1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
68 /* This context's GFP mask */
71 /* Allocation order */
75 * Nodemask of nodes allowed by the caller. If NULL, all nodes
81 * The memory cgroup that hit its limit and as a result is the
82 * primary target of this reclaim invocation.
84 struct mem_cgroup *target_mem_cgroup;
86 /* Scan (total_size >> priority) pages at once */
89 /* The highest zone to isolate pages for reclaim from */
90 enum zone_type reclaim_idx;
92 /* Writepage batching in laptop mode; RECLAIM_WRITE */
93 unsigned int may_writepage:1;
95 /* Can mapped pages be reclaimed? */
96 unsigned int may_unmap:1;
98 /* Can pages be swapped as part of reclaim? */
99 unsigned int may_swap:1;
102 * Cgroups are not reclaimed below their configured memory.low,
103 * unless we threaten to OOM. If any cgroups are skipped due to
104 * memory.low and nothing was reclaimed, go back for memory.low.
106 unsigned int memcg_low_reclaim:1;
107 unsigned int memcg_low_skipped:1;
109 unsigned int hibernation_mode:1;
111 /* One of the zones is ready for compaction */
112 unsigned int compaction_ready:1;
114 /* Incremented by the number of inactive pages that were scanned */
115 unsigned long nr_scanned;
117 /* Number of pages freed so far during a call to shrink_zones() */
118 unsigned long nr_reclaimed;
121 #ifdef ARCH_HAS_PREFETCH
122 #define prefetch_prev_lru_page(_page, _base, _field) \
124 if ((_page)->lru.prev != _base) { \
127 prev = lru_to_page(&(_page->lru)); \
128 prefetch(&prev->_field); \
132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
135 #ifdef ARCH_HAS_PREFETCHW
136 #define prefetchw_prev_lru_page(_page, _base, _field) \
138 if ((_page)->lru.prev != _base) { \
141 prev = lru_to_page(&(_page->lru)); \
142 prefetchw(&prev->_field); \
146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
150 * From 0 .. 100. Higher means more swappy.
152 int vm_swappiness = 60;
154 * The total number of pages which are beyond the high watermark within all
157 unsigned long vm_total_pages;
159 static LIST_HEAD(shrinker_list);
160 static DECLARE_RWSEM(shrinker_rwsem);
163 static bool global_reclaim(struct scan_control *sc)
165 return !sc->target_mem_cgroup;
169 * sane_reclaim - is the usual dirty throttling mechanism operational?
170 * @sc: scan_control in question
172 * The normal page dirty throttling mechanism in balance_dirty_pages() is
173 * completely broken with the legacy memcg and direct stalling in
174 * shrink_page_list() is used for throttling instead, which lacks all the
175 * niceties such as fairness, adaptive pausing, bandwidth proportional
176 * allocation and configurability.
178 * This function tests whether the vmscan currently in progress can assume
179 * that the normal dirty throttling mechanism is operational.
181 static bool sane_reclaim(struct scan_control *sc)
183 struct mem_cgroup *memcg = sc->target_mem_cgroup;
187 #ifdef CONFIG_CGROUP_WRITEBACK
188 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
194 static bool global_reclaim(struct scan_control *sc)
199 static bool sane_reclaim(struct scan_control *sc)
206 * This misses isolated pages which are not accounted for to save counters.
207 * As the data only determines if reclaim or compaction continues, it is
208 * not expected that isolated pages will be a dominating factor.
210 unsigned long zone_reclaimable_pages(struct zone *zone)
214 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
215 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
216 if (get_nr_swap_pages() > 0)
217 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
218 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
224 * lruvec_lru_size - Returns the number of pages on the given LRU list.
225 * @lruvec: lru vector
227 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
229 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
231 unsigned long lru_size;
234 if (!mem_cgroup_disabled())
235 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
237 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
239 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
240 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
243 if (!managed_zone(zone))
246 if (!mem_cgroup_disabled())
247 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
249 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
250 NR_ZONE_LRU_BASE + lru);
251 lru_size -= min(size, lru_size);
259 * Add a shrinker callback to be called from the vm.
261 int register_shrinker(struct shrinker *shrinker)
263 size_t size = sizeof(*shrinker->nr_deferred);
265 if (shrinker->flags & SHRINKER_NUMA_AWARE)
268 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
269 if (!shrinker->nr_deferred)
272 down_write(&shrinker_rwsem);
273 list_add_tail(&shrinker->list, &shrinker_list);
274 up_write(&shrinker_rwsem);
277 EXPORT_SYMBOL(register_shrinker);
282 void unregister_shrinker(struct shrinker *shrinker)
284 if (!shrinker->nr_deferred)
286 down_write(&shrinker_rwsem);
287 list_del(&shrinker->list);
288 up_write(&shrinker_rwsem);
289 kfree(shrinker->nr_deferred);
290 shrinker->nr_deferred = NULL;
292 EXPORT_SYMBOL(unregister_shrinker);
294 #define SHRINK_BATCH 128
296 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
297 struct shrinker *shrinker, int priority)
299 unsigned long freed = 0;
300 unsigned long long delta;
305 int nid = shrinkctl->nid;
306 long batch_size = shrinker->batch ? shrinker->batch
308 long scanned = 0, next_deferred;
310 freeable = shrinker->count_objects(shrinker, shrinkctl);
315 * copy the current shrinker scan count into a local variable
316 * and zero it so that other concurrent shrinker invocations
317 * don't also do this scanning work.
319 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
322 delta = freeable >> priority;
324 do_div(delta, shrinker->seeks);
326 if (total_scan < 0) {
327 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
328 shrinker->scan_objects, total_scan);
329 total_scan = freeable;
332 next_deferred = total_scan;
335 * We need to avoid excessive windup on filesystem shrinkers
336 * due to large numbers of GFP_NOFS allocations causing the
337 * shrinkers to return -1 all the time. This results in a large
338 * nr being built up so when a shrink that can do some work
339 * comes along it empties the entire cache due to nr >>>
340 * freeable. This is bad for sustaining a working set in
343 * Hence only allow the shrinker to scan the entire cache when
344 * a large delta change is calculated directly.
346 if (delta < freeable / 4)
347 total_scan = min(total_scan, freeable / 2);
350 * Avoid risking looping forever due to too large nr value:
351 * never try to free more than twice the estimate number of
354 if (total_scan > freeable * 2)
355 total_scan = freeable * 2;
357 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
358 freeable, delta, total_scan, priority);
361 * Normally, we should not scan less than batch_size objects in one
362 * pass to avoid too frequent shrinker calls, but if the slab has less
363 * than batch_size objects in total and we are really tight on memory,
364 * we will try to reclaim all available objects, otherwise we can end
365 * up failing allocations although there are plenty of reclaimable
366 * objects spread over several slabs with usage less than the
369 * We detect the "tight on memory" situations by looking at the total
370 * number of objects we want to scan (total_scan). If it is greater
371 * than the total number of objects on slab (freeable), we must be
372 * scanning at high prio and therefore should try to reclaim as much as
375 while (total_scan >= batch_size ||
376 total_scan >= freeable) {
378 unsigned long nr_to_scan = min(batch_size, total_scan);
380 shrinkctl->nr_to_scan = nr_to_scan;
381 shrinkctl->nr_scanned = nr_to_scan;
382 ret = shrinker->scan_objects(shrinker, shrinkctl);
383 if (ret == SHRINK_STOP)
387 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
388 total_scan -= shrinkctl->nr_scanned;
389 scanned += shrinkctl->nr_scanned;
394 if (next_deferred >= scanned)
395 next_deferred -= scanned;
399 * move the unused scan count back into the shrinker in a
400 * manner that handles concurrent updates. If we exhausted the
401 * scan, there is no need to do an update.
403 if (next_deferred > 0)
404 new_nr = atomic_long_add_return(next_deferred,
405 &shrinker->nr_deferred[nid]);
407 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
409 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
414 * shrink_slab - shrink slab caches
415 * @gfp_mask: allocation context
416 * @nid: node whose slab caches to target
417 * @memcg: memory cgroup whose slab caches to target
418 * @priority: the reclaim priority
420 * Call the shrink functions to age shrinkable caches.
422 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
423 * unaware shrinkers will receive a node id of 0 instead.
425 * @memcg specifies the memory cgroup to target. If it is not NULL,
426 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
427 * objects from the memory cgroup specified. Otherwise, only unaware
428 * shrinkers are called.
430 * @priority is sc->priority, we take the number of objects and >> by priority
431 * in order to get the scan target.
433 * Returns the number of reclaimed slab objects.
435 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
436 struct mem_cgroup *memcg,
439 struct shrinker *shrinker;
440 unsigned long freed = 0;
442 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
445 if (!down_read_trylock(&shrinker_rwsem)) {
447 * If we would return 0, our callers would understand that we
448 * have nothing else to shrink and give up trying. By returning
449 * 1 we keep it going and assume we'll be able to shrink next
456 list_for_each_entry(shrinker, &shrinker_list, list) {
457 struct shrink_control sc = {
458 .gfp_mask = gfp_mask,
464 * If kernel memory accounting is disabled, we ignore
465 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
466 * passing NULL for memcg.
468 if (memcg_kmem_enabled() &&
469 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
472 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
475 freed += do_shrink_slab(&sc, shrinker, priority);
477 * Bail out if someone want to register a new shrinker to
478 * prevent the regsitration from being stalled for long periods
479 * by parallel ongoing shrinking.
481 if (rwsem_is_contended(&shrinker_rwsem)) {
487 up_read(&shrinker_rwsem);
493 void drop_slab_node(int nid)
498 struct mem_cgroup *memcg = NULL;
502 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
503 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
504 } while (freed > 10);
511 for_each_online_node(nid)
515 static inline int is_page_cache_freeable(struct page *page)
518 * A freeable page cache page is referenced only by the caller
519 * that isolated the page, the page cache radix tree and
520 * optional buffer heads at page->private.
522 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
524 return page_count(page) - page_has_private(page) == 1 + radix_pins;
527 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
529 if (current->flags & PF_SWAPWRITE)
531 if (!inode_write_congested(inode))
533 if (inode_to_bdi(inode) == current->backing_dev_info)
539 * We detected a synchronous write error writing a page out. Probably
540 * -ENOSPC. We need to propagate that into the address_space for a subsequent
541 * fsync(), msync() or close().
543 * The tricky part is that after writepage we cannot touch the mapping: nothing
544 * prevents it from being freed up. But we have a ref on the page and once
545 * that page is locked, the mapping is pinned.
547 * We're allowed to run sleeping lock_page() here because we know the caller has
550 static void handle_write_error(struct address_space *mapping,
551 struct page *page, int error)
554 if (page_mapping(page) == mapping)
555 mapping_set_error(mapping, error);
559 /* possible outcome of pageout() */
561 /* failed to write page out, page is locked */
563 /* move page to the active list, page is locked */
565 /* page has been sent to the disk successfully, page is unlocked */
567 /* page is clean and locked */
572 * pageout is called by shrink_page_list() for each dirty page.
573 * Calls ->writepage().
575 static pageout_t pageout(struct page *page, struct address_space *mapping,
576 struct scan_control *sc)
579 * If the page is dirty, only perform writeback if that write
580 * will be non-blocking. To prevent this allocation from being
581 * stalled by pagecache activity. But note that there may be
582 * stalls if we need to run get_block(). We could test
583 * PagePrivate for that.
585 * If this process is currently in __generic_file_write_iter() against
586 * this page's queue, we can perform writeback even if that
589 * If the page is swapcache, write it back even if that would
590 * block, for some throttling. This happens by accident, because
591 * swap_backing_dev_info is bust: it doesn't reflect the
592 * congestion state of the swapdevs. Easy to fix, if needed.
594 if (!is_page_cache_freeable(page))
598 * Some data journaling orphaned pages can have
599 * page->mapping == NULL while being dirty with clean buffers.
601 if (page_has_private(page)) {
602 if (try_to_free_buffers(page)) {
603 ClearPageDirty(page);
604 pr_info("%s: orphaned page\n", __func__);
610 if (mapping->a_ops->writepage == NULL)
611 return PAGE_ACTIVATE;
612 if (!may_write_to_inode(mapping->host, sc))
615 if (clear_page_dirty_for_io(page)) {
617 struct writeback_control wbc = {
618 .sync_mode = WB_SYNC_NONE,
619 .nr_to_write = SWAP_CLUSTER_MAX,
621 .range_end = LLONG_MAX,
625 SetPageReclaim(page);
626 res = mapping->a_ops->writepage(page, &wbc);
628 handle_write_error(mapping, page, res);
629 if (res == AOP_WRITEPAGE_ACTIVATE) {
630 ClearPageReclaim(page);
631 return PAGE_ACTIVATE;
634 if (!PageWriteback(page)) {
635 /* synchronous write or broken a_ops? */
636 ClearPageReclaim(page);
638 trace_mm_vmscan_writepage(page);
639 inc_node_page_state(page, NR_VMSCAN_WRITE);
647 * Same as remove_mapping, but if the page is removed from the mapping, it
648 * gets returned with a refcount of 0.
650 static int __remove_mapping(struct address_space *mapping, struct page *page,
656 BUG_ON(!PageLocked(page));
657 BUG_ON(mapping != page_mapping(page));
659 spin_lock_irqsave(&mapping->tree_lock, flags);
661 * The non racy check for a busy page.
663 * Must be careful with the order of the tests. When someone has
664 * a ref to the page, it may be possible that they dirty it then
665 * drop the reference. So if PageDirty is tested before page_count
666 * here, then the following race may occur:
668 * get_user_pages(&page);
669 * [user mapping goes away]
671 * !PageDirty(page) [good]
672 * SetPageDirty(page);
674 * !page_count(page) [good, discard it]
676 * [oops, our write_to data is lost]
678 * Reversing the order of the tests ensures such a situation cannot
679 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
680 * load is not satisfied before that of page->_refcount.
682 * Note that if SetPageDirty is always performed via set_page_dirty,
683 * and thus under tree_lock, then this ordering is not required.
685 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
686 refcount = 1 + HPAGE_PMD_NR;
689 if (!page_ref_freeze(page, refcount))
691 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
692 if (unlikely(PageDirty(page))) {
693 page_ref_unfreeze(page, refcount);
697 if (PageSwapCache(page)) {
698 swp_entry_t swap = { .val = page_private(page) };
699 mem_cgroup_swapout(page, swap);
700 __delete_from_swap_cache(page);
701 spin_unlock_irqrestore(&mapping->tree_lock, flags);
702 put_swap_page(page, swap);
704 void (*freepage)(struct page *);
707 freepage = mapping->a_ops->freepage;
709 * Remember a shadow entry for reclaimed file cache in
710 * order to detect refaults, thus thrashing, later on.
712 * But don't store shadows in an address space that is
713 * already exiting. This is not just an optizimation,
714 * inode reclaim needs to empty out the radix tree or
715 * the nodes are lost. Don't plant shadows behind its
718 * We also don't store shadows for DAX mappings because the
719 * only page cache pages found in these are zero pages
720 * covering holes, and because we don't want to mix DAX
721 * exceptional entries and shadow exceptional entries in the
724 if (reclaimed && page_is_file_cache(page) &&
725 !mapping_exiting(mapping) && !dax_mapping(mapping))
726 shadow = workingset_eviction(mapping, page);
727 __delete_from_page_cache(page, shadow);
728 spin_unlock_irqrestore(&mapping->tree_lock, flags);
730 if (freepage != NULL)
737 spin_unlock_irqrestore(&mapping->tree_lock, flags);
742 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
743 * someone else has a ref on the page, abort and return 0. If it was
744 * successfully detached, return 1. Assumes the caller has a single ref on
747 int remove_mapping(struct address_space *mapping, struct page *page)
749 if (__remove_mapping(mapping, page, false)) {
751 * Unfreezing the refcount with 1 rather than 2 effectively
752 * drops the pagecache ref for us without requiring another
755 page_ref_unfreeze(page, 1);
762 * putback_lru_page - put previously isolated page onto appropriate LRU list
763 * @page: page to be put back to appropriate lru list
765 * Add previously isolated @page to appropriate LRU list.
766 * Page may still be unevictable for other reasons.
768 * lru_lock must not be held, interrupts must be enabled.
770 void putback_lru_page(struct page *page)
773 int was_unevictable = PageUnevictable(page);
775 VM_BUG_ON_PAGE(PageLRU(page), page);
778 ClearPageUnevictable(page);
780 if (page_evictable(page)) {
782 * For evictable pages, we can use the cache.
783 * In event of a race, worst case is we end up with an
784 * unevictable page on [in]active list.
785 * We know how to handle that.
787 is_unevictable = false;
791 * Put unevictable pages directly on zone's unevictable
794 is_unevictable = true;
795 add_page_to_unevictable_list(page);
797 * When racing with an mlock or AS_UNEVICTABLE clearing
798 * (page is unlocked) make sure that if the other thread
799 * does not observe our setting of PG_lru and fails
800 * isolation/check_move_unevictable_pages,
801 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
802 * the page back to the evictable list.
804 * The other side is TestClearPageMlocked() or shmem_lock().
810 * page's status can change while we move it among lru. If an evictable
811 * page is on unevictable list, it never be freed. To avoid that,
812 * check after we added it to the list, again.
814 if (is_unevictable && page_evictable(page)) {
815 if (!isolate_lru_page(page)) {
819 /* This means someone else dropped this page from LRU
820 * So, it will be freed or putback to LRU again. There is
821 * nothing to do here.
825 if (was_unevictable && !is_unevictable)
826 count_vm_event(UNEVICTABLE_PGRESCUED);
827 else if (!was_unevictable && is_unevictable)
828 count_vm_event(UNEVICTABLE_PGCULLED);
830 put_page(page); /* drop ref from isolate */
833 enum page_references {
835 PAGEREF_RECLAIM_CLEAN,
840 static enum page_references page_check_references(struct page *page,
841 struct scan_control *sc)
843 int referenced_ptes, referenced_page;
844 unsigned long vm_flags;
846 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
848 referenced_page = TestClearPageReferenced(page);
851 * Mlock lost the isolation race with us. Let try_to_unmap()
852 * move the page to the unevictable list.
854 if (vm_flags & VM_LOCKED)
855 return PAGEREF_RECLAIM;
857 if (referenced_ptes) {
858 if (PageSwapBacked(page))
859 return PAGEREF_ACTIVATE;
861 * All mapped pages start out with page table
862 * references from the instantiating fault, so we need
863 * to look twice if a mapped file page is used more
866 * Mark it and spare it for another trip around the
867 * inactive list. Another page table reference will
868 * lead to its activation.
870 * Note: the mark is set for activated pages as well
871 * so that recently deactivated but used pages are
874 SetPageReferenced(page);
876 if (referenced_page || referenced_ptes > 1)
877 return PAGEREF_ACTIVATE;
880 * Activate file-backed executable pages after first usage.
882 if (vm_flags & VM_EXEC)
883 return PAGEREF_ACTIVATE;
888 /* Reclaim if clean, defer dirty pages to writeback */
889 if (referenced_page && !PageSwapBacked(page))
890 return PAGEREF_RECLAIM_CLEAN;
892 return PAGEREF_RECLAIM;
895 /* Check if a page is dirty or under writeback */
896 static void page_check_dirty_writeback(struct page *page,
897 bool *dirty, bool *writeback)
899 struct address_space *mapping;
902 * Anonymous pages are not handled by flushers and must be written
903 * from reclaim context. Do not stall reclaim based on them
905 if (!page_is_file_cache(page) ||
906 (PageAnon(page) && !PageSwapBacked(page))) {
912 /* By default assume that the page flags are accurate */
913 *dirty = PageDirty(page);
914 *writeback = PageWriteback(page);
916 /* Verify dirty/writeback state if the filesystem supports it */
917 if (!page_has_private(page))
920 mapping = page_mapping(page);
921 if (mapping && mapping->a_ops->is_dirty_writeback)
922 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
925 struct reclaim_stat {
927 unsigned nr_unqueued_dirty;
928 unsigned nr_congested;
929 unsigned nr_writeback;
930 unsigned nr_immediate;
931 unsigned nr_activate;
932 unsigned nr_ref_keep;
933 unsigned nr_unmap_fail;
937 * shrink_page_list() returns the number of reclaimed pages
939 static unsigned long shrink_page_list(struct list_head *page_list,
940 struct pglist_data *pgdat,
941 struct scan_control *sc,
942 enum ttu_flags ttu_flags,
943 struct reclaim_stat *stat,
946 LIST_HEAD(ret_pages);
947 LIST_HEAD(free_pages);
949 unsigned nr_unqueued_dirty = 0;
950 unsigned nr_dirty = 0;
951 unsigned nr_congested = 0;
952 unsigned nr_reclaimed = 0;
953 unsigned nr_writeback = 0;
954 unsigned nr_immediate = 0;
955 unsigned nr_ref_keep = 0;
956 unsigned nr_unmap_fail = 0;
960 while (!list_empty(page_list)) {
961 struct address_space *mapping;
964 enum page_references references = PAGEREF_RECLAIM_CLEAN;
965 bool dirty, writeback;
969 page = lru_to_page(page_list);
970 list_del(&page->lru);
972 if (!trylock_page(page))
975 VM_BUG_ON_PAGE(PageActive(page), page);
979 if (unlikely(!page_evictable(page)))
980 goto activate_locked;
982 if (!sc->may_unmap && page_mapped(page))
985 /* Double the slab pressure for mapped and swapcache pages */
986 if ((page_mapped(page) || PageSwapCache(page)) &&
987 !(PageAnon(page) && !PageSwapBacked(page)))
990 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
991 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
994 * The number of dirty pages determines if a zone is marked
995 * reclaim_congested which affects wait_iff_congested. kswapd
996 * will stall and start writing pages if the tail of the LRU
997 * is all dirty unqueued pages.
999 page_check_dirty_writeback(page, &dirty, &writeback);
1000 if (dirty || writeback)
1003 if (dirty && !writeback)
1004 nr_unqueued_dirty++;
1007 * Treat this page as congested if the underlying BDI is or if
1008 * pages are cycling through the LRU so quickly that the
1009 * pages marked for immediate reclaim are making it to the
1010 * end of the LRU a second time.
1012 mapping = page_mapping(page);
1013 if (((dirty || writeback) && mapping &&
1014 inode_write_congested(mapping->host)) ||
1015 (writeback && PageReclaim(page)))
1019 * If a page at the tail of the LRU is under writeback, there
1020 * are three cases to consider.
1022 * 1) If reclaim is encountering an excessive number of pages
1023 * under writeback and this page is both under writeback and
1024 * PageReclaim then it indicates that pages are being queued
1025 * for IO but are being recycled through the LRU before the
1026 * IO can complete. Waiting on the page itself risks an
1027 * indefinite stall if it is impossible to writeback the
1028 * page due to IO error or disconnected storage so instead
1029 * note that the LRU is being scanned too quickly and the
1030 * caller can stall after page list has been processed.
1032 * 2) Global or new memcg reclaim encounters a page that is
1033 * not marked for immediate reclaim, or the caller does not
1034 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1035 * not to fs). In this case mark the page for immediate
1036 * reclaim and continue scanning.
1038 * Require may_enter_fs because we would wait on fs, which
1039 * may not have submitted IO yet. And the loop driver might
1040 * enter reclaim, and deadlock if it waits on a page for
1041 * which it is needed to do the write (loop masks off
1042 * __GFP_IO|__GFP_FS for this reason); but more thought
1043 * would probably show more reasons.
1045 * 3) Legacy memcg encounters a page that is already marked
1046 * PageReclaim. memcg does not have any dirty pages
1047 * throttling so we could easily OOM just because too many
1048 * pages are in writeback and there is nothing else to
1049 * reclaim. Wait for the writeback to complete.
1051 * In cases 1) and 2) we activate the pages to get them out of
1052 * the way while we continue scanning for clean pages on the
1053 * inactive list and refilling from the active list. The
1054 * observation here is that waiting for disk writes is more
1055 * expensive than potentially causing reloads down the line.
1056 * Since they're marked for immediate reclaim, they won't put
1057 * memory pressure on the cache working set any longer than it
1058 * takes to write them to disk.
1060 if (PageWriteback(page)) {
1062 if (current_is_kswapd() &&
1063 PageReclaim(page) &&
1064 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1066 goto activate_locked;
1069 } else if (sane_reclaim(sc) ||
1070 !PageReclaim(page) || !may_enter_fs) {
1072 * This is slightly racy - end_page_writeback()
1073 * might have just cleared PageReclaim, then
1074 * setting PageReclaim here end up interpreted
1075 * as PageReadahead - but that does not matter
1076 * enough to care. What we do want is for this
1077 * page to have PageReclaim set next time memcg
1078 * reclaim reaches the tests above, so it will
1079 * then wait_on_page_writeback() to avoid OOM;
1080 * and it's also appropriate in global reclaim.
1082 SetPageReclaim(page);
1084 goto activate_locked;
1089 wait_on_page_writeback(page);
1090 /* then go back and try same page again */
1091 list_add_tail(&page->lru, page_list);
1097 references = page_check_references(page, sc);
1099 switch (references) {
1100 case PAGEREF_ACTIVATE:
1101 goto activate_locked;
1105 case PAGEREF_RECLAIM:
1106 case PAGEREF_RECLAIM_CLEAN:
1107 ; /* try to reclaim the page below */
1111 * Anonymous process memory has backing store?
1112 * Try to allocate it some swap space here.
1113 * Lazyfree page could be freed directly
1115 if (PageAnon(page) && PageSwapBacked(page)) {
1116 if (!PageSwapCache(page)) {
1117 if (!(sc->gfp_mask & __GFP_IO))
1119 if (PageTransHuge(page)) {
1120 /* cannot split THP, skip it */
1121 if (!can_split_huge_page(page, NULL))
1122 goto activate_locked;
1124 * Split pages without a PMD map right
1125 * away. Chances are some or all of the
1126 * tail pages can be freed without IO.
1128 if (!compound_mapcount(page) &&
1129 split_huge_page_to_list(page,
1131 goto activate_locked;
1133 if (!add_to_swap(page)) {
1134 if (!PageTransHuge(page))
1135 goto activate_locked;
1136 /* Fallback to swap normal pages */
1137 if (split_huge_page_to_list(page,
1139 goto activate_locked;
1140 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1141 count_vm_event(THP_SWPOUT_FALLBACK);
1143 if (!add_to_swap(page))
1144 goto activate_locked;
1149 /* Adding to swap updated mapping */
1150 mapping = page_mapping(page);
1152 } else if (unlikely(PageTransHuge(page))) {
1153 /* Split file THP */
1154 if (split_huge_page_to_list(page, page_list))
1159 * The page is mapped into the page tables of one or more
1160 * processes. Try to unmap it here.
1162 if (page_mapped(page)) {
1163 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1165 if (unlikely(PageTransHuge(page)))
1166 flags |= TTU_SPLIT_HUGE_PMD;
1167 if (!try_to_unmap(page, flags)) {
1169 goto activate_locked;
1173 if (PageDirty(page)) {
1175 * Only kswapd can writeback filesystem pages
1176 * to avoid risk of stack overflow. But avoid
1177 * injecting inefficient single-page IO into
1178 * flusher writeback as much as possible: only
1179 * write pages when we've encountered many
1180 * dirty pages, and when we've already scanned
1181 * the rest of the LRU for clean pages and see
1182 * the same dirty pages again (PageReclaim).
1184 if (page_is_file_cache(page) &&
1185 (!current_is_kswapd() || !PageReclaim(page) ||
1186 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1188 * Immediately reclaim when written back.
1189 * Similar in principal to deactivate_page()
1190 * except we already have the page isolated
1191 * and know it's dirty
1193 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1194 SetPageReclaim(page);
1196 goto activate_locked;
1199 if (references == PAGEREF_RECLAIM_CLEAN)
1203 if (!sc->may_writepage)
1207 * Page is dirty. Flush the TLB if a writable entry
1208 * potentially exists to avoid CPU writes after IO
1209 * starts and then write it out here.
1211 try_to_unmap_flush_dirty();
1212 switch (pageout(page, mapping, sc)) {
1216 goto activate_locked;
1218 if (PageWriteback(page))
1220 if (PageDirty(page))
1224 * A synchronous write - probably a ramdisk. Go
1225 * ahead and try to reclaim the page.
1227 if (!trylock_page(page))
1229 if (PageDirty(page) || PageWriteback(page))
1231 mapping = page_mapping(page);
1233 ; /* try to free the page below */
1238 * If the page has buffers, try to free the buffer mappings
1239 * associated with this page. If we succeed we try to free
1242 * We do this even if the page is PageDirty().
1243 * try_to_release_page() does not perform I/O, but it is
1244 * possible for a page to have PageDirty set, but it is actually
1245 * clean (all its buffers are clean). This happens if the
1246 * buffers were written out directly, with submit_bh(). ext3
1247 * will do this, as well as the blockdev mapping.
1248 * try_to_release_page() will discover that cleanness and will
1249 * drop the buffers and mark the page clean - it can be freed.
1251 * Rarely, pages can have buffers and no ->mapping. These are
1252 * the pages which were not successfully invalidated in
1253 * truncate_complete_page(). We try to drop those buffers here
1254 * and if that worked, and the page is no longer mapped into
1255 * process address space (page_count == 1) it can be freed.
1256 * Otherwise, leave the page on the LRU so it is swappable.
1258 if (page_has_private(page)) {
1259 if (!try_to_release_page(page, sc->gfp_mask))
1260 goto activate_locked;
1261 if (!mapping && page_count(page) == 1) {
1263 if (put_page_testzero(page))
1267 * rare race with speculative reference.
1268 * the speculative reference will free
1269 * this page shortly, so we may
1270 * increment nr_reclaimed here (and
1271 * leave it off the LRU).
1279 if (PageAnon(page) && !PageSwapBacked(page)) {
1280 /* follow __remove_mapping for reference */
1281 if (!page_ref_freeze(page, 1))
1283 if (PageDirty(page)) {
1284 page_ref_unfreeze(page, 1);
1288 count_vm_event(PGLAZYFREED);
1289 count_memcg_page_event(page, PGLAZYFREED);
1290 } else if (!mapping || !__remove_mapping(mapping, page, true))
1293 * At this point, we have no other references and there is
1294 * no way to pick any more up (removed from LRU, removed
1295 * from pagecache). Can use non-atomic bitops now (and
1296 * we obviously don't have to worry about waking up a process
1297 * waiting on the page lock, because there are no references.
1299 __ClearPageLocked(page);
1304 * Is there need to periodically free_page_list? It would
1305 * appear not as the counts should be low
1307 if (unlikely(PageTransHuge(page))) {
1308 mem_cgroup_uncharge(page);
1309 (*get_compound_page_dtor(page))(page);
1311 list_add(&page->lru, &free_pages);
1315 /* Not a candidate for swapping, so reclaim swap space. */
1316 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1318 try_to_free_swap(page);
1319 VM_BUG_ON_PAGE(PageActive(page), page);
1320 if (!PageMlocked(page)) {
1321 SetPageActive(page);
1323 count_memcg_page_event(page, PGACTIVATE);
1328 list_add(&page->lru, &ret_pages);
1329 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1332 mem_cgroup_uncharge_list(&free_pages);
1333 try_to_unmap_flush();
1334 free_unref_page_list(&free_pages);
1336 list_splice(&ret_pages, page_list);
1337 count_vm_events(PGACTIVATE, pgactivate);
1340 stat->nr_dirty = nr_dirty;
1341 stat->nr_congested = nr_congested;
1342 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1343 stat->nr_writeback = nr_writeback;
1344 stat->nr_immediate = nr_immediate;
1345 stat->nr_activate = pgactivate;
1346 stat->nr_ref_keep = nr_ref_keep;
1347 stat->nr_unmap_fail = nr_unmap_fail;
1349 return nr_reclaimed;
1352 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1353 struct list_head *page_list)
1355 struct scan_control sc = {
1356 .gfp_mask = GFP_KERNEL,
1357 .priority = DEF_PRIORITY,
1361 struct page *page, *next;
1362 LIST_HEAD(clean_pages);
1364 list_for_each_entry_safe(page, next, page_list, lru) {
1365 if (page_is_file_cache(page) && !PageDirty(page) &&
1366 !__PageMovable(page)) {
1367 ClearPageActive(page);
1368 list_move(&page->lru, &clean_pages);
1372 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1373 TTU_IGNORE_ACCESS, NULL, true);
1374 list_splice(&clean_pages, page_list);
1375 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1380 * Attempt to remove the specified page from its LRU. Only take this page
1381 * if it is of the appropriate PageActive status. Pages which are being
1382 * freed elsewhere are also ignored.
1384 * page: page to consider
1385 * mode: one of the LRU isolation modes defined above
1387 * returns 0 on success, -ve errno on failure.
1389 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1393 /* Only take pages on the LRU. */
1397 /* Compaction should not handle unevictable pages but CMA can do so */
1398 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1404 * To minimise LRU disruption, the caller can indicate that it only
1405 * wants to isolate pages it will be able to operate on without
1406 * blocking - clean pages for the most part.
1408 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1409 * that it is possible to migrate without blocking
1411 if (mode & ISOLATE_ASYNC_MIGRATE) {
1412 /* All the caller can do on PageWriteback is block */
1413 if (PageWriteback(page))
1416 if (PageDirty(page)) {
1417 struct address_space *mapping;
1420 * Only pages without mappings or that have a
1421 * ->migratepage callback are possible to migrate
1424 mapping = page_mapping(page);
1425 if (mapping && !mapping->a_ops->migratepage)
1430 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1433 if (likely(get_page_unless_zero(page))) {
1435 * Be careful not to clear PageLRU until after we're
1436 * sure the page is not being freed elsewhere -- the
1437 * page release code relies on it.
1448 * Update LRU sizes after isolating pages. The LRU size updates must
1449 * be complete before mem_cgroup_update_lru_size due to a santity check.
1451 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1452 enum lru_list lru, unsigned long *nr_zone_taken)
1456 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1457 if (!nr_zone_taken[zid])
1460 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1462 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1469 * zone_lru_lock is heavily contended. Some of the functions that
1470 * shrink the lists perform better by taking out a batch of pages
1471 * and working on them outside the LRU lock.
1473 * For pagecache intensive workloads, this function is the hottest
1474 * spot in the kernel (apart from copy_*_user functions).
1476 * Appropriate locks must be held before calling this function.
1478 * @nr_to_scan: The number of eligible pages to look through on the list.
1479 * @lruvec: The LRU vector to pull pages from.
1480 * @dst: The temp list to put pages on to.
1481 * @nr_scanned: The number of pages that were scanned.
1482 * @sc: The scan_control struct for this reclaim session
1483 * @mode: One of the LRU isolation modes
1484 * @lru: LRU list id for isolating
1486 * returns how many pages were moved onto *@dst.
1488 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1489 struct lruvec *lruvec, struct list_head *dst,
1490 unsigned long *nr_scanned, struct scan_control *sc,
1491 isolate_mode_t mode, enum lru_list lru)
1493 struct list_head *src = &lruvec->lists[lru];
1494 unsigned long nr_taken = 0;
1495 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1496 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1497 unsigned long skipped = 0;
1498 unsigned long scan, total_scan, nr_pages;
1499 LIST_HEAD(pages_skipped);
1502 for (total_scan = 0;
1503 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1507 page = lru_to_page(src);
1508 prefetchw_prev_lru_page(page, src, flags);
1510 VM_BUG_ON_PAGE(!PageLRU(page), page);
1512 if (page_zonenum(page) > sc->reclaim_idx) {
1513 list_move(&page->lru, &pages_skipped);
1514 nr_skipped[page_zonenum(page)]++;
1519 * Do not count skipped pages because that makes the function
1520 * return with no isolated pages if the LRU mostly contains
1521 * ineligible pages. This causes the VM to not reclaim any
1522 * pages, triggering a premature OOM.
1525 switch (__isolate_lru_page(page, mode)) {
1527 nr_pages = hpage_nr_pages(page);
1528 nr_taken += nr_pages;
1529 nr_zone_taken[page_zonenum(page)] += nr_pages;
1530 list_move(&page->lru, dst);
1534 /* else it is being freed elsewhere */
1535 list_move(&page->lru, src);
1544 * Splice any skipped pages to the start of the LRU list. Note that
1545 * this disrupts the LRU order when reclaiming for lower zones but
1546 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1547 * scanning would soon rescan the same pages to skip and put the
1548 * system at risk of premature OOM.
1550 if (!list_empty(&pages_skipped)) {
1553 list_splice(&pages_skipped, src);
1554 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1555 if (!nr_skipped[zid])
1558 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1559 skipped += nr_skipped[zid];
1562 *nr_scanned = total_scan;
1563 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1564 total_scan, skipped, nr_taken, mode, lru);
1565 update_lru_sizes(lruvec, lru, nr_zone_taken);
1570 * isolate_lru_page - tries to isolate a page from its LRU list
1571 * @page: page to isolate from its LRU list
1573 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1574 * vmstat statistic corresponding to whatever LRU list the page was on.
1576 * Returns 0 if the page was removed from an LRU list.
1577 * Returns -EBUSY if the page was not on an LRU list.
1579 * The returned page will have PageLRU() cleared. If it was found on
1580 * the active list, it will have PageActive set. If it was found on
1581 * the unevictable list, it will have the PageUnevictable bit set. That flag
1582 * may need to be cleared by the caller before letting the page go.
1584 * The vmstat statistic corresponding to the list on which the page was
1585 * found will be decremented.
1588 * (1) Must be called with an elevated refcount on the page. This is a
1589 * fundamentnal difference from isolate_lru_pages (which is called
1590 * without a stable reference).
1591 * (2) the lru_lock must not be held.
1592 * (3) interrupts must be enabled.
1594 int isolate_lru_page(struct page *page)
1598 VM_BUG_ON_PAGE(!page_count(page), page);
1599 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1601 if (PageLRU(page)) {
1602 struct zone *zone = page_zone(page);
1603 struct lruvec *lruvec;
1605 spin_lock_irq(zone_lru_lock(zone));
1606 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1607 if (PageLRU(page)) {
1608 int lru = page_lru(page);
1611 del_page_from_lru_list(page, lruvec, lru);
1614 spin_unlock_irq(zone_lru_lock(zone));
1620 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1621 * then get resheduled. When there are massive number of tasks doing page
1622 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1623 * the LRU list will go small and be scanned faster than necessary, leading to
1624 * unnecessary swapping, thrashing and OOM.
1626 static int too_many_isolated(struct pglist_data *pgdat, int file,
1627 struct scan_control *sc)
1629 unsigned long inactive, isolated;
1631 if (current_is_kswapd())
1634 if (!sane_reclaim(sc))
1638 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1639 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1641 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1642 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1646 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1647 * won't get blocked by normal direct-reclaimers, forming a circular
1650 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1653 return isolated > inactive;
1656 static noinline_for_stack void
1657 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1659 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1660 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1661 LIST_HEAD(pages_to_free);
1664 * Put back any unfreeable pages.
1666 while (!list_empty(page_list)) {
1667 struct page *page = lru_to_page(page_list);
1670 VM_BUG_ON_PAGE(PageLRU(page), page);
1671 list_del(&page->lru);
1672 if (unlikely(!page_evictable(page))) {
1673 spin_unlock_irq(&pgdat->lru_lock);
1674 putback_lru_page(page);
1675 spin_lock_irq(&pgdat->lru_lock);
1679 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1682 lru = page_lru(page);
1683 add_page_to_lru_list(page, lruvec, lru);
1685 if (is_active_lru(lru)) {
1686 int file = is_file_lru(lru);
1687 int numpages = hpage_nr_pages(page);
1688 reclaim_stat->recent_rotated[file] += numpages;
1690 if (put_page_testzero(page)) {
1691 __ClearPageLRU(page);
1692 __ClearPageActive(page);
1693 del_page_from_lru_list(page, lruvec, lru);
1695 if (unlikely(PageCompound(page))) {
1696 spin_unlock_irq(&pgdat->lru_lock);
1697 mem_cgroup_uncharge(page);
1698 (*get_compound_page_dtor(page))(page);
1699 spin_lock_irq(&pgdat->lru_lock);
1701 list_add(&page->lru, &pages_to_free);
1706 * To save our caller's stack, now use input list for pages to free.
1708 list_splice(&pages_to_free, page_list);
1712 * If a kernel thread (such as nfsd for loop-back mounts) services
1713 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1714 * In that case we should only throttle if the backing device it is
1715 * writing to is congested. In other cases it is safe to throttle.
1717 static int current_may_throttle(void)
1719 return !(current->flags & PF_LESS_THROTTLE) ||
1720 current->backing_dev_info == NULL ||
1721 bdi_write_congested(current->backing_dev_info);
1725 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1726 * of reclaimed pages
1728 static noinline_for_stack unsigned long
1729 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1730 struct scan_control *sc, enum lru_list lru)
1732 LIST_HEAD(page_list);
1733 unsigned long nr_scanned;
1734 unsigned long nr_reclaimed = 0;
1735 unsigned long nr_taken;
1736 struct reclaim_stat stat = {};
1737 isolate_mode_t isolate_mode = 0;
1738 int file = is_file_lru(lru);
1739 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1740 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1741 bool stalled = false;
1743 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1747 /* wait a bit for the reclaimer. */
1751 /* We are about to die and free our memory. Return now. */
1752 if (fatal_signal_pending(current))
1753 return SWAP_CLUSTER_MAX;
1759 isolate_mode |= ISOLATE_UNMAPPED;
1761 spin_lock_irq(&pgdat->lru_lock);
1763 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1764 &nr_scanned, sc, isolate_mode, lru);
1766 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1767 reclaim_stat->recent_scanned[file] += nr_taken;
1769 if (current_is_kswapd()) {
1770 if (global_reclaim(sc))
1771 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1772 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1775 if (global_reclaim(sc))
1776 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1777 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1780 spin_unlock_irq(&pgdat->lru_lock);
1785 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1788 spin_lock_irq(&pgdat->lru_lock);
1790 if (current_is_kswapd()) {
1791 if (global_reclaim(sc))
1792 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1793 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1796 if (global_reclaim(sc))
1797 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1798 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1802 putback_inactive_pages(lruvec, &page_list);
1804 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1806 spin_unlock_irq(&pgdat->lru_lock);
1808 mem_cgroup_uncharge_list(&page_list);
1809 free_unref_page_list(&page_list);
1812 * If reclaim is isolating dirty pages under writeback, it implies
1813 * that the long-lived page allocation rate is exceeding the page
1814 * laundering rate. Either the global limits are not being effective
1815 * at throttling processes due to the page distribution throughout
1816 * zones or there is heavy usage of a slow backing device. The
1817 * only option is to throttle from reclaim context which is not ideal
1818 * as there is no guarantee the dirtying process is throttled in the
1819 * same way balance_dirty_pages() manages.
1821 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1822 * of pages under pages flagged for immediate reclaim and stall if any
1823 * are encountered in the nr_immediate check below.
1825 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1826 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1829 * Legacy memcg will stall in page writeback so avoid forcibly
1832 if (sane_reclaim(sc)) {
1834 * Tag a zone as congested if all the dirty pages scanned were
1835 * backed by a congested BDI and wait_iff_congested will stall.
1837 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1838 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1841 * If dirty pages are scanned that are not queued for IO, it
1842 * implies that flushers are not doing their job. This can
1843 * happen when memory pressure pushes dirty pages to the end of
1844 * the LRU before the dirty limits are breached and the dirty
1845 * data has expired. It can also happen when the proportion of
1846 * dirty pages grows not through writes but through memory
1847 * pressure reclaiming all the clean cache. And in some cases,
1848 * the flushers simply cannot keep up with the allocation
1849 * rate. Nudge the flusher threads in case they are asleep, but
1850 * also allow kswapd to start writing pages during reclaim.
1852 if (stat.nr_unqueued_dirty == nr_taken) {
1853 wakeup_flusher_threads(WB_REASON_VMSCAN);
1854 set_bit(PGDAT_DIRTY, &pgdat->flags);
1858 * If kswapd scans pages marked marked for immediate
1859 * reclaim and under writeback (nr_immediate), it implies
1860 * that pages are cycling through the LRU faster than
1861 * they are written so also forcibly stall.
1863 if (stat.nr_immediate && current_may_throttle())
1864 congestion_wait(BLK_RW_ASYNC, HZ/10);
1868 * Stall direct reclaim for IO completions if underlying BDIs or zone
1869 * is congested. Allow kswapd to continue until it starts encountering
1870 * unqueued dirty pages or cycling through the LRU too quickly.
1872 if (!sc->hibernation_mode && !current_is_kswapd() &&
1873 current_may_throttle())
1874 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1876 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1877 nr_scanned, nr_reclaimed,
1878 stat.nr_dirty, stat.nr_writeback,
1879 stat.nr_congested, stat.nr_immediate,
1880 stat.nr_activate, stat.nr_ref_keep,
1882 sc->priority, file);
1883 return nr_reclaimed;
1887 * This moves pages from the active list to the inactive list.
1889 * We move them the other way if the page is referenced by one or more
1890 * processes, from rmap.
1892 * If the pages are mostly unmapped, the processing is fast and it is
1893 * appropriate to hold zone_lru_lock across the whole operation. But if
1894 * the pages are mapped, the processing is slow (page_referenced()) so we
1895 * should drop zone_lru_lock around each page. It's impossible to balance
1896 * this, so instead we remove the pages from the LRU while processing them.
1897 * It is safe to rely on PG_active against the non-LRU pages in here because
1898 * nobody will play with that bit on a non-LRU page.
1900 * The downside is that we have to touch page->_refcount against each page.
1901 * But we had to alter page->flags anyway.
1903 * Returns the number of pages moved to the given lru.
1906 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1907 struct list_head *list,
1908 struct list_head *pages_to_free,
1911 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1916 while (!list_empty(list)) {
1917 page = lru_to_page(list);
1918 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1920 VM_BUG_ON_PAGE(PageLRU(page), page);
1923 nr_pages = hpage_nr_pages(page);
1924 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1925 list_move(&page->lru, &lruvec->lists[lru]);
1927 if (put_page_testzero(page)) {
1928 __ClearPageLRU(page);
1929 __ClearPageActive(page);
1930 del_page_from_lru_list(page, lruvec, lru);
1932 if (unlikely(PageCompound(page))) {
1933 spin_unlock_irq(&pgdat->lru_lock);
1934 mem_cgroup_uncharge(page);
1935 (*get_compound_page_dtor(page))(page);
1936 spin_lock_irq(&pgdat->lru_lock);
1938 list_add(&page->lru, pages_to_free);
1940 nr_moved += nr_pages;
1944 if (!is_active_lru(lru)) {
1945 __count_vm_events(PGDEACTIVATE, nr_moved);
1946 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1953 static void shrink_active_list(unsigned long nr_to_scan,
1954 struct lruvec *lruvec,
1955 struct scan_control *sc,
1958 unsigned long nr_taken;
1959 unsigned long nr_scanned;
1960 unsigned long vm_flags;
1961 LIST_HEAD(l_hold); /* The pages which were snipped off */
1962 LIST_HEAD(l_active);
1963 LIST_HEAD(l_inactive);
1965 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1966 unsigned nr_deactivate, nr_activate;
1967 unsigned nr_rotated = 0;
1968 isolate_mode_t isolate_mode = 0;
1969 int file = is_file_lru(lru);
1970 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1975 isolate_mode |= ISOLATE_UNMAPPED;
1977 spin_lock_irq(&pgdat->lru_lock);
1979 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1980 &nr_scanned, sc, isolate_mode, lru);
1982 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983 reclaim_stat->recent_scanned[file] += nr_taken;
1985 __count_vm_events(PGREFILL, nr_scanned);
1986 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
1988 spin_unlock_irq(&pgdat->lru_lock);
1990 while (!list_empty(&l_hold)) {
1992 page = lru_to_page(&l_hold);
1993 list_del(&page->lru);
1995 if (unlikely(!page_evictable(page))) {
1996 putback_lru_page(page);
2000 if (unlikely(buffer_heads_over_limit)) {
2001 if (page_has_private(page) && trylock_page(page)) {
2002 if (page_has_private(page))
2003 try_to_release_page(page, 0);
2008 if (page_referenced(page, 0, sc->target_mem_cgroup,
2010 nr_rotated += hpage_nr_pages(page);
2012 * Identify referenced, file-backed active pages and
2013 * give them one more trip around the active list. So
2014 * that executable code get better chances to stay in
2015 * memory under moderate memory pressure. Anon pages
2016 * are not likely to be evicted by use-once streaming
2017 * IO, plus JVM can create lots of anon VM_EXEC pages,
2018 * so we ignore them here.
2020 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2021 list_add(&page->lru, &l_active);
2026 ClearPageActive(page); /* we are de-activating */
2027 list_add(&page->lru, &l_inactive);
2031 * Move pages back to the lru list.
2033 spin_lock_irq(&pgdat->lru_lock);
2035 * Count referenced pages from currently used mappings as rotated,
2036 * even though only some of them are actually re-activated. This
2037 * helps balance scan pressure between file and anonymous pages in
2040 reclaim_stat->recent_rotated[file] += nr_rotated;
2042 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2043 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2044 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2045 spin_unlock_irq(&pgdat->lru_lock);
2047 mem_cgroup_uncharge_list(&l_hold);
2048 free_unref_page_list(&l_hold);
2049 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2050 nr_deactivate, nr_rotated, sc->priority, file);
2054 * The inactive anon list should be small enough that the VM never has
2055 * to do too much work.
2057 * The inactive file list should be small enough to leave most memory
2058 * to the established workingset on the scan-resistant active list,
2059 * but large enough to avoid thrashing the aggregate readahead window.
2061 * Both inactive lists should also be large enough that each inactive
2062 * page has a chance to be referenced again before it is reclaimed.
2064 * If that fails and refaulting is observed, the inactive list grows.
2066 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2067 * on this LRU, maintained by the pageout code. An inactive_ratio
2068 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2071 * memory ratio inactive
2072 * -------------------------------------
2081 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2082 struct mem_cgroup *memcg,
2083 struct scan_control *sc, bool actual_reclaim)
2085 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2086 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2087 enum lru_list inactive_lru = file * LRU_FILE;
2088 unsigned long inactive, active;
2089 unsigned long inactive_ratio;
2090 unsigned long refaults;
2094 * If we don't have swap space, anonymous page deactivation
2097 if (!file && !total_swap_pages)
2100 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2101 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2104 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2106 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2109 * When refaults are being observed, it means a new workingset
2110 * is being established. Disable active list protection to get
2111 * rid of the stale workingset quickly.
2113 if (file && actual_reclaim && lruvec->refaults != refaults) {
2116 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2118 inactive_ratio = int_sqrt(10 * gb);
2124 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2125 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2126 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2127 inactive_ratio, file);
2129 return inactive * inactive_ratio < active;
2132 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2133 struct lruvec *lruvec, struct mem_cgroup *memcg,
2134 struct scan_control *sc)
2136 if (is_active_lru(lru)) {
2137 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2139 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2143 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2154 * Determine how aggressively the anon and file LRU lists should be
2155 * scanned. The relative value of each set of LRU lists is determined
2156 * by looking at the fraction of the pages scanned we did rotate back
2157 * onto the active list instead of evict.
2159 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2160 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2162 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2163 struct scan_control *sc, unsigned long *nr,
2164 unsigned long *lru_pages)
2166 int swappiness = mem_cgroup_swappiness(memcg);
2167 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2169 u64 denominator = 0; /* gcc */
2170 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2171 unsigned long anon_prio, file_prio;
2172 enum scan_balance scan_balance;
2173 unsigned long anon, file;
2174 unsigned long ap, fp;
2177 /* If we have no swap space, do not bother scanning anon pages. */
2178 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2179 scan_balance = SCAN_FILE;
2184 * Global reclaim will swap to prevent OOM even with no
2185 * swappiness, but memcg users want to use this knob to
2186 * disable swapping for individual groups completely when
2187 * using the memory controller's swap limit feature would be
2190 if (!global_reclaim(sc) && !swappiness) {
2191 scan_balance = SCAN_FILE;
2196 * Do not apply any pressure balancing cleverness when the
2197 * system is close to OOM, scan both anon and file equally
2198 * (unless the swappiness setting disagrees with swapping).
2200 if (!sc->priority && swappiness) {
2201 scan_balance = SCAN_EQUAL;
2206 * Prevent the reclaimer from falling into the cache trap: as
2207 * cache pages start out inactive, every cache fault will tip
2208 * the scan balance towards the file LRU. And as the file LRU
2209 * shrinks, so does the window for rotation from references.
2210 * This means we have a runaway feedback loop where a tiny
2211 * thrashing file LRU becomes infinitely more attractive than
2212 * anon pages. Try to detect this based on file LRU size.
2214 if (global_reclaim(sc)) {
2215 unsigned long pgdatfile;
2216 unsigned long pgdatfree;
2218 unsigned long total_high_wmark = 0;
2220 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2221 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2222 node_page_state(pgdat, NR_INACTIVE_FILE);
2224 for (z = 0; z < MAX_NR_ZONES; z++) {
2225 struct zone *zone = &pgdat->node_zones[z];
2226 if (!managed_zone(zone))
2229 total_high_wmark += high_wmark_pages(zone);
2232 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2234 * Force SCAN_ANON if there are enough inactive
2235 * anonymous pages on the LRU in eligible zones.
2236 * Otherwise, the small LRU gets thrashed.
2238 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2239 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2241 scan_balance = SCAN_ANON;
2248 * If there is enough inactive page cache, i.e. if the size of the
2249 * inactive list is greater than that of the active list *and* the
2250 * inactive list actually has some pages to scan on this priority, we
2251 * do not reclaim anything from the anonymous working set right now.
2252 * Without the second condition we could end up never scanning an
2253 * lruvec even if it has plenty of old anonymous pages unless the
2254 * system is under heavy pressure.
2256 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2257 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2258 scan_balance = SCAN_FILE;
2262 scan_balance = SCAN_FRACT;
2265 * With swappiness at 100, anonymous and file have the same priority.
2266 * This scanning priority is essentially the inverse of IO cost.
2268 anon_prio = swappiness;
2269 file_prio = 200 - anon_prio;
2272 * OK, so we have swap space and a fair amount of page cache
2273 * pages. We use the recently rotated / recently scanned
2274 * ratios to determine how valuable each cache is.
2276 * Because workloads change over time (and to avoid overflow)
2277 * we keep these statistics as a floating average, which ends
2278 * up weighing recent references more than old ones.
2280 * anon in [0], file in [1]
2283 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2284 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2285 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2286 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2288 spin_lock_irq(&pgdat->lru_lock);
2289 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2290 reclaim_stat->recent_scanned[0] /= 2;
2291 reclaim_stat->recent_rotated[0] /= 2;
2294 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2295 reclaim_stat->recent_scanned[1] /= 2;
2296 reclaim_stat->recent_rotated[1] /= 2;
2300 * The amount of pressure on anon vs file pages is inversely
2301 * proportional to the fraction of recently scanned pages on
2302 * each list that were recently referenced and in active use.
2304 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2305 ap /= reclaim_stat->recent_rotated[0] + 1;
2307 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2308 fp /= reclaim_stat->recent_rotated[1] + 1;
2309 spin_unlock_irq(&pgdat->lru_lock);
2313 denominator = ap + fp + 1;
2316 for_each_evictable_lru(lru) {
2317 int file = is_file_lru(lru);
2321 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2322 scan = size >> sc->priority;
2324 * If the cgroup's already been deleted, make sure to
2325 * scrape out the remaining cache.
2327 if (!scan && !mem_cgroup_online(memcg))
2328 scan = min(size, SWAP_CLUSTER_MAX);
2330 switch (scan_balance) {
2332 /* Scan lists relative to size */
2336 * Scan types proportional to swappiness and
2337 * their relative recent reclaim efficiency.
2339 scan = div64_u64(scan * fraction[file],
2344 /* Scan one type exclusively */
2345 if ((scan_balance == SCAN_FILE) != file) {
2351 /* Look ma, no brain */
2361 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2363 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2364 struct scan_control *sc, unsigned long *lru_pages)
2366 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2367 unsigned long nr[NR_LRU_LISTS];
2368 unsigned long targets[NR_LRU_LISTS];
2369 unsigned long nr_to_scan;
2371 unsigned long nr_reclaimed = 0;
2372 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2373 struct blk_plug plug;
2376 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2378 /* Record the original scan target for proportional adjustments later */
2379 memcpy(targets, nr, sizeof(nr));
2382 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2383 * event that can occur when there is little memory pressure e.g.
2384 * multiple streaming readers/writers. Hence, we do not abort scanning
2385 * when the requested number of pages are reclaimed when scanning at
2386 * DEF_PRIORITY on the assumption that the fact we are direct
2387 * reclaiming implies that kswapd is not keeping up and it is best to
2388 * do a batch of work at once. For memcg reclaim one check is made to
2389 * abort proportional reclaim if either the file or anon lru has already
2390 * dropped to zero at the first pass.
2392 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2393 sc->priority == DEF_PRIORITY);
2395 blk_start_plug(&plug);
2396 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2397 nr[LRU_INACTIVE_FILE]) {
2398 unsigned long nr_anon, nr_file, percentage;
2399 unsigned long nr_scanned;
2401 for_each_evictable_lru(lru) {
2403 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2404 nr[lru] -= nr_to_scan;
2406 nr_reclaimed += shrink_list(lru, nr_to_scan,
2413 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2417 * For kswapd and memcg, reclaim at least the number of pages
2418 * requested. Ensure that the anon and file LRUs are scanned
2419 * proportionally what was requested by get_scan_count(). We
2420 * stop reclaiming one LRU and reduce the amount scanning
2421 * proportional to the original scan target.
2423 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2424 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2427 * It's just vindictive to attack the larger once the smaller
2428 * has gone to zero. And given the way we stop scanning the
2429 * smaller below, this makes sure that we only make one nudge
2430 * towards proportionality once we've got nr_to_reclaim.
2432 if (!nr_file || !nr_anon)
2435 if (nr_file > nr_anon) {
2436 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2437 targets[LRU_ACTIVE_ANON] + 1;
2439 percentage = nr_anon * 100 / scan_target;
2441 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2442 targets[LRU_ACTIVE_FILE] + 1;
2444 percentage = nr_file * 100 / scan_target;
2447 /* Stop scanning the smaller of the LRU */
2449 nr[lru + LRU_ACTIVE] = 0;
2452 * Recalculate the other LRU scan count based on its original
2453 * scan target and the percentage scanning already complete
2455 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2456 nr_scanned = targets[lru] - nr[lru];
2457 nr[lru] = targets[lru] * (100 - percentage) / 100;
2458 nr[lru] -= min(nr[lru], nr_scanned);
2461 nr_scanned = targets[lru] - nr[lru];
2462 nr[lru] = targets[lru] * (100 - percentage) / 100;
2463 nr[lru] -= min(nr[lru], nr_scanned);
2465 scan_adjusted = true;
2467 blk_finish_plug(&plug);
2468 sc->nr_reclaimed += nr_reclaimed;
2471 * Even if we did not try to evict anon pages at all, we want to
2472 * rebalance the anon lru active/inactive ratio.
2474 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2475 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2476 sc, LRU_ACTIVE_ANON);
2479 /* Use reclaim/compaction for costly allocs or under memory pressure */
2480 static bool in_reclaim_compaction(struct scan_control *sc)
2482 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2483 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2484 sc->priority < DEF_PRIORITY - 2))
2491 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2492 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2493 * true if more pages should be reclaimed such that when the page allocator
2494 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2495 * It will give up earlier than that if there is difficulty reclaiming pages.
2497 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2498 unsigned long nr_reclaimed,
2499 unsigned long nr_scanned,
2500 struct scan_control *sc)
2502 unsigned long pages_for_compaction;
2503 unsigned long inactive_lru_pages;
2506 /* If not in reclaim/compaction mode, stop */
2507 if (!in_reclaim_compaction(sc))
2510 /* Consider stopping depending on scan and reclaim activity */
2511 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2513 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2514 * full LRU list has been scanned and we are still failing
2515 * to reclaim pages. This full LRU scan is potentially
2516 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2518 if (!nr_reclaimed && !nr_scanned)
2522 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2523 * fail without consequence, stop if we failed to reclaim
2524 * any pages from the last SWAP_CLUSTER_MAX number of
2525 * pages that were scanned. This will return to the
2526 * caller faster at the risk reclaim/compaction and
2527 * the resulting allocation attempt fails
2534 * If we have not reclaimed enough pages for compaction and the
2535 * inactive lists are large enough, continue reclaiming
2537 pages_for_compaction = compact_gap(sc->order);
2538 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2539 if (get_nr_swap_pages() > 0)
2540 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2541 if (sc->nr_reclaimed < pages_for_compaction &&
2542 inactive_lru_pages > pages_for_compaction)
2545 /* If compaction would go ahead or the allocation would succeed, stop */
2546 for (z = 0; z <= sc->reclaim_idx; z++) {
2547 struct zone *zone = &pgdat->node_zones[z];
2548 if (!managed_zone(zone))
2551 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2552 case COMPACT_SUCCESS:
2553 case COMPACT_CONTINUE:
2556 /* check next zone */
2563 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2565 struct reclaim_state *reclaim_state = current->reclaim_state;
2566 unsigned long nr_reclaimed, nr_scanned;
2567 bool reclaimable = false;
2570 struct mem_cgroup *root = sc->target_mem_cgroup;
2571 struct mem_cgroup_reclaim_cookie reclaim = {
2573 .priority = sc->priority,
2575 unsigned long node_lru_pages = 0;
2576 struct mem_cgroup *memcg;
2578 nr_reclaimed = sc->nr_reclaimed;
2579 nr_scanned = sc->nr_scanned;
2581 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2583 unsigned long lru_pages;
2584 unsigned long reclaimed;
2585 unsigned long scanned;
2587 if (mem_cgroup_low(root, memcg)) {
2588 if (!sc->memcg_low_reclaim) {
2589 sc->memcg_low_skipped = 1;
2592 mem_cgroup_event(memcg, MEMCG_LOW);
2595 reclaimed = sc->nr_reclaimed;
2596 scanned = sc->nr_scanned;
2597 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2598 node_lru_pages += lru_pages;
2601 shrink_slab(sc->gfp_mask, pgdat->node_id,
2602 memcg, sc->priority);
2604 /* Record the group's reclaim efficiency */
2605 vmpressure(sc->gfp_mask, memcg, false,
2606 sc->nr_scanned - scanned,
2607 sc->nr_reclaimed - reclaimed);
2610 * Direct reclaim and kswapd have to scan all memory
2611 * cgroups to fulfill the overall scan target for the
2614 * Limit reclaim, on the other hand, only cares about
2615 * nr_to_reclaim pages to be reclaimed and it will
2616 * retry with decreasing priority if one round over the
2617 * whole hierarchy is not sufficient.
2619 if (!global_reclaim(sc) &&
2620 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2621 mem_cgroup_iter_break(root, memcg);
2624 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2626 if (global_reclaim(sc))
2627 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2630 if (reclaim_state) {
2631 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2632 reclaim_state->reclaimed_slab = 0;
2635 /* Record the subtree's reclaim efficiency */
2636 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2637 sc->nr_scanned - nr_scanned,
2638 sc->nr_reclaimed - nr_reclaimed);
2640 if (sc->nr_reclaimed - nr_reclaimed)
2643 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2644 sc->nr_scanned - nr_scanned, sc));
2647 * Kswapd gives up on balancing particular nodes after too
2648 * many failures to reclaim anything from them and goes to
2649 * sleep. On reclaim progress, reset the failure counter. A
2650 * successful direct reclaim run will revive a dormant kswapd.
2653 pgdat->kswapd_failures = 0;
2659 * Returns true if compaction should go ahead for a costly-order request, or
2660 * the allocation would already succeed without compaction. Return false if we
2661 * should reclaim first.
2663 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2665 unsigned long watermark;
2666 enum compact_result suitable;
2668 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2669 if (suitable == COMPACT_SUCCESS)
2670 /* Allocation should succeed already. Don't reclaim. */
2672 if (suitable == COMPACT_SKIPPED)
2673 /* Compaction cannot yet proceed. Do reclaim. */
2677 * Compaction is already possible, but it takes time to run and there
2678 * are potentially other callers using the pages just freed. So proceed
2679 * with reclaim to make a buffer of free pages available to give
2680 * compaction a reasonable chance of completing and allocating the page.
2681 * Note that we won't actually reclaim the whole buffer in one attempt
2682 * as the target watermark in should_continue_reclaim() is lower. But if
2683 * we are already above the high+gap watermark, don't reclaim at all.
2685 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2687 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2691 * This is the direct reclaim path, for page-allocating processes. We only
2692 * try to reclaim pages from zones which will satisfy the caller's allocation
2695 * If a zone is deemed to be full of pinned pages then just give it a light
2696 * scan then give up on it.
2698 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2702 unsigned long nr_soft_reclaimed;
2703 unsigned long nr_soft_scanned;
2705 pg_data_t *last_pgdat = NULL;
2708 * If the number of buffer_heads in the machine exceeds the maximum
2709 * allowed level, force direct reclaim to scan the highmem zone as
2710 * highmem pages could be pinning lowmem pages storing buffer_heads
2712 orig_mask = sc->gfp_mask;
2713 if (buffer_heads_over_limit) {
2714 sc->gfp_mask |= __GFP_HIGHMEM;
2715 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2718 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2719 sc->reclaim_idx, sc->nodemask) {
2721 * Take care memory controller reclaiming has small influence
2724 if (global_reclaim(sc)) {
2725 if (!cpuset_zone_allowed(zone,
2726 GFP_KERNEL | __GFP_HARDWALL))
2730 * If we already have plenty of memory free for
2731 * compaction in this zone, don't free any more.
2732 * Even though compaction is invoked for any
2733 * non-zero order, only frequent costly order
2734 * reclamation is disruptive enough to become a
2735 * noticeable problem, like transparent huge
2738 if (IS_ENABLED(CONFIG_COMPACTION) &&
2739 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2740 compaction_ready(zone, sc)) {
2741 sc->compaction_ready = true;
2746 * Shrink each node in the zonelist once. If the
2747 * zonelist is ordered by zone (not the default) then a
2748 * node may be shrunk multiple times but in that case
2749 * the user prefers lower zones being preserved.
2751 if (zone->zone_pgdat == last_pgdat)
2755 * This steals pages from memory cgroups over softlimit
2756 * and returns the number of reclaimed pages and
2757 * scanned pages. This works for global memory pressure
2758 * and balancing, not for a memcg's limit.
2760 nr_soft_scanned = 0;
2761 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2762 sc->order, sc->gfp_mask,
2764 sc->nr_reclaimed += nr_soft_reclaimed;
2765 sc->nr_scanned += nr_soft_scanned;
2766 /* need some check for avoid more shrink_zone() */
2769 /* See comment about same check for global reclaim above */
2770 if (zone->zone_pgdat == last_pgdat)
2772 last_pgdat = zone->zone_pgdat;
2773 shrink_node(zone->zone_pgdat, sc);
2777 * Restore to original mask to avoid the impact on the caller if we
2778 * promoted it to __GFP_HIGHMEM.
2780 sc->gfp_mask = orig_mask;
2783 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2785 struct mem_cgroup *memcg;
2787 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2789 unsigned long refaults;
2790 struct lruvec *lruvec;
2793 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2795 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2797 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2798 lruvec->refaults = refaults;
2799 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2803 * This is the main entry point to direct page reclaim.
2805 * If a full scan of the inactive list fails to free enough memory then we
2806 * are "out of memory" and something needs to be killed.
2808 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2809 * high - the zone may be full of dirty or under-writeback pages, which this
2810 * caller can't do much about. We kick the writeback threads and take explicit
2811 * naps in the hope that some of these pages can be written. But if the
2812 * allocating task holds filesystem locks which prevent writeout this might not
2813 * work, and the allocation attempt will fail.
2815 * returns: 0, if no pages reclaimed
2816 * else, the number of pages reclaimed
2818 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2819 struct scan_control *sc)
2821 int initial_priority = sc->priority;
2822 pg_data_t *last_pgdat;
2826 delayacct_freepages_start();
2828 if (global_reclaim(sc))
2829 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2832 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2835 shrink_zones(zonelist, sc);
2837 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2840 if (sc->compaction_ready)
2844 * If we're getting trouble reclaiming, start doing
2845 * writepage even in laptop mode.
2847 if (sc->priority < DEF_PRIORITY - 2)
2848 sc->may_writepage = 1;
2849 } while (--sc->priority >= 0);
2852 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2854 if (zone->zone_pgdat == last_pgdat)
2856 last_pgdat = zone->zone_pgdat;
2857 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2860 delayacct_freepages_end();
2862 if (sc->nr_reclaimed)
2863 return sc->nr_reclaimed;
2865 /* Aborted reclaim to try compaction? don't OOM, then */
2866 if (sc->compaction_ready)
2869 /* Untapped cgroup reserves? Don't OOM, retry. */
2870 if (sc->memcg_low_skipped) {
2871 sc->priority = initial_priority;
2872 sc->memcg_low_reclaim = 1;
2873 sc->memcg_low_skipped = 0;
2880 static bool allow_direct_reclaim(pg_data_t *pgdat)
2883 unsigned long pfmemalloc_reserve = 0;
2884 unsigned long free_pages = 0;
2888 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2891 for (i = 0; i <= ZONE_NORMAL; i++) {
2892 zone = &pgdat->node_zones[i];
2893 if (!managed_zone(zone))
2896 if (!zone_reclaimable_pages(zone))
2899 pfmemalloc_reserve += min_wmark_pages(zone);
2900 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2903 /* If there are no reserves (unexpected config) then do not throttle */
2904 if (!pfmemalloc_reserve)
2907 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2909 /* kswapd must be awake if processes are being throttled */
2910 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2911 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2912 (enum zone_type)ZONE_NORMAL);
2913 wake_up_interruptible(&pgdat->kswapd_wait);
2920 * Throttle direct reclaimers if backing storage is backed by the network
2921 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2922 * depleted. kswapd will continue to make progress and wake the processes
2923 * when the low watermark is reached.
2925 * Returns true if a fatal signal was delivered during throttling. If this
2926 * happens, the page allocator should not consider triggering the OOM killer.
2928 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2929 nodemask_t *nodemask)
2933 pg_data_t *pgdat = NULL;
2936 * Kernel threads should not be throttled as they may be indirectly
2937 * responsible for cleaning pages necessary for reclaim to make forward
2938 * progress. kjournald for example may enter direct reclaim while
2939 * committing a transaction where throttling it could forcing other
2940 * processes to block on log_wait_commit().
2942 if (current->flags & PF_KTHREAD)
2946 * If a fatal signal is pending, this process should not throttle.
2947 * It should return quickly so it can exit and free its memory
2949 if (fatal_signal_pending(current))
2953 * Check if the pfmemalloc reserves are ok by finding the first node
2954 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2955 * GFP_KERNEL will be required for allocating network buffers when
2956 * swapping over the network so ZONE_HIGHMEM is unusable.
2958 * Throttling is based on the first usable node and throttled processes
2959 * wait on a queue until kswapd makes progress and wakes them. There
2960 * is an affinity then between processes waking up and where reclaim
2961 * progress has been made assuming the process wakes on the same node.
2962 * More importantly, processes running on remote nodes will not compete
2963 * for remote pfmemalloc reserves and processes on different nodes
2964 * should make reasonable progress.
2966 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2967 gfp_zone(gfp_mask), nodemask) {
2968 if (zone_idx(zone) > ZONE_NORMAL)
2971 /* Throttle based on the first usable node */
2972 pgdat = zone->zone_pgdat;
2973 if (allow_direct_reclaim(pgdat))
2978 /* If no zone was usable by the allocation flags then do not throttle */
2982 /* Account for the throttling */
2983 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2986 * If the caller cannot enter the filesystem, it's possible that it
2987 * is due to the caller holding an FS lock or performing a journal
2988 * transaction in the case of a filesystem like ext[3|4]. In this case,
2989 * it is not safe to block on pfmemalloc_wait as kswapd could be
2990 * blocked waiting on the same lock. Instead, throttle for up to a
2991 * second before continuing.
2993 if (!(gfp_mask & __GFP_FS)) {
2994 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2995 allow_direct_reclaim(pgdat), HZ);
3000 /* Throttle until kswapd wakes the process */
3001 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3002 allow_direct_reclaim(pgdat));
3005 if (fatal_signal_pending(current))
3012 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3013 gfp_t gfp_mask, nodemask_t *nodemask)
3015 unsigned long nr_reclaimed;
3016 struct scan_control sc = {
3017 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3018 .gfp_mask = current_gfp_context(gfp_mask),
3019 .reclaim_idx = gfp_zone(gfp_mask),
3021 .nodemask = nodemask,
3022 .priority = DEF_PRIORITY,
3023 .may_writepage = !laptop_mode,
3029 * Do not enter reclaim if fatal signal was delivered while throttled.
3030 * 1 is returned so that the page allocator does not OOM kill at this
3033 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3036 trace_mm_vmscan_direct_reclaim_begin(order,
3041 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3043 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3045 return nr_reclaimed;
3050 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3051 gfp_t gfp_mask, bool noswap,
3053 unsigned long *nr_scanned)
3055 struct scan_control sc = {
3056 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3057 .target_mem_cgroup = memcg,
3058 .may_writepage = !laptop_mode,
3060 .reclaim_idx = MAX_NR_ZONES - 1,
3061 .may_swap = !noswap,
3063 unsigned long lru_pages;
3065 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3066 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3068 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3074 * NOTE: Although we can get the priority field, using it
3075 * here is not a good idea, since it limits the pages we can scan.
3076 * if we don't reclaim here, the shrink_node from balance_pgdat
3077 * will pick up pages from other mem cgroup's as well. We hack
3078 * the priority and make it zero.
3080 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3082 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3084 *nr_scanned = sc.nr_scanned;
3085 return sc.nr_reclaimed;
3088 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3089 unsigned long nr_pages,
3093 struct zonelist *zonelist;
3094 unsigned long nr_reclaimed;
3096 unsigned int noreclaim_flag;
3097 struct scan_control sc = {
3098 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3099 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3100 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3101 .reclaim_idx = MAX_NR_ZONES - 1,
3102 .target_mem_cgroup = memcg,
3103 .priority = DEF_PRIORITY,
3104 .may_writepage = !laptop_mode,
3106 .may_swap = may_swap,
3110 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3111 * take care of from where we get pages. So the node where we start the
3112 * scan does not need to be the current node.
3114 nid = mem_cgroup_select_victim_node(memcg);
3116 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3118 trace_mm_vmscan_memcg_reclaim_begin(0,
3123 noreclaim_flag = memalloc_noreclaim_save();
3124 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3125 memalloc_noreclaim_restore(noreclaim_flag);
3127 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3129 return nr_reclaimed;
3133 static void age_active_anon(struct pglist_data *pgdat,
3134 struct scan_control *sc)
3136 struct mem_cgroup *memcg;
3138 if (!total_swap_pages)
3141 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3143 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3145 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3146 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3147 sc, LRU_ACTIVE_ANON);
3149 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3154 * Returns true if there is an eligible zone balanced for the request order
3157 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3160 unsigned long mark = -1;
3163 for (i = 0; i <= classzone_idx; i++) {
3164 zone = pgdat->node_zones + i;
3166 if (!managed_zone(zone))
3169 mark = high_wmark_pages(zone);
3170 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3175 * If a node has no populated zone within classzone_idx, it does not
3176 * need balancing by definition. This can happen if a zone-restricted
3177 * allocation tries to wake a remote kswapd.
3185 /* Clear pgdat state for congested, dirty or under writeback. */
3186 static void clear_pgdat_congested(pg_data_t *pgdat)
3188 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3189 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3190 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3194 * Prepare kswapd for sleeping. This verifies that there are no processes
3195 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3197 * Returns true if kswapd is ready to sleep
3199 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3202 * The throttled processes are normally woken up in balance_pgdat() as
3203 * soon as allow_direct_reclaim() is true. But there is a potential
3204 * race between when kswapd checks the watermarks and a process gets
3205 * throttled. There is also a potential race if processes get
3206 * throttled, kswapd wakes, a large process exits thereby balancing the
3207 * zones, which causes kswapd to exit balance_pgdat() before reaching
3208 * the wake up checks. If kswapd is going to sleep, no process should
3209 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3210 * the wake up is premature, processes will wake kswapd and get
3211 * throttled again. The difference from wake ups in balance_pgdat() is
3212 * that here we are under prepare_to_wait().
3214 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3215 wake_up_all(&pgdat->pfmemalloc_wait);
3217 /* Hopeless node, leave it to direct reclaim */
3218 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3221 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3222 clear_pgdat_congested(pgdat);
3230 * kswapd shrinks a node of pages that are at or below the highest usable
3231 * zone that is currently unbalanced.
3233 * Returns true if kswapd scanned at least the requested number of pages to
3234 * reclaim or if the lack of progress was due to pages under writeback.
3235 * This is used to determine if the scanning priority needs to be raised.
3237 static bool kswapd_shrink_node(pg_data_t *pgdat,
3238 struct scan_control *sc)
3243 /* Reclaim a number of pages proportional to the number of zones */
3244 sc->nr_to_reclaim = 0;
3245 for (z = 0; z <= sc->reclaim_idx; z++) {
3246 zone = pgdat->node_zones + z;
3247 if (!managed_zone(zone))
3250 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3254 * Historically care was taken to put equal pressure on all zones but
3255 * now pressure is applied based on node LRU order.
3257 shrink_node(pgdat, sc);
3260 * Fragmentation may mean that the system cannot be rebalanced for
3261 * high-order allocations. If twice the allocation size has been
3262 * reclaimed then recheck watermarks only at order-0 to prevent
3263 * excessive reclaim. Assume that a process requested a high-order
3264 * can direct reclaim/compact.
3266 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3269 return sc->nr_scanned >= sc->nr_to_reclaim;
3273 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3274 * that are eligible for use by the caller until at least one zone is
3277 * Returns the order kswapd finished reclaiming at.
3279 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3280 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3281 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3282 * or lower is eligible for reclaim until at least one usable zone is
3285 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3288 unsigned long nr_soft_reclaimed;
3289 unsigned long nr_soft_scanned;
3291 struct scan_control sc = {
3292 .gfp_mask = GFP_KERNEL,
3294 .priority = DEF_PRIORITY,
3295 .may_writepage = !laptop_mode,
3299 count_vm_event(PAGEOUTRUN);
3302 unsigned long nr_reclaimed = sc.nr_reclaimed;
3303 bool raise_priority = true;
3305 sc.reclaim_idx = classzone_idx;
3308 * If the number of buffer_heads exceeds the maximum allowed
3309 * then consider reclaiming from all zones. This has a dual
3310 * purpose -- on 64-bit systems it is expected that
3311 * buffer_heads are stripped during active rotation. On 32-bit
3312 * systems, highmem pages can pin lowmem memory and shrinking
3313 * buffers can relieve lowmem pressure. Reclaim may still not
3314 * go ahead if all eligible zones for the original allocation
3315 * request are balanced to avoid excessive reclaim from kswapd.
3317 if (buffer_heads_over_limit) {
3318 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3319 zone = pgdat->node_zones + i;
3320 if (!managed_zone(zone))
3329 * Only reclaim if there are no eligible zones. Note that
3330 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3333 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3337 * Do some background aging of the anon list, to give
3338 * pages a chance to be referenced before reclaiming. All
3339 * pages are rotated regardless of classzone as this is
3340 * about consistent aging.
3342 age_active_anon(pgdat, &sc);
3345 * If we're getting trouble reclaiming, start doing writepage
3346 * even in laptop mode.
3348 if (sc.priority < DEF_PRIORITY - 2)
3349 sc.may_writepage = 1;
3351 /* Call soft limit reclaim before calling shrink_node. */
3353 nr_soft_scanned = 0;
3354 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3355 sc.gfp_mask, &nr_soft_scanned);
3356 sc.nr_reclaimed += nr_soft_reclaimed;
3359 * There should be no need to raise the scanning priority if
3360 * enough pages are already being scanned that that high
3361 * watermark would be met at 100% efficiency.
3363 if (kswapd_shrink_node(pgdat, &sc))
3364 raise_priority = false;
3367 * If the low watermark is met there is no need for processes
3368 * to be throttled on pfmemalloc_wait as they should not be
3369 * able to safely make forward progress. Wake them
3371 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3372 allow_direct_reclaim(pgdat))
3373 wake_up_all(&pgdat->pfmemalloc_wait);
3375 /* Check if kswapd should be suspending */
3376 if (try_to_freeze() || kthread_should_stop())
3380 * Raise priority if scanning rate is too low or there was no
3381 * progress in reclaiming pages
3383 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3384 if (raise_priority || !nr_reclaimed)
3386 } while (sc.priority >= 1);
3388 if (!sc.nr_reclaimed)
3389 pgdat->kswapd_failures++;
3392 snapshot_refaults(NULL, pgdat);
3394 * Return the order kswapd stopped reclaiming at as
3395 * prepare_kswapd_sleep() takes it into account. If another caller
3396 * entered the allocator slow path while kswapd was awake, order will
3397 * remain at the higher level.
3403 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3404 * allocation request woke kswapd for. When kswapd has not woken recently,
3405 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3406 * given classzone and returns it or the highest classzone index kswapd
3407 * was recently woke for.
3409 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3410 enum zone_type classzone_idx)
3412 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3413 return classzone_idx;
3415 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3418 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3419 unsigned int classzone_idx)
3424 if (freezing(current) || kthread_should_stop())
3427 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3430 * Try to sleep for a short interval. Note that kcompactd will only be
3431 * woken if it is possible to sleep for a short interval. This is
3432 * deliberate on the assumption that if reclaim cannot keep an
3433 * eligible zone balanced that it's also unlikely that compaction will
3436 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3438 * Compaction records what page blocks it recently failed to
3439 * isolate pages from and skips them in the future scanning.
3440 * When kswapd is going to sleep, it is reasonable to assume
3441 * that pages and compaction may succeed so reset the cache.
3443 reset_isolation_suitable(pgdat);
3446 * We have freed the memory, now we should compact it to make
3447 * allocation of the requested order possible.
3449 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3451 remaining = schedule_timeout(HZ/10);
3454 * If woken prematurely then reset kswapd_classzone_idx and
3455 * order. The values will either be from a wakeup request or
3456 * the previous request that slept prematurely.
3459 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3460 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3463 finish_wait(&pgdat->kswapd_wait, &wait);
3464 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3468 * After a short sleep, check if it was a premature sleep. If not, then
3469 * go fully to sleep until explicitly woken up.
3472 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3473 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3476 * vmstat counters are not perfectly accurate and the estimated
3477 * value for counters such as NR_FREE_PAGES can deviate from the
3478 * true value by nr_online_cpus * threshold. To avoid the zone
3479 * watermarks being breached while under pressure, we reduce the
3480 * per-cpu vmstat threshold while kswapd is awake and restore
3481 * them before going back to sleep.
3483 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3485 if (!kthread_should_stop())
3488 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3491 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3493 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3495 finish_wait(&pgdat->kswapd_wait, &wait);
3499 * The background pageout daemon, started as a kernel thread
3500 * from the init process.
3502 * This basically trickles out pages so that we have _some_
3503 * free memory available even if there is no other activity
3504 * that frees anything up. This is needed for things like routing
3505 * etc, where we otherwise might have all activity going on in
3506 * asynchronous contexts that cannot page things out.
3508 * If there are applications that are active memory-allocators
3509 * (most normal use), this basically shouldn't matter.
3511 static int kswapd(void *p)
3513 unsigned int alloc_order, reclaim_order;
3514 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3515 pg_data_t *pgdat = (pg_data_t*)p;
3516 struct task_struct *tsk = current;
3518 struct reclaim_state reclaim_state = {
3519 .reclaimed_slab = 0,
3521 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3523 if (!cpumask_empty(cpumask))
3524 set_cpus_allowed_ptr(tsk, cpumask);
3525 current->reclaim_state = &reclaim_state;
3528 * Tell the memory management that we're a "memory allocator",
3529 * and that if we need more memory we should get access to it
3530 * regardless (see "__alloc_pages()"). "kswapd" should
3531 * never get caught in the normal page freeing logic.
3533 * (Kswapd normally doesn't need memory anyway, but sometimes
3534 * you need a small amount of memory in order to be able to
3535 * page out something else, and this flag essentially protects
3536 * us from recursively trying to free more memory as we're
3537 * trying to free the first piece of memory in the first place).
3539 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3542 pgdat->kswapd_order = 0;
3543 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3547 alloc_order = reclaim_order = pgdat->kswapd_order;
3548 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3551 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3554 /* Read the new order and classzone_idx */
3555 alloc_order = reclaim_order = pgdat->kswapd_order;
3556 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3557 pgdat->kswapd_order = 0;
3558 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3560 ret = try_to_freeze();
3561 if (kthread_should_stop())
3565 * We can speed up thawing tasks if we don't call balance_pgdat
3566 * after returning from the refrigerator
3572 * Reclaim begins at the requested order but if a high-order
3573 * reclaim fails then kswapd falls back to reclaiming for
3574 * order-0. If that happens, kswapd will consider sleeping
3575 * for the order it finished reclaiming at (reclaim_order)
3576 * but kcompactd is woken to compact for the original
3577 * request (alloc_order).
3579 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3581 fs_reclaim_acquire(GFP_KERNEL);
3582 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3583 fs_reclaim_release(GFP_KERNEL);
3584 if (reclaim_order < alloc_order)
3585 goto kswapd_try_sleep;
3588 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3589 current->reclaim_state = NULL;
3595 * A zone is low on free memory, so wake its kswapd task to service it.
3597 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3601 if (!managed_zone(zone))
3604 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3606 pgdat = zone->zone_pgdat;
3607 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3609 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3610 if (!waitqueue_active(&pgdat->kswapd_wait))
3613 /* Hopeless node, leave it to direct reclaim */
3614 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3617 if (pgdat_balanced(pgdat, order, classzone_idx))
3620 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
3621 wake_up_interruptible(&pgdat->kswapd_wait);
3624 #ifdef CONFIG_HIBERNATION
3626 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3629 * Rather than trying to age LRUs the aim is to preserve the overall
3630 * LRU order by reclaiming preferentially
3631 * inactive > active > active referenced > active mapped
3633 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3635 struct reclaim_state reclaim_state;
3636 struct scan_control sc = {
3637 .nr_to_reclaim = nr_to_reclaim,
3638 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3639 .reclaim_idx = MAX_NR_ZONES - 1,
3640 .priority = DEF_PRIORITY,
3644 .hibernation_mode = 1,
3646 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3647 struct task_struct *p = current;
3648 unsigned long nr_reclaimed;
3649 unsigned int noreclaim_flag;
3651 noreclaim_flag = memalloc_noreclaim_save();
3652 fs_reclaim_acquire(sc.gfp_mask);
3653 reclaim_state.reclaimed_slab = 0;
3654 p->reclaim_state = &reclaim_state;
3656 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3658 p->reclaim_state = NULL;
3659 fs_reclaim_release(sc.gfp_mask);
3660 memalloc_noreclaim_restore(noreclaim_flag);
3662 return nr_reclaimed;
3664 #endif /* CONFIG_HIBERNATION */
3666 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3667 not required for correctness. So if the last cpu in a node goes
3668 away, we get changed to run anywhere: as the first one comes back,
3669 restore their cpu bindings. */
3670 static int kswapd_cpu_online(unsigned int cpu)
3674 for_each_node_state(nid, N_MEMORY) {
3675 pg_data_t *pgdat = NODE_DATA(nid);
3676 const struct cpumask *mask;
3678 mask = cpumask_of_node(pgdat->node_id);
3680 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3681 /* One of our CPUs online: restore mask */
3682 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3688 * This kswapd start function will be called by init and node-hot-add.
3689 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3691 int kswapd_run(int nid)
3693 pg_data_t *pgdat = NODE_DATA(nid);
3699 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3700 if (IS_ERR(pgdat->kswapd)) {
3701 /* failure at boot is fatal */
3702 BUG_ON(system_state < SYSTEM_RUNNING);
3703 pr_err("Failed to start kswapd on node %d\n", nid);
3704 ret = PTR_ERR(pgdat->kswapd);
3705 pgdat->kswapd = NULL;
3711 * Called by memory hotplug when all memory in a node is offlined. Caller must
3712 * hold mem_hotplug_begin/end().
3714 void kswapd_stop(int nid)
3716 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3719 kthread_stop(kswapd);
3720 NODE_DATA(nid)->kswapd = NULL;
3724 static int __init kswapd_init(void)
3729 for_each_node_state(nid, N_MEMORY)
3731 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3732 "mm/vmscan:online", kswapd_cpu_online,
3738 module_init(kswapd_init)
3744 * If non-zero call node_reclaim when the number of free pages falls below
3747 int node_reclaim_mode __read_mostly;
3749 #define RECLAIM_OFF 0
3750 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3751 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3752 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3755 * Priority for NODE_RECLAIM. This determines the fraction of pages
3756 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3759 #define NODE_RECLAIM_PRIORITY 4
3762 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3765 int sysctl_min_unmapped_ratio = 1;
3768 * If the number of slab pages in a zone grows beyond this percentage then
3769 * slab reclaim needs to occur.
3771 int sysctl_min_slab_ratio = 5;
3773 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3775 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3776 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3777 node_page_state(pgdat, NR_ACTIVE_FILE);
3780 * It's possible for there to be more file mapped pages than
3781 * accounted for by the pages on the file LRU lists because
3782 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3784 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3787 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3788 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3790 unsigned long nr_pagecache_reclaimable;
3791 unsigned long delta = 0;
3794 * If RECLAIM_UNMAP is set, then all file pages are considered
3795 * potentially reclaimable. Otherwise, we have to worry about
3796 * pages like swapcache and node_unmapped_file_pages() provides
3799 if (node_reclaim_mode & RECLAIM_UNMAP)
3800 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3802 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3804 /* If we can't clean pages, remove dirty pages from consideration */
3805 if (!(node_reclaim_mode & RECLAIM_WRITE))
3806 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3808 /* Watch for any possible underflows due to delta */
3809 if (unlikely(delta > nr_pagecache_reclaimable))
3810 delta = nr_pagecache_reclaimable;
3812 return nr_pagecache_reclaimable - delta;
3816 * Try to free up some pages from this node through reclaim.
3818 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3820 /* Minimum pages needed in order to stay on node */
3821 const unsigned long nr_pages = 1 << order;
3822 struct task_struct *p = current;
3823 struct reclaim_state reclaim_state;
3824 unsigned int noreclaim_flag;
3825 struct scan_control sc = {
3826 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3827 .gfp_mask = current_gfp_context(gfp_mask),
3829 .priority = NODE_RECLAIM_PRIORITY,
3830 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3831 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3833 .reclaim_idx = gfp_zone(gfp_mask),
3838 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3839 * and we also need to be able to write out pages for RECLAIM_WRITE
3840 * and RECLAIM_UNMAP.
3842 noreclaim_flag = memalloc_noreclaim_save();
3843 p->flags |= PF_SWAPWRITE;
3844 fs_reclaim_acquire(sc.gfp_mask);
3845 reclaim_state.reclaimed_slab = 0;
3846 p->reclaim_state = &reclaim_state;
3848 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3850 * Free memory by calling shrink zone with increasing
3851 * priorities until we have enough memory freed.
3854 shrink_node(pgdat, &sc);
3855 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3858 p->reclaim_state = NULL;
3859 fs_reclaim_release(gfp_mask);
3860 current->flags &= ~PF_SWAPWRITE;
3861 memalloc_noreclaim_restore(noreclaim_flag);
3862 return sc.nr_reclaimed >= nr_pages;
3865 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3870 * Node reclaim reclaims unmapped file backed pages and
3871 * slab pages if we are over the defined limits.
3873 * A small portion of unmapped file backed pages is needed for
3874 * file I/O otherwise pages read by file I/O will be immediately
3875 * thrown out if the node is overallocated. So we do not reclaim
3876 * if less than a specified percentage of the node is used by
3877 * unmapped file backed pages.
3879 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3880 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3881 return NODE_RECLAIM_FULL;
3884 * Do not scan if the allocation should not be delayed.
3886 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3887 return NODE_RECLAIM_NOSCAN;
3890 * Only run node reclaim on the local node or on nodes that do not
3891 * have associated processors. This will favor the local processor
3892 * over remote processors and spread off node memory allocations
3893 * as wide as possible.
3895 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3896 return NODE_RECLAIM_NOSCAN;
3898 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3899 return NODE_RECLAIM_NOSCAN;
3901 ret = __node_reclaim(pgdat, gfp_mask, order);
3902 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3905 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3912 * page_evictable - test whether a page is evictable
3913 * @page: the page to test
3915 * Test whether page is evictable--i.e., should be placed on active/inactive
3916 * lists vs unevictable list.
3918 * Reasons page might not be evictable:
3919 * (1) page's mapping marked unevictable
3920 * (2) page is part of an mlocked VMA
3923 int page_evictable(struct page *page)
3925 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3930 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3931 * @pages: array of pages to check
3932 * @nr_pages: number of pages to check
3934 * Checks pages for evictability and moves them to the appropriate lru list.
3936 * This function is only used for SysV IPC SHM_UNLOCK.
3938 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3940 struct lruvec *lruvec;
3941 struct pglist_data *pgdat = NULL;
3946 for (i = 0; i < nr_pages; i++) {
3947 struct page *page = pages[i];
3948 struct pglist_data *pagepgdat = page_pgdat(page);
3951 if (pagepgdat != pgdat) {
3953 spin_unlock_irq(&pgdat->lru_lock);
3955 spin_lock_irq(&pgdat->lru_lock);
3957 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3959 if (!PageLRU(page) || !PageUnevictable(page))
3962 if (page_evictable(page)) {
3963 enum lru_list lru = page_lru_base_type(page);
3965 VM_BUG_ON_PAGE(PageActive(page), page);
3966 ClearPageUnevictable(page);
3967 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3968 add_page_to_lru_list(page, lruvec, lru);
3974 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3975 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3976 spin_unlock_irq(&pgdat->lru_lock);
3979 #endif /* CONFIG_SHMEM */