite loop in
3693 * extreme circumstances.
3695 if (nr_boost_reclaim && !nr_reclaimed)
3698 if (raise_priority || !nr_reclaimed)
3700 } while (sc.priority >= 1);
3702 if (!sc.nr_reclaimed)
3703 pgdat->kswapd_failures++;
3706 /* If reclaim was boosted, account for the reclaim done in this pass */
3708 unsigned long flags;
3710 for (i = 0; i <= classzone_idx; i++) {
3711 if (!zone_boosts[i])
3714 /* Increments are under the zone lock */
3715 zone = pgdat->node_zones + i;
3716 spin_lock_irqsave(&zone->lock, flags);
3717 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3718 spin_unlock_irqrestore(&zone->lock, flags);
3722 * As there is now likely space, wakeup kcompact to defragment
3725 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3728 snapshot_refaults(NULL, pgdat);
3729 __fs_reclaim_release();
3730 psi_memstall_leave(&pflags);
3732 * Return the order kswapd stopped reclaiming at as
3733 * prepare_kswapd_sleep() takes it into account. If another caller
3734 * entered the allocator slow path while kswapd was awake, order will
3735 * remain at the higher level.
3741 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3742 * allocation request woke kswapd for. When kswapd has not woken recently,
3743 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3744 * given classzone and returns it or the highest classzone index kswapd
3745 * was recently woke for.
3747 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3748 enum zone_type classzone_idx)
3750 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3751 return classzone_idx;
3753 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3756 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3757 unsigned int classzone_idx)
3762 if (freezing(current) || kthread_should_stop())
3765 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3768 * Try to sleep for a short interval. Note that kcompactd will only be
3769 * woken if it is possible to sleep for a short interval. This is
3770 * deliberate on the assumption that if reclaim cannot keep an
3771 * eligible zone balanced that it's also unlikely that compaction will
3774 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3776 * Compaction records what page blocks it recently failed to
3777 * isolate pages from and skips them in the future scanning.
3778 * When kswapd is going to sleep, it is reasonable to assume
3779 * that pages and compaction may succeed so reset the cache.
3781 reset_isolation_suitable(pgdat);
3784 * We have freed the memory, now we should compact it to make
3785 * allocation of the requested order possible.
3787 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3789 remaining = schedule_timeout(HZ/10);
3792 * If woken prematurely then reset kswapd_classzone_idx and
3793 * order. The values will either be from a wakeup request or
3794 * the previous request that slept prematurely.
3797 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3798 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3801 finish_wait(&pgdat->kswapd_wait, &wait);
3802 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3806 * After a short sleep, check if it was a premature sleep. If not, then
3807 * go fully to sleep until explicitly woken up.
3810 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3811 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3814 * vmstat counters are not perfectly accurate and the estimated
3815 * value for counters such as NR_FREE_PAGES can deviate from the
3816 * true value by nr_online_cpus * threshold. To avoid the zone
3817 * watermarks being breached while under pressure, we reduce the
3818 * per-cpu vmstat threshold while kswapd is awake and restore
3819 * them before going back to sleep.
3821 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3823 if (!kthread_should_stop())
3826 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3829 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3831 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3833 finish_wait(&pgdat->kswapd_wait, &wait);
3837 * The background pageout daemon, started as a kernel thread
3838 * from the init process.
3840 * This basically trickles out pages so that we have _some_
3841 * free memory available even if there is no other activity
3842 * that frees anything up. This is needed for things like routing
3843 * etc, where we otherwise might have all activity going on in
3844 * asynchronous contexts that cannot page things out.
3846 * If there are applications that are active memory-allocators
3847 * (most normal use), this basically shouldn't matter.
3849 static int kswapd(void *p)
3851 unsigned int alloc_order, reclaim_order;
3852 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3853 pg_data_t *pgdat = (pg_data_t*)p;
3854 struct task_struct *tsk = current;
3856 struct reclaim_state reclaim_state = {
3857 .reclaimed_slab = 0,
3859 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3861 if (!cpumask_empty(cpumask))
3862 set_cpus_allowed_ptr(tsk, cpumask);
3863 current->reclaim_state = &reclaim_state;
3866 * Tell the memory management that we're a "memory allocator",
3867 * and that if we need more memory we should get access to it
3868 * regardless (see "__alloc_pages()"). "kswapd" should
3869 * never get caught in the normal page freeing logic.
3871 * (Kswapd normally doesn't need memory anyway, but sometimes
3872 * you need a small amount of memory in order to be able to
3873 * page out something else, and this flag essentially protects
3874 * us from recursively trying to free more memory as we're
3875 * trying to free the first piece of memory in the first place).
3877 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3880 pgdat->kswapd_order = 0;
3881 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3885 alloc_order = reclaim_order = pgdat->kswapd_order;
3886 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3889 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3892 /* Read the new order and classzone_idx */
3893 alloc_order = reclaim_order = pgdat->kswapd_order;
3894 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3895 pgdat->kswapd_order = 0;
3896 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3898 ret = try_to_freeze();
3899 if (kthread_should_stop())
3903 * We can speed up thawing tasks if we don't call balance_pgdat
3904 * after returning from the refrigerator
3910 * Reclaim begins at the requested order but if a high-order
3911 * reclaim fails then kswapd falls back to reclaiming for
3912 * order-0. If that happens, kswapd will consider sleeping
3913 * for the order it finished reclaiming at (reclaim_order)
3914 * but kcompactd is woken to compact for the original
3915 * request (alloc_order).
3917 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3919 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3920 if (reclaim_order < alloc_order)
3921 goto kswapd_try_sleep;
3924 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3925 current->reclaim_state = NULL;
3931 * A zone is low on free memory or too fragmented for high-order memory. If
3932 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3933 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3934 * has failed or is not needed, still wake up kcompactd if only compaction is
3937 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3938 enum zone_type classzone_idx)
3942 if (!managed_zone(zone))
3945 if (!cpuset_zone_allowed(zone, gfp_flags))
3947 pgdat = zone->zone_pgdat;
3948 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3950 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3951 if (!waitqueue_active(&pgdat->kswapd_wait))
3954 /* Hopeless node, leave it to direct reclaim if possible */
3955 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3956 (pgdat_balanced(pgdat, order, classzone_idx) &&
3957 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3959 * There may be plenty of free memory available, but it's too
3960 * fragmented for high-order allocations. Wake up kcompactd
3961 * and rely on compaction_suitable() to determine if it's
3962 * needed. If it fails, it will defer subsequent attempts to
3963 * ratelimit its work.
3965 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3966 wakeup_kcompactd(pgdat, order, classzone_idx);
3970 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3972 wake_up_interruptible(&pgdat->kswapd_wait);
3975 #ifdef CONFIG_HIBERNATION
3977 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3980 * Rather than trying to age LRUs the aim is to preserve the overall
3981 * LRU order by reclaiming preferentially
3982 * inactive > active > active referenced > active mapped
3984 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3986 struct reclaim_state reclaim_state;
3987 struct scan_control sc = {
3988 .nr_to_reclaim = nr_to_reclaim,
3989 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3990 .reclaim_idx = MAX_NR_ZONES - 1,
3991 .priority = DEF_PRIORITY,
3995 .hibernation_mode = 1,
3997 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3998 struct task_struct *p = current;
3999 unsigned long nr_reclaimed;
4000 unsigned int noreclaim_flag;
4002 fs_reclaim_acquire(sc.gfp_mask);
4003 noreclaim_flag = memalloc_noreclaim_save();
4004 reclaim_state.reclaimed_slab = 0;
4005 p->reclaim_state = &reclaim_state;
4007 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4009 p->reclaim_state = NULL;
4010 memalloc_noreclaim_restore(noreclaim_flag);
4011 fs_reclaim_release(sc.gfp_mask);
4013 return nr_reclaimed;
4015 #endif /* CONFIG_HIBERNATION */
4017 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4018 not required for correctness. So if the last cpu in a node goes
4019 away, we get changed to run anywhere: as the first one comes back,
4020 restore their cpu bindings. */
4021 static int kswapd_cpu_online(unsigned int cpu)
4025 for_each_node_state(nid, N_MEMORY) {
4026 pg_data_t *pgdat = NODE_DATA(nid);
4027 const struct cpumask *mask;
4029 mask = cpumask_of_node(pgdat->node_id);
4031 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4032 /* One of our CPUs online: restore mask */
4033 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4039 * This kswapd start function will be called by init and node-hot-add.
4040 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4042 int kswapd_run(int nid)
4044 pg_data_t *pgdat = NODE_DATA(nid);
4050 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4051 if (IS_ERR(pgdat->kswapd)) {
4052 /* failure at boot is fatal */
4053 BUG_ON(system_state < SYSTEM_RUNNING);
4054 pr_err("Failed to start kswapd on node %d\n", nid);
4055 ret = PTR_ERR(pgdat->kswapd);
4056 pgdat->kswapd = NULL;
4062 * Called by memory hotplug when all memory in a node is offlined. Caller must
4063 * hold mem_hotplug_begin/end().
4065 void kswapd_stop(int nid)
4067 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4070 kthread_stop(kswapd);
4071 NODE_DATA(nid)->kswapd = NULL;
4075 static int __init kswapd_init(void)
4080 for_each_node_state(nid, N_MEMORY)
4082 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4083 "mm/vmscan:online", kswapd_cpu_online,
4089 module_init(kswapd_init)
4095 * If non-zero call node_reclaim when the number of free pages falls below
4098 int node_reclaim_mode __read_mostly;
4100 #define RECLAIM_OFF 0
4101 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4102 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4103 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4106 * Priority for NODE_RECLAIM. This determines the fraction of pages
4107 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4110 #define NODE_RECLAIM_PRIORITY 4
4113 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4116 int sysctl_min_unmapped_ratio = 1;
4119 * If the number of slab pages in a zone grows beyond this percentage then
4120 * slab reclaim needs to occur.
4122 int sysctl_min_slab_ratio = 5;
4124 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4126 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4127 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4128 node_page_state(pgdat, NR_ACTIVE_FILE);
4131 * It's possible for there to be more file mapped pages than
4132 * accounted for by the pages on the file LRU lists because
4133 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4135 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4138 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4139 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4141 unsigned long nr_pagecache_reclaimable;
4142 unsigned long delta = 0;
4145 * If RECLAIM_UNMAP is set, then all file pages are considered
4146 * potentially reclaimable. Otherwise, we have to worry about
4147 * pages like swapcache and node_unmapped_file_pages() provides
4150 if (node_reclaim_mode & RECLAIM_UNMAP)
4151 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4153 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4155 /* If we can't clean pages, remove dirty pages from consideration */
4156 if (!(node_reclaim_mode & RECLAIM_WRITE))
4157 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4159 /* Watch for any possible underflows due to delta */
4160 if (unlikely(delta > nr_pagecache_reclaimable))
4161 delta = nr_pagecache_reclaimable;
4163 return nr_pagecache_reclaimable - delta;
4167 * Try to free up some pages from this node through reclaim.
4169 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4171 /* Minimum pages needed in order to stay on node */
4172 const unsigned long nr_pages = 1 << order;
4173 struct task_struct *p = current;
4174 struct reclaim_state reclaim_state;
4175 unsigned int noreclaim_flag;
4176 struct scan_control sc = {
4177 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4178 .gfp_mask = current_gfp_context(gfp_mask),
4180 .priority = NODE_RECLAIM_PRIORITY,
4181 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4182 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4184 .reclaim_idx = gfp_zone(gfp_mask),
4188 fs_reclaim_acquire(sc.gfp_mask);
4190 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4191 * and we also need to be able to write out pages for RECLAIM_WRITE
4192 * and RECLAIM_UNMAP.
4194 noreclaim_flag = memalloc_noreclaim_save();
4195 p->flags |= PF_SWAPWRITE;
4196 reclaim_state.reclaimed_slab = 0;
4197 p->reclaim_state = &reclaim_state;
4199 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4201 * Free memory by calling shrink node with increasing
4202 * priorities until we have enough memory freed.
4205 shrink_node(pgdat, &sc);
4206 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4209 p->reclaim_state = NULL;
4210 current->flags &= ~PF_SWAPWRITE;
4211 memalloc_noreclaim_restore(noreclaim_flag);
4212 fs_reclaim_release(sc.gfp_mask);
4213 return sc.nr_reclaimed >= nr_pages;
4216 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4221 * Node reclaim reclaims unmapped file backed pages and
4222 * slab pages if we are over the defined limits.
4224 * A small portion of unmapped file backed pages is needed for
4225 * file I/O otherwise pages read by file I/O will be immediately
4226 * thrown out if the node is overallocated. So we do not reclaim
4227 * if less than a specified percentage of the node is used by
4228 * unmapped file backed pages.
4230 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4231 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4232 return NODE_RECLAIM_FULL;
4235 * Do not scan if the allocation should not be delayed.
4237 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4238 return NODE_RECLAIM_NOSCAN;
4241 * Only run node reclaim on the local node or on nodes that do not
4242 * have associated processors. This will favor the local processor
4243 * over remote processors and spread off node memory allocations
4244 * as wide as possible.
4246 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4247 return NODE_RECLAIM_NOSCAN;
4249 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4250 return NODE_RECLAIM_NOSCAN;
4252 ret = __node_reclaim(pgdat, gfp_mask, order);
4253 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4256 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4263 * page_evictable - test whether a page is evictable
4264 * @page: the page to test
4266 * Test whether page is evictable--i.e., should be placed on active/inactive
4267 * lists vs unevictable list.
4269 * Reasons page might not be evictable:
4270 * (1) page's mapping marked unevictable
4271 * (2) page is part of an mlocked VMA
4274 int page_evictable(struct page *page)
4278 /* Prevent address_space of inode and swap cache from being freed */
4280 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4286 * check_move_unevictable_pages - check pages for evictability and move to
4287 * appropriate zone lru list
4288 * @pvec: pagevec with lru pages to check
4290 * Checks pages for evictability, if an evictable page is in the unevictable
4291 * lru list, moves it to the appropriate evictable lru list. This function
4292 * should be only used for lru pages.
4294 void check_move_unevictable_pages(struct pagevec *pvec)
4296 struct lruvec *lruvec;
4297 struct pglist_data *pgdat = NULL;
4302 for (i = 0; i < pvec->nr; i++) {
4303 struct page *page = pvec->pages[i];
4304 struct pglist_data *pagepgdat = page_pgdat(page);
4307 if (pagepgdat != pgdat) {
4309 spin_unlock_irq(&pgdat->lru_lock);
4311 spin_lock_irq(&pgdat->lru_lock);
4313 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4315 if (!PageLRU(page) || !PageUnevictable(page))
4318 if (page_evictable(page)) {
4319 enum lru_list lru = page_lru_base_type(page);
4321 VM_BUG_ON_PAGE(PageActive(page), page);
4322 ClearPageUnevictable(page);
4323 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4324 add_page_to_lru_list(page, lruvec, lru);
4330 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4331 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4332 spin_unlock_irq(&pgdat->lru_lock);
4335 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);