4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
53 #include <linux/swapops.h>
54 #include <linux/balloon_compaction.h>
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/vmscan.h>
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
65 /* This context's GFP mask */
68 /* Allocation order */
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup *target_mem_cgroup;
83 /* Scan (total_size >> priority) pages at once */
86 unsigned int may_writepage:1;
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
94 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash:1;
97 unsigned int hibernation_mode:1;
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned;
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed;
109 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
111 #ifdef ARCH_HAS_PREFETCH
112 #define prefetch_prev_lru_page(_page, _base, _field) \
114 if ((_page)->lru.prev != _base) { \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetch(&prev->_field); \
122 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125 #ifdef ARCH_HAS_PREFETCHW
126 #define prefetchw_prev_lru_page(_page, _base, _field) \
128 if ((_page)->lru.prev != _base) { \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetchw(&prev->_field); \
136 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
140 * From 0 .. 100. Higher means more swappy.
142 int vm_swappiness = 60;
144 * The total number of pages which are beyond the high watermark within all
147 unsigned long vm_total_pages;
149 static LIST_HEAD(shrinker_list);
150 static DECLARE_RWSEM(shrinker_rwsem);
153 static bool global_reclaim(struct scan_control *sc)
155 return !sc->target_mem_cgroup;
159 * sane_reclaim - is the usual dirty throttling mechanism operational?
160 * @sc: scan_control in question
162 * The normal page dirty throttling mechanism in balance_dirty_pages() is
163 * completely broken with the legacy memcg and direct stalling in
164 * shrink_page_list() is used for throttling instead, which lacks all the
165 * niceties such as fairness, adaptive pausing, bandwidth proportional
166 * allocation and configurability.
168 * This function tests whether the vmscan currently in progress can assume
169 * that the normal dirty throttling mechanism is operational.
171 static bool sane_reclaim(struct scan_control *sc)
173 struct mem_cgroup *memcg = sc->target_mem_cgroup;
177 #ifdef CONFIG_CGROUP_WRITEBACK
178 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
184 static bool global_reclaim(struct scan_control *sc)
189 static bool sane_reclaim(struct scan_control *sc)
195 static unsigned long zone_reclaimable_pages(struct zone *zone)
199 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
200 zone_page_state(zone, NR_INACTIVE_FILE);
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
204 zone_page_state(zone, NR_INACTIVE_ANON);
209 bool zone_reclaimable(struct zone *zone)
211 return zone_page_state(zone, NR_PAGES_SCANNED) <
212 zone_reclaimable_pages(zone) * 6;
215 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
217 if (!mem_cgroup_disabled())
218 return mem_cgroup_get_lru_size(lruvec, lru);
220 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
224 * Add a shrinker callback to be called from the vm.
226 int register_shrinker(struct shrinker *shrinker)
228 size_t size = sizeof(*shrinker->nr_deferred);
231 * If we only have one possible node in the system anyway, save
232 * ourselves the trouble and disable NUMA aware behavior. This way we
233 * will save memory and some small loop time later.
235 if (nr_node_ids == 1)
236 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
238 if (shrinker->flags & SHRINKER_NUMA_AWARE)
241 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
242 if (!shrinker->nr_deferred)
245 down_write(&shrinker_rwsem);
246 list_add_tail(&shrinker->list, &shrinker_list);
247 up_write(&shrinker_rwsem);
250 EXPORT_SYMBOL(register_shrinker);
255 void unregister_shrinker(struct shrinker *shrinker)
257 down_write(&shrinker_rwsem);
258 list_del(&shrinker->list);
259 up_write(&shrinker_rwsem);
260 kfree(shrinker->nr_deferred);
262 EXPORT_SYMBOL(unregister_shrinker);
264 #define SHRINK_BATCH 128
266 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
267 struct shrinker *shrinker,
268 unsigned long nr_scanned,
269 unsigned long nr_eligible)
271 unsigned long freed = 0;
272 unsigned long long delta;
277 int nid = shrinkctl->nid;
278 long batch_size = shrinker->batch ? shrinker->batch
281 freeable = shrinker->count_objects(shrinker, shrinkctl);
286 * copy the current shrinker scan count into a local variable
287 * and zero it so that other concurrent shrinker invocations
288 * don't also do this scanning work.
290 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
293 delta = (4 * nr_scanned) / shrinker->seeks;
295 do_div(delta, nr_eligible + 1);
297 if (total_scan < 0) {
298 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
299 shrinker->scan_objects, total_scan);
300 total_scan = freeable;
304 * We need to avoid excessive windup on filesystem shrinkers
305 * due to large numbers of GFP_NOFS allocations causing the
306 * shrinkers to return -1 all the time. This results in a large
307 * nr being built up so when a shrink that can do some work
308 * comes along it empties the entire cache due to nr >>>
309 * freeable. This is bad for sustaining a working set in
312 * Hence only allow the shrinker to scan the entire cache when
313 * a large delta change is calculated directly.
315 if (delta < freeable / 4)
316 total_scan = min(total_scan, freeable / 2);
319 * Avoid risking looping forever due to too large nr value:
320 * never try to free more than twice the estimate number of
323 if (total_scan > freeable * 2)
324 total_scan = freeable * 2;
326 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
327 nr_scanned, nr_eligible,
328 freeable, delta, total_scan);
331 * Normally, we should not scan less than batch_size objects in one
332 * pass to avoid too frequent shrinker calls, but if the slab has less
333 * than batch_size objects in total and we are really tight on memory,
334 * we will try to reclaim all available objects, otherwise we can end
335 * up failing allocations although there are plenty of reclaimable
336 * objects spread over several slabs with usage less than the
339 * We detect the "tight on memory" situations by looking at the total
340 * number of objects we want to scan (total_scan). If it is greater
341 * than the total number of objects on slab (freeable), we must be
342 * scanning at high prio and therefore should try to reclaim as much as
345 while (total_scan >= batch_size ||
346 total_scan >= freeable) {
348 unsigned long nr_to_scan = min(batch_size, total_scan);
350 shrinkctl->nr_to_scan = nr_to_scan;
351 ret = shrinker->scan_objects(shrinker, shrinkctl);
352 if (ret == SHRINK_STOP)
356 count_vm_events(SLABS_SCANNED, nr_to_scan);
357 total_scan -= nr_to_scan;
363 * move the unused scan count back into the shrinker in a
364 * manner that handles concurrent updates. If we exhausted the
365 * scan, there is no need to do an update.
368 new_nr = atomic_long_add_return(total_scan,
369 &shrinker->nr_deferred[nid]);
371 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
373 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
378 * shrink_slab - shrink slab caches
379 * @gfp_mask: allocation context
380 * @nid: node whose slab caches to target
381 * @memcg: memory cgroup whose slab caches to target
382 * @nr_scanned: pressure numerator
383 * @nr_eligible: pressure denominator
385 * Call the shrink functions to age shrinkable caches.
387 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
388 * unaware shrinkers will receive a node id of 0 instead.
390 * @memcg specifies the memory cgroup to target. If it is not NULL,
391 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
392 * objects from the memory cgroup specified. Otherwise all shrinkers
393 * are called, and memcg aware shrinkers are supposed to scan the
396 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
397 * the available objects should be scanned. Page reclaim for example
398 * passes the number of pages scanned and the number of pages on the
399 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
400 * when it encountered mapped pages. The ratio is further biased by
401 * the ->seeks setting of the shrink function, which indicates the
402 * cost to recreate an object relative to that of an LRU page.
404 * Returns the number of reclaimed slab objects.
406 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
407 struct mem_cgroup *memcg,
408 unsigned long nr_scanned,
409 unsigned long nr_eligible)
411 struct shrinker *shrinker;
412 unsigned long freed = 0;
414 if (memcg && !memcg_kmem_is_active(memcg))
418 nr_scanned = SWAP_CLUSTER_MAX;
420 if (!down_read_trylock(&shrinker_rwsem)) {
422 * If we would return 0, our callers would understand that we
423 * have nothing else to shrink and give up trying. By returning
424 * 1 we keep it going and assume we'll be able to shrink next
431 list_for_each_entry(shrinker, &shrinker_list, list) {
432 struct shrink_control sc = {
433 .gfp_mask = gfp_mask,
438 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
441 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
444 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
447 up_read(&shrinker_rwsem);
453 void drop_slab_node(int nid)
458 struct mem_cgroup *memcg = NULL;
462 freed += shrink_slab(GFP_KERNEL, nid, memcg,
464 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
465 } while (freed > 10);
472 for_each_online_node(nid)
476 static inline int is_page_cache_freeable(struct page *page)
479 * A freeable page cache page is referenced only by the caller
480 * that isolated the page, the page cache radix tree and
481 * optional buffer heads at page->private.
483 return page_count(page) - page_has_private(page) == 2;
486 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
488 if (current->flags & PF_SWAPWRITE)
490 if (!inode_write_congested(inode))
492 if (inode_to_bdi(inode) == current->backing_dev_info)
498 * We detected a synchronous write error writing a page out. Probably
499 * -ENOSPC. We need to propagate that into the address_space for a subsequent
500 * fsync(), msync() or close().
502 * The tricky part is that after writepage we cannot touch the mapping: nothing
503 * prevents it from being freed up. But we have a ref on the page and once
504 * that page is locked, the mapping is pinned.
506 * We're allowed to run sleeping lock_page() here because we know the caller has
509 static void handle_write_error(struct address_space *mapping,
510 struct page *page, int error)
513 if (page_mapping(page) == mapping)
514 mapping_set_error(mapping, error);
518 /* possible outcome of pageout() */
520 /* failed to write page out, page is locked */
522 /* move page to the active list, page is locked */
524 /* page has been sent to the disk successfully, page is unlocked */
526 /* page is clean and locked */
531 * pageout is called by shrink_page_list() for each dirty page.
532 * Calls ->writepage().
534 static pageout_t pageout(struct page *page, struct address_space *mapping,
535 struct scan_control *sc)
538 * If the page is dirty, only perform writeback if that write
539 * will be non-blocking. To prevent this allocation from being
540 * stalled by pagecache activity. But note that there may be
541 * stalls if we need to run get_block(). We could test
542 * PagePrivate for that.
544 * If this process is currently in __generic_file_write_iter() against
545 * this page's queue, we can perform writeback even if that
548 * If the page is swapcache, write it back even if that would
549 * block, for some throttling. This happens by accident, because
550 * swap_backing_dev_info is bust: it doesn't reflect the
551 * congestion state of the swapdevs. Easy to fix, if needed.
553 if (!is_page_cache_freeable(page))
557 * Some data journaling orphaned pages can have
558 * page->mapping == NULL while being dirty with clean buffers.
560 if (page_has_private(page)) {
561 if (try_to_free_buffers(page)) {
562 ClearPageDirty(page);
563 pr_info("%s: orphaned page\n", __func__);
569 if (mapping->a_ops->writepage == NULL)
570 return PAGE_ACTIVATE;
571 if (!may_write_to_inode(mapping->host, sc))
574 if (clear_page_dirty_for_io(page)) {
576 struct writeback_control wbc = {
577 .sync_mode = WB_SYNC_NONE,
578 .nr_to_write = SWAP_CLUSTER_MAX,
580 .range_end = LLONG_MAX,
584 SetPageReclaim(page);
585 res = mapping->a_ops->writepage(page, &wbc);
587 handle_write_error(mapping, page, res);
588 if (res == AOP_WRITEPAGE_ACTIVATE) {
589 ClearPageReclaim(page);
590 return PAGE_ACTIVATE;
593 if (!PageWriteback(page)) {
594 /* synchronous write or broken a_ops? */
595 ClearPageReclaim(page);
597 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
598 inc_zone_page_state(page, NR_VMSCAN_WRITE);
606 * Same as remove_mapping, but if the page is removed from the mapping, it
607 * gets returned with a refcount of 0.
609 static int __remove_mapping(struct address_space *mapping, struct page *page,
613 struct mem_cgroup *memcg;
615 BUG_ON(!PageLocked(page));
616 BUG_ON(mapping != page_mapping(page));
618 memcg = mem_cgroup_begin_page_stat(page);
619 spin_lock_irqsave(&mapping->tree_lock, flags);
621 * The non racy check for a busy page.
623 * Must be careful with the order of the tests. When someone has
624 * a ref to the page, it may be possible that they dirty it then
625 * drop the reference. So if PageDirty is tested before page_count
626 * here, then the following race may occur:
628 * get_user_pages(&page);
629 * [user mapping goes away]
631 * !PageDirty(page) [good]
632 * SetPageDirty(page);
634 * !page_count(page) [good, discard it]
636 * [oops, our write_to data is lost]
638 * Reversing the order of the tests ensures such a situation cannot
639 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
640 * load is not satisfied before that of page->_count.
642 * Note that if SetPageDirty is always performed via set_page_dirty,
643 * and thus under tree_lock, then this ordering is not required.
645 if (!page_freeze_refs(page, 2))
647 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
648 if (unlikely(PageDirty(page))) {
649 page_unfreeze_refs(page, 2);
653 if (PageSwapCache(page)) {
654 swp_entry_t swap = { .val = page_private(page) };
655 mem_cgroup_swapout(page, swap);
656 __delete_from_swap_cache(page);
657 spin_unlock_irqrestore(&mapping->tree_lock, flags);
658 mem_cgroup_end_page_stat(memcg);
659 swapcache_free(swap);
661 void (*freepage)(struct page *);
664 freepage = mapping->a_ops->freepage;
666 * Remember a shadow entry for reclaimed file cache in
667 * order to detect refaults, thus thrashing, later on.
669 * But don't store shadows in an address space that is
670 * already exiting. This is not just an optizimation,
671 * inode reclaim needs to empty out the radix tree or
672 * the nodes are lost. Don't plant shadows behind its
675 if (reclaimed && page_is_file_cache(page) &&
676 !mapping_exiting(mapping))
677 shadow = workingset_eviction(mapping, page);
678 __delete_from_page_cache(page, shadow, memcg);
679 spin_unlock_irqrestore(&mapping->tree_lock, flags);
680 mem_cgroup_end_page_stat(memcg);
682 if (freepage != NULL)
689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
690 mem_cgroup_end_page_stat(memcg);
695 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
696 * someone else has a ref on the page, abort and return 0. If it was
697 * successfully detached, return 1. Assumes the caller has a single ref on
700 int remove_mapping(struct address_space *mapping, struct page *page)
702 if (__remove_mapping(mapping, page, false)) {
704 * Unfreezing the refcount with 1 rather than 2 effectively
705 * drops the pagecache ref for us without requiring another
708 page_unfreeze_refs(page, 1);
715 * putback_lru_page - put previously isolated page onto appropriate LRU list
716 * @page: page to be put back to appropriate lru list
718 * Add previously isolated @page to appropriate LRU list.
719 * Page may still be unevictable for other reasons.
721 * lru_lock must not be held, interrupts must be enabled.
723 void putback_lru_page(struct page *page)
726 int was_unevictable = PageUnevictable(page);
728 VM_BUG_ON_PAGE(PageLRU(page), page);
731 ClearPageUnevictable(page);
733 if (page_evictable(page)) {
735 * For evictable pages, we can use the cache.
736 * In event of a race, worst case is we end up with an
737 * unevictable page on [in]active list.
738 * We know how to handle that.
740 is_unevictable = false;
744 * Put unevictable pages directly on zone's unevictable
747 is_unevictable = true;
748 add_page_to_unevictable_list(page);
750 * When racing with an mlock or AS_UNEVICTABLE clearing
751 * (page is unlocked) make sure that if the other thread
752 * does not observe our setting of PG_lru and fails
753 * isolation/check_move_unevictable_pages,
754 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
755 * the page back to the evictable list.
757 * The other side is TestClearPageMlocked() or shmem_lock().
763 * page's status can change while we move it among lru. If an evictable
764 * page is on unevictable list, it never be freed. To avoid that,
765 * check after we added it to the list, again.
767 if (is_unevictable && page_evictable(page)) {
768 if (!isolate_lru_page(page)) {
772 /* This means someone else dropped this page from LRU
773 * So, it will be freed or putback to LRU again. There is
774 * nothing to do here.
778 if (was_unevictable && !is_unevictable)
779 count_vm_event(UNEVICTABLE_PGRESCUED);
780 else if (!was_unevictable && is_unevictable)
781 count_vm_event(UNEVICTABLE_PGCULLED);
783 put_page(page); /* drop ref from isolate */
786 enum page_references {
788 PAGEREF_RECLAIM_CLEAN,
793 static enum page_references page_check_references(struct page *page,
794 struct scan_control *sc)
796 int referenced_ptes, referenced_page;
797 unsigned long vm_flags;
799 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
801 referenced_page = TestClearPageReferenced(page);
804 * Mlock lost the isolation race with us. Let try_to_unmap()
805 * move the page to the unevictable list.
807 if (vm_flags & VM_LOCKED)
808 return PAGEREF_RECLAIM;
810 if (referenced_ptes) {
811 if (PageSwapBacked(page))
812 return PAGEREF_ACTIVATE;
814 * All mapped pages start out with page table
815 * references from the instantiating fault, so we need
816 * to look twice if a mapped file page is used more
819 * Mark it and spare it for another trip around the
820 * inactive list. Another page table reference will
821 * lead to its activation.
823 * Note: the mark is set for activated pages as well
824 * so that recently deactivated but used pages are
827 SetPageReferenced(page);
829 if (referenced_page || referenced_ptes > 1)
830 return PAGEREF_ACTIVATE;
833 * Activate file-backed executable pages after first usage.
835 if (vm_flags & VM_EXEC)
836 return PAGEREF_ACTIVATE;
841 /* Reclaim if clean, defer dirty pages to writeback */
842 if (referenced_page && !PageSwapBacked(page))
843 return PAGEREF_RECLAIM_CLEAN;
845 return PAGEREF_RECLAIM;
848 /* Check if a page is dirty or under writeback */
849 static void page_check_dirty_writeback(struct page *page,
850 bool *dirty, bool *writeback)
852 struct address_space *mapping;
855 * Anonymous pages are not handled by flushers and must be written
856 * from reclaim context. Do not stall reclaim based on them
858 if (!page_is_file_cache(page)) {
864 /* By default assume that the page flags are accurate */
865 *dirty = PageDirty(page);
866 *writeback = PageWriteback(page);
868 /* Verify dirty/writeback state if the filesystem supports it */
869 if (!page_has_private(page))
872 mapping = page_mapping(page);
873 if (mapping && mapping->a_ops->is_dirty_writeback)
874 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
878 * shrink_page_list() returns the number of reclaimed pages
880 static unsigned long shrink_page_list(struct list_head *page_list,
882 struct scan_control *sc,
883 enum ttu_flags ttu_flags,
884 unsigned long *ret_nr_dirty,
885 unsigned long *ret_nr_unqueued_dirty,
886 unsigned long *ret_nr_congested,
887 unsigned long *ret_nr_writeback,
888 unsigned long *ret_nr_immediate,
891 LIST_HEAD(ret_pages);
892 LIST_HEAD(free_pages);
894 unsigned long nr_unqueued_dirty = 0;
895 unsigned long nr_dirty = 0;
896 unsigned long nr_congested = 0;
897 unsigned long nr_reclaimed = 0;
898 unsigned long nr_writeback = 0;
899 unsigned long nr_immediate = 0;
903 while (!list_empty(page_list)) {
904 struct address_space *mapping;
907 enum page_references references = PAGEREF_RECLAIM_CLEAN;
908 bool dirty, writeback;
912 page = lru_to_page(page_list);
913 list_del(&page->lru);
915 if (!trylock_page(page))
918 VM_BUG_ON_PAGE(PageActive(page), page);
919 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
923 if (unlikely(!page_evictable(page)))
926 if (!sc->may_unmap && page_mapped(page))
929 /* Double the slab pressure for mapped and swapcache pages */
930 if (page_mapped(page) || PageSwapCache(page))
933 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
934 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
937 * The number of dirty pages determines if a zone is marked
938 * reclaim_congested which affects wait_iff_congested. kswapd
939 * will stall and start writing pages if the tail of the LRU
940 * is all dirty unqueued pages.
942 page_check_dirty_writeback(page, &dirty, &writeback);
943 if (dirty || writeback)
946 if (dirty && !writeback)
950 * Treat this page as congested if the underlying BDI is or if
951 * pages are cycling through the LRU so quickly that the
952 * pages marked for immediate reclaim are making it to the
953 * end of the LRU a second time.
955 mapping = page_mapping(page);
956 if (((dirty || writeback) && mapping &&
957 inode_write_congested(mapping->host)) ||
958 (writeback && PageReclaim(page)))
962 * If a page at the tail of the LRU is under writeback, there
963 * are three cases to consider.
965 * 1) If reclaim is encountering an excessive number of pages
966 * under writeback and this page is both under writeback and
967 * PageReclaim then it indicates that pages are being queued
968 * for IO but are being recycled through the LRU before the
969 * IO can complete. Waiting on the page itself risks an
970 * indefinite stall if it is impossible to writeback the
971 * page due to IO error or disconnected storage so instead
972 * note that the LRU is being scanned too quickly and the
973 * caller can stall after page list has been processed.
975 * 2) Global or new memcg reclaim encounters a page that is
976 * not marked for immediate reclaim, or the caller does not
977 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
978 * not to fs). In this case mark the page for immediate
979 * reclaim and continue scanning.
981 * Require may_enter_fs because we would wait on fs, which
982 * may not have submitted IO yet. And the loop driver might
983 * enter reclaim, and deadlock if it waits on a page for
984 * which it is needed to do the write (loop masks off
985 * __GFP_IO|__GFP_FS for this reason); but more thought
986 * would probably show more reasons.
988 * 3) Legacy memcg encounters a page that is already marked
989 * PageReclaim. memcg does not have any dirty pages
990 * throttling so we could easily OOM just because too many
991 * pages are in writeback and there is nothing else to
992 * reclaim. Wait for the writeback to complete.
994 if (PageWriteback(page)) {
996 if (current_is_kswapd() &&
998 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1003 } else if (sane_reclaim(sc) ||
1004 !PageReclaim(page) || !may_enter_fs) {
1006 * This is slightly racy - end_page_writeback()
1007 * might have just cleared PageReclaim, then
1008 * setting PageReclaim here end up interpreted
1009 * as PageReadahead - but that does not matter
1010 * enough to care. What we do want is for this
1011 * page to have PageReclaim set next time memcg
1012 * reclaim reaches the tests above, so it will
1013 * then wait_on_page_writeback() to avoid OOM;
1014 * and it's also appropriate in global reclaim.
1016 SetPageReclaim(page);
1023 wait_on_page_writeback(page);
1024 /* then go back and try same page again */
1025 list_add_tail(&page->lru, page_list);
1031 references = page_check_references(page, sc);
1033 switch (references) {
1034 case PAGEREF_ACTIVATE:
1035 goto activate_locked;
1038 case PAGEREF_RECLAIM:
1039 case PAGEREF_RECLAIM_CLEAN:
1040 ; /* try to reclaim the page below */
1044 * Anonymous process memory has backing store?
1045 * Try to allocate it some swap space here.
1047 if (PageAnon(page) && !PageSwapCache(page)) {
1048 if (!(sc->gfp_mask & __GFP_IO))
1050 if (!add_to_swap(page, page_list))
1051 goto activate_locked;
1054 /* Adding to swap updated mapping */
1055 mapping = page_mapping(page);
1059 * The page is mapped into the page tables of one or more
1060 * processes. Try to unmap it here.
1062 if (page_mapped(page) && mapping) {
1063 switch (try_to_unmap(page,
1064 ttu_flags|TTU_BATCH_FLUSH)) {
1066 goto activate_locked;
1072 ; /* try to free the page below */
1076 if (PageDirty(page)) {
1078 * Only kswapd can writeback filesystem pages to
1079 * avoid risk of stack overflow but only writeback
1080 * if many dirty pages have been encountered.
1082 if (page_is_file_cache(page) &&
1083 (!current_is_kswapd() ||
1084 !test_bit(ZONE_DIRTY, &zone->flags))) {
1086 * Immediately reclaim when written back.
1087 * Similar in principal to deactivate_page()
1088 * except we already have the page isolated
1089 * and know it's dirty
1091 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1092 SetPageReclaim(page);
1097 if (references == PAGEREF_RECLAIM_CLEAN)
1101 if (!sc->may_writepage)
1105 * Page is dirty. Flush the TLB if a writable entry
1106 * potentially exists to avoid CPU writes after IO
1107 * starts and then write it out here.
1109 try_to_unmap_flush_dirty();
1110 switch (pageout(page, mapping, sc)) {
1114 goto activate_locked;
1116 if (PageWriteback(page))
1118 if (PageDirty(page))
1122 * A synchronous write - probably a ramdisk. Go
1123 * ahead and try to reclaim the page.
1125 if (!trylock_page(page))
1127 if (PageDirty(page) || PageWriteback(page))
1129 mapping = page_mapping(page);
1131 ; /* try to free the page below */
1136 * If the page has buffers, try to free the buffer mappings
1137 * associated with this page. If we succeed we try to free
1140 * We do this even if the page is PageDirty().
1141 * try_to_release_page() does not perform I/O, but it is
1142 * possible for a page to have PageDirty set, but it is actually
1143 * clean (all its buffers are clean). This happens if the
1144 * buffers were written out directly, with submit_bh(). ext3
1145 * will do this, as well as the blockdev mapping.
1146 * try_to_release_page() will discover that cleanness and will
1147 * drop the buffers and mark the page clean - it can be freed.
1149 * Rarely, pages can have buffers and no ->mapping. These are
1150 * the pages which were not successfully invalidated in
1151 * truncate_complete_page(). We try to drop those buffers here
1152 * and if that worked, and the page is no longer mapped into
1153 * process address space (page_count == 1) it can be freed.
1154 * Otherwise, leave the page on the LRU so it is swappable.
1156 if (page_has_private(page)) {
1157 if (!try_to_release_page(page, sc->gfp_mask))
1158 goto activate_locked;
1159 if (!mapping && page_count(page) == 1) {
1161 if (put_page_testzero(page))
1165 * rare race with speculative reference.
1166 * the speculative reference will free
1167 * this page shortly, so we may
1168 * increment nr_reclaimed here (and
1169 * leave it off the LRU).
1177 if (!mapping || !__remove_mapping(mapping, page, true))
1181 * At this point, we have no other references and there is
1182 * no way to pick any more up (removed from LRU, removed
1183 * from pagecache). Can use non-atomic bitops now (and
1184 * we obviously don't have to worry about waking up a process
1185 * waiting on the page lock, because there are no references.
1187 __clear_page_locked(page);
1192 * Is there need to periodically free_page_list? It would
1193 * appear not as the counts should be low
1195 list_add(&page->lru, &free_pages);
1199 if (PageSwapCache(page))
1200 try_to_free_swap(page);
1202 list_add(&page->lru, &ret_pages);
1206 /* Not a candidate for swapping, so reclaim swap space. */
1207 if (PageSwapCache(page) && vm_swap_full())
1208 try_to_free_swap(page);
1209 VM_BUG_ON_PAGE(PageActive(page), page);
1210 SetPageActive(page);
1215 list_add(&page->lru, &ret_pages);
1216 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1219 mem_cgroup_uncharge_list(&free_pages);
1220 try_to_unmap_flush();
1221 free_hot_cold_page_list(&free_pages, true);
1223 list_splice(&ret_pages, page_list);
1224 count_vm_events(PGACTIVATE, pgactivate);
1226 *ret_nr_dirty += nr_dirty;
1227 *ret_nr_congested += nr_congested;
1228 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1229 *ret_nr_writeback += nr_writeback;
1230 *ret_nr_immediate += nr_immediate;
1231 return nr_reclaimed;
1234 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1235 struct list_head *page_list)
1237 struct scan_control sc = {
1238 .gfp_mask = GFP_KERNEL,
1239 .priority = DEF_PRIORITY,
1242 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1243 struct page *page, *next;
1244 LIST_HEAD(clean_pages);
1246 list_for_each_entry_safe(page, next, page_list, lru) {
1247 if (page_is_file_cache(page) && !PageDirty(page) &&
1248 !isolated_balloon_page(page)) {
1249 ClearPageActive(page);
1250 list_move(&page->lru, &clean_pages);
1254 ret = shrink_page_list(&clean_pages, zone, &sc,
1255 TTU_UNMAP|TTU_IGNORE_ACCESS,
1256 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1257 list_splice(&clean_pages, page_list);
1258 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1263 * Attempt to remove the specified page from its LRU. Only take this page
1264 * if it is of the appropriate PageActive status. Pages which are being
1265 * freed elsewhere are also ignored.
1267 * page: page to consider
1268 * mode: one of the LRU isolation modes defined above
1270 * returns 0 on success, -ve errno on failure.
1272 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1276 /* Only take pages on the LRU. */
1280 /* Compaction should not handle unevictable pages but CMA can do so */
1281 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1287 * To minimise LRU disruption, the caller can indicate that it only
1288 * wants to isolate pages it will be able to operate on without
1289 * blocking - clean pages for the most part.
1291 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1292 * is used by reclaim when it is cannot write to backing storage
1294 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1295 * that it is possible to migrate without blocking
1297 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1298 /* All the caller can do on PageWriteback is block */
1299 if (PageWriteback(page))
1302 if (PageDirty(page)) {
1303 struct address_space *mapping;
1305 /* ISOLATE_CLEAN means only clean pages */
1306 if (mode & ISOLATE_CLEAN)
1310 * Only pages without mappings or that have a
1311 * ->migratepage callback are possible to migrate
1314 mapping = page_mapping(page);
1315 if (mapping && !mapping->a_ops->migratepage)
1320 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1323 if (likely(get_page_unless_zero(page))) {
1325 * Be careful not to clear PageLRU until after we're
1326 * sure the page is not being freed elsewhere -- the
1327 * page release code relies on it.
1337 * zone->lru_lock is heavily contended. Some of the functions that
1338 * shrink the lists perform better by taking out a batch of pages
1339 * and working on them outside the LRU lock.
1341 * For pagecache intensive workloads, this function is the hottest
1342 * spot in the kernel (apart from copy_*_user functions).
1344 * Appropriate locks must be held before calling this function.
1346 * @nr_to_scan: The number of pages to look through on the list.
1347 * @lruvec: The LRU vector to pull pages from.
1348 * @dst: The temp list to put pages on to.
1349 * @nr_scanned: The number of pages that were scanned.
1350 * @sc: The scan_control struct for this reclaim session
1351 * @mode: One of the LRU isolation modes
1352 * @lru: LRU list id for isolating
1354 * returns how many pages were moved onto *@dst.
1356 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1357 struct lruvec *lruvec, struct list_head *dst,
1358 unsigned long *nr_scanned, struct scan_control *sc,
1359 isolate_mode_t mode, enum lru_list lru)
1361 struct list_head *src = &lruvec->lists[lru];
1362 unsigned long nr_taken = 0;
1365 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1366 !list_empty(src); scan++) {
1370 page = lru_to_page(src);
1371 prefetchw_prev_lru_page(page, src, flags);
1373 VM_BUG_ON_PAGE(!PageLRU(page), page);
1375 switch (__isolate_lru_page(page, mode)) {
1377 nr_pages = hpage_nr_pages(page);
1378 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1379 list_move(&page->lru, dst);
1380 nr_taken += nr_pages;
1384 /* else it is being freed elsewhere */
1385 list_move(&page->lru, src);
1394 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1395 nr_taken, mode, is_file_lru(lru));
1400 * isolate_lru_page - tries to isolate a page from its LRU list
1401 * @page: page to isolate from its LRU list
1403 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1404 * vmstat statistic corresponding to whatever LRU list the page was on.
1406 * Returns 0 if the page was removed from an LRU list.
1407 * Returns -EBUSY if the page was not on an LRU list.
1409 * The returned page will have PageLRU() cleared. If it was found on
1410 * the active list, it will have PageActive set. If it was found on
1411 * the unevictable list, it will have the PageUnevictable bit set. That flag
1412 * may need to be cleared by the caller before letting the page go.
1414 * The vmstat statistic corresponding to the list on which the page was
1415 * found will be decremented.
1418 * (1) Must be called with an elevated refcount on the page. This is a
1419 * fundamentnal difference from isolate_lru_pages (which is called
1420 * without a stable reference).
1421 * (2) the lru_lock must not be held.
1422 * (3) interrupts must be enabled.
1424 int isolate_lru_page(struct page *page)
1428 VM_BUG_ON_PAGE(!page_count(page), page);
1430 if (PageLRU(page)) {
1431 struct zone *zone = page_zone(page);
1432 struct lruvec *lruvec;
1434 spin_lock_irq(&zone->lru_lock);
1435 lruvec = mem_cgroup_page_lruvec(page, zone);
1436 if (PageLRU(page)) {
1437 int lru = page_lru(page);
1440 del_page_from_lru_list(page, lruvec, lru);
1443 spin_unlock_irq(&zone->lru_lock);
1449 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1450 * then get resheduled. When there are massive number of tasks doing page
1451 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1452 * the LRU list will go small and be scanned faster than necessary, leading to
1453 * unnecessary swapping, thrashing and OOM.
1455 static int too_many_isolated(struct zone *zone, int file,
1456 struct scan_control *sc)
1458 unsigned long inactive, isolated;
1460 if (current_is_kswapd())
1463 if (!sane_reclaim(sc))
1467 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1468 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1470 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1471 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1475 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1476 * won't get blocked by normal direct-reclaimers, forming a circular
1479 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1482 return isolated > inactive;
1485 static noinline_for_stack void
1486 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1488 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1489 struct zone *zone = lruvec_zone(lruvec);
1490 LIST_HEAD(pages_to_free);
1493 * Put back any unfreeable pages.
1495 while (!list_empty(page_list)) {
1496 struct page *page = lru_to_page(page_list);
1499 VM_BUG_ON_PAGE(PageLRU(page), page);
1500 list_del(&page->lru);
1501 if (unlikely(!page_evictable(page))) {
1502 spin_unlock_irq(&zone->lru_lock);
1503 putback_lru_page(page);
1504 spin_lock_irq(&zone->lru_lock);
1508 lruvec = mem_cgroup_page_lruvec(page, zone);
1511 lru = page_lru(page);
1512 add_page_to_lru_list(page, lruvec, lru);
1514 if (is_active_lru(lru)) {
1515 int file = is_file_lru(lru);
1516 int numpages = hpage_nr_pages(page);
1517 reclaim_stat->recent_rotated[file] += numpages;
1519 if (put_page_testzero(page)) {
1520 __ClearPageLRU(page);
1521 __ClearPageActive(page);
1522 del_page_from_lru_list(page, lruvec, lru);
1524 if (unlikely(PageCompound(page))) {
1525 spin_unlock_irq(&zone->lru_lock);
1526 mem_cgroup_uncharge(page);
1527 (*get_compound_page_dtor(page))(page);
1528 spin_lock_irq(&zone->lru_lock);
1530 list_add(&page->lru, &pages_to_free);
1535 * To save our caller's stack, now use input list for pages to free.
1537 list_splice(&pages_to_free, page_list);
1541 * If a kernel thread (such as nfsd for loop-back mounts) services
1542 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1543 * In that case we should only throttle if the backing device it is
1544 * writing to is congested. In other cases it is safe to throttle.
1546 static int current_may_throttle(void)
1548 return !(current->flags & PF_LESS_THROTTLE) ||
1549 current->backing_dev_info == NULL ||
1550 bdi_write_congested(current->backing_dev_info);
1554 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1555 * of reclaimed pages
1557 static noinline_for_stack unsigned long
1558 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1559 struct scan_control *sc, enum lru_list lru)
1561 LIST_HEAD(page_list);
1562 unsigned long nr_scanned;
1563 unsigned long nr_reclaimed = 0;
1564 unsigned long nr_taken;
1565 unsigned long nr_dirty = 0;
1566 unsigned long nr_congested = 0;
1567 unsigned long nr_unqueued_dirty = 0;
1568 unsigned long nr_writeback = 0;
1569 unsigned long nr_immediate = 0;
1570 isolate_mode_t isolate_mode = 0;
1571 int file = is_file_lru(lru);
1572 struct zone *zone = lruvec_zone(lruvec);
1573 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1575 while (unlikely(too_many_isolated(zone, file, sc))) {
1576 congestion_wait(BLK_RW_ASYNC, HZ/10);
1578 /* We are about to die and free our memory. Return now. */
1579 if (fatal_signal_pending(current))
1580 return SWAP_CLUSTER_MAX;
1586 isolate_mode |= ISOLATE_UNMAPPED;
1587 if (!sc->may_writepage)
1588 isolate_mode |= ISOLATE_CLEAN;
1590 spin_lock_irq(&zone->lru_lock);
1592 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1593 &nr_scanned, sc, isolate_mode, lru);
1595 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1596 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1598 if (global_reclaim(sc)) {
1599 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1600 if (current_is_kswapd())
1601 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1603 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1605 spin_unlock_irq(&zone->lru_lock);
1610 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1611 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1612 &nr_writeback, &nr_immediate,
1615 spin_lock_irq(&zone->lru_lock);
1617 reclaim_stat->recent_scanned[file] += nr_taken;
1619 if (global_reclaim(sc)) {
1620 if (current_is_kswapd())
1621 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1624 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1628 putback_inactive_pages(lruvec, &page_list);
1630 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1632 spin_unlock_irq(&zone->lru_lock);
1634 mem_cgroup_uncharge_list(&page_list);
1635 free_hot_cold_page_list(&page_list, true);
1638 * If reclaim is isolating dirty pages under writeback, it implies
1639 * that the long-lived page allocation rate is exceeding the page
1640 * laundering rate. Either the global limits are not being effective
1641 * at throttling processes due to the page distribution throughout
1642 * zones or there is heavy usage of a slow backing device. The
1643 * only option is to throttle from reclaim context which is not ideal
1644 * as there is no guarantee the dirtying process is throttled in the
1645 * same way balance_dirty_pages() manages.
1647 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1648 * of pages under pages flagged for immediate reclaim and stall if any
1649 * are encountered in the nr_immediate check below.
1651 if (nr_writeback && nr_writeback == nr_taken)
1652 set_bit(ZONE_WRITEBACK, &zone->flags);
1655 * Legacy memcg will stall in page writeback so avoid forcibly
1658 if (sane_reclaim(sc)) {
1660 * Tag a zone as congested if all the dirty pages scanned were
1661 * backed by a congested BDI and wait_iff_congested will stall.
1663 if (nr_dirty && nr_dirty == nr_congested)
1664 set_bit(ZONE_CONGESTED, &zone->flags);
1667 * If dirty pages are scanned that are not queued for IO, it
1668 * implies that flushers are not keeping up. In this case, flag
1669 * the zone ZONE_DIRTY and kswapd will start writing pages from
1672 if (nr_unqueued_dirty == nr_taken)
1673 set_bit(ZONE_DIRTY, &zone->flags);
1676 * If kswapd scans pages marked marked for immediate
1677 * reclaim and under writeback (nr_immediate), it implies
1678 * that pages are cycling through the LRU faster than
1679 * they are written so also forcibly stall.
1681 if (nr_immediate && current_may_throttle())
1682 congestion_wait(BLK_RW_ASYNC, HZ/10);
1686 * Stall direct reclaim for IO completions if underlying BDIs or zone
1687 * is congested. Allow kswapd to continue until it starts encountering
1688 * unqueued dirty pages or cycling through the LRU too quickly.
1690 if (!sc->hibernation_mode && !current_is_kswapd() &&
1691 current_may_throttle())
1692 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1694 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1696 nr_scanned, nr_reclaimed,
1698 trace_shrink_flags(file));
1699 return nr_reclaimed;
1703 * This moves pages from the active list to the inactive list.
1705 * We move them the other way if the page is referenced by one or more
1706 * processes, from rmap.
1708 * If the pages are mostly unmapped, the processing is fast and it is
1709 * appropriate to hold zone->lru_lock across the whole operation. But if
1710 * the pages are mapped, the processing is slow (page_referenced()) so we
1711 * should drop zone->lru_lock around each page. It's impossible to balance
1712 * this, so instead we remove the pages from the LRU while processing them.
1713 * It is safe to rely on PG_active against the non-LRU pages in here because
1714 * nobody will play with that bit on a non-LRU page.
1716 * The downside is that we have to touch page->_count against each page.
1717 * But we had to alter page->flags anyway.
1720 static void move_active_pages_to_lru(struct lruvec *lruvec,
1721 struct list_head *list,
1722 struct list_head *pages_to_free,
1725 struct zone *zone = lruvec_zone(lruvec);
1726 unsigned long pgmoved = 0;
1730 while (!list_empty(list)) {
1731 page = lru_to_page(list);
1732 lruvec = mem_cgroup_page_lruvec(page, zone);
1734 VM_BUG_ON_PAGE(PageLRU(page), page);
1737 nr_pages = hpage_nr_pages(page);
1738 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1739 list_move(&page->lru, &lruvec->lists[lru]);
1740 pgmoved += nr_pages;
1742 if (put_page_testzero(page)) {
1743 __ClearPageLRU(page);
1744 __ClearPageActive(page);
1745 del_page_from_lru_list(page, lruvec, lru);
1747 if (unlikely(PageCompound(page))) {
1748 spin_unlock_irq(&zone->lru_lock);
1749 mem_cgroup_uncharge(page);
1750 (*get_compound_page_dtor(page))(page);
1751 spin_lock_irq(&zone->lru_lock);
1753 list_add(&page->lru, pages_to_free);
1756 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1757 if (!is_active_lru(lru))
1758 __count_vm_events(PGDEACTIVATE, pgmoved);
1761 static void shrink_active_list(unsigned long nr_to_scan,
1762 struct lruvec *lruvec,
1763 struct scan_control *sc,
1766 unsigned long nr_taken;
1767 unsigned long nr_scanned;
1768 unsigned long vm_flags;
1769 LIST_HEAD(l_hold); /* The pages which were snipped off */
1770 LIST_HEAD(l_active);
1771 LIST_HEAD(l_inactive);
1773 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1774 unsigned long nr_rotated = 0;
1775 isolate_mode_t isolate_mode = 0;
1776 int file = is_file_lru(lru);
1777 struct zone *zone = lruvec_zone(lruvec);
1782 isolate_mode |= ISOLATE_UNMAPPED;
1783 if (!sc->may_writepage)
1784 isolate_mode |= ISOLATE_CLEAN;
1786 spin_lock_irq(&zone->lru_lock);
1788 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1789 &nr_scanned, sc, isolate_mode, lru);
1790 if (global_reclaim(sc))
1791 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1793 reclaim_stat->recent_scanned[file] += nr_taken;
1795 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1796 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1797 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1798 spin_unlock_irq(&zone->lru_lock);
1800 while (!list_empty(&l_hold)) {
1802 page = lru_to_page(&l_hold);
1803 list_del(&page->lru);
1805 if (unlikely(!page_evictable(page))) {
1806 putback_lru_page(page);
1810 if (unlikely(buffer_heads_over_limit)) {
1811 if (page_has_private(page) && trylock_page(page)) {
1812 if (page_has_private(page))
1813 try_to_release_page(page, 0);
1818 if (page_referenced(page, 0, sc->target_mem_cgroup,
1820 nr_rotated += hpage_nr_pages(page);
1822 * Identify referenced, file-backed active pages and
1823 * give them one more trip around the active list. So
1824 * that executable code get better chances to stay in
1825 * memory under moderate memory pressure. Anon pages
1826 * are not likely to be evicted by use-once streaming
1827 * IO, plus JVM can create lots of anon VM_EXEC pages,
1828 * so we ignore them here.
1830 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1831 list_add(&page->lru, &l_active);
1836 ClearPageActive(page); /* we are de-activating */
1837 list_add(&page->lru, &l_inactive);
1841 * Move pages back to the lru list.
1843 spin_lock_irq(&zone->lru_lock);
1845 * Count referenced pages from currently used mappings as rotated,
1846 * even though only some of them are actually re-activated. This
1847 * helps balance scan pressure between file and anonymous pages in
1850 reclaim_stat->recent_rotated[file] += nr_rotated;
1852 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1853 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1854 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1855 spin_unlock_irq(&zone->lru_lock);
1857 mem_cgroup_uncharge_list(&l_hold);
1858 free_hot_cold_page_list(&l_hold, true);
1862 static bool inactive_anon_is_low_global(struct zone *zone)
1864 unsigned long active, inactive;
1866 active = zone_page_state(zone, NR_ACTIVE_ANON);
1867 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1869 return inactive * zone->inactive_ratio < active;
1873 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1874 * @lruvec: LRU vector to check
1876 * Returns true if the zone does not have enough inactive anon pages,
1877 * meaning some active anon pages need to be deactivated.
1879 static bool inactive_anon_is_low(struct lruvec *lruvec)
1882 * If we don't have swap space, anonymous page deactivation
1885 if (!total_swap_pages)
1888 if (!mem_cgroup_disabled())
1889 return mem_cgroup_inactive_anon_is_low(lruvec);
1891 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1894 static inline bool inactive_anon_is_low(struct lruvec *lruvec)
1901 * inactive_file_is_low - check if file pages need to be deactivated
1902 * @lruvec: LRU vector to check
1904 * When the system is doing streaming IO, memory pressure here
1905 * ensures that active file pages get deactivated, until more
1906 * than half of the file pages are on the inactive list.
1908 * Once we get to that situation, protect the system's working
1909 * set from being evicted by disabling active file page aging.
1911 * This uses a different ratio than the anonymous pages, because
1912 * the page cache uses a use-once replacement algorithm.
1914 static bool inactive_file_is_low(struct lruvec *lruvec)
1916 unsigned long inactive;
1917 unsigned long active;
1919 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1920 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1922 return active > inactive;
1925 static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1927 if (is_file_lru(lru))
1928 return inactive_file_is_low(lruvec);
1930 return inactive_anon_is_low(lruvec);
1933 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1934 struct lruvec *lruvec, struct scan_control *sc)
1936 if (is_active_lru(lru)) {
1937 if (inactive_list_is_low(lruvec, lru))
1938 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1942 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1953 * Determine how aggressively the anon and file LRU lists should be
1954 * scanned. The relative value of each set of LRU lists is determined
1955 * by looking at the fraction of the pages scanned we did rotate back
1956 * onto the active list instead of evict.
1958 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1959 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1961 static void get_scan_count(struct lruvec *lruvec, int swappiness,
1962 struct scan_control *sc, unsigned long *nr,
1963 unsigned long *lru_pages)
1965 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1967 u64 denominator = 0; /* gcc */
1968 struct zone *zone = lruvec_zone(lruvec);
1969 unsigned long anon_prio, file_prio;
1970 enum scan_balance scan_balance;
1971 unsigned long anon, file;
1972 bool force_scan = false;
1973 unsigned long ap, fp;
1979 * If the zone or memcg is small, nr[l] can be 0. This
1980 * results in no scanning on this priority and a potential
1981 * priority drop. Global direct reclaim can go to the next
1982 * zone and tends to have no problems. Global kswapd is for
1983 * zone balancing and it needs to scan a minimum amount. When
1984 * reclaiming for a memcg, a priority drop can cause high
1985 * latencies, so it's better to scan a minimum amount there as
1988 if (current_is_kswapd()) {
1989 if (!zone_reclaimable(zone))
1991 if (!mem_cgroup_lruvec_online(lruvec))
1994 if (!global_reclaim(sc))
1997 /* If we have no swap space, do not bother scanning anon pages. */
1998 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1999 scan_balance = SCAN_FILE;
2004 * Global reclaim will swap to prevent OOM even with no
2005 * swappiness, but memcg users want to use this knob to
2006 * disable swapping for individual groups completely when
2007 * using the memory controller's swap limit feature would be
2010 if (!global_reclaim(sc) && !swappiness) {
2011 scan_balance = SCAN_FILE;
2016 * Do not apply any pressure balancing cleverness when the
2017 * system is close to OOM, scan both anon and file equally
2018 * (unless the swappiness setting disagrees with swapping).
2020 if (!sc->priority && swappiness) {
2021 scan_balance = SCAN_EQUAL;
2026 * Prevent the reclaimer from falling into the cache trap: as
2027 * cache pages start out inactive, every cache fault will tip
2028 * the scan balance towards the file LRU. And as the file LRU
2029 * shrinks, so does the window for rotation from references.
2030 * This means we have a runaway feedback loop where a tiny
2031 * thrashing file LRU becomes infinitely more attractive than
2032 * anon pages. Try to detect this based on file LRU size.
2034 if (global_reclaim(sc)) {
2035 unsigned long zonefile;
2036 unsigned long zonefree;
2038 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2039 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2040 zone_page_state(zone, NR_INACTIVE_FILE);
2042 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2043 scan_balance = SCAN_ANON;
2049 * There is enough inactive page cache, do not reclaim
2050 * anything from the anonymous working set right now.
2052 if (!inactive_file_is_low(lruvec)) {
2053 scan_balance = SCAN_FILE;
2057 scan_balance = SCAN_FRACT;
2060 * With swappiness at 100, anonymous and file have the same priority.
2061 * This scanning priority is essentially the inverse of IO cost.
2063 anon_prio = swappiness;
2064 file_prio = 200 - anon_prio;
2067 * OK, so we have swap space and a fair amount of page cache
2068 * pages. We use the recently rotated / recently scanned
2069 * ratios to determine how valuable each cache is.
2071 * Because workloads change over time (and to avoid overflow)
2072 * we keep these statistics as a floating average, which ends
2073 * up weighing recent references more than old ones.
2075 * anon in [0], file in [1]
2078 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2079 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2080 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2081 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2083 spin_lock_irq(&zone->lru_lock);
2084 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2085 reclaim_stat->recent_scanned[0] /= 2;
2086 reclaim_stat->recent_rotated[0] /= 2;
2089 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2090 reclaim_stat->recent_scanned[1] /= 2;
2091 reclaim_stat->recent_rotated[1] /= 2;
2095 * The amount of pressure on anon vs file pages is inversely
2096 * proportional to the fraction of recently scanned pages on
2097 * each list that were recently referenced and in active use.
2099 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2100 ap /= reclaim_stat->recent_rotated[0] + 1;
2102 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2103 fp /= reclaim_stat->recent_rotated[1] + 1;
2104 spin_unlock_irq(&zone->lru_lock);
2108 denominator = ap + fp + 1;
2110 some_scanned = false;
2111 /* Only use force_scan on second pass. */
2112 for (pass = 0; !some_scanned && pass < 2; pass++) {
2114 for_each_evictable_lru(lru) {
2115 int file = is_file_lru(lru);
2119 size = get_lru_size(lruvec, lru);
2120 scan = size >> sc->priority;
2122 if (!scan && pass && force_scan)
2123 scan = min(size, SWAP_CLUSTER_MAX);
2125 switch (scan_balance) {
2127 /* Scan lists relative to size */
2131 * Scan types proportional to swappiness and
2132 * their relative recent reclaim efficiency.
2134 scan = div64_u64(scan * fraction[file],
2139 /* Scan one type exclusively */
2140 if ((scan_balance == SCAN_FILE) != file) {
2146 /* Look ma, no brain */
2154 * Skip the second pass and don't force_scan,
2155 * if we found something to scan.
2157 some_scanned |= !!scan;
2163 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2165 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2166 struct scan_control *sc, unsigned long *lru_pages)
2168 unsigned long nr[NR_LRU_LISTS];
2169 unsigned long targets[NR_LRU_LISTS];
2170 unsigned long nr_to_scan;
2172 unsigned long nr_reclaimed = 0;
2173 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2174 struct blk_plug plug;
2177 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2179 /* Record the original scan target for proportional adjustments later */
2180 memcpy(targets, nr, sizeof(nr));
2183 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2184 * event that can occur when there is little memory pressure e.g.
2185 * multiple streaming readers/writers. Hence, we do not abort scanning
2186 * when the requested number of pages are reclaimed when scanning at
2187 * DEF_PRIORITY on the assumption that the fact we are direct
2188 * reclaiming implies that kswapd is not keeping up and it is best to
2189 * do a batch of work at once. For memcg reclaim one check is made to
2190 * abort proportional reclaim if either the file or anon lru has already
2191 * dropped to zero at the first pass.
2193 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2194 sc->priority == DEF_PRIORITY);
2196 blk_start_plug(&plug);
2197 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2198 nr[LRU_INACTIVE_FILE]) {
2199 unsigned long nr_anon, nr_file, percentage;
2200 unsigned long nr_scanned;
2202 for_each_evictable_lru(lru) {
2204 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2205 nr[lru] -= nr_to_scan;
2207 nr_reclaimed += shrink_list(lru, nr_to_scan,
2212 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2216 * For kswapd and memcg, reclaim at least the number of pages
2217 * requested. Ensure that the anon and file LRUs are scanned
2218 * proportionally what was requested by get_scan_count(). We
2219 * stop reclaiming one LRU and reduce the amount scanning
2220 * proportional to the original scan target.
2222 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2223 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2226 * It's just vindictive to attack the larger once the smaller
2227 * has gone to zero. And given the way we stop scanning the
2228 * smaller below, this makes sure that we only make one nudge
2229 * towards proportionality once we've got nr_to_reclaim.
2231 if (!nr_file || !nr_anon)
2234 if (nr_file > nr_anon) {
2235 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2236 targets[LRU_ACTIVE_ANON] + 1;
2238 percentage = nr_anon * 100 / scan_target;
2240 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2241 targets[LRU_ACTIVE_FILE] + 1;
2243 percentage = nr_file * 100 / scan_target;
2246 /* Stop scanning the smaller of the LRU */
2248 nr[lru + LRU_ACTIVE] = 0;
2251 * Recalculate the other LRU scan count based on its original
2252 * scan target and the percentage scanning already complete
2254 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2255 nr_scanned = targets[lru] - nr[lru];
2256 nr[lru] = targets[lru] * (100 - percentage) / 100;
2257 nr[lru] -= min(nr[lru], nr_scanned);
2260 nr_scanned = targets[lru] - nr[lru];
2261 nr[lru] = targets[lru] * (100 - percentage) / 100;
2262 nr[lru] -= min(nr[lru], nr_scanned);
2264 scan_adjusted = true;
2266 blk_finish_plug(&plug);
2267 sc->nr_reclaimed += nr_reclaimed;
2270 * Even if we did not try to evict anon pages at all, we want to
2271 * rebalance the anon lru active/inactive ratio.
2273 if (inactive_anon_is_low(lruvec))
2274 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2275 sc, LRU_ACTIVE_ANON);
2277 throttle_vm_writeout(sc->gfp_mask);
2280 /* Use reclaim/compaction for costly allocs or under memory pressure */
2281 static bool in_reclaim_compaction(struct scan_control *sc)
2283 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2284 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2285 sc->priority < DEF_PRIORITY - 2))
2292 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2293 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2294 * true if more pages should be reclaimed such that when the page allocator
2295 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2296 * It will give up earlier than that if there is difficulty reclaiming pages.
2298 static inline bool should_continue_reclaim(struct zone *zone,
2299 unsigned long nr_reclaimed,
2300 unsigned long nr_scanned,
2301 struct scan_control *sc)
2303 unsigned long pages_for_compaction;
2304 unsigned long inactive_lru_pages;
2306 /* If not in reclaim/compaction mode, stop */
2307 if (!in_reclaim_compaction(sc))
2310 /* Consider stopping depending on scan and reclaim activity */
2311 if (sc->gfp_mask & __GFP_REPEAT) {
2313 * For __GFP_REPEAT allocations, stop reclaiming if the
2314 * full LRU list has been scanned and we are still failing
2315 * to reclaim pages. This full LRU scan is potentially
2316 * expensive but a __GFP_REPEAT caller really wants to succeed
2318 if (!nr_reclaimed && !nr_scanned)
2322 * For non-__GFP_REPEAT allocations which can presumably
2323 * fail without consequence, stop if we failed to reclaim
2324 * any pages from the last SWAP_CLUSTER_MAX number of
2325 * pages that were scanned. This will return to the
2326 * caller faster at the risk reclaim/compaction and
2327 * the resulting allocation attempt fails
2334 * If we have not reclaimed enough pages for compaction and the
2335 * inactive lists are large enough, continue reclaiming
2337 pages_for_compaction = (2UL << sc->order);
2338 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2339 if (get_nr_swap_pages() > 0)
2340 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2341 if (sc->nr_reclaimed < pages_for_compaction &&
2342 inactive_lru_pages > pages_for_compaction)
2345 /* If compaction would go ahead or the allocation would succeed, stop */
2346 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2347 case COMPACT_PARTIAL:
2348 case COMPACT_CONTINUE:
2355 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2358 struct reclaim_state *reclaim_state = current->reclaim_state;
2359 unsigned long nr_reclaimed, nr_scanned;
2360 bool reclaimable = false;
2363 struct mem_cgroup *root = sc->target_mem_cgroup;
2364 struct mem_cgroup_reclaim_cookie reclaim = {
2366 .priority = sc->priority,
2368 unsigned long zone_lru_pages = 0;
2369 struct mem_cgroup *memcg;
2371 nr_reclaimed = sc->nr_reclaimed;
2372 nr_scanned = sc->nr_scanned;
2374 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2376 unsigned long lru_pages;
2377 unsigned long scanned;
2378 struct lruvec *lruvec;
2381 if (mem_cgroup_low(root, memcg)) {
2382 if (!sc->may_thrash)
2384 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2387 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2388 swappiness = mem_cgroup_swappiness(memcg);
2389 scanned = sc->nr_scanned;
2391 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2392 zone_lru_pages += lru_pages;
2394 if (memcg && is_classzone)
2395 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2396 memcg, sc->nr_scanned - scanned,
2400 * Direct reclaim and kswapd have to scan all memory
2401 * cgroups to fulfill the overall scan target for the
2404 * Limit reclaim, on the other hand, only cares about
2405 * nr_to_reclaim pages to be reclaimed and it will
2406 * retry with decreasing priority if one round over the
2407 * whole hierarchy is not sufficient.
2409 if (!global_reclaim(sc) &&
2410 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2411 mem_cgroup_iter_break(root, memcg);
2414 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2417 * Shrink the slab caches in the same proportion that
2418 * the eligible LRU pages were scanned.
2420 if (global_reclaim(sc) && is_classzone)
2421 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2422 sc->nr_scanned - nr_scanned,
2425 if (reclaim_state) {
2426 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2427 reclaim_state->reclaimed_slab = 0;
2430 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2431 sc->nr_scanned - nr_scanned,
2432 sc->nr_reclaimed - nr_reclaimed);
2434 if (sc->nr_reclaimed - nr_reclaimed)
2437 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2438 sc->nr_scanned - nr_scanned, sc));
2444 * Returns true if compaction should go ahead for a high-order request, or
2445 * the high-order allocation would succeed without compaction.
2447 static inline bool compaction_ready(struct zone *zone, int order)
2449 unsigned long balance_gap, watermark;
2453 * Compaction takes time to run and there are potentially other
2454 * callers using the pages just freed. Continue reclaiming until
2455 * there is a buffer of free pages available to give compaction
2456 * a reasonable chance of completing and allocating the page
2458 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2459 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2460 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2461 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2464 * If compaction is deferred, reclaim up to a point where
2465 * compaction will have a chance of success when re-enabled
2467 if (compaction_deferred(zone, order))
2468 return watermark_ok;
2471 * If compaction is not ready to start and allocation is not likely
2472 * to succeed without it, then keep reclaiming.
2474 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2477 return watermark_ok;
2481 * This is the direct reclaim path, for page-allocating processes. We only
2482 * try to reclaim pages from zones which will satisfy the caller's allocation
2485 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2487 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2489 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2490 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2491 * zone defense algorithm.
2493 * If a zone is deemed to be full of pinned pages then just give it a light
2494 * scan then give up on it.
2496 * Returns true if a zone was reclaimable.
2498 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2502 unsigned long nr_soft_reclaimed;
2503 unsigned long nr_soft_scanned;
2505 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2506 bool reclaimable = false;
2509 * If the number of buffer_heads in the machine exceeds the maximum
2510 * allowed level, force direct reclaim to scan the highmem zone as
2511 * highmem pages could be pinning lowmem pages storing buffer_heads
2513 orig_mask = sc->gfp_mask;
2514 if (buffer_heads_over_limit)
2515 sc->gfp_mask |= __GFP_HIGHMEM;
2517 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2518 gfp_zone(sc->gfp_mask), sc->nodemask) {
2519 enum zone_type classzone_idx;
2521 if (!populated_zone(zone))
2524 classzone_idx = requested_highidx;
2525 while (!populated_zone(zone->zone_pgdat->node_zones +
2530 * Take care memory controller reclaiming has small influence
2533 if (global_reclaim(sc)) {
2534 if (!cpuset_zone_allowed(zone,
2535 GFP_KERNEL | __GFP_HARDWALL))
2538 if (sc->priority != DEF_PRIORITY &&
2539 !zone_reclaimable(zone))
2540 continue; /* Let kswapd poll it */
2543 * If we already have plenty of memory free for
2544 * compaction in this zone, don't free any more.
2545 * Even though compaction is invoked for any
2546 * non-zero order, only frequent costly order
2547 * reclamation is disruptive enough to become a
2548 * noticeable problem, like transparent huge
2551 if (IS_ENABLED(CONFIG_COMPACTION) &&
2552 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2553 zonelist_zone_idx(z) <= requested_highidx &&
2554 compaction_ready(zone, sc->order)) {
2555 sc->compaction_ready = true;
2560 * This steals pages from memory cgroups over softlimit
2561 * and returns the number of reclaimed pages and
2562 * scanned pages. This works for global memory pressure
2563 * and balancing, not for a memcg's limit.
2565 nr_soft_scanned = 0;
2566 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2567 sc->order, sc->gfp_mask,
2569 sc->nr_reclaimed += nr_soft_reclaimed;
2570 sc->nr_scanned += nr_soft_scanned;
2571 if (nr_soft_reclaimed)
2573 /* need some check for avoid more shrink_zone() */
2576 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2579 if (global_reclaim(sc) &&
2580 !reclaimable && zone_reclaimable(zone))
2585 * Restore to original mask to avoid the impact on the caller if we
2586 * promoted it to __GFP_HIGHMEM.
2588 sc->gfp_mask = orig_mask;
2594 * This is the main entry point to direct page reclaim.
2596 * If a full scan of the inactive list fails to free enough memory then we
2597 * are "out of memory" and something needs to be killed.
2599 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2600 * high - the zone may be full of dirty or under-writeback pages, which this
2601 * caller can't do much about. We kick the writeback threads and take explicit
2602 * naps in the hope that some of these pages can be written. But if the
2603 * allocating task holds filesystem locks which prevent writeout this might not
2604 * work, and the allocation attempt will fail.
2606 * returns: 0, if no pages reclaimed
2607 * else, the number of pages reclaimed
2609 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2610 struct scan_control *sc)
2612 int initial_priority = sc->priority;
2613 unsigned long total_scanned = 0;
2614 unsigned long writeback_threshold;
2615 bool zones_reclaimable;
2617 delayacct_freepages_start();
2619 if (global_reclaim(sc))
2620 count_vm_event(ALLOCSTALL);
2623 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2626 zones_reclaimable = shrink_zones(zonelist, sc);
2628 total_scanned += sc->nr_scanned;
2629 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2632 if (sc->compaction_ready)
2636 * If we're getting trouble reclaiming, start doing
2637 * writepage even in laptop mode.
2639 if (sc->priority < DEF_PRIORITY - 2)
2640 sc->may_writepage = 1;
2643 * Try to write back as many pages as we just scanned. This
2644 * tends to cause slow streaming writers to write data to the
2645 * disk smoothly, at the dirtying rate, which is nice. But
2646 * that's undesirable in laptop mode, where we *want* lumpy
2647 * writeout. So in laptop mode, write out the whole world.
2649 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2650 if (total_scanned > writeback_threshold) {
2651 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2652 WB_REASON_TRY_TO_FREE_PAGES);
2653 sc->may_writepage = 1;
2655 } while (--sc->priority >= 0);
2657 delayacct_freepages_end();
2659 if (sc->nr_reclaimed)
2660 return sc->nr_reclaimed;
2662 /* Aborted reclaim to try compaction? don't OOM, then */
2663 if (sc->compaction_ready)
2666 /* Untapped cgroup reserves? Don't OOM, retry. */
2667 if (!sc->may_thrash) {
2668 sc->priority = initial_priority;
2673 /* Any of the zones still reclaimable? Don't OOM. */
2674 if (zones_reclaimable)
2680 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2683 unsigned long pfmemalloc_reserve = 0;
2684 unsigned long free_pages = 0;
2688 for (i = 0; i <= ZONE_NORMAL; i++) {
2689 zone = &pgdat->node_zones[i];
2690 if (!populated_zone(zone) ||
2691 zone_reclaimable_pages(zone) == 0)
2694 pfmemalloc_reserve += min_wmark_pages(zone);
2695 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2698 /* If there are no reserves (unexpected config) then do not throttle */
2699 if (!pfmemalloc_reserve)
2702 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2704 /* kswapd must be awake if processes are being throttled */
2705 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2706 pgdat->classzone_idx = min(pgdat->classzone_idx,
2707 (enum zone_type)ZONE_NORMAL);
2708 wake_up_interruptible(&pgdat->kswapd_wait);
2715 * Throttle direct reclaimers if backing storage is backed by the network
2716 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2717 * depleted. kswapd will continue to make progress and wake the processes
2718 * when the low watermark is reached.
2720 * Returns true if a fatal signal was delivered during throttling. If this
2721 * happens, the page allocator should not consider triggering the OOM killer.
2723 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2724 nodemask_t *nodemask)
2728 pg_data_t *pgdat = NULL;
2731 * Kernel threads should not be throttled as they may be indirectly
2732 * responsible for cleaning pages necessary for reclaim to make forward
2733 * progress. kjournald for example may enter direct reclaim while
2734 * committing a transaction where throttling it could forcing other
2735 * processes to block on log_wait_commit().
2737 if (current->flags & PF_KTHREAD)
2741 * If a fatal signal is pending, this process should not throttle.
2742 * It should return quickly so it can exit and free its memory
2744 if (fatal_signal_pending(current))
2748 * Check if the pfmemalloc reserves are ok by finding the first node
2749 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2750 * GFP_KERNEL will be required for allocating network buffers when
2751 * swapping over the network so ZONE_HIGHMEM is unusable.
2753 * Throttling is based on the first usable node and throttled processes
2754 * wait on a queue until kswapd makes progress and wakes them. There
2755 * is an affinity then between processes waking up and where reclaim
2756 * progress has been made assuming the process wakes on the same node.
2757 * More importantly, processes running on remote nodes will not compete
2758 * for remote pfmemalloc reserves and processes on different nodes
2759 * should make reasonable progress.
2761 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2762 gfp_zone(gfp_mask), nodemask) {
2763 if (zone_idx(zone) > ZONE_NORMAL)
2766 /* Throttle based on the first usable node */
2767 pgdat = zone->zone_pgdat;
2768 if (pfmemalloc_watermark_ok(pgdat))
2773 /* If no zone was usable by the allocation flags then do not throttle */
2777 /* Account for the throttling */
2778 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2781 * If the caller cannot enter the filesystem, it's possible that it
2782 * is due to the caller holding an FS lock or performing a journal
2783 * transaction in the case of a filesystem like ext[3|4]. In this case,
2784 * it is not safe to block on pfmemalloc_wait as kswapd could be
2785 * blocked waiting on the same lock. Instead, throttle for up to a
2786 * second before continuing.
2788 if (!(gfp_mask & __GFP_FS)) {
2789 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2790 pfmemalloc_watermark_ok(pgdat), HZ);
2795 /* Throttle until kswapd wakes the process */
2796 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2797 pfmemalloc_watermark_ok(pgdat));
2800 if (fatal_signal_pending(current))
2807 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2808 gfp_t gfp_mask, nodemask_t *nodemask)
2810 unsigned long nr_reclaimed;
2811 struct scan_control sc = {
2812 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2813 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2815 .nodemask = nodemask,
2816 .priority = DEF_PRIORITY,
2817 .may_writepage = !laptop_mode,
2823 * Do not enter reclaim if fatal signal was delivered while throttled.
2824 * 1 is returned so that the page allocator does not OOM kill at this
2827 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2830 trace_mm_vmscan_direct_reclaim_begin(order,
2834 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2836 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2838 return nr_reclaimed;
2843 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2844 gfp_t gfp_mask, bool noswap,
2846 unsigned long *nr_scanned)
2848 struct scan_control sc = {
2849 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2850 .target_mem_cgroup = memcg,
2851 .may_writepage = !laptop_mode,
2853 .may_swap = !noswap,
2855 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2856 int swappiness = mem_cgroup_swappiness(memcg);
2857 unsigned long lru_pages;
2859 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2860 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2862 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2867 * NOTE: Although we can get the priority field, using it
2868 * here is not a good idea, since it limits the pages we can scan.
2869 * if we don't reclaim here, the shrink_zone from balance_pgdat
2870 * will pick up pages from other mem cgroup's as well. We hack
2871 * the priority and make it zero.
2873 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
2875 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2877 *nr_scanned = sc.nr_scanned;
2878 return sc.nr_reclaimed;
2881 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2882 unsigned long nr_pages,
2886 struct zonelist *zonelist;
2887 unsigned long nr_reclaimed;
2889 struct scan_control sc = {
2890 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2891 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2892 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2893 .target_mem_cgroup = memcg,
2894 .priority = DEF_PRIORITY,
2895 .may_writepage = !laptop_mode,
2897 .may_swap = may_swap,
2901 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2902 * take care of from where we get pages. So the node where we start the
2903 * scan does not need to be the current node.
2905 nid = mem_cgroup_select_victim_node(memcg);
2907 zonelist = NODE_DATA(nid)->node_zonelists;
2909 trace_mm_vmscan_memcg_reclaim_begin(0,
2913 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2915 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2917 return nr_reclaimed;
2921 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2923 struct mem_cgroup *memcg;
2925 if (!total_swap_pages)
2928 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2930 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2932 if (inactive_anon_is_low(lruvec))
2933 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2934 sc, LRU_ACTIVE_ANON);
2936 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2940 static bool zone_balanced(struct zone *zone, int order,
2941 unsigned long balance_gap, int classzone_idx)
2943 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2944 balance_gap, classzone_idx))
2947 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2948 order, 0, classzone_idx) == COMPACT_SKIPPED)
2955 * pgdat_balanced() is used when checking if a node is balanced.
2957 * For order-0, all zones must be balanced!
2959 * For high-order allocations only zones that meet watermarks and are in a
2960 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2961 * total of balanced pages must be at least 25% of the zones allowed by
2962 * classzone_idx for the node to be considered balanced. Forcing all zones to
2963 * be balanced for high orders can cause excessive reclaim when there are
2965 * The choice of 25% is due to
2966 * o a 16M DMA zone that is balanced will not balance a zone on any
2967 * reasonable sized machine
2968 * o On all other machines, the top zone must be at least a reasonable
2969 * percentage of the middle zones. For example, on 32-bit x86, highmem
2970 * would need to be at least 256M for it to be balance a whole node.
2971 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2972 * to balance a node on its own. These seemed like reasonable ratios.
2974 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2976 unsigned long managed_pages = 0;
2977 unsigned long balanced_pages = 0;
2980 /* Check the watermark levels */
2981 for (i = 0; i <= classzone_idx; i++) {
2982 struct zone *zone = pgdat->node_zones + i;
2984 if (!populated_zone(zone))
2987 managed_pages += zone->managed_pages;
2990 * A special case here:
2992 * balance_pgdat() skips over all_unreclaimable after
2993 * DEF_PRIORITY. Effectively, it considers them balanced so
2994 * they must be considered balanced here as well!
2996 if (!zone_reclaimable(zone)) {
2997 balanced_pages += zone->managed_pages;
3001 if (zone_balanced(zone, order, 0, i))
3002 balanced_pages += zone->managed_pages;
3008 return balanced_pages >= (managed_pages >> 2);
3014 * Prepare kswapd for sleeping. This verifies that there are no processes
3015 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3017 * Returns true if kswapd is ready to sleep
3019 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3022 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3027 * The throttled processes are normally woken up in balance_pgdat() as
3028 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3029 * race between when kswapd checks the watermarks and a process gets
3030 * throttled. There is also a potential race if processes get
3031 * throttled, kswapd wakes, a large process exits thereby balancing the
3032 * zones, which causes kswapd to exit balance_pgdat() before reaching
3033 * the wake up checks. If kswapd is going to sleep, no process should
3034 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3035 * the wake up is premature, processes will wake kswapd and get
3036 * throttled again. The difference from wake ups in balance_pgdat() is
3037 * that here we are under prepare_to_wait().
3039 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3040 wake_up_all(&pgdat->pfmemalloc_wait);
3042 return pgdat_balanced(pgdat, order, classzone_idx);
3046 * kswapd shrinks the zone by the number of pages required to reach
3047 * the high watermark.
3049 * Returns true if kswapd scanned at least the requested number of pages to
3050 * reclaim or if the lack of progress was due to pages under writeback.
3051 * This is used to determine if the scanning priority needs to be raised.
3053 static bool kswapd_shrink_zone(struct zone *zone,
3055 struct scan_control *sc,
3056 unsigned long *nr_attempted)
3058 int testorder = sc->order;
3059 unsigned long balance_gap;
3060 bool lowmem_pressure;
3062 /* Reclaim above the high watermark. */
3063 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3066 * Kswapd reclaims only single pages with compaction enabled. Trying
3067 * too hard to reclaim until contiguous free pages have become
3068 * available can hurt performance by evicting too much useful data
3069 * from memory. Do not reclaim more than needed for compaction.
3071 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3072 compaction_suitable(zone, sc->order, 0, classzone_idx)
3077 * We put equal pressure on every zone, unless one zone has way too
3078 * many pages free already. The "too many pages" is defined as the
3079 * high wmark plus a "gap" where the gap is either the low
3080 * watermark or 1% of the zone, whichever is smaller.
3082 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3083 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3086 * If there is no low memory pressure or the zone is balanced then no
3087 * reclaim is necessary
3089 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3090 if (!lowmem_pressure && zone_balanced(zone, testorder,
3091 balance_gap, classzone_idx))
3094 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3096 /* Account for the number of pages attempted to reclaim */
3097 *nr_attempted += sc->nr_to_reclaim;
3099 clear_bit(ZONE_WRITEBACK, &zone->flags);
3102 * If a zone reaches its high watermark, consider it to be no longer
3103 * congested. It's possible there are dirty pages backed by congested
3104 * BDIs but as pressure is relieved, speculatively avoid congestion
3107 if (zone_reclaimable(zone) &&
3108 zone_balanced(zone, testorder, 0, classzone_idx)) {
3109 clear_bit(ZONE_CONGESTED, &zone->flags);
3110 clear_bit(ZONE_DIRTY, &zone->flags);
3113 return sc->nr_scanned >= sc->nr_to_reclaim;
3117 * For kswapd, balance_pgdat() will work across all this node's zones until
3118 * they are all at high_wmark_pages(zone).
3120 * Returns the final order kswapd was reclaiming at
3122 * There is special handling here for zones which are full of pinned pages.
3123 * This can happen if the pages are all mlocked, or if they are all used by
3124 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3125 * What we do is to detect the case where all pages in the zone have been
3126 * scanned twice and there has been zero successful reclaim. Mark the zone as
3127 * dead and from now on, only perform a short scan. Basically we're polling
3128 * the zone for when the problem goes away.
3130 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3131 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3132 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3133 * lower zones regardless of the number of free pages in the lower zones. This
3134 * interoperates with the page allocator fallback scheme to ensure that aging
3135 * of pages is balanced across the zones.
3137 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3141 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3142 unsigned long nr_soft_reclaimed;
3143 unsigned long nr_soft_scanned;
3144 struct scan_control sc = {
3145 .gfp_mask = GFP_KERNEL,
3147 .priority = DEF_PRIORITY,
3148 .may_writepage = !laptop_mode,
3152 count_vm_event(PAGEOUTRUN);
3155 unsigned long nr_attempted = 0;
3156 bool raise_priority = true;
3157 bool pgdat_needs_compaction = (order > 0);
3159 sc.nr_reclaimed = 0;
3162 * Scan in the highmem->dma direction for the highest
3163 * zone which needs scanning
3165 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3166 struct zone *zone = pgdat->node_zones + i;
3168 if (!populated_zone(zone))
3171 if (sc.priority != DEF_PRIORITY &&
3172 !zone_reclaimable(zone))
3176 * Do some background aging of the anon list, to give
3177 * pages a chance to be referenced before reclaiming.
3179 age_active_anon(zone, &sc);
3182 * If the number of buffer_heads in the machine
3183 * exceeds the maximum allowed level and this node
3184 * has a highmem zone, force kswapd to reclaim from
3185 * it to relieve lowmem pressure.
3187 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3192 if (!zone_balanced(zone, order, 0, 0)) {
3197 * If balanced, clear the dirty and congested
3200 clear_bit(ZONE_CONGESTED, &zone->flags);
3201 clear_bit(ZONE_DIRTY, &zone->flags);
3208 for (i = 0; i <= end_zone; i++) {
3209 struct zone *zone = pgdat->node_zones + i;
3211 if (!populated_zone(zone))
3215 * If any zone is currently balanced then kswapd will
3216 * not call compaction as it is expected that the
3217 * necessary pages are already available.
3219 if (pgdat_needs_compaction &&
3220 zone_watermark_ok(zone, order,
3221 low_wmark_pages(zone),
3223 pgdat_needs_compaction = false;
3227 * If we're getting trouble reclaiming, start doing writepage
3228 * even in laptop mode.
3230 if (sc.priority < DEF_PRIORITY - 2)
3231 sc.may_writepage = 1;
3234 * Now scan the zone in the dma->highmem direction, stopping
3235 * at the last zone which needs scanning.
3237 * We do this because the page allocator works in the opposite
3238 * direction. This prevents the page allocator from allocating
3239 * pages behind kswapd's direction of progress, which would
3240 * cause too much scanning of the lower zones.
3242 for (i = 0; i <= end_zone; i++) {
3243 struct zone *zone = pgdat->node_zones + i;
3245 if (!populated_zone(zone))
3248 if (sc.priority != DEF_PRIORITY &&
3249 !zone_reclaimable(zone))
3254 nr_soft_scanned = 0;
3256 * Call soft limit reclaim before calling shrink_zone.
3258 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3261 sc.nr_reclaimed += nr_soft_reclaimed;
3264 * There should be no need to raise the scanning
3265 * priority if enough pages are already being scanned
3266 * that that high watermark would be met at 100%
3269 if (kswapd_shrink_zone(zone, end_zone,
3270 &sc, &nr_attempted))
3271 raise_priority = false;
3275 * If the low watermark is met there is no need for processes
3276 * to be throttled on pfmemalloc_wait as they should not be
3277 * able to safely make forward progress. Wake them
3279 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3280 pfmemalloc_watermark_ok(pgdat))
3281 wake_up_all(&pgdat->pfmemalloc_wait);
3284 * Fragmentation may mean that the system cannot be rebalanced
3285 * for high-order allocations in all zones. If twice the
3286 * allocation size has been reclaimed and the zones are still
3287 * not balanced then recheck the watermarks at order-0 to
3288 * prevent kswapd reclaiming excessively. Assume that a
3289 * process requested a high-order can direct reclaim/compact.
3291 if (order && sc.nr_reclaimed >= 2UL << order)
3292 order = sc.order = 0;
3294 /* Check if kswapd should be suspending */
3295 if (try_to_freeze() || kthread_should_stop())
3299 * Compact if necessary and kswapd is reclaiming at least the
3300 * high watermark number of pages as requsted
3302 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3303 compact_pgdat(pgdat, order);
3306 * Raise priority if scanning rate is too low or there was no
3307 * progress in reclaiming pages
3309 if (raise_priority || !sc.nr_reclaimed)
3311 } while (sc.priority >= 1 &&
3312 !pgdat_balanced(pgdat, order, *classzone_idx));
3316 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3317 * makes a decision on the order we were last reclaiming at. However,
3318 * if another caller entered the allocator slow path while kswapd
3319 * was awake, order will remain at the higher level
3321 *classzone_idx = end_zone;
3325 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3330 if (freezing(current) || kthread_should_stop())
3333 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3335 /* Try to sleep for a short interval */
3336 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3337 remaining = schedule_timeout(HZ/10);
3338 finish_wait(&pgdat->kswapd_wait, &wait);
3339 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3343 * After a short sleep, check if it was a premature sleep. If not, then
3344 * go fully to sleep until explicitly woken up.
3346 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3347 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3350 * vmstat counters are not perfectly accurate and the estimated
3351 * value for counters such as NR_FREE_PAGES can deviate from the
3352 * true value by nr_online_cpus * threshold. To avoid the zone
3353 * watermarks being breached while under pressure, we reduce the
3354 * per-cpu vmstat threshold while kswapd is awake and restore
3355 * them before going back to sleep.
3357 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3360 * Compaction records what page blocks it recently failed to
3361 * isolate pages from and skips them in the future scanning.
3362 * When kswapd is going to sleep, it is reasonable to assume
3363 * that pages and compaction may succeed so reset the cache.
3365 reset_isolation_suitable(pgdat);
3367 if (!kthread_should_stop())
3370 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3373 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3375 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3377 finish_wait(&pgdat->kswapd_wait, &wait);
3381 * The background pageout daemon, started as a kernel thread
3382 * from the init process.
3384 * This basically trickles out pages so that we have _some_
3385 * free memory available even if there is no other activity
3386 * that frees anything up. This is needed for things like routing
3387 * etc, where we otherwise might have all activity going on in
3388 * asynchronous contexts that cannot page things out.
3390 * If there are applications that are active memory-allocators
3391 * (most normal use), this basically shouldn't matter.
3393 static int kswapd(void *p)
3395 unsigned long order, new_order;
3396 unsigned balanced_order;
3397 int classzone_idx, new_classzone_idx;
3398 int balanced_classzone_idx;
3399 pg_data_t *pgdat = (pg_data_t*)p;
3400 struct task_struct *tsk = current;
3402 struct reclaim_state reclaim_state = {
3403 .reclaimed_slab = 0,
3405 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3407 lockdep_set_current_reclaim_state(GFP_KERNEL);
3409 if (!cpumask_empty(cpumask))
3410 set_cpus_allowed_ptr(tsk, cpumask);
3411 current->reclaim_state = &reclaim_state;
3414 * Tell the memory management that we're a "memory allocator",
3415 * and that if we need more memory we should get access to it
3416 * regardless (see "__alloc_pages()"). "kswapd" should
3417 * never get caught in the normal page freeing logic.
3419 * (Kswapd normally doesn't need memory anyway, but sometimes
3420 * you need a small amount of memory in order to be able to
3421 * page out something else, and this flag essentially protects
3422 * us from recursively trying to free more memory as we're
3423 * trying to free the first piece of memory in the first place).
3425 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3428 order = new_order = 0;
3430 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3431 balanced_classzone_idx = classzone_idx;
3436 * If the last balance_pgdat was unsuccessful it's unlikely a
3437 * new request of a similar or harder type will succeed soon
3438 * so consider going to sleep on the basis we reclaimed at
3440 if (balanced_classzone_idx >= new_classzone_idx &&
3441 balanced_order == new_order) {
3442 new_order = pgdat->kswapd_max_order;
3443 new_classzone_idx = pgdat->classzone_idx;
3444 pgdat->kswapd_max_order = 0;
3445 pgdat->classzone_idx = pgdat->nr_zones - 1;
3448 if (order < new_order || classzone_idx > new_classzone_idx) {
3450 * Don't sleep if someone wants a larger 'order'
3451 * allocation or has tigher zone constraints
3454 classzone_idx = new_classzone_idx;
3456 kswapd_try_to_sleep(pgdat, balanced_order,
3457 balanced_classzone_idx);
3458 order = pgdat->kswapd_max_order;
3459 classzone_idx = pgdat->classzone_idx;
3461 new_classzone_idx = classzone_idx;
3462 pgdat->kswapd_max_order = 0;
3463 pgdat->classzone_idx = pgdat->nr_zones - 1;
3466 ret = try_to_freeze();
3467 if (kthread_should_stop())
3471 * We can speed up thawing tasks if we don't call balance_pgdat
3472 * after returning from the refrigerator
3475 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3476 balanced_classzone_idx = classzone_idx;
3477 balanced_order = balance_pgdat(pgdat, order,
3478 &balanced_classzone_idx);
3482 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3483 current->reclaim_state = NULL;
3484 lockdep_clear_current_reclaim_state();
3490 * A zone is low on free memory, so wake its kswapd task to service it.
3492 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3496 if (!populated_zone(zone))
3499 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3501 pgdat = zone->zone_pgdat;
3502 if (pgdat->kswapd_max_order < order) {
3503 pgdat->kswapd_max_order = order;
3504 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3506 if (!waitqueue_active(&pgdat->kswapd_wait))
3508 if (zone_balanced(zone, order, 0, 0))
3511 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3512 wake_up_interruptible(&pgdat->kswapd_wait);
3515 #ifdef CONFIG_HIBERNATION
3517 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3520 * Rather than trying to age LRUs the aim is to preserve the overall
3521 * LRU order by reclaiming preferentially
3522 * inactive > active > active referenced > active mapped
3524 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3526 struct reclaim_state reclaim_state;
3527 struct scan_control sc = {
3528 .nr_to_reclaim = nr_to_reclaim,
3529 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3530 .priority = DEF_PRIORITY,
3534 .hibernation_mode = 1,
3536 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3537 struct task_struct *p = current;
3538 unsigned long nr_reclaimed;
3540 p->flags |= PF_MEMALLOC;
3541 lockdep_set_current_reclaim_state(sc.gfp_mask);
3542 reclaim_state.reclaimed_slab = 0;
3543 p->reclaim_state = &reclaim_state;
3545 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3547 p->reclaim_state = NULL;
3548 lockdep_clear_current_reclaim_state();
3549 p->flags &= ~PF_MEMALLOC;
3551 return nr_reclaimed;
3553 #endif /* CONFIG_HIBERNATION */
3555 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3556 not required for correctness. So if the last cpu in a node goes
3557 away, we get changed to run anywhere: as the first one comes back,
3558 restore their cpu bindings. */
3559 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3564 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3565 for_each_node_state(nid, N_MEMORY) {
3566 pg_data_t *pgdat = NODE_DATA(nid);
3567 const struct cpumask *mask;
3569 mask = cpumask_of_node(pgdat->node_id);
3571 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3572 /* One of our CPUs online: restore mask */
3573 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3580 * This kswapd start function will be called by init and node-hot-add.
3581 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3583 int kswapd_run(int nid)
3585 pg_data_t *pgdat = NODE_DATA(nid);
3591 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3592 if (IS_ERR(pgdat->kswapd)) {
3593 /* failure at boot is fatal */
3594 BUG_ON(system_state == SYSTEM_BOOTING);
3595 pr_err("Failed to start kswapd on node %d\n", nid);
3596 ret = PTR_ERR(pgdat->kswapd);
3597 pgdat->kswapd = NULL;
3603 * Called by memory hotplug when all memory in a node is offlined. Caller must
3604 * hold mem_hotplug_begin/end().
3606 void kswapd_stop(int nid)
3608 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3611 kthread_stop(kswapd);
3612 NODE_DATA(nid)->kswapd = NULL;
3616 static int __init kswapd_init(void)
3621 for_each_node_state(nid, N_MEMORY)
3623 hotcpu_notifier(cpu_callback, 0);
3627 module_init(kswapd_init)
3633 * If non-zero call zone_reclaim when the number of free pages falls below
3636 int zone_reclaim_mode __read_mostly;
3638 #define RECLAIM_OFF 0
3639 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3640 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3641 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3644 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3645 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3648 #define ZONE_RECLAIM_PRIORITY 4
3651 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3654 int sysctl_min_unmapped_ratio = 1;
3657 * If the number of slab pages in a zone grows beyond this percentage then
3658 * slab reclaim needs to occur.
3660 int sysctl_min_slab_ratio = 5;
3662 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3664 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3665 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3666 zone_page_state(zone, NR_ACTIVE_FILE);
3669 * It's possible for there to be more file mapped pages than
3670 * accounted for by the pages on the file LRU lists because
3671 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3673 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3676 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3677 static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3679 unsigned long nr_pagecache_reclaimable;
3680 unsigned long delta = 0;
3683 * If RECLAIM_UNMAP is set, then all file pages are considered
3684 * potentially reclaimable. Otherwise, we have to worry about
3685 * pages like swapcache and zone_unmapped_file_pages() provides
3688 if (zone_reclaim_mode & RECLAIM_UNMAP)
3689 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3691 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3693 /* If we can't clean pages, remove dirty pages from consideration */
3694 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3695 delta += zone_page_state(zone, NR_FILE_DIRTY);
3697 /* Watch for any possible underflows due to delta */
3698 if (unlikely(delta > nr_pagecache_reclaimable))
3699 delta = nr_pagecache_reclaimable;
3701 return nr_pagecache_reclaimable - delta;
3705 * Try to free up some pages from this zone through reclaim.
3707 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3709 /* Minimum pages needed in order to stay on node */
3710 const unsigned long nr_pages = 1 << order;
3711 struct task_struct *p = current;
3712 struct reclaim_state reclaim_state;
3713 struct scan_control sc = {
3714 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3715 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3717 .priority = ZONE_RECLAIM_PRIORITY,
3718 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3719 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3725 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3726 * and we also need to be able to write out pages for RECLAIM_WRITE
3727 * and RECLAIM_UNMAP.
3729 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3730 lockdep_set_current_reclaim_state(gfp_mask);
3731 reclaim_state.reclaimed_slab = 0;
3732 p->reclaim_state = &reclaim_state;
3734 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3736 * Free memory by calling shrink zone with increasing
3737 * priorities until we have enough memory freed.
3740 shrink_zone(zone, &sc, true);
3741 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3744 p->reclaim_state = NULL;
3745 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3746 lockdep_clear_current_reclaim_state();
3747 return sc.nr_reclaimed >= nr_pages;
3750 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3756 * Zone reclaim reclaims unmapped file backed pages and
3757 * slab pages if we are over the defined limits.
3759 * A small portion of unmapped file backed pages is needed for
3760 * file I/O otherwise pages read by file I/O will be immediately
3761 * thrown out if the zone is overallocated. So we do not reclaim
3762 * if less than a specified percentage of the zone is used by
3763 * unmapped file backed pages.
3765 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3766 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3767 return ZONE_RECLAIM_FULL;
3769 if (!zone_reclaimable(zone))
3770 return ZONE_RECLAIM_FULL;
3773 * Do not scan if the allocation should not be delayed.
3775 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3776 return ZONE_RECLAIM_NOSCAN;
3779 * Only run zone reclaim on the local zone or on zones that do not
3780 * have associated processors. This will favor the local processor
3781 * over remote processors and spread off node memory allocations
3782 * as wide as possible.
3784 node_id = zone_to_nid(zone);
3785 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3786 return ZONE_RECLAIM_NOSCAN;
3788 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3789 return ZONE_RECLAIM_NOSCAN;
3791 ret = __zone_reclaim(zone, gfp_mask, order);
3792 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3795 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3802 * page_evictable - test whether a page is evictable
3803 * @page: the page to test
3805 * Test whether page is evictable--i.e., should be placed on active/inactive
3806 * lists vs unevictable list.
3808 * Reasons page might not be evictable:
3809 * (1) page's mapping marked unevictable
3810 * (2) page is part of an mlocked VMA
3813 int page_evictable(struct page *page)
3815 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3820 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3821 * @pages: array of pages to check
3822 * @nr_pages: number of pages to check
3824 * Checks pages for evictability and moves them to the appropriate lru list.
3826 * This function is only used for SysV IPC SHM_UNLOCK.
3828 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3830 struct lruvec *lruvec;
3831 struct zone *zone = NULL;
3836 for (i = 0; i < nr_pages; i++) {
3837 struct page *page = pages[i];
3838 struct zone *pagezone;
3841 pagezone = page_zone(page);
3842 if (pagezone != zone) {
3844 spin_unlock_irq(&zone->lru_lock);
3846 spin_lock_irq(&zone->lru_lock);
3848 lruvec = mem_cgroup_page_lruvec(page, zone);
3850 if (!PageLRU(page) || !PageUnevictable(page))
3853 if (page_evictable(page)) {
3854 enum lru_list lru = page_lru_base_type(page);
3856 VM_BUG_ON_PAGE(PageActive(page), page);
3857 ClearPageUnevictable(page);
3858 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3859 add_page_to_lru_list(page, lruvec, lru);
3865 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3866 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3867 spin_unlock_irq(&zone->lru_lock);
3870 #endif /* CONFIG_SHMEM */