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 = lruvec_page_state_local(lruvec, NR_LRU_BASE + 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 unsigned int 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: %pS 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(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);
1109 unsigned nr_reclaimed = 0;
1110 unsigned pgactivate = 0;
1112 memset(stat, 0, sizeof(*stat));
1115 while (!list_empty(page_list)) {
1116 struct address_space *mapping;
1119 enum page_references references = PAGEREF_RECLAIM_CLEAN;
1120 bool dirty, writeback;
1124 page = lru_to_page(page_list);
1125 list_del(&page->lru);
1127 if (!trylock_page(page))
1130 VM_BUG_ON_PAGE(PageActive(page), page);
1134 if (unlikely(!page_evictable(page)))
1135 goto activate_locked;
1137 if (!sc->may_unmap && page_mapped(page))
1140 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1141 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1144 * The number of dirty pages determines if a node is marked
1145 * reclaim_congested which affects wait_iff_congested. kswapd
1146 * will stall and start writing pages if the tail of the LRU
1147 * is all dirty unqueued pages.
1149 page_check_dirty_writeback(page, &dirty, &writeback);
1150 if (dirty || writeback)
1153 if (dirty && !writeback)
1154 stat->nr_unqueued_dirty++;
1157 * Treat this page as congested if the underlying BDI is or if
1158 * pages are cycling through the LRU so quickly that the
1159 * pages marked for immediate reclaim are making it to the
1160 * end of the LRU a second time.
1162 mapping = page_mapping(page);
1163 if (((dirty || writeback) && mapping &&
1164 inode_write_congested(mapping->host)) ||
1165 (writeback && PageReclaim(page)))
1166 stat->nr_congested++;
1169 * If a page at the tail of the LRU is under writeback, there
1170 * are three cases to consider.
1172 * 1) If reclaim is encountering an excessive number of pages
1173 * under writeback and this page is both under writeback and
1174 * PageReclaim then it indicates that pages are being queued
1175 * for IO but are being recycled through the LRU before the
1176 * IO can complete. Waiting on the page itself risks an
1177 * indefinite stall if it is impossible to writeback the
1178 * page due to IO error or disconnected storage so instead
1179 * note that the LRU is being scanned too quickly and the
1180 * caller can stall after page list has been processed.
1182 * 2) Global or new memcg reclaim encounters a page that is
1183 * not marked for immediate reclaim, or the caller does not
1184 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1185 * not to fs). In this case mark the page for immediate
1186 * reclaim and continue scanning.
1188 * Require may_enter_fs because we would wait on fs, which
1189 * may not have submitted IO yet. And the loop driver might
1190 * enter reclaim, and deadlock if it waits on a page for
1191 * which it is needed to do the write (loop masks off
1192 * __GFP_IO|__GFP_FS for this reason); but more thought
1193 * would probably show more reasons.
1195 * 3) Legacy memcg encounters a page that is already marked
1196 * PageReclaim. memcg does not have any dirty pages
1197 * throttling so we could easily OOM just because too many
1198 * pages are in writeback and there is nothing else to
1199 * reclaim. Wait for the writeback to complete.
1201 * In cases 1) and 2) we activate the pages to get them out of
1202 * the way while we continue scanning for clean pages on the
1203 * inactive list and refilling from the active list. The
1204 * observation here is that waiting for disk writes is more
1205 * expensive than potentially causing reloads down the line.
1206 * Since they're marked for immediate reclaim, they won't put
1207 * memory pressure on the cache working set any longer than it
1208 * takes to write them to disk.
1210 if (PageWriteback(page)) {
1212 if (current_is_kswapd() &&
1213 PageReclaim(page) &&
1214 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1215 stat->nr_immediate++;
1216 goto activate_locked;
1219 } else if (sane_reclaim(sc) ||
1220 !PageReclaim(page) || !may_enter_fs) {
1222 * This is slightly racy - end_page_writeback()
1223 * might have just cleared PageReclaim, then
1224 * setting PageReclaim here end up interpreted
1225 * as PageReadahead - but that does not matter
1226 * enough to care. What we do want is for this
1227 * page to have PageReclaim set next time memcg
1228 * reclaim reaches the tests above, so it will
1229 * then wait_on_page_writeback() to avoid OOM;
1230 * and it's also appropriate in global reclaim.
1232 SetPageReclaim(page);
1233 stat->nr_writeback++;
1234 goto activate_locked;
1239 wait_on_page_writeback(page);
1240 /* then go back and try same page again */
1241 list_add_tail(&page->lru, page_list);
1247 references = page_check_references(page, sc);
1249 switch (references) {
1250 case PAGEREF_ACTIVATE:
1251 goto activate_locked;
1253 stat->nr_ref_keep++;
1255 case PAGEREF_RECLAIM:
1256 case PAGEREF_RECLAIM_CLEAN:
1257 ; /* try to reclaim the page below */
1261 * Anonymous process memory has backing store?
1262 * Try to allocate it some swap space here.
1263 * Lazyfree page could be freed directly
1265 if (PageAnon(page) && PageSwapBacked(page)) {
1266 if (!PageSwapCache(page)) {
1267 if (!(sc->gfp_mask & __GFP_IO))
1269 if (PageTransHuge(page)) {
1270 /* cannot split THP, skip it */
1271 if (!can_split_huge_page(page, NULL))
1272 goto activate_locked;
1274 * Split pages without a PMD map right
1275 * away. Chances are some or all of the
1276 * tail pages can be freed without IO.
1278 if (!compound_mapcount(page) &&
1279 split_huge_page_to_list(page,
1281 goto activate_locked;
1283 if (!add_to_swap(page)) {
1284 if (!PageTransHuge(page))
1285 goto activate_locked;
1286 /* Fallback to swap normal pages */
1287 if (split_huge_page_to_list(page,
1289 goto activate_locked;
1290 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1291 count_vm_event(THP_SWPOUT_FALLBACK);
1293 if (!add_to_swap(page))
1294 goto activate_locked;
1299 /* Adding to swap updated mapping */
1300 mapping = page_mapping(page);
1302 } else if (unlikely(PageTransHuge(page))) {
1303 /* Split file THP */
1304 if (split_huge_page_to_list(page, page_list))
1309 * The page is mapped into the page tables of one or more
1310 * processes. Try to unmap it here.
1312 if (page_mapped(page)) {
1313 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1315 if (unlikely(PageTransHuge(page)))
1316 flags |= TTU_SPLIT_HUGE_PMD;
1317 if (!try_to_unmap(page, flags)) {
1318 stat->nr_unmap_fail++;
1319 goto activate_locked;
1323 if (PageDirty(page)) {
1325 * Only kswapd can writeback filesystem pages
1326 * to avoid risk of stack overflow. But avoid
1327 * injecting inefficient single-page IO into
1328 * flusher writeback as much as possible: only
1329 * write pages when we've encountered many
1330 * dirty pages, and when we've already scanned
1331 * the rest of the LRU for clean pages and see
1332 * the same dirty pages again (PageReclaim).
1334 if (page_is_file_cache(page) &&
1335 (!current_is_kswapd() || !PageReclaim(page) ||
1336 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1338 * Immediately reclaim when written back.
1339 * Similar in principal to deactivate_page()
1340 * except we already have the page isolated
1341 * and know it's dirty
1343 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1344 SetPageReclaim(page);
1346 goto activate_locked;
1349 if (references == PAGEREF_RECLAIM_CLEAN)
1353 if (!sc->may_writepage)
1357 * Page is dirty. Flush the TLB if a writable entry
1358 * potentially exists to avoid CPU writes after IO
1359 * starts and then write it out here.
1361 try_to_unmap_flush_dirty();
1362 switch (pageout(page, mapping, sc)) {
1366 goto activate_locked;
1368 if (PageWriteback(page))
1370 if (PageDirty(page))
1374 * A synchronous write - probably a ramdisk. Go
1375 * ahead and try to reclaim the page.
1377 if (!trylock_page(page))
1379 if (PageDirty(page) || PageWriteback(page))
1381 mapping = page_mapping(page);
1383 ; /* try to free the page below */
1388 * If the page has buffers, try to free the buffer mappings
1389 * associated with this page. If we succeed we try to free
1392 * We do this even if the page is PageDirty().
1393 * try_to_release_page() does not perform I/O, but it is
1394 * possible for a page to have PageDirty set, but it is actually
1395 * clean (all its buffers are clean). This happens if the
1396 * buffers were written out directly, with submit_bh(). ext3
1397 * will do this, as well as the blockdev mapping.
1398 * try_to_release_page() will discover that cleanness and will
1399 * drop the buffers and mark the page clean - it can be freed.
1401 * Rarely, pages can have buffers and no ->mapping. These are
1402 * the pages which were not successfully invalidated in
1403 * truncate_complete_page(). We try to drop those buffers here
1404 * and if that worked, and the page is no longer mapped into
1405 * process address space (page_count == 1) it can be freed.
1406 * Otherwise, leave the page on the LRU so it is swappable.
1408 if (page_has_private(page)) {
1409 if (!try_to_release_page(page, sc->gfp_mask))
1410 goto activate_locked;
1411 if (!mapping && page_count(page) == 1) {
1413 if (put_page_testzero(page))
1417 * rare race with speculative reference.
1418 * the speculative reference will free
1419 * this page shortly, so we may
1420 * increment nr_reclaimed here (and
1421 * leave it off the LRU).
1429 if (PageAnon(page) && !PageSwapBacked(page)) {
1430 /* follow __remove_mapping for reference */
1431 if (!page_ref_freeze(page, 1))
1433 if (PageDirty(page)) {
1434 page_ref_unfreeze(page, 1);
1438 count_vm_event(PGLAZYFREED);
1439 count_memcg_page_event(page, PGLAZYFREED);
1440 } else if (!mapping || !__remove_mapping(mapping, page, true))
1448 * Is there need to periodically free_page_list? It would
1449 * appear not as the counts should be low
1451 if (unlikely(PageTransHuge(page))) {
1452 mem_cgroup_uncharge(page);
1453 (*get_compound_page_dtor(page))(page);
1455 list_add(&page->lru, &free_pages);
1459 /* Not a candidate for swapping, so reclaim swap space. */
1460 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1462 try_to_free_swap(page);
1463 VM_BUG_ON_PAGE(PageActive(page), page);
1464 if (!PageMlocked(page)) {
1465 int type = page_is_file_cache(page);
1466 SetPageActive(page);
1468 stat->nr_activate[type] += hpage_nr_pages(page);
1469 count_memcg_page_event(page, PGACTIVATE);
1474 list_add(&page->lru, &ret_pages);
1475 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1478 mem_cgroup_uncharge_list(&free_pages);
1479 try_to_unmap_flush();
1480 free_unref_page_list(&free_pages);
1482 list_splice(&ret_pages, page_list);
1483 count_vm_events(PGACTIVATE, pgactivate);
1485 return nr_reclaimed;
1488 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1489 struct list_head *page_list)
1491 struct scan_control sc = {
1492 .gfp_mask = GFP_KERNEL,
1493 .priority = DEF_PRIORITY,
1496 struct reclaim_stat dummy_stat;
1498 struct page *page, *next;
1499 LIST_HEAD(clean_pages);
1501 list_for_each_entry_safe(page, next, page_list, lru) {
1502 if (page_is_file_cache(page) && !PageDirty(page) &&
1503 !__PageMovable(page) && !PageUnevictable(page)) {
1504 ClearPageActive(page);
1505 list_move(&page->lru, &clean_pages);
1509 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1510 TTU_IGNORE_ACCESS, &dummy_stat, true);
1511 list_splice(&clean_pages, page_list);
1512 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1517 * Attempt to remove the specified page from its LRU. Only take this page
1518 * if it is of the appropriate PageActive status. Pages which are being
1519 * freed elsewhere are also ignored.
1521 * page: page to consider
1522 * mode: one of the LRU isolation modes defined above
1524 * returns 0 on success, -ve errno on failure.
1526 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1530 /* Only take pages on the LRU. */
1534 /* Compaction should not handle unevictable pages but CMA can do so */
1535 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1541 * To minimise LRU disruption, the caller can indicate that it only
1542 * wants to isolate pages it will be able to operate on without
1543 * blocking - clean pages for the most part.
1545 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1546 * that it is possible to migrate without blocking
1548 if (mode & ISOLATE_ASYNC_MIGRATE) {
1549 /* All the caller can do on PageWriteback is block */
1550 if (PageWriteback(page))
1553 if (PageDirty(page)) {
1554 struct address_space *mapping;
1558 * Only pages without mappings or that have a
1559 * ->migratepage callback are possible to migrate
1560 * without blocking. However, we can be racing with
1561 * truncation so it's necessary to lock the page
1562 * to stabilise the mapping as truncation holds
1563 * the page lock until after the page is removed
1564 * from the page cache.
1566 if (!trylock_page(page))
1569 mapping = page_mapping(page);
1570 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1577 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1580 if (likely(get_page_unless_zero(page))) {
1582 * Be careful not to clear PageLRU until after we're
1583 * sure the page is not being freed elsewhere -- the
1584 * page release code relies on it.
1595 * Update LRU sizes after isolating pages. The LRU size updates must
1596 * be complete before mem_cgroup_update_lru_size due to a santity check.
1598 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1599 enum lru_list lru, unsigned long *nr_zone_taken)
1603 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1604 if (!nr_zone_taken[zid])
1607 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1609 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1616 * pgdat->lru_lock is heavily contended. Some of the functions that
1617 * shrink the lists perform better by taking out a batch of pages
1618 * and working on them outside the LRU lock.
1620 * For pagecache intensive workloads, this function is the hottest
1621 * spot in the kernel (apart from copy_*_user functions).
1623 * Appropriate locks must be held before calling this function.
1625 * @nr_to_scan: The number of eligible pages to look through on the list.
1626 * @lruvec: The LRU vector to pull pages from.
1627 * @dst: The temp list to put pages on to.
1628 * @nr_scanned: The number of pages that were scanned.
1629 * @sc: The scan_control struct for this reclaim session
1630 * @mode: One of the LRU isolation modes
1631 * @lru: LRU list id for isolating
1633 * returns how many pages were moved onto *@dst.
1635 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1636 struct lruvec *lruvec, struct list_head *dst,
1637 unsigned long *nr_scanned, struct scan_control *sc,
1640 struct list_head *src = &lruvec->lists[lru];
1641 unsigned long nr_taken = 0;
1642 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1643 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1644 unsigned long skipped = 0;
1645 unsigned long scan, total_scan, nr_pages;
1646 LIST_HEAD(pages_skipped);
1647 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1650 for (total_scan = 0;
1651 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1655 page = lru_to_page(src);
1656 prefetchw_prev_lru_page(page, src, flags);
1658 VM_BUG_ON_PAGE(!PageLRU(page), page);
1660 if (page_zonenum(page) > sc->reclaim_idx) {
1661 list_move(&page->lru, &pages_skipped);
1662 nr_skipped[page_zonenum(page)]++;
1667 * Do not count skipped pages because that makes the function
1668 * return with no isolated pages if the LRU mostly contains
1669 * ineligible pages. This causes the VM to not reclaim any
1670 * pages, triggering a premature OOM.
1673 switch (__isolate_lru_page(page, mode)) {
1675 nr_pages = hpage_nr_pages(page);
1676 nr_taken += nr_pages;
1677 nr_zone_taken[page_zonenum(page)] += nr_pages;
1678 list_move(&page->lru, dst);
1682 /* else it is being freed elsewhere */
1683 list_move(&page->lru, src);
1692 * Splice any skipped pages to the start of the LRU list. Note that
1693 * this disrupts the LRU order when reclaiming for lower zones but
1694 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1695 * scanning would soon rescan the same pages to skip and put the
1696 * system at risk of premature OOM.
1698 if (!list_empty(&pages_skipped)) {
1701 list_splice(&pages_skipped, src);
1702 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1703 if (!nr_skipped[zid])
1706 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1707 skipped += nr_skipped[zid];
1710 *nr_scanned = total_scan;
1711 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1712 total_scan, skipped, nr_taken, mode, lru);
1713 update_lru_sizes(lruvec, lru, nr_zone_taken);
1718 * isolate_lru_page - tries to isolate a page from its LRU list
1719 * @page: page to isolate from its LRU list
1721 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1722 * vmstat statistic corresponding to whatever LRU list the page was on.
1724 * Returns 0 if the page was removed from an LRU list.
1725 * Returns -EBUSY if the page was not on an LRU list.
1727 * The returned page will have PageLRU() cleared. If it was found on
1728 * the active list, it will have PageActive set. If it was found on
1729 * the unevictable list, it will have the PageUnevictable bit set. That flag
1730 * may need to be cleared by the caller before letting the page go.
1732 * The vmstat statistic corresponding to the list on which the page was
1733 * found will be decremented.
1737 * (1) Must be called with an elevated refcount on the page. This is a
1738 * fundamentnal difference from isolate_lru_pages (which is called
1739 * without a stable reference).
1740 * (2) the lru_lock must not be held.
1741 * (3) interrupts must be enabled.
1743 int isolate_lru_page(struct page *page)
1747 VM_BUG_ON_PAGE(!page_count(page), page);
1748 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1750 if (PageLRU(page)) {
1751 pg_data_t *pgdat = page_pgdat(page);
1752 struct lruvec *lruvec;
1754 spin_lock_irq(&pgdat->lru_lock);
1755 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1756 if (PageLRU(page)) {
1757 int lru = page_lru(page);
1760 del_page_from_lru_list(page, lruvec, lru);
1763 spin_unlock_irq(&pgdat->lru_lock);
1769 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1770 * then get resheduled. When there are massive number of tasks doing page
1771 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1772 * the LRU list will go small and be scanned faster than necessary, leading to
1773 * unnecessary swapping, thrashing and OOM.
1775 static int too_many_isolated(struct pglist_data *pgdat, int file,
1776 struct scan_control *sc)
1778 unsigned long inactive, isolated;
1780 if (current_is_kswapd())
1783 if (!sane_reclaim(sc))
1787 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1788 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1790 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1791 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1795 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1796 * won't get blocked by normal direct-reclaimers, forming a circular
1799 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1802 return isolated > inactive;
1806 * This moves pages from @list to corresponding LRU list.
1808 * We move them the other way if the page is referenced by one or more
1809 * processes, from rmap.
1811 * If the pages are mostly unmapped, the processing is fast and it is
1812 * appropriate to hold zone_lru_lock across the whole operation. But if
1813 * the pages are mapped, the processing is slow (page_referenced()) so we
1814 * should drop zone_lru_lock around each page. It's impossible to balance
1815 * this, so instead we remove the pages from the LRU while processing them.
1816 * It is safe to rely on PG_active against the non-LRU pages in here because
1817 * nobody will play with that bit on a non-LRU page.
1819 * The downside is that we have to touch page->_refcount against each page.
1820 * But we had to alter page->flags anyway.
1822 * Returns the number of pages moved to the given lruvec.
1825 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1826 struct list_head *list)
1828 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1829 int nr_pages, nr_moved = 0;
1830 LIST_HEAD(pages_to_free);
1834 while (!list_empty(list)) {
1835 page = lru_to_page(list);
1836 VM_BUG_ON_PAGE(PageLRU(page), page);
1837 if (unlikely(!page_evictable(page))) {
1838 list_del(&page->lru);
1839 spin_unlock_irq(&pgdat->lru_lock);
1840 putback_lru_page(page);
1841 spin_lock_irq(&pgdat->lru_lock);
1844 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1847 lru = page_lru(page);
1849 nr_pages = hpage_nr_pages(page);
1850 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1851 list_move(&page->lru, &lruvec->lists[lru]);
1853 if (put_page_testzero(page)) {
1854 __ClearPageLRU(page);
1855 __ClearPageActive(page);
1856 del_page_from_lru_list(page, lruvec, lru);
1858 if (unlikely(PageCompound(page))) {
1859 spin_unlock_irq(&pgdat->lru_lock);
1860 mem_cgroup_uncharge(page);
1861 (*get_compound_page_dtor(page))(page);
1862 spin_lock_irq(&pgdat->lru_lock);
1864 list_add(&page->lru, &pages_to_free);
1866 nr_moved += nr_pages;
1871 * To save our caller's stack, now use input list for pages to free.
1873 list_splice(&pages_to_free, list);
1879 * If a kernel thread (such as nfsd for loop-back mounts) services
1880 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1881 * In that case we should only throttle if the backing device it is
1882 * writing to is congested. In other cases it is safe to throttle.
1884 static int current_may_throttle(void)
1886 return !(current->flags & PF_LESS_THROTTLE) ||
1887 current->backing_dev_info == NULL ||
1888 bdi_write_congested(current->backing_dev_info);
1892 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1893 * of reclaimed pages
1895 static noinline_for_stack unsigned long
1896 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1897 struct scan_control *sc, enum lru_list lru)
1899 LIST_HEAD(page_list);
1900 unsigned long nr_scanned;
1901 unsigned long nr_reclaimed = 0;
1902 unsigned long nr_taken;
1903 struct reclaim_stat stat;
1904 int file = is_file_lru(lru);
1905 enum vm_event_item item;
1906 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1907 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1908 bool stalled = false;
1910 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1914 /* wait a bit for the reclaimer. */
1918 /* We are about to die and free our memory. Return now. */
1919 if (fatal_signal_pending(current))
1920 return SWAP_CLUSTER_MAX;
1925 spin_lock_irq(&pgdat->lru_lock);
1927 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1928 &nr_scanned, sc, lru);
1930 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1931 reclaim_stat->recent_scanned[file] += nr_taken;
1933 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1934 if (global_reclaim(sc))
1935 __count_vm_events(item, nr_scanned);
1936 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1937 spin_unlock_irq(&pgdat->lru_lock);
1942 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1945 spin_lock_irq(&pgdat->lru_lock);
1947 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1948 if (global_reclaim(sc))
1949 __count_vm_events(item, nr_reclaimed);
1950 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1951 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1952 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1954 move_pages_to_lru(lruvec, &page_list);
1956 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1958 spin_unlock_irq(&pgdat->lru_lock);
1960 mem_cgroup_uncharge_list(&page_list);
1961 free_unref_page_list(&page_list);
1964 * If dirty pages are scanned that are not queued for IO, it
1965 * implies that flushers are not doing their job. This can
1966 * happen when memory pressure pushes dirty pages to the end of
1967 * the LRU before the dirty limits are breached and the dirty
1968 * data has expired. It can also happen when the proportion of
1969 * dirty pages grows not through writes but through memory
1970 * pressure reclaiming all the clean cache. And in some cases,
1971 * the flushers simply cannot keep up with the allocation
1972 * rate. Nudge the flusher threads in case they are asleep.
1974 if (stat.nr_unqueued_dirty == nr_taken)
1975 wakeup_flusher_threads(WB_REASON_VMSCAN);
1977 sc->nr.dirty += stat.nr_dirty;
1978 sc->nr.congested += stat.nr_congested;
1979 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1980 sc->nr.writeback += stat.nr_writeback;
1981 sc->nr.immediate += stat.nr_immediate;
1982 sc->nr.taken += nr_taken;
1984 sc->nr.file_taken += nr_taken;
1986 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1987 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1988 return nr_reclaimed;
1991 static void shrink_active_list(unsigned long nr_to_scan,
1992 struct lruvec *lruvec,
1993 struct scan_control *sc,
1996 unsigned long nr_taken;
1997 unsigned long nr_scanned;
1998 unsigned long vm_flags;
1999 LIST_HEAD(l_hold); /* The pages which were snipped off */
2000 LIST_HEAD(l_active);
2001 LIST_HEAD(l_inactive);
2003 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2004 unsigned nr_deactivate, nr_activate;
2005 unsigned nr_rotated = 0;
2006 int file = is_file_lru(lru);
2007 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2011 spin_lock_irq(&pgdat->lru_lock);
2013 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2014 &nr_scanned, sc, lru);
2016 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2017 reclaim_stat->recent_scanned[file] += nr_taken;
2019 __count_vm_events(PGREFILL, nr_scanned);
2020 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2022 spin_unlock_irq(&pgdat->lru_lock);
2024 while (!list_empty(&l_hold)) {
2026 page = lru_to_page(&l_hold);
2027 list_del(&page->lru);
2029 if (unlikely(!page_evictable(page))) {
2030 putback_lru_page(page);
2034 if (unlikely(buffer_heads_over_limit)) {
2035 if (page_has_private(page) && trylock_page(page)) {
2036 if (page_has_private(page))
2037 try_to_release_page(page, 0);
2042 if (page_referenced(page, 0, sc->target_mem_cgroup,
2044 nr_rotated += hpage_nr_pages(page);
2046 * Identify referenced, file-backed active pages and
2047 * give them one more trip around the active list. So
2048 * that executable code get better chances to stay in
2049 * memory under moderate memory pressure. Anon pages
2050 * are not likely to be evicted by use-once streaming
2051 * IO, plus JVM can create lots of anon VM_EXEC pages,
2052 * so we ignore them here.
2054 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2055 list_add(&page->lru, &l_active);
2060 ClearPageActive(page); /* we are de-activating */
2061 SetPageWorkingset(page);
2062 list_add(&page->lru, &l_inactive);
2066 * Move pages back to the lru list.
2068 spin_lock_irq(&pgdat->lru_lock);
2070 * Count referenced pages from currently used mappings as rotated,
2071 * even though only some of them are actually re-activated. This
2072 * helps balance scan pressure between file and anonymous pages in
2075 reclaim_stat->recent_rotated[file] += nr_rotated;
2077 nr_activate = move_pages_to_lru(lruvec, &l_active);
2078 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2079 /* Keep all free pages in l_active list */
2080 list_splice(&l_inactive, &l_active);
2082 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2083 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2085 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2086 spin_unlock_irq(&pgdat->lru_lock);
2088 mem_cgroup_uncharge_list(&l_active);
2089 free_unref_page_list(&l_active);
2090 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2091 nr_deactivate, nr_rotated, sc->priority, file);
2095 * The inactive anon list should be small enough that the VM never has
2096 * to do too much work.
2098 * The inactive file list should be small enough to leave most memory
2099 * to the established workingset on the scan-resistant active list,
2100 * but large enough to avoid thrashing the aggregate readahead window.
2102 * Both inactive lists should also be large enough that each inactive
2103 * page has a chance to be referenced again before it is reclaimed.
2105 * If that fails and refaulting is observed, the inactive list grows.
2107 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2108 * on this LRU, maintained by the pageout code. An inactive_ratio
2109 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2112 * memory ratio inactive
2113 * -------------------------------------
2122 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2123 struct scan_control *sc, bool trace)
2125 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2126 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2127 enum lru_list inactive_lru = file * LRU_FILE;
2128 unsigned long inactive, active;
2129 unsigned long inactive_ratio;
2130 unsigned long refaults;
2134 * If we don't have swap space, anonymous page deactivation
2137 if (!file && !total_swap_pages)
2140 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2141 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2144 * When refaults are being observed, it means a new workingset
2145 * is being established. Disable active list protection to get
2146 * rid of the stale workingset quickly.
2148 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2149 if (file && lruvec->refaults != refaults) {
2152 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2154 inactive_ratio = int_sqrt(10 * gb);
2160 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2161 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2162 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2163 inactive_ratio, file);
2165 return inactive * inactive_ratio < active;
2168 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2169 struct lruvec *lruvec, struct scan_control *sc)
2171 if (is_active_lru(lru)) {
2172 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2173 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2177 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2188 * Determine how aggressively the anon and file LRU lists should be
2189 * scanned. The relative value of each set of LRU lists is determined
2190 * by looking at the fraction of the pages scanned we did rotate back
2191 * onto the active list instead of evict.
2193 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2194 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2196 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2197 struct scan_control *sc, unsigned long *nr,
2198 unsigned long *lru_pages)
2200 int swappiness = mem_cgroup_swappiness(memcg);
2201 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2203 u64 denominator = 0; /* gcc */
2204 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2205 unsigned long anon_prio, file_prio;
2206 enum scan_balance scan_balance;
2207 unsigned long anon, file;
2208 unsigned long ap, fp;
2211 /* If we have no swap space, do not bother scanning anon pages. */
2212 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2213 scan_balance = SCAN_FILE;
2218 * Global reclaim will swap to prevent OOM even with no
2219 * swappiness, but memcg users want to use this knob to
2220 * disable swapping for individual groups completely when
2221 * using the memory controller's swap limit feature would be
2224 if (!global_reclaim(sc) && !swappiness) {
2225 scan_balance = SCAN_FILE;
2230 * Do not apply any pressure balancing cleverness when the
2231 * system is close to OOM, scan both anon and file equally
2232 * (unless the swappiness setting disagrees with swapping).
2234 if (!sc->priority && swappiness) {
2235 scan_balance = SCAN_EQUAL;
2240 * Prevent the reclaimer from falling into the cache trap: as
2241 * cache pages start out inactive, every cache fault will tip
2242 * the scan balance towards the file LRU. And as the file LRU
2243 * shrinks, so does the window for rotation from references.
2244 * This means we have a runaway feedback loop where a tiny
2245 * thrashing file LRU becomes infinitely more attractive than
2246 * anon pages. Try to detect this based on file LRU size.
2248 if (global_reclaim(sc)) {
2249 unsigned long pgdatfile;
2250 unsigned long pgdatfree;
2252 unsigned long total_high_wmark = 0;
2254 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2255 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2256 node_page_state(pgdat, NR_INACTIVE_FILE);
2258 for (z = 0; z < MAX_NR_ZONES; z++) {
2259 struct zone *zone = &pgdat->node_zones[z];
2260 if (!managed_zone(zone))
2263 total_high_wmark += high_wmark_pages(zone);
2266 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2268 * Force SCAN_ANON if there are enough inactive
2269 * anonymous pages on the LRU in eligible zones.
2270 * Otherwise, the small LRU gets thrashed.
2272 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2273 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2275 scan_balance = SCAN_ANON;
2282 * If there is enough inactive page cache, i.e. if the size of the
2283 * inactive list is greater than that of the active list *and* the
2284 * inactive list actually has some pages to scan on this priority, we
2285 * do not reclaim anything from the anonymous working set right now.
2286 * Without the second condition we could end up never scanning an
2287 * lruvec even if it has plenty of old anonymous pages unless the
2288 * system is under heavy pressure.
2290 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2291 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2292 scan_balance = SCAN_FILE;
2296 scan_balance = SCAN_FRACT;
2299 * With swappiness at 100, anonymous and file have the same priority.
2300 * This scanning priority is essentially the inverse of IO cost.
2302 anon_prio = swappiness;
2303 file_prio = 200 - anon_prio;
2306 * OK, so we have swap space and a fair amount of page cache
2307 * pages. We use the recently rotated / recently scanned
2308 * ratios to determine how valuable each cache is.
2310 * Because workloads change over time (and to avoid overflow)
2311 * we keep these statistics as a floating average, which ends
2312 * up weighing recent references more than old ones.
2314 * anon in [0], file in [1]
2317 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2318 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2319 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2320 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2322 spin_lock_irq(&pgdat->lru_lock);
2323 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2324 reclaim_stat->recent_scanned[0] /= 2;
2325 reclaim_stat->recent_rotated[0] /= 2;
2328 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2329 reclaim_stat->recent_scanned[1] /= 2;
2330 reclaim_stat->recent_rotated[1] /= 2;
2334 * The amount of pressure on anon vs file pages is inversely
2335 * proportional to the fraction of recently scanned pages on
2336 * each list that were recently referenced and in active use.
2338 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2339 ap /= reclaim_stat->recent_rotated[0] + 1;
2341 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2342 fp /= reclaim_stat->recent_rotated[1] + 1;
2343 spin_unlock_irq(&pgdat->lru_lock);
2347 denominator = ap + fp + 1;
2350 for_each_evictable_lru(lru) {
2351 int file = is_file_lru(lru);
2355 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2356 scan = size >> sc->priority;
2358 * If the cgroup's already been deleted, make sure to
2359 * scrape out the remaining cache.
2361 if (!scan && !mem_cgroup_online(memcg))
2362 scan = min(size, SWAP_CLUSTER_MAX);
2364 switch (scan_balance) {
2366 /* Scan lists relative to size */
2370 * Scan types proportional to swappiness and
2371 * their relative recent reclaim efficiency.
2372 * Make sure we don't miss the last page
2373 * because of a round-off error.
2375 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2380 /* Scan one type exclusively */
2381 if ((scan_balance == SCAN_FILE) != file) {
2387 /* Look ma, no brain */
2397 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2399 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2400 struct scan_control *sc, unsigned long *lru_pages)
2402 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2403 unsigned long nr[NR_LRU_LISTS];
2404 unsigned long targets[NR_LRU_LISTS];
2405 unsigned long nr_to_scan;
2407 unsigned long nr_reclaimed = 0;
2408 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2409 struct blk_plug plug;
2412 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2414 /* Record the original scan target for proportional adjustments later */
2415 memcpy(targets, nr, sizeof(nr));
2418 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2419 * event that can occur when there is little memory pressure e.g.
2420 * multiple streaming readers/writers. Hence, we do not abort scanning
2421 * when the requested number of pages are reclaimed when scanning at
2422 * DEF_PRIORITY on the assumption that the fact we are direct
2423 * reclaiming implies that kswapd is not keeping up and it is best to
2424 * do a batch of work at once. For memcg reclaim one check is made to
2425 * abort proportional reclaim if either the file or anon lru has already
2426 * dropped to zero at the first pass.
2428 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2429 sc->priority == DEF_PRIORITY);
2431 blk_start_plug(&plug);
2432 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2433 nr[LRU_INACTIVE_FILE]) {
2434 unsigned long nr_anon, nr_file, percentage;
2435 unsigned long nr_scanned;
2437 for_each_evictable_lru(lru) {
2439 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2440 nr[lru] -= nr_to_scan;
2442 nr_reclaimed += shrink_list(lru, nr_to_scan,
2449 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2453 * For kswapd and memcg, reclaim at least the number of pages
2454 * requested. Ensure that the anon and file LRUs are scanned
2455 * proportionally what was requested by get_scan_count(). We
2456 * stop reclaiming one LRU and reduce the amount scanning
2457 * proportional to the original scan target.
2459 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2460 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2463 * It's just vindictive to attack the larger once the smaller
2464 * has gone to zero. And given the way we stop scanning the
2465 * smaller below, this makes sure that we only make one nudge
2466 * towards proportionality once we've got nr_to_reclaim.
2468 if (!nr_file || !nr_anon)
2471 if (nr_file > nr_anon) {
2472 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2473 targets[LRU_ACTIVE_ANON] + 1;
2475 percentage = nr_anon * 100 / scan_target;
2477 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2478 targets[LRU_ACTIVE_FILE] + 1;
2480 percentage = nr_file * 100 / scan_target;
2483 /* Stop scanning the smaller of the LRU */
2485 nr[lru + LRU_ACTIVE] = 0;
2488 * Recalculate the other LRU scan count based on its original
2489 * scan target and the percentage scanning already complete
2491 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2492 nr_scanned = targets[lru] - nr[lru];
2493 nr[lru] = targets[lru] * (100 - percentage) / 100;
2494 nr[lru] -= min(nr[lru], nr_scanned);
2497 nr_scanned = targets[lru] - nr[lru];
2498 nr[lru] = targets[lru] * (100 - percentage) / 100;
2499 nr[lru] -= min(nr[lru], nr_scanned);
2501 scan_adjusted = true;
2503 blk_finish_plug(&plug);
2504 sc->nr_reclaimed += nr_reclaimed;
2507 * Even if we did not try to evict anon pages at all, we want to
2508 * rebalance the anon lru active/inactive ratio.
2510 if (inactive_list_is_low(lruvec, false, sc, true))
2511 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2512 sc, LRU_ACTIVE_ANON);
2515 /* Use reclaim/compaction for costly allocs or under memory pressure */
2516 static bool in_reclaim_compaction(struct scan_control *sc)
2518 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2519 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2520 sc->priority < DEF_PRIORITY - 2))
2527 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2528 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2529 * true if more pages should be reclaimed such that when the page allocator
2530 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2531 * It will give up earlier than that if there is difficulty reclaiming pages.
2533 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2534 unsigned long nr_reclaimed,
2535 unsigned long nr_scanned,
2536 struct scan_control *sc)
2538 unsigned long pages_for_compaction;
2539 unsigned long inactive_lru_pages;
2542 /* If not in reclaim/compaction mode, stop */
2543 if (!in_reclaim_compaction(sc))
2546 /* Consider stopping depending on scan and reclaim activity */
2547 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2549 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2550 * full LRU list has been scanned and we are still failing
2551 * to reclaim pages. This full LRU scan is potentially
2552 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2554 if (!nr_reclaimed && !nr_scanned)
2558 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2559 * fail without consequence, stop if we failed to reclaim
2560 * any pages from the last SWAP_CLUSTER_MAX number of
2561 * pages that were scanned. This will return to the
2562 * caller faster at the risk reclaim/compaction and
2563 * the resulting allocation attempt fails
2570 * If we have not reclaimed enough pages for compaction and the
2571 * inactive lists are large enough, continue reclaiming
2573 pages_for_compaction = compact_gap(sc->order);
2574 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2575 if (get_nr_swap_pages() > 0)
2576 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2577 if (sc->nr_reclaimed < pages_for_compaction &&
2578 inactive_lru_pages > pages_for_compaction)
2581 /* If compaction would go ahead or the allocation would succeed, stop */
2582 for (z = 0; z <= sc->reclaim_idx; z++) {
2583 struct zone *zone = &pgdat->node_zones[z];
2584 if (!managed_zone(zone))
2587 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2588 case COMPACT_SUCCESS:
2589 case COMPACT_CONTINUE:
2592 /* check next zone */
2599 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2601 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2602 (memcg && memcg_congested(pgdat, memcg));
2605 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2607 struct reclaim_state *reclaim_state = current->reclaim_state;
2608 unsigned long nr_reclaimed, nr_scanned;
2609 bool reclaimable = false;
2612 struct mem_cgroup *root = sc->target_mem_cgroup;
2613 struct mem_cgroup_reclaim_cookie reclaim = {
2615 .priority = sc->priority,
2617 unsigned long node_lru_pages = 0;
2618 struct mem_cgroup *memcg;
2620 memset(&sc->nr, 0, sizeof(sc->nr));
2622 nr_reclaimed = sc->nr_reclaimed;
2623 nr_scanned = sc->nr_scanned;
2625 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2627 unsigned long lru_pages;
2628 unsigned long reclaimed;
2629 unsigned long scanned;
2631 switch (mem_cgroup_protected(root, memcg)) {
2632 case MEMCG_PROT_MIN:
2635 * If there is no reclaimable memory, OOM.
2638 case MEMCG_PROT_LOW:
2641 * Respect the protection only as long as
2642 * there is an unprotected supply
2643 * of reclaimable memory from other cgroups.
2645 if (!sc->memcg_low_reclaim) {
2646 sc->memcg_low_skipped = 1;
2649 memcg_memory_event(memcg, MEMCG_LOW);
2651 case MEMCG_PROT_NONE:
2655 reclaimed = sc->nr_reclaimed;
2656 scanned = sc->nr_scanned;
2657 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2658 node_lru_pages += lru_pages;
2660 if (sc->may_shrinkslab) {
2661 shrink_slab(sc->gfp_mask, pgdat->node_id,
2662 memcg, sc->priority);
2665 /* Record the group's reclaim efficiency */
2666 vmpressure(sc->gfp_mask, memcg, false,
2667 sc->nr_scanned - scanned,
2668 sc->nr_reclaimed - reclaimed);
2671 * Kswapd have to scan all memory cgroups to fulfill
2672 * the overall scan target for the node.
2674 * Limit reclaim, on the other hand, only cares about
2675 * nr_to_reclaim pages to be reclaimed and it will
2676 * retry with decreasing priority if one round over the
2677 * whole hierarchy is not sufficient.
2679 if (!current_is_kswapd() &&
2680 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2681 mem_cgroup_iter_break(root, memcg);
2684 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2686 if (reclaim_state) {
2687 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2688 reclaim_state->reclaimed_slab = 0;
2691 /* Record the subtree's reclaim efficiency */
2692 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2693 sc->nr_scanned - nr_scanned,
2694 sc->nr_reclaimed - nr_reclaimed);
2696 if (sc->nr_reclaimed - nr_reclaimed)
2699 if (current_is_kswapd()) {
2701 * If reclaim is isolating dirty pages under writeback,
2702 * it implies that the long-lived page allocation rate
2703 * is exceeding the page laundering rate. Either the
2704 * global limits are not being effective at throttling
2705 * processes due to the page distribution throughout
2706 * zones or there is heavy usage of a slow backing
2707 * device. The only option is to throttle from reclaim
2708 * context which is not ideal as there is no guarantee
2709 * the dirtying process is throttled in the same way
2710 * balance_dirty_pages() manages.
2712 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2713 * count the number of pages under pages flagged for
2714 * immediate reclaim and stall if any are encountered
2715 * in the nr_immediate check below.
2717 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2718 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2721 * Tag a node as congested if all the dirty pages
2722 * scanned were backed by a congested BDI and
2723 * wait_iff_congested will stall.
2725 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2726 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2728 /* Allow kswapd to start writing pages during reclaim.*/
2729 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2730 set_bit(PGDAT_DIRTY, &pgdat->flags);
2733 * If kswapd scans pages marked marked for immediate
2734 * reclaim and under writeback (nr_immediate), it
2735 * implies that pages are cycling through the LRU
2736 * faster than they are written so also forcibly stall.
2738 if (sc->nr.immediate)
2739 congestion_wait(BLK_RW_ASYNC, HZ/10);
2743 * Legacy memcg will stall in page writeback so avoid forcibly
2744 * stalling in wait_iff_congested().
2746 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2747 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2748 set_memcg_congestion(pgdat, root, true);
2751 * Stall direct reclaim for IO completions if underlying BDIs
2752 * and node is congested. Allow kswapd to continue until it
2753 * starts encountering unqueued dirty pages or cycling through
2754 * the LRU too quickly.
2756 if (!sc->hibernation_mode && !current_is_kswapd() &&
2757 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2758 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2760 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2761 sc->nr_scanned - nr_scanned, sc));
2764 * Kswapd gives up on balancing particular nodes after too
2765 * many failures to reclaim anything from them and goes to
2766 * sleep. On reclaim progress, reset the failure counter. A
2767 * successful direct reclaim run will revive a dormant kswapd.
2770 pgdat->kswapd_failures = 0;
2776 * Returns true if compaction should go ahead for a costly-order request, or
2777 * the allocation would already succeed without compaction. Return false if we
2778 * should reclaim first.
2780 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2782 unsigned long watermark;
2783 enum compact_result suitable;
2785 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2786 if (suitable == COMPACT_SUCCESS)
2787 /* Allocation should succeed already. Don't reclaim. */
2789 if (suitable == COMPACT_SKIPPED)
2790 /* Compaction cannot yet proceed. Do reclaim. */
2794 * Compaction is already possible, but it takes time to run and there
2795 * are potentially other callers using the pages just freed. So proceed
2796 * with reclaim to make a buffer of free pages available to give
2797 * compaction a reasonable chance of completing and allocating the page.
2798 * Note that we won't actually reclaim the whole buffer in one attempt
2799 * as the target watermark in should_continue_reclaim() is lower. But if
2800 * we are already above the high+gap watermark, don't reclaim at all.
2802 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2804 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2808 * This is the direct reclaim path, for page-allocating processes. We only
2809 * try to reclaim pages from zones which will satisfy the caller's allocation
2812 * If a zone is deemed to be full of pinned pages then just give it a light
2813 * scan then give up on it.
2815 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2819 unsigned long nr_soft_reclaimed;
2820 unsigned long nr_soft_scanned;
2822 pg_data_t *last_pgdat = NULL;
2825 * If the number of buffer_heads in the machine exceeds the maximum
2826 * allowed level, force direct reclaim to scan the highmem zone as
2827 * highmem pages could be pinning lowmem pages storing buffer_heads
2829 orig_mask = sc->gfp_mask;
2830 if (buffer_heads_over_limit) {
2831 sc->gfp_mask |= __GFP_HIGHMEM;
2832 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2835 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2836 sc->reclaim_idx, sc->nodemask) {
2838 * Take care memory controller reclaiming has small influence
2841 if (global_reclaim(sc)) {
2842 if (!cpuset_zone_allowed(zone,
2843 GFP_KERNEL | __GFP_HARDWALL))
2847 * If we already have plenty of memory free for
2848 * compaction in this zone, don't free any more.
2849 * Even though compaction is invoked for any
2850 * non-zero order, only frequent costly order
2851 * reclamation is disruptive enough to become a
2852 * noticeable problem, like transparent huge
2855 if (IS_ENABLED(CONFIG_COMPACTION) &&
2856 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2857 compaction_ready(zone, sc)) {
2858 sc->compaction_ready = true;
2863 * Shrink each node in the zonelist once. If the
2864 * zonelist is ordered by zone (not the default) then a
2865 * node may be shrunk multiple times but in that case
2866 * the user prefers lower zones being preserved.
2868 if (zone->zone_pgdat == last_pgdat)
2872 * This steals pages from memory cgroups over softlimit
2873 * and returns the number of reclaimed pages and
2874 * scanned pages. This works for global memory pressure
2875 * and balancing, not for a memcg's limit.
2877 nr_soft_scanned = 0;
2878 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2879 sc->order, sc->gfp_mask,
2881 sc->nr_reclaimed += nr_soft_reclaimed;
2882 sc->nr_scanned += nr_soft_scanned;
2883 /* need some check for avoid more shrink_zone() */
2886 /* See comment about same check for global reclaim above */
2887 if (zone->zone_pgdat == last_pgdat)
2889 last_pgdat = zone->zone_pgdat;
2890 shrink_node(zone->zone_pgdat, sc);
2894 * Restore to original mask to avoid the impact on the caller if we
2895 * promoted it to __GFP_HIGHMEM.
2897 sc->gfp_mask = orig_mask;
2900 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2902 struct mem_cgroup *memcg;
2904 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2906 unsigned long refaults;
2907 struct lruvec *lruvec;
2909 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2910 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2911 lruvec->refaults = refaults;
2912 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2916 * This is the main entry point to direct page reclaim.
2918 * If a full scan of the inactive list fails to free enough memory then we
2919 * are "out of memory" and something needs to be killed.
2921 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2922 * high - the zone may be full of dirty or under-writeback pages, which this
2923 * caller can't do much about. We kick the writeback threads and take explicit
2924 * naps in the hope that some of these pages can be written. But if the
2925 * allocating task holds filesystem locks which prevent writeout this might not
2926 * work, and the allocation attempt will fail.
2928 * returns: 0, if no pages reclaimed
2929 * else, the number of pages reclaimed
2931 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2932 struct scan_control *sc)
2934 int initial_priority = sc->priority;
2935 pg_data_t *last_pgdat;
2939 delayacct_freepages_start();
2941 if (global_reclaim(sc))
2942 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2945 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2948 shrink_zones(zonelist, sc);
2950 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2953 if (sc->compaction_ready)
2957 * If we're getting trouble reclaiming, start doing
2958 * writepage even in laptop mode.
2960 if (sc->priority < DEF_PRIORITY - 2)
2961 sc->may_writepage = 1;
2962 } while (--sc->priority >= 0);
2965 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2967 if (zone->zone_pgdat == last_pgdat)
2969 last_pgdat = zone->zone_pgdat;
2970 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2971 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
2974 delayacct_freepages_end();
2976 if (sc->nr_reclaimed)
2977 return sc->nr_reclaimed;
2979 /* Aborted reclaim to try compaction? don't OOM, then */
2980 if (sc->compaction_ready)
2983 /* Untapped cgroup reserves? Don't OOM, retry. */
2984 if (sc->memcg_low_skipped) {
2985 sc->priority = initial_priority;
2986 sc->memcg_low_reclaim = 1;
2987 sc->memcg_low_skipped = 0;
2994 static bool allow_direct_reclaim(pg_data_t *pgdat)
2997 unsigned long pfmemalloc_reserve = 0;
2998 unsigned long free_pages = 0;
3002 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3005 for (i = 0; i <= ZONE_NORMAL; i++) {
3006 zone = &pgdat->node_zones[i];
3007 if (!managed_zone(zone))
3010 if (!zone_reclaimable_pages(zone))
3013 pfmemalloc_reserve += min_wmark_pages(zone);
3014 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3017 /* If there are no reserves (unexpected config) then do not throttle */
3018 if (!pfmemalloc_reserve)
3021 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3023 /* kswapd must be awake if processes are being throttled */
3024 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3025 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3026 (enum zone_type)ZONE_NORMAL);
3027 wake_up_interruptible(&pgdat->kswapd_wait);
3034 * Throttle direct reclaimers if backing storage is backed by the network
3035 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3036 * depleted. kswapd will continue to make progress and wake the processes
3037 * when the low watermark is reached.
3039 * Returns true if a fatal signal was delivered during throttling. If this
3040 * happens, the page allocator should not consider triggering the OOM killer.
3042 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3043 nodemask_t *nodemask)
3047 pg_data_t *pgdat = NULL;
3050 * Kernel threads should not be throttled as they may be indirectly
3051 * responsible for cleaning pages necessary for reclaim to make forward
3052 * progress. kjournald for example may enter direct reclaim while
3053 * committing a transaction where throttling it could forcing other
3054 * processes to block on log_wait_commit().
3056 if (current->flags & PF_KTHREAD)
3060 * If a fatal signal is pending, this process should not throttle.
3061 * It should return quickly so it can exit and free its memory
3063 if (fatal_signal_pending(current))
3067 * Check if the pfmemalloc reserves are ok by finding the first node
3068 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3069 * GFP_KERNEL will be required for allocating network buffers when
3070 * swapping over the network so ZONE_HIGHMEM is unusable.
3072 * Throttling is based on the first usable node and throttled processes
3073 * wait on a queue until kswapd makes progress and wakes them. There
3074 * is an affinity then between processes waking up and where reclaim
3075 * progress has been made assuming the process wakes on the same node.
3076 * More importantly, processes running on remote nodes will not compete
3077 * for remote pfmemalloc reserves and processes on different nodes
3078 * should make reasonable progress.
3080 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3081 gfp_zone(gfp_mask), nodemask) {
3082 if (zone_idx(zone) > ZONE_NORMAL)
3085 /* Throttle based on the first usable node */
3086 pgdat = zone->zone_pgdat;
3087 if (allow_direct_reclaim(pgdat))
3092 /* If no zone was usable by the allocation flags then do not throttle */
3096 /* Account for the throttling */
3097 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3100 * If the caller cannot enter the filesystem, it's possible that it
3101 * is due to the caller holding an FS lock or performing a journal
3102 * transaction in the case of a filesystem like ext[3|4]. In this case,
3103 * it is not safe to block on pfmemalloc_wait as kswapd could be
3104 * blocked waiting on the same lock. Instead, throttle for up to a
3105 * second before continuing.
3107 if (!(gfp_mask & __GFP_FS)) {
3108 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3109 allow_direct_reclaim(pgdat), HZ);
3114 /* Throttle until kswapd wakes the process */
3115 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3116 allow_direct_reclaim(pgdat));
3119 if (fatal_signal_pending(current))
3126 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3127 gfp_t gfp_mask, nodemask_t *nodemask)
3129 unsigned long nr_reclaimed;
3130 struct scan_control sc = {
3131 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3132 .gfp_mask = current_gfp_context(gfp_mask),
3133 .reclaim_idx = gfp_zone(gfp_mask),
3135 .nodemask = nodemask,
3136 .priority = DEF_PRIORITY,
3137 .may_writepage = !laptop_mode,
3140 .may_shrinkslab = 1,
3144 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3145 * Confirm they are large enough for max values.
3147 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3148 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3149 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3152 * Do not enter reclaim if fatal signal was delivered while throttled.
3153 * 1 is returned so that the page allocator does not OOM kill at this
3156 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3159 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3161 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3163 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3165 return nr_reclaimed;
3170 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3171 gfp_t gfp_mask, bool noswap,
3173 unsigned long *nr_scanned)
3175 struct scan_control sc = {
3176 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3177 .target_mem_cgroup = memcg,
3178 .may_writepage = !laptop_mode,
3180 .reclaim_idx = MAX_NR_ZONES - 1,
3181 .may_swap = !noswap,
3182 .may_shrinkslab = 1,
3184 unsigned long lru_pages;
3186 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3187 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3189 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3193 * NOTE: Although we can get the priority field, using it
3194 * here is not a good idea, since it limits the pages we can scan.
3195 * if we don't reclaim here, the shrink_node from balance_pgdat
3196 * will pick up pages from other mem cgroup's as well. We hack
3197 * the priority and make it zero.
3199 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3201 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3203 *nr_scanned = sc.nr_scanned;
3204 return sc.nr_reclaimed;
3207 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3208 unsigned long nr_pages,
3212 struct zonelist *zonelist;
3213 unsigned long nr_reclaimed;
3214 unsigned long pflags;
3216 unsigned int noreclaim_flag;
3217 struct scan_control sc = {
3218 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3219 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3220 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3221 .reclaim_idx = MAX_NR_ZONES - 1,
3222 .target_mem_cgroup = memcg,
3223 .priority = DEF_PRIORITY,
3224 .may_writepage = !laptop_mode,
3226 .may_swap = may_swap,
3227 .may_shrinkslab = 1,
3231 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3232 * take care of from where we get pages. So the node where we start the
3233 * scan does not need to be the current node.
3235 nid = mem_cgroup_select_victim_node(memcg);
3237 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3239 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3241 psi_memstall_enter(&pflags);
3242 noreclaim_flag = memalloc_noreclaim_save();
3244 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3246 memalloc_noreclaim_restore(noreclaim_flag);
3247 psi_memstall_leave(&pflags);
3249 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3251 return nr_reclaimed;
3255 static void age_active_anon(struct pglist_data *pgdat,
3256 struct scan_control *sc)
3258 struct mem_cgroup *memcg;
3260 if (!total_swap_pages)
3263 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3265 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3267 if (inactive_list_is_low(lruvec, false, sc, true))
3268 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3269 sc, LRU_ACTIVE_ANON);
3271 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3275 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3281 * Check for watermark boosts top-down as the higher zones
3282 * are more likely to be boosted. Both watermarks and boosts
3283 * should not be checked at the time time as reclaim would
3284 * start prematurely when there is no boosting and a lower
3287 for (i = classzone_idx; i >= 0; i--) {
3288 zone = pgdat->node_zones + i;
3289 if (!managed_zone(zone))
3292 if (zone->watermark_boost)
3300 * Returns true if there is an eligible zone balanced for the request order
3303 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3306 unsigned long mark = -1;
3310 * Check watermarks bottom-up as lower zones are more likely to
3313 for (i = 0; i <= classzone_idx; i++) {
3314 zone = pgdat->node_zones + i;
3316 if (!managed_zone(zone))
3319 mark = high_wmark_pages(zone);
3320 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3325 * If a node has no populated zone within classzone_idx, it does not
3326 * need balancing by definition. This can happen if a zone-restricted
3327 * allocation tries to wake a remote kswapd.
3335 /* Clear pgdat state for congested, dirty or under writeback. */
3336 static void clear_pgdat_congested(pg_data_t *pgdat)
3338 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3339 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3340 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3344 * Prepare kswapd for sleeping. This verifies that there are no processes
3345 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3347 * Returns true if kswapd is ready to sleep
3349 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3352 * The throttled processes are normally woken up in balance_pgdat() as
3353 * soon as allow_direct_reclaim() is true. But there is a potential
3354 * race between when kswapd checks the watermarks and a process gets
3355 * throttled. There is also a potential race if processes get
3356 * throttled, kswapd wakes, a large process exits thereby balancing the
3357 * zones, which causes kswapd to exit balance_pgdat() before reaching
3358 * the wake up checks. If kswapd is going to sleep, no process should
3359 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3360 * the wake up is premature, processes will wake kswapd and get
3361 * throttled again. The difference from wake ups in balance_pgdat() is
3362 * that here we are under prepare_to_wait().
3364 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3365 wake_up_all(&pgdat->pfmemalloc_wait);
3367 /* Hopeless node, leave it to direct reclaim */
3368 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3371 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3372 clear_pgdat_congested(pgdat);
3380 * kswapd shrinks a node of pages that are at or below the highest usable
3381 * zone that is currently unbalanced.
3383 * Returns true if kswapd scanned at least the requested number of pages to
3384 * reclaim or if the lack of progress was due to pages under writeback.
3385 * This is used to determine if the scanning priority needs to be raised.
3387 static bool kswapd_shrink_node(pg_data_t *pgdat,
3388 struct scan_control *sc)
3393 /* Reclaim a number of pages proportional to the number of zones */
3394 sc->nr_to_reclaim = 0;
3395 for (z = 0; z <= sc->reclaim_idx; z++) {
3396 zone = pgdat->node_zones + z;
3397 if (!managed_zone(zone))
3400 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3404 * Historically care was taken to put equal pressure on all zones but
3405 * now pressure is applied based on node LRU order.
3407 shrink_node(pgdat, sc);
3410 * Fragmentation may mean that the system cannot be rebalanced for
3411 * high-order allocations. If twice the allocation size has been
3412 * reclaimed then recheck watermarks only at order-0 to prevent
3413 * excessive reclaim. Assume that a process requested a high-order
3414 * can direct reclaim/compact.
3416 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3419 return sc->nr_scanned >= sc->nr_to_reclaim;
3423 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3424 * that are eligible for use by the caller until at least one zone is
3427 * Returns the order kswapd finished reclaiming at.
3429 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3430 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3431 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3432 * or lower is eligible for reclaim until at least one usable zone is
3435 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3438 unsigned long nr_soft_reclaimed;
3439 unsigned long nr_soft_scanned;
3440 unsigned long pflags;
3441 unsigned long nr_boost_reclaim;
3442 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3445 struct scan_control sc = {
3446 .gfp_mask = GFP_KERNEL,
3451 psi_memstall_enter(&pflags);
3452 __fs_reclaim_acquire();
3454 count_vm_event(PAGEOUTRUN);
3457 * Account for the reclaim boost. Note that the zone boost is left in
3458 * place so that parallel allocations that are near the watermark will
3459 * stall or direct reclaim until kswapd is finished.
3461 nr_boost_reclaim = 0;
3462 for (i = 0; i <= classzone_idx; i++) {
3463 zone = pgdat->node_zones + i;
3464 if (!managed_zone(zone))
3467 nr_boost_reclaim += zone->watermark_boost;
3468 zone_boosts[i] = zone->watermark_boost;
3470 boosted = nr_boost_reclaim;
3473 sc.priority = DEF_PRIORITY;
3475 unsigned long nr_reclaimed = sc.nr_reclaimed;
3476 bool raise_priority = true;
3480 sc.reclaim_idx = classzone_idx;
3483 * If the number of buffer_heads exceeds the maximum allowed
3484 * then consider reclaiming from all zones. This has a dual
3485 * purpose -- on 64-bit systems it is expected that
3486 * buffer_heads are stripped during active rotation. On 32-bit
3487 * systems, highmem pages can pin lowmem memory and shrinking
3488 * buffers can relieve lowmem pressure. Reclaim may still not
3489 * go ahead if all eligible zones for the original allocation
3490 * request are balanced to avoid excessive reclaim from kswapd.
3492 if (buffer_heads_over_limit) {
3493 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3494 zone = pgdat->node_zones + i;
3495 if (!managed_zone(zone))
3504 * If the pgdat is imbalanced then ignore boosting and preserve
3505 * the watermarks for a later time and restart. Note that the
3506 * zone watermarks will be still reset at the end of balancing
3507 * on the grounds that the normal reclaim should be enough to
3508 * re-evaluate if boosting is required when kswapd next wakes.
3510 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3511 if (!balanced && nr_boost_reclaim) {
3512 nr_boost_reclaim = 0;
3517 * If boosting is not active then only reclaim if there are no
3518 * eligible zones. Note that sc.reclaim_idx is not used as
3519 * buffer_heads_over_limit may have adjusted it.
3521 if (!nr_boost_reclaim && balanced)
3524 /* Limit the priority of boosting to avoid reclaim writeback */
3525 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3526 raise_priority = false;
3529 * Do not writeback or swap pages for boosted reclaim. The
3530 * intent is to relieve pressure not issue sub-optimal IO
3531 * from reclaim context. If no pages are reclaimed, the
3532 * reclaim will be aborted.
3534 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3535 sc.may_swap = !nr_boost_reclaim;
3536 sc.may_shrinkslab = !nr_boost_reclaim;
3539 * Do some background aging of the anon list, to give
3540 * pages a chance to be referenced before reclaiming. All
3541 * pages are rotated regardless of classzone as this is
3542 * about consistent aging.
3544 age_active_anon(pgdat, &sc);
3547 * If we're getting trouble reclaiming, start doing writepage
3548 * even in laptop mode.
3550 if (sc.priority < DEF_PRIORITY - 2)
3551 sc.may_writepage = 1;
3553 /* Call soft limit reclaim before calling shrink_node. */
3555 nr_soft_scanned = 0;
3556 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3557 sc.gfp_mask, &nr_soft_scanned);
3558 sc.nr_reclaimed += nr_soft_reclaimed;
3561 * There should be no need to raise the scanning priority if
3562 * enough pages are already being scanned that that high
3563 * watermark would be met at 100% efficiency.
3565 if (kswapd_shrink_node(pgdat, &sc))
3566 raise_priority = false;
3569 * If the low watermark is met there is no need for processes
3570 * to be throttled on pfmemalloc_wait as they should not be
3571 * able to safely make forward progress. Wake them
3573 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3574 allow_direct_reclaim(pgdat))
3575 wake_up_all(&pgdat->pfmemalloc_wait);
3577 /* Check if kswapd should be suspending */
3578 __fs_reclaim_release();
3579 ret = try_to_freeze();
3580 __fs_reclaim_acquire();
3581 if (ret || kthread_should_stop())
3585 * Raise priority if scanning rate is too low or there was no
3586 * progress in reclaiming pages
3588 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3589 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3592 * If reclaim made no progress for a boost, stop reclaim as
3593 * IO cannot be queued and it could be an infinite loop in
3594 * extreme circumstances.
3596 if (nr_boost_reclaim && !nr_reclaimed)
3599 if (raise_priority || !nr_reclaimed)
3601 } while (sc.priority >= 1);
3603 if (!sc.nr_reclaimed)
3604 pgdat->kswapd_failures++;
3607 /* If reclaim was boosted, account for the reclaim done in this pass */
3609 unsigned long flags;
3611 for (i = 0; i <= classzone_idx; i++) {
3612 if (!zone_boosts[i])
3615 /* Increments are under the zone lock */
3616 zone = pgdat->node_zones + i;
3617 spin_lock_irqsave(&zone->lock, flags);
3618 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3619 spin_unlock_irqrestore(&zone->lock, flags);
3623 * As there is now likely space, wakeup kcompact to defragment
3626 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3629 snapshot_refaults(NULL, pgdat);
3630 __fs_reclaim_release();
3631 psi_memstall_leave(&pflags);
3633 * Return the order kswapd stopped reclaiming at as
3634 * prepare_kswapd_sleep() takes it into account. If another caller
3635 * entered the allocator slow path while kswapd was awake, order will
3636 * remain at the higher level.
3642 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3643 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3644 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3645 * after previous reclaim attempt (node is still unbalanced). In that case
3646 * return the zone index of the previous kswapd reclaim cycle.
3648 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3649 enum zone_type prev_classzone_idx)
3651 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3652 return prev_classzone_idx;
3653 return pgdat->kswapd_classzone_idx;
3656 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3657 unsigned int classzone_idx)
3662 if (freezing(current) || kthread_should_stop())
3665 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3668 * Try to sleep for a short interval. Note that kcompactd will only be
3669 * woken if it is possible to sleep for a short interval. This is
3670 * deliberate on the assumption that if reclaim cannot keep an
3671 * eligible zone balanced that it's also unlikely that compaction will
3674 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3676 * Compaction records what page blocks it recently failed to
3677 * isolate pages from and skips them in the future scanning.
3678 * When kswapd is going to sleep, it is reasonable to assume
3679 * that pages and compaction may succeed so reset the cache.
3681 reset_isolation_suitable(pgdat);
3684 * We have freed the memory, now we should compact it to make
3685 * allocation of the requested order possible.
3687 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3689 remaining = schedule_timeout(HZ/10);
3692 * If woken prematurely then reset kswapd_classzone_idx and
3693 * order. The values will either be from a wakeup request or
3694 * the previous request that slept prematurely.
3697 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3698 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3701 finish_wait(&pgdat->kswapd_wait, &wait);
3702 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3706 * After a short sleep, check if it was a premature sleep. If not, then
3707 * go fully to sleep until explicitly woken up.
3710 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3711 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3714 * vmstat counters are not perfectly accurate and the estimated
3715 * value for counters such as NR_FREE_PAGES can deviate from the
3716 * true value by nr_online_cpus * threshold. To avoid the zone
3717 * watermarks being breached while under pressure, we reduce the
3718 * per-cpu vmstat threshold while kswapd is awake and restore
3719 * them before going back to sleep.
3721 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3723 if (!kthread_should_stop())
3726 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3729 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3731 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3733 finish_wait(&pgdat->kswapd_wait, &wait);
3737 * The background pageout daemon, started as a kernel thread
3738 * from the init process.
3740 * This basically trickles out pages so that we have _some_
3741 * free memory available even if there is no other activity
3742 * that frees anything up. This is needed for things like routing
3743 * etc, where we otherwise might have all activity going on in
3744 * asynchronous contexts that cannot page things out.
3746 * If there are applications that are active memory-allocators
3747 * (most normal use), this basically shouldn't matter.
3749 static int kswapd(void *p)
3751 unsigned int alloc_order, reclaim_order;
3752 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3753 pg_data_t *pgdat = (pg_data_t*)p;
3754 struct task_struct *tsk = current;
3756 struct reclaim_state reclaim_state = {
3757 .reclaimed_slab = 0,
3759 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3761 if (!cpumask_empty(cpumask))
3762 set_cpus_allowed_ptr(tsk, cpumask);
3763 current->reclaim_state = &reclaim_state;
3766 * Tell the memory management that we're a "memory allocator",
3767 * and that if we need more memory we should get access to it
3768 * regardless (see "__alloc_pages()"). "kswapd" should
3769 * never get caught in the normal page freeing logic.
3771 * (Kswapd normally doesn't need memory anyway, but sometimes
3772 * you need a small amount of memory in order to be able to
3773 * page out something else, and this flag essentially protects
3774 * us from recursively trying to free more memory as we're
3775 * trying to free the first piece of memory in the first place).
3777 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3780 pgdat->kswapd_order = 0;
3781 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3785 alloc_order = reclaim_order = pgdat->kswapd_order;
3786 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3789 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3792 /* Read the new order and classzone_idx */
3793 alloc_order = reclaim_order = pgdat->kswapd_order;
3794 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3795 pgdat->kswapd_order = 0;
3796 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3798 ret = try_to_freeze();
3799 if (kthread_should_stop())
3803 * We can speed up thawing tasks if we don't call balance_pgdat
3804 * after returning from the refrigerator
3810 * Reclaim begins at the requested order but if a high-order
3811 * reclaim fails then kswapd falls back to reclaiming for
3812 * order-0. If that happens, kswapd will consider sleeping
3813 * for the order it finished reclaiming at (reclaim_order)
3814 * but kcompactd is woken to compact for the original
3815 * request (alloc_order).
3817 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3819 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3820 if (reclaim_order < alloc_order)
3821 goto kswapd_try_sleep;
3824 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3825 current->reclaim_state = NULL;
3831 * A zone is low on free memory or too fragmented for high-order memory. If
3832 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3833 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3834 * has failed or is not needed, still wake up kcompactd if only compaction is
3837 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3838 enum zone_type classzone_idx)
3842 if (!managed_zone(zone))
3845 if (!cpuset_zone_allowed(zone, gfp_flags))
3847 pgdat = zone->zone_pgdat;
3849 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3850 pgdat->kswapd_classzone_idx = classzone_idx;
3852 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3854 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3855 if (!waitqueue_active(&pgdat->kswapd_wait))
3858 /* Hopeless node, leave it to direct reclaim if possible */
3859 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3860 (pgdat_balanced(pgdat, order, classzone_idx) &&
3861 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3863 * There may be plenty of free memory available, but it's too
3864 * fragmented for high-order allocations. Wake up kcompactd
3865 * and rely on compaction_suitable() to determine if it's
3866 * needed. If it fails, it will defer subsequent attempts to
3867 * ratelimit its work.
3869 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3870 wakeup_kcompactd(pgdat, order, classzone_idx);
3874 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3876 wake_up_interruptible(&pgdat->kswapd_wait);
3879 #ifdef CONFIG_HIBERNATION
3881 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3884 * Rather than trying to age LRUs the aim is to preserve the overall
3885 * LRU order by reclaiming preferentially
3886 * inactive > active > active referenced > active mapped
3888 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3890 struct reclaim_state reclaim_state;
3891 struct scan_control sc = {
3892 .nr_to_reclaim = nr_to_reclaim,
3893 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3894 .reclaim_idx = MAX_NR_ZONES - 1,
3895 .priority = DEF_PRIORITY,
3899 .hibernation_mode = 1,
3901 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3902 struct task_struct *p = current;
3903 unsigned long nr_reclaimed;
3904 unsigned int noreclaim_flag;
3906 fs_reclaim_acquire(sc.gfp_mask);
3907 noreclaim_flag = memalloc_noreclaim_save();
3908 reclaim_state.reclaimed_slab = 0;
3909 p->reclaim_state = &reclaim_state;
3911 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3913 p->reclaim_state = NULL;
3914 memalloc_noreclaim_restore(noreclaim_flag);
3915 fs_reclaim_release(sc.gfp_mask);
3917 return nr_reclaimed;
3919 #endif /* CONFIG_HIBERNATION */
3921 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3922 not required for correctness. So if the last cpu in a node goes
3923 away, we get changed to run anywhere: as the first one comes back,
3924 restore their cpu bindings. */
3925 static int kswapd_cpu_online(unsigned int cpu)
3929 for_each_node_state(nid, N_MEMORY) {
3930 pg_data_t *pgdat = NODE_DATA(nid);
3931 const struct cpumask *mask;
3933 mask = cpumask_of_node(pgdat->node_id);
3935 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3936 /* One of our CPUs online: restore mask */
3937 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3943 * This kswapd start function will be called by init and node-hot-add.
3944 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3946 int kswapd_run(int nid)
3948 pg_data_t *pgdat = NODE_DATA(nid);
3954 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3955 if (IS_ERR(pgdat->kswapd)) {
3956 /* failure at boot is fatal */
3957 BUG_ON(system_state < SYSTEM_RUNNING);
3958 pr_err("Failed to start kswapd on node %d\n", nid);
3959 ret = PTR_ERR(pgdat->kswapd);
3960 pgdat->kswapd = NULL;
3966 * Called by memory hotplug when all memory in a node is offlined. Caller must
3967 * hold mem_hotplug_begin/end().
3969 void kswapd_stop(int nid)
3971 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3974 kthread_stop(kswapd);
3975 NODE_DATA(nid)->kswapd = NULL;
3979 static int __init kswapd_init(void)
3984 for_each_node_state(nid, N_MEMORY)
3986 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3987 "mm/vmscan:online", kswapd_cpu_online,
3993 module_init(kswapd_init)
3999 * If non-zero call node_reclaim when the number of free pages falls below
4002 int node_reclaim_mode __read_mostly;
4004 #define RECLAIM_OFF 0
4005 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4006 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4007 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4010 * Priority for NODE_RECLAIM. This determines the fraction of pages
4011 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4014 #define NODE_RECLAIM_PRIORITY 4
4017 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4020 int sysctl_min_unmapped_ratio = 1;
4023 * If the number of slab pages in a zone grows beyond this percentage then
4024 * slab reclaim needs to occur.
4026 int sysctl_min_slab_ratio = 5;
4028 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4030 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4031 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4032 node_page_state(pgdat, NR_ACTIVE_FILE);
4035 * It's possible for there to be more file mapped pages than
4036 * accounted for by the pages on the file LRU lists because
4037 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4039 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4042 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4043 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4045 unsigned long nr_pagecache_reclaimable;
4046 unsigned long delta = 0;
4049 * If RECLAIM_UNMAP is set, then all file pages are considered
4050 * potentially reclaimable. Otherwise, we have to worry about
4051 * pages like swapcache and node_unmapped_file_pages() provides
4054 if (node_reclaim_mode & RECLAIM_UNMAP)
4055 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4057 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4059 /* If we can't clean pages, remove dirty pages from consideration */
4060 if (!(node_reclaim_mode & RECLAIM_WRITE))
4061 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4063 /* Watch for any possible underflows due to delta */
4064 if (unlikely(delta > nr_pagecache_reclaimable))
4065 delta = nr_pagecache_reclaimable;
4067 return nr_pagecache_reclaimable - delta;
4071 * Try to free up some pages from this node through reclaim.
4073 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4075 /* Minimum pages needed in order to stay on node */
4076 const unsigned long nr_pages = 1 << order;
4077 struct task_struct *p = current;
4078 struct reclaim_state reclaim_state;
4079 unsigned int noreclaim_flag;
4080 struct scan_control sc = {
4081 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4082 .gfp_mask = current_gfp_context(gfp_mask),
4084 .priority = NODE_RECLAIM_PRIORITY,
4085 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4086 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4088 .reclaim_idx = gfp_zone(gfp_mask),
4091 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4095 fs_reclaim_acquire(sc.gfp_mask);
4097 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4098 * and we also need to be able to write out pages for RECLAIM_WRITE
4099 * and RECLAIM_UNMAP.
4101 noreclaim_flag = memalloc_noreclaim_save();
4102 p->flags |= PF_SWAPWRITE;
4103 reclaim_state.reclaimed_slab = 0;
4104 p->reclaim_state = &reclaim_state;
4106 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4108 * Free memory by calling shrink node with increasing
4109 * priorities until we have enough memory freed.
4112 shrink_node(pgdat, &sc);
4113 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4116 p->reclaim_state = NULL;
4117 current->flags &= ~PF_SWAPWRITE;
4118 memalloc_noreclaim_restore(noreclaim_flag);
4119 fs_reclaim_release(sc.gfp_mask);
4121 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4123 return sc.nr_reclaimed >= nr_pages;
4126 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4131 * Node reclaim reclaims unmapped file backed pages and
4132 * slab pages if we are over the defined limits.
4134 * A small portion of unmapped file backed pages is needed for
4135 * file I/O otherwise pages read by file I/O will be immediately
4136 * thrown out if the node is overallocated. So we do not reclaim
4137 * if less than a specified percentage of the node is used by
4138 * unmapped file backed pages.
4140 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4141 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4142 return NODE_RECLAIM_FULL;
4145 * Do not scan if the allocation should not be delayed.
4147 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4148 return NODE_RECLAIM_NOSCAN;
4151 * Only run node reclaim on the local node or on nodes that do not
4152 * have associated processors. This will favor the local processor
4153 * over remote processors and spread off node memory allocations
4154 * as wide as possible.
4156 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4157 return NODE_RECLAIM_NOSCAN;
4159 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4160 return NODE_RECLAIM_NOSCAN;
4162 ret = __node_reclaim(pgdat, gfp_mask, order);
4163 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4166 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4173 * page_evictable - test whether a page is evictable
4174 * @page: the page to test
4176 * Test whether page is evictable--i.e., should be placed on active/inactive
4177 * lists vs unevictable list.
4179 * Reasons page might not be evictable:
4180 * (1) page's mapping marked unevictable
4181 * (2) page is part of an mlocked VMA
4184 int page_evictable(struct page *page)
4188 /* Prevent address_space of inode and swap cache from being freed */
4190 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4196 * check_move_unevictable_pages - check pages for evictability and move to
4197 * appropriate zone lru list
4198 * @pvec: pagevec with lru pages to check
4200 * Checks pages for evictability, if an evictable page is in the unevictable
4201 * lru list, moves it to the appropriate evictable lru list. This function
4202 * should be only used for lru pages.
4204 void check_move_unevictable_pages(struct pagevec *pvec)
4206 struct lruvec *lruvec;
4207 struct pglist_data *pgdat = NULL;
4212 for (i = 0; i < pvec->nr; i++) {
4213 struct page *page = pvec->pages[i];
4214 struct pglist_data *pagepgdat = page_pgdat(page);
4217 if (pagepgdat != pgdat) {
4219 spin_unlock_irq(&pgdat->lru_lock);
4221 spin_lock_irq(&pgdat->lru_lock);
4223 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4225 if (!PageLRU(page) || !PageUnevictable(page))
4228 if (page_evictable(page)) {
4229 enum lru_list lru = page_lru_base_type(page);
4231 VM_BUG_ON_PAGE(PageActive(page), page);
4232 ClearPageUnevictable(page);
4233 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4234 add_page_to_lru_list(page, lruvec, lru);
4240 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4241 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4242 spin_unlock_irq(&pgdat->lru_lock);
4245 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);