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;
1421 * Only pages without mappings or that have a
1422 * ->migratepage callback are possible to migrate
1423 * without blocking. However, we can be racing with
1424 * truncation so it's necessary to lock the page
1425 * to stabilise the mapping as truncation holds
1426 * the page lock until after the page is removed
1427 * from the page cache.
1429 if (!trylock_page(page))
1432 mapping = page_mapping(page);
1433 migrate_dirty = mapping && mapping->a_ops->migratepage;
1440 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1443 if (likely(get_page_unless_zero(page))) {
1445 * Be careful not to clear PageLRU until after we're
1446 * sure the page is not being freed elsewhere -- the
1447 * page release code relies on it.
1458 * Update LRU sizes after isolating pages. The LRU size updates must
1459 * be complete before mem_cgroup_update_lru_size due to a santity check.
1461 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1462 enum lru_list lru, unsigned long *nr_zone_taken)
1466 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1467 if (!nr_zone_taken[zid])
1470 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1472 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1479 * zone_lru_lock is heavily contended. Some of the functions that
1480 * shrink the lists perform better by taking out a batch of pages
1481 * and working on them outside the LRU lock.
1483 * For pagecache intensive workloads, this function is the hottest
1484 * spot in the kernel (apart from copy_*_user functions).
1486 * Appropriate locks must be held before calling this function.
1488 * @nr_to_scan: The number of eligible pages to look through on the list.
1489 * @lruvec: The LRU vector to pull pages from.
1490 * @dst: The temp list to put pages on to.
1491 * @nr_scanned: The number of pages that were scanned.
1492 * @sc: The scan_control struct for this reclaim session
1493 * @mode: One of the LRU isolation modes
1494 * @lru: LRU list id for isolating
1496 * returns how many pages were moved onto *@dst.
1498 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1499 struct lruvec *lruvec, struct list_head *dst,
1500 unsigned long *nr_scanned, struct scan_control *sc,
1501 isolate_mode_t mode, enum lru_list lru)
1503 struct list_head *src = &lruvec->lists[lru];
1504 unsigned long nr_taken = 0;
1505 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1506 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1507 unsigned long skipped = 0;
1508 unsigned long scan, total_scan, nr_pages;
1509 LIST_HEAD(pages_skipped);
1512 for (total_scan = 0;
1513 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1517 page = lru_to_page(src);
1518 prefetchw_prev_lru_page(page, src, flags);
1520 VM_BUG_ON_PAGE(!PageLRU(page), page);
1522 if (page_zonenum(page) > sc->reclaim_idx) {
1523 list_move(&page->lru, &pages_skipped);
1524 nr_skipped[page_zonenum(page)]++;
1529 * Do not count skipped pages because that makes the function
1530 * return with no isolated pages if the LRU mostly contains
1531 * ineligible pages. This causes the VM to not reclaim any
1532 * pages, triggering a premature OOM.
1535 switch (__isolate_lru_page(page, mode)) {
1537 nr_pages = hpage_nr_pages(page);
1538 nr_taken += nr_pages;
1539 nr_zone_taken[page_zonenum(page)] += nr_pages;
1540 list_move(&page->lru, dst);
1544 /* else it is being freed elsewhere */
1545 list_move(&page->lru, src);
1554 * Splice any skipped pages to the start of the LRU list. Note that
1555 * this disrupts the LRU order when reclaiming for lower zones but
1556 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1557 * scanning would soon rescan the same pages to skip and put the
1558 * system at risk of premature OOM.
1560 if (!list_empty(&pages_skipped)) {
1563 list_splice(&pages_skipped, src);
1564 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1565 if (!nr_skipped[zid])
1568 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1569 skipped += nr_skipped[zid];
1572 *nr_scanned = total_scan;
1573 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1574 total_scan, skipped, nr_taken, mode, lru);
1575 update_lru_sizes(lruvec, lru, nr_zone_taken);
1580 * isolate_lru_page - tries to isolate a page from its LRU list
1581 * @page: page to isolate from its LRU list
1583 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1584 * vmstat statistic corresponding to whatever LRU list the page was on.
1586 * Returns 0 if the page was removed from an LRU list.
1587 * Returns -EBUSY if the page was not on an LRU list.
1589 * The returned page will have PageLRU() cleared. If it was found on
1590 * the active list, it will have PageActive set. If it was found on
1591 * the unevictable list, it will have the PageUnevictable bit set. That flag
1592 * may need to be cleared by the caller before letting the page go.
1594 * The vmstat statistic corresponding to the list on which the page was
1595 * found will be decremented.
1598 * (1) Must be called with an elevated refcount on the page. This is a
1599 * fundamentnal difference from isolate_lru_pages (which is called
1600 * without a stable reference).
1601 * (2) the lru_lock must not be held.
1602 * (3) interrupts must be enabled.
1604 int isolate_lru_page(struct page *page)
1608 VM_BUG_ON_PAGE(!page_count(page), page);
1609 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1611 if (PageLRU(page)) {
1612 struct zone *zone = page_zone(page);
1613 struct lruvec *lruvec;
1615 spin_lock_irq(zone_lru_lock(zone));
1616 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1617 if (PageLRU(page)) {
1618 int lru = page_lru(page);
1621 del_page_from_lru_list(page, lruvec, lru);
1624 spin_unlock_irq(zone_lru_lock(zone));
1630 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1631 * then get resheduled. When there are massive number of tasks doing page
1632 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1633 * the LRU list will go small and be scanned faster than necessary, leading to
1634 * unnecessary swapping, thrashing and OOM.
1636 static int too_many_isolated(struct pglist_data *pgdat, int file,
1637 struct scan_control *sc)
1639 unsigned long inactive, isolated;
1641 if (current_is_kswapd())
1644 if (!sane_reclaim(sc))
1648 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1649 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1651 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1652 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1656 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1657 * won't get blocked by normal direct-reclaimers, forming a circular
1660 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1663 return isolated > inactive;
1666 static noinline_for_stack void
1667 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1669 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1670 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1671 LIST_HEAD(pages_to_free);
1674 * Put back any unfreeable pages.
1676 while (!list_empty(page_list)) {
1677 struct page *page = lru_to_page(page_list);
1680 VM_BUG_ON_PAGE(PageLRU(page), page);
1681 list_del(&page->lru);
1682 if (unlikely(!page_evictable(page))) {
1683 spin_unlock_irq(&pgdat->lru_lock);
1684 putback_lru_page(page);
1685 spin_lock_irq(&pgdat->lru_lock);
1689 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1692 lru = page_lru(page);
1693 add_page_to_lru_list(page, lruvec, lru);
1695 if (is_active_lru(lru)) {
1696 int file = is_file_lru(lru);
1697 int numpages = hpage_nr_pages(page);
1698 reclaim_stat->recent_rotated[file] += numpages;
1700 if (put_page_testzero(page)) {
1701 __ClearPageLRU(page);
1702 __ClearPageActive(page);
1703 del_page_from_lru_list(page, lruvec, lru);
1705 if (unlikely(PageCompound(page))) {
1706 spin_unlock_irq(&pgdat->lru_lock);
1707 mem_cgroup_uncharge(page);
1708 (*get_compound_page_dtor(page))(page);
1709 spin_lock_irq(&pgdat->lru_lock);
1711 list_add(&page->lru, &pages_to_free);
1716 * To save our caller's stack, now use input list for pages to free.
1718 list_splice(&pages_to_free, page_list);
1722 * If a kernel thread (such as nfsd for loop-back mounts) services
1723 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1724 * In that case we should only throttle if the backing device it is
1725 * writing to is congested. In other cases it is safe to throttle.
1727 static int current_may_throttle(void)
1729 return !(current->flags & PF_LESS_THROTTLE) ||
1730 current->backing_dev_info == NULL ||
1731 bdi_write_congested(current->backing_dev_info);
1735 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1736 * of reclaimed pages
1738 static noinline_for_stack unsigned long
1739 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1740 struct scan_control *sc, enum lru_list lru)
1742 LIST_HEAD(page_list);
1743 unsigned long nr_scanned;
1744 unsigned long nr_reclaimed = 0;
1745 unsigned long nr_taken;
1746 struct reclaim_stat stat = {};
1747 isolate_mode_t isolate_mode = 0;
1748 int file = is_file_lru(lru);
1749 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1750 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1751 bool stalled = false;
1753 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1757 /* wait a bit for the reclaimer. */
1761 /* We are about to die and free our memory. Return now. */
1762 if (fatal_signal_pending(current))
1763 return SWAP_CLUSTER_MAX;
1769 isolate_mode |= ISOLATE_UNMAPPED;
1771 spin_lock_irq(&pgdat->lru_lock);
1773 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1774 &nr_scanned, sc, isolate_mode, lru);
1776 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1777 reclaim_stat->recent_scanned[file] += nr_taken;
1779 if (current_is_kswapd()) {
1780 if (global_reclaim(sc))
1781 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1782 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1785 if (global_reclaim(sc))
1786 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1787 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1790 spin_unlock_irq(&pgdat->lru_lock);
1795 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1798 spin_lock_irq(&pgdat->lru_lock);
1800 if (current_is_kswapd()) {
1801 if (global_reclaim(sc))
1802 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1803 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1806 if (global_reclaim(sc))
1807 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1808 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1812 putback_inactive_pages(lruvec, &page_list);
1814 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1816 spin_unlock_irq(&pgdat->lru_lock);
1818 mem_cgroup_uncharge_list(&page_list);
1819 free_unref_page_list(&page_list);
1822 * If reclaim is isolating dirty pages under writeback, it implies
1823 * that the long-lived page allocation rate is exceeding the page
1824 * laundering rate. Either the global limits are not being effective
1825 * at throttling processes due to the page distribution throughout
1826 * zones or there is heavy usage of a slow backing device. The
1827 * only option is to throttle from reclaim context which is not ideal
1828 * as there is no guarantee the dirtying process is throttled in the
1829 * same way balance_dirty_pages() manages.
1831 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1832 * of pages under pages flagged for immediate reclaim and stall if any
1833 * are encountered in the nr_immediate check below.
1835 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1836 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1839 * Legacy memcg will stall in page writeback so avoid forcibly
1842 if (sane_reclaim(sc)) {
1844 * Tag a zone as congested if all the dirty pages scanned were
1845 * backed by a congested BDI and wait_iff_congested will stall.
1847 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1848 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1851 * If dirty pages are scanned that are not queued for IO, it
1852 * implies that flushers are not doing their job. This can
1853 * happen when memory pressure pushes dirty pages to the end of
1854 * the LRU before the dirty limits are breached and the dirty
1855 * data has expired. It can also happen when the proportion of
1856 * dirty pages grows not through writes but through memory
1857 * pressure reclaiming all the clean cache. And in some cases,
1858 * the flushers simply cannot keep up with the allocation
1859 * rate. Nudge the flusher threads in case they are asleep, but
1860 * also allow kswapd to start writing pages during reclaim.
1862 if (stat.nr_unqueued_dirty == nr_taken) {
1863 wakeup_flusher_threads(WB_REASON_VMSCAN);
1864 set_bit(PGDAT_DIRTY, &pgdat->flags);
1868 * If kswapd scans pages marked marked for immediate
1869 * reclaim and under writeback (nr_immediate), it implies
1870 * that pages are cycling through the LRU faster than
1871 * they are written so also forcibly stall.
1873 if (stat.nr_immediate && current_may_throttle())
1874 congestion_wait(BLK_RW_ASYNC, HZ/10);
1878 * Stall direct reclaim for IO completions if underlying BDIs or zone
1879 * is congested. Allow kswapd to continue until it starts encountering
1880 * unqueued dirty pages or cycling through the LRU too quickly.
1882 if (!sc->hibernation_mode && !current_is_kswapd() &&
1883 current_may_throttle())
1884 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1886 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1887 nr_scanned, nr_reclaimed,
1888 stat.nr_dirty, stat.nr_writeback,
1889 stat.nr_congested, stat.nr_immediate,
1890 stat.nr_activate, stat.nr_ref_keep,
1892 sc->priority, file);
1893 return nr_reclaimed;
1897 * This moves pages from the active list to the inactive list.
1899 * We move them the other way if the page is referenced by one or more
1900 * processes, from rmap.
1902 * If the pages are mostly unmapped, the processing is fast and it is
1903 * appropriate to hold zone_lru_lock across the whole operation. But if
1904 * the pages are mapped, the processing is slow (page_referenced()) so we
1905 * should drop zone_lru_lock around each page. It's impossible to balance
1906 * this, so instead we remove the pages from the LRU while processing them.
1907 * It is safe to rely on PG_active against the non-LRU pages in here because
1908 * nobody will play with that bit on a non-LRU page.
1910 * The downside is that we have to touch page->_refcount against each page.
1911 * But we had to alter page->flags anyway.
1913 * Returns the number of pages moved to the given lru.
1916 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1917 struct list_head *list,
1918 struct list_head *pages_to_free,
1921 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1926 while (!list_empty(list)) {
1927 page = lru_to_page(list);
1928 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1930 VM_BUG_ON_PAGE(PageLRU(page), page);
1933 nr_pages = hpage_nr_pages(page);
1934 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1935 list_move(&page->lru, &lruvec->lists[lru]);
1937 if (put_page_testzero(page)) {
1938 __ClearPageLRU(page);
1939 __ClearPageActive(page);
1940 del_page_from_lru_list(page, lruvec, lru);
1942 if (unlikely(PageCompound(page))) {
1943 spin_unlock_irq(&pgdat->lru_lock);
1944 mem_cgroup_uncharge(page);
1945 (*get_compound_page_dtor(page))(page);
1946 spin_lock_irq(&pgdat->lru_lock);
1948 list_add(&page->lru, pages_to_free);
1950 nr_moved += nr_pages;
1954 if (!is_active_lru(lru)) {
1955 __count_vm_events(PGDEACTIVATE, nr_moved);
1956 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1963 static void shrink_active_list(unsigned long nr_to_scan,
1964 struct lruvec *lruvec,
1965 struct scan_control *sc,
1968 unsigned long nr_taken;
1969 unsigned long nr_scanned;
1970 unsigned long vm_flags;
1971 LIST_HEAD(l_hold); /* The pages which were snipped off */
1972 LIST_HEAD(l_active);
1973 LIST_HEAD(l_inactive);
1975 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1976 unsigned nr_deactivate, nr_activate;
1977 unsigned nr_rotated = 0;
1978 isolate_mode_t isolate_mode = 0;
1979 int file = is_file_lru(lru);
1980 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1985 isolate_mode |= ISOLATE_UNMAPPED;
1987 spin_lock_irq(&pgdat->lru_lock);
1989 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1990 &nr_scanned, sc, isolate_mode, lru);
1992 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1993 reclaim_stat->recent_scanned[file] += nr_taken;
1995 __count_vm_events(PGREFILL, nr_scanned);
1996 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
1998 spin_unlock_irq(&pgdat->lru_lock);
2000 while (!list_empty(&l_hold)) {
2002 page = lru_to_page(&l_hold);
2003 list_del(&page->lru);
2005 if (unlikely(!page_evictable(page))) {
2006 putback_lru_page(page);
2010 if (unlikely(buffer_heads_over_limit)) {
2011 if (page_has_private(page) && trylock_page(page)) {
2012 if (page_has_private(page))
2013 try_to_release_page(page, 0);
2018 if (page_referenced(page, 0, sc->target_mem_cgroup,
2020 nr_rotated += hpage_nr_pages(page);
2022 * Identify referenced, file-backed active pages and
2023 * give them one more trip around the active list. So
2024 * that executable code get better chances to stay in
2025 * memory under moderate memory pressure. Anon pages
2026 * are not likely to be evicted by use-once streaming
2027 * IO, plus JVM can create lots of anon VM_EXEC pages,
2028 * so we ignore them here.
2030 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2031 list_add(&page->lru, &l_active);
2036 ClearPageActive(page); /* we are de-activating */
2037 list_add(&page->lru, &l_inactive);
2041 * Move pages back to the lru list.
2043 spin_lock_irq(&pgdat->lru_lock);
2045 * Count referenced pages from currently used mappings as rotated,
2046 * even though only some of them are actually re-activated. This
2047 * helps balance scan pressure between file and anonymous pages in
2050 reclaim_stat->recent_rotated[file] += nr_rotated;
2052 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2053 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2054 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2055 spin_unlock_irq(&pgdat->lru_lock);
2057 mem_cgroup_uncharge_list(&l_hold);
2058 free_unref_page_list(&l_hold);
2059 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2060 nr_deactivate, nr_rotated, sc->priority, file);
2064 * The inactive anon list should be small enough that the VM never has
2065 * to do too much work.
2067 * The inactive file list should be small enough to leave most memory
2068 * to the established workingset on the scan-resistant active list,
2069 * but large enough to avoid thrashing the aggregate readahead window.
2071 * Both inactive lists should also be large enough that each inactive
2072 * page has a chance to be referenced again before it is reclaimed.
2074 * If that fails and refaulting is observed, the inactive list grows.
2076 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2077 * on this LRU, maintained by the pageout code. An inactive_ratio
2078 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2081 * memory ratio inactive
2082 * -------------------------------------
2091 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2092 struct mem_cgroup *memcg,
2093 struct scan_control *sc, bool actual_reclaim)
2095 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2096 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2097 enum lru_list inactive_lru = file * LRU_FILE;
2098 unsigned long inactive, active;
2099 unsigned long inactive_ratio;
2100 unsigned long refaults;
2104 * If we don't have swap space, anonymous page deactivation
2107 if (!file && !total_swap_pages)
2110 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2111 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2114 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2116 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2119 * When refaults are being observed, it means a new workingset
2120 * is being established. Disable active list protection to get
2121 * rid of the stale workingset quickly.
2123 if (file && actual_reclaim && lruvec->refaults != refaults) {
2126 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2128 inactive_ratio = int_sqrt(10 * gb);
2134 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2135 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2136 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2137 inactive_ratio, file);
2139 return inactive * inactive_ratio < active;
2142 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2143 struct lruvec *lruvec, struct mem_cgroup *memcg,
2144 struct scan_control *sc)
2146 if (is_active_lru(lru)) {
2147 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2149 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2153 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2164 * Determine how aggressively the anon and file LRU lists should be
2165 * scanned. The relative value of each set of LRU lists is determined
2166 * by looking at the fraction of the pages scanned we did rotate back
2167 * onto the active list instead of evict.
2169 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2170 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2172 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2173 struct scan_control *sc, unsigned long *nr,
2174 unsigned long *lru_pages)
2176 int swappiness = mem_cgroup_swappiness(memcg);
2177 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2179 u64 denominator = 0; /* gcc */
2180 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2181 unsigned long anon_prio, file_prio;
2182 enum scan_balance scan_balance;
2183 unsigned long anon, file;
2184 unsigned long ap, fp;
2187 /* If we have no swap space, do not bother scanning anon pages. */
2188 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2189 scan_balance = SCAN_FILE;
2194 * Global reclaim will swap to prevent OOM even with no
2195 * swappiness, but memcg users want to use this knob to
2196 * disable swapping for individual groups completely when
2197 * using the memory controller's swap limit feature would be
2200 if (!global_reclaim(sc) && !swappiness) {
2201 scan_balance = SCAN_FILE;
2206 * Do not apply any pressure balancing cleverness when the
2207 * system is close to OOM, scan both anon and file equally
2208 * (unless the swappiness setting disagrees with swapping).
2210 if (!sc->priority && swappiness) {
2211 scan_balance = SCAN_EQUAL;
2216 * Prevent the reclaimer from falling into the cache trap: as
2217 * cache pages start out inactive, every cache fault will tip
2218 * the scan balance towards the file LRU. And as the file LRU
2219 * shrinks, so does the window for rotation from references.
2220 * This means we have a runaway feedback loop where a tiny
2221 * thrashing file LRU becomes infinitely more attractive than
2222 * anon pages. Try to detect this based on file LRU size.
2224 if (global_reclaim(sc)) {
2225 unsigned long pgdatfile;
2226 unsigned long pgdatfree;
2228 unsigned long total_high_wmark = 0;
2230 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2231 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2232 node_page_state(pgdat, NR_INACTIVE_FILE);
2234 for (z = 0; z < MAX_NR_ZONES; z++) {
2235 struct zone *zone = &pgdat->node_zones[z];
2236 if (!managed_zone(zone))
2239 total_high_wmark += high_wmark_pages(zone);
2242 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2244 * Force SCAN_ANON if there are enough inactive
2245 * anonymous pages on the LRU in eligible zones.
2246 * Otherwise, the small LRU gets thrashed.
2248 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2249 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2251 scan_balance = SCAN_ANON;
2258 * If there is enough inactive page cache, i.e. if the size of the
2259 * inactive list is greater than that of the active list *and* the
2260 * inactive list actually has some pages to scan on this priority, we
2261 * do not reclaim anything from the anonymous working set right now.
2262 * Without the second condition we could end up never scanning an
2263 * lruvec even if it has plenty of old anonymous pages unless the
2264 * system is under heavy pressure.
2266 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2267 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2268 scan_balance = SCAN_FILE;
2272 scan_balance = SCAN_FRACT;
2275 * With swappiness at 100, anonymous and file have the same priority.
2276 * This scanning priority is essentially the inverse of IO cost.
2278 anon_prio = swappiness;
2279 file_prio = 200 - anon_prio;
2282 * OK, so we have swap space and a fair amount of page cache
2283 * pages. We use the recently rotated / recently scanned
2284 * ratios to determine how valuable each cache is.
2286 * Because workloads change over time (and to avoid overflow)
2287 * we keep these statistics as a floating average, which ends
2288 * up weighing recent references more than old ones.
2290 * anon in [0], file in [1]
2293 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2294 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2295 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2296 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2298 spin_lock_irq(&pgdat->lru_lock);
2299 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2300 reclaim_stat->recent_scanned[0] /= 2;
2301 reclaim_stat->recent_rotated[0] /= 2;
2304 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2305 reclaim_stat->recent_scanned[1] /= 2;
2306 reclaim_stat->recent_rotated[1] /= 2;
2310 * The amount of pressure on anon vs file pages is inversely
2311 * proportional to the fraction of recently scanned pages on
2312 * each list that were recently referenced and in active use.
2314 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2315 ap /= reclaim_stat->recent_rotated[0] + 1;
2317 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2318 fp /= reclaim_stat->recent_rotated[1] + 1;
2319 spin_unlock_irq(&pgdat->lru_lock);
2323 denominator = ap + fp + 1;
2326 for_each_evictable_lru(lru) {
2327 int file = is_file_lru(lru);
2331 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2332 scan = size >> sc->priority;
2334 * If the cgroup's already been deleted, make sure to
2335 * scrape out the remaining cache.
2337 if (!scan && !mem_cgroup_online(memcg))
2338 scan = min(size, SWAP_CLUSTER_MAX);
2340 switch (scan_balance) {
2342 /* Scan lists relative to size */
2346 * Scan types proportional to swappiness and
2347 * their relative recent reclaim efficiency.
2349 scan = div64_u64(scan * fraction[file],
2354 /* Scan one type exclusively */
2355 if ((scan_balance == SCAN_FILE) != file) {
2361 /* Look ma, no brain */
2371 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2373 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2374 struct scan_control *sc, unsigned long *lru_pages)
2376 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2377 unsigned long nr[NR_LRU_LISTS];
2378 unsigned long targets[NR_LRU_LISTS];
2379 unsigned long nr_to_scan;
2381 unsigned long nr_reclaimed = 0;
2382 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2383 struct blk_plug plug;
2386 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2388 /* Record the original scan target for proportional adjustments later */
2389 memcpy(targets, nr, sizeof(nr));
2392 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2393 * event that can occur when there is little memory pressure e.g.
2394 * multiple streaming readers/writers. Hence, we do not abort scanning
2395 * when the requested number of pages are reclaimed when scanning at
2396 * DEF_PRIORITY on the assumption that the fact we are direct
2397 * reclaiming implies that kswapd is not keeping up and it is best to
2398 * do a batch of work at once. For memcg reclaim one check is made to
2399 * abort proportional reclaim if either the file or anon lru has already
2400 * dropped to zero at the first pass.
2402 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2403 sc->priority == DEF_PRIORITY);
2405 blk_start_plug(&plug);
2406 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2407 nr[LRU_INACTIVE_FILE]) {
2408 unsigned long nr_anon, nr_file, percentage;
2409 unsigned long nr_scanned;
2411 for_each_evictable_lru(lru) {
2413 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2414 nr[lru] -= nr_to_scan;
2416 nr_reclaimed += shrink_list(lru, nr_to_scan,
2423 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2427 * For kswapd and memcg, reclaim at least the number of pages
2428 * requested. Ensure that the anon and file LRUs are scanned
2429 * proportionally what was requested by get_scan_count(). We
2430 * stop reclaiming one LRU and reduce the amount scanning
2431 * proportional to the original scan target.
2433 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2434 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2437 * It's just vindictive to attack the larger once the smaller
2438 * has gone to zero. And given the way we stop scanning the
2439 * smaller below, this makes sure that we only make one nudge
2440 * towards proportionality once we've got nr_to_reclaim.
2442 if (!nr_file || !nr_anon)
2445 if (nr_file > nr_anon) {
2446 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2447 targets[LRU_ACTIVE_ANON] + 1;
2449 percentage = nr_anon * 100 / scan_target;
2451 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2452 targets[LRU_ACTIVE_FILE] + 1;
2454 percentage = nr_file * 100 / scan_target;
2457 /* Stop scanning the smaller of the LRU */
2459 nr[lru + LRU_ACTIVE] = 0;
2462 * Recalculate the other LRU scan count based on its original
2463 * scan target and the percentage scanning already complete
2465 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2466 nr_scanned = targets[lru] - nr[lru];
2467 nr[lru] = targets[lru] * (100 - percentage) / 100;
2468 nr[lru] -= min(nr[lru], nr_scanned);
2471 nr_scanned = targets[lru] - nr[lru];
2472 nr[lru] = targets[lru] * (100 - percentage) / 100;
2473 nr[lru] -= min(nr[lru], nr_scanned);
2475 scan_adjusted = true;
2477 blk_finish_plug(&plug);
2478 sc->nr_reclaimed += nr_reclaimed;
2481 * Even if we did not try to evict anon pages at all, we want to
2482 * rebalance the anon lru active/inactive ratio.
2484 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2485 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2486 sc, LRU_ACTIVE_ANON);
2489 /* Use reclaim/compaction for costly allocs or under memory pressure */
2490 static bool in_reclaim_compaction(struct scan_control *sc)
2492 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2493 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2494 sc->priority < DEF_PRIORITY - 2))
2501 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2502 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2503 * true if more pages should be reclaimed such that when the page allocator
2504 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2505 * It will give up earlier than that if there is difficulty reclaiming pages.
2507 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2508 unsigned long nr_reclaimed,
2509 unsigned long nr_scanned,
2510 struct scan_control *sc)
2512 unsigned long pages_for_compaction;
2513 unsigned long inactive_lru_pages;
2516 /* If not in reclaim/compaction mode, stop */
2517 if (!in_reclaim_compaction(sc))
2520 /* Consider stopping depending on scan and reclaim activity */
2521 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2523 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2524 * full LRU list has been scanned and we are still failing
2525 * to reclaim pages. This full LRU scan is potentially
2526 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2528 if (!nr_reclaimed && !nr_scanned)
2532 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2533 * fail without consequence, stop if we failed to reclaim
2534 * any pages from the last SWAP_CLUSTER_MAX number of
2535 * pages that were scanned. This will return to the
2536 * caller faster at the risk reclaim/compaction and
2537 * the resulting allocation attempt fails
2544 * If we have not reclaimed enough pages for compaction and the
2545 * inactive lists are large enough, continue reclaiming
2547 pages_for_compaction = compact_gap(sc->order);
2548 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2549 if (get_nr_swap_pages() > 0)
2550 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2551 if (sc->nr_reclaimed < pages_for_compaction &&
2552 inactive_lru_pages > pages_for_compaction)
2555 /* If compaction would go ahead or the allocation would succeed, stop */
2556 for (z = 0; z <= sc->reclaim_idx; z++) {
2557 struct zone *zone = &pgdat->node_zones[z];
2558 if (!managed_zone(zone))
2561 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2562 case COMPACT_SUCCESS:
2563 case COMPACT_CONTINUE:
2566 /* check next zone */
2573 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2575 struct reclaim_state *reclaim_state = current->reclaim_state;
2576 unsigned long nr_reclaimed, nr_scanned;
2577 bool reclaimable = false;
2580 struct mem_cgroup *root = sc->target_mem_cgroup;
2581 struct mem_cgroup_reclaim_cookie reclaim = {
2583 .priority = sc->priority,
2585 unsigned long node_lru_pages = 0;
2586 struct mem_cgroup *memcg;
2588 nr_reclaimed = sc->nr_reclaimed;
2589 nr_scanned = sc->nr_scanned;
2591 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2593 unsigned long lru_pages;
2594 unsigned long reclaimed;
2595 unsigned long scanned;
2597 if (mem_cgroup_low(root, memcg)) {
2598 if (!sc->memcg_low_reclaim) {
2599 sc->memcg_low_skipped = 1;
2602 mem_cgroup_event(memcg, MEMCG_LOW);
2605 reclaimed = sc->nr_reclaimed;
2606 scanned = sc->nr_scanned;
2607 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2608 node_lru_pages += lru_pages;
2611 shrink_slab(sc->gfp_mask, pgdat->node_id,
2612 memcg, sc->priority);
2614 /* Record the group's reclaim efficiency */
2615 vmpressure(sc->gfp_mask, memcg, false,
2616 sc->nr_scanned - scanned,
2617 sc->nr_reclaimed - reclaimed);
2620 * Direct reclaim and kswapd have to scan all memory
2621 * cgroups to fulfill the overall scan target for the
2624 * Limit reclaim, on the other hand, only cares about
2625 * nr_to_reclaim pages to be reclaimed and it will
2626 * retry with decreasing priority if one round over the
2627 * whole hierarchy is not sufficient.
2629 if (!global_reclaim(sc) &&
2630 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2631 mem_cgroup_iter_break(root, memcg);
2634 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2636 if (global_reclaim(sc))
2637 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2640 if (reclaim_state) {
2641 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2642 reclaim_state->reclaimed_slab = 0;
2645 /* Record the subtree's reclaim efficiency */
2646 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2647 sc->nr_scanned - nr_scanned,
2648 sc->nr_reclaimed - nr_reclaimed);
2650 if (sc->nr_reclaimed - nr_reclaimed)
2653 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2654 sc->nr_scanned - nr_scanned, sc));
2657 * Kswapd gives up on balancing particular nodes after too
2658 * many failures to reclaim anything from them and goes to
2659 * sleep. On reclaim progress, reset the failure counter. A
2660 * successful direct reclaim run will revive a dormant kswapd.
2663 pgdat->kswapd_failures = 0;
2669 * Returns true if compaction should go ahead for a costly-order request, or
2670 * the allocation would already succeed without compaction. Return false if we
2671 * should reclaim first.
2673 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2675 unsigned long watermark;
2676 enum compact_result suitable;
2678 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2679 if (suitable == COMPACT_SUCCESS)
2680 /* Allocation should succeed already. Don't reclaim. */
2682 if (suitable == COMPACT_SKIPPED)
2683 /* Compaction cannot yet proceed. Do reclaim. */
2687 * Compaction is already possible, but it takes time to run and there
2688 * are potentially other callers using the pages just freed. So proceed
2689 * with reclaim to make a buffer of free pages available to give
2690 * compaction a reasonable chance of completing and allocating the page.
2691 * Note that we won't actually reclaim the whole buffer in one attempt
2692 * as the target watermark in should_continue_reclaim() is lower. But if
2693 * we are already above the high+gap watermark, don't reclaim at all.
2695 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2697 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2701 * This is the direct reclaim path, for page-allocating processes. We only
2702 * try to reclaim pages from zones which will satisfy the caller's allocation
2705 * If a zone is deemed to be full of pinned pages then just give it a light
2706 * scan then give up on it.
2708 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2712 unsigned long nr_soft_reclaimed;
2713 unsigned long nr_soft_scanned;
2715 pg_data_t *last_pgdat = NULL;
2718 * If the number of buffer_heads in the machine exceeds the maximum
2719 * allowed level, force direct reclaim to scan the highmem zone as
2720 * highmem pages could be pinning lowmem pages storing buffer_heads
2722 orig_mask = sc->gfp_mask;
2723 if (buffer_heads_over_limit) {
2724 sc->gfp_mask |= __GFP_HIGHMEM;
2725 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2728 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2729 sc->reclaim_idx, sc->nodemask) {
2731 * Take care memory controller reclaiming has small influence
2734 if (global_reclaim(sc)) {
2735 if (!cpuset_zone_allowed(zone,
2736 GFP_KERNEL | __GFP_HARDWALL))
2740 * If we already have plenty of memory free for
2741 * compaction in this zone, don't free any more.
2742 * Even though compaction is invoked for any
2743 * non-zero order, only frequent costly order
2744 * reclamation is disruptive enough to become a
2745 * noticeable problem, like transparent huge
2748 if (IS_ENABLED(CONFIG_COMPACTION) &&
2749 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2750 compaction_ready(zone, sc)) {
2751 sc->compaction_ready = true;
2756 * Shrink each node in the zonelist once. If the
2757 * zonelist is ordered by zone (not the default) then a
2758 * node may be shrunk multiple times but in that case
2759 * the user prefers lower zones being preserved.
2761 if (zone->zone_pgdat == last_pgdat)
2765 * This steals pages from memory cgroups over softlimit
2766 * and returns the number of reclaimed pages and
2767 * scanned pages. This works for global memory pressure
2768 * and balancing, not for a memcg's limit.
2770 nr_soft_scanned = 0;
2771 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2772 sc->order, sc->gfp_mask,
2774 sc->nr_reclaimed += nr_soft_reclaimed;
2775 sc->nr_scanned += nr_soft_scanned;
2776 /* need some check for avoid more shrink_zone() */
2779 /* See comment about same check for global reclaim above */
2780 if (zone->zone_pgdat == last_pgdat)
2782 last_pgdat = zone->zone_pgdat;
2783 shrink_node(zone->zone_pgdat, sc);
2787 * Restore to original mask to avoid the impact on the caller if we
2788 * promoted it to __GFP_HIGHMEM.
2790 sc->gfp_mask = orig_mask;
2793 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2795 struct mem_cgroup *memcg;
2797 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2799 unsigned long refaults;
2800 struct lruvec *lruvec;
2803 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2805 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2807 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2808 lruvec->refaults = refaults;
2809 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2813 * This is the main entry point to direct page reclaim.
2815 * If a full scan of the inactive list fails to free enough memory then we
2816 * are "out of memory" and something needs to be killed.
2818 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2819 * high - the zone may be full of dirty or under-writeback pages, which this
2820 * caller can't do much about. We kick the writeback threads and take explicit
2821 * naps in the hope that some of these pages can be written. But if the
2822 * allocating task holds filesystem locks which prevent writeout this might not
2823 * work, and the allocation attempt will fail.
2825 * returns: 0, if no pages reclaimed
2826 * else, the number of pages reclaimed
2828 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2829 struct scan_control *sc)
2831 int initial_priority = sc->priority;
2832 pg_data_t *last_pgdat;
2836 delayacct_freepages_start();
2838 if (global_reclaim(sc))
2839 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2842 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2845 shrink_zones(zonelist, sc);
2847 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2850 if (sc->compaction_ready)
2854 * If we're getting trouble reclaiming, start doing
2855 * writepage even in laptop mode.
2857 if (sc->priority < DEF_PRIORITY - 2)
2858 sc->may_writepage = 1;
2859 } while (--sc->priority >= 0);
2862 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2864 if (zone->zone_pgdat == last_pgdat)
2866 last_pgdat = zone->zone_pgdat;
2867 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2870 delayacct_freepages_end();
2872 if (sc->nr_reclaimed)
2873 return sc->nr_reclaimed;
2875 /* Aborted reclaim to try compaction? don't OOM, then */
2876 if (sc->compaction_ready)
2879 /* Untapped cgroup reserves? Don't OOM, retry. */
2880 if (sc->memcg_low_skipped) {
2881 sc->priority = initial_priority;
2882 sc->memcg_low_reclaim = 1;
2883 sc->memcg_low_skipped = 0;
2890 static bool allow_direct_reclaim(pg_data_t *pgdat)
2893 unsigned long pfmemalloc_reserve = 0;
2894 unsigned long free_pages = 0;
2898 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2901 for (i = 0; i <= ZONE_NORMAL; i++) {
2902 zone = &pgdat->node_zones[i];
2903 if (!managed_zone(zone))
2906 if (!zone_reclaimable_pages(zone))
2909 pfmemalloc_reserve += min_wmark_pages(zone);
2910 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2913 /* If there are no reserves (unexpected config) then do not throttle */
2914 if (!pfmemalloc_reserve)
2917 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2919 /* kswapd must be awake if processes are being throttled */
2920 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2921 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2922 (enum zone_type)ZONE_NORMAL);
2923 wake_up_interruptible(&pgdat->kswapd_wait);
2930 * Throttle direct reclaimers if backing storage is backed by the network
2931 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2932 * depleted. kswapd will continue to make progress and wake the processes
2933 * when the low watermark is reached.
2935 * Returns true if a fatal signal was delivered during throttling. If this
2936 * happens, the page allocator should not consider triggering the OOM killer.
2938 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2939 nodemask_t *nodemask)
2943 pg_data_t *pgdat = NULL;
2946 * Kernel threads should not be throttled as they may be indirectly
2947 * responsible for cleaning pages necessary for reclaim to make forward
2948 * progress. kjournald for example may enter direct reclaim while
2949 * committing a transaction where throttling it could forcing other
2950 * processes to block on log_wait_commit().
2952 if (current->flags & PF_KTHREAD)
2956 * If a fatal signal is pending, this process should not throttle.
2957 * It should return quickly so it can exit and free its memory
2959 if (fatal_signal_pending(current))
2963 * Check if the pfmemalloc reserves are ok by finding the first node
2964 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2965 * GFP_KERNEL will be required for allocating network buffers when
2966 * swapping over the network so ZONE_HIGHMEM is unusable.
2968 * Throttling is based on the first usable node and throttled processes
2969 * wait on a queue until kswapd makes progress and wakes them. There
2970 * is an affinity then between processes waking up and where reclaim
2971 * progress has been made assuming the process wakes on the same node.
2972 * More importantly, processes running on remote nodes will not compete
2973 * for remote pfmemalloc reserves and processes on different nodes
2974 * should make reasonable progress.
2976 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2977 gfp_zone(gfp_mask), nodemask) {
2978 if (zone_idx(zone) > ZONE_NORMAL)
2981 /* Throttle based on the first usable node */
2982 pgdat = zone->zone_pgdat;
2983 if (allow_direct_reclaim(pgdat))
2988 /* If no zone was usable by the allocation flags then do not throttle */
2992 /* Account for the throttling */
2993 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2996 * If the caller cannot enter the filesystem, it's possible that it
2997 * is due to the caller holding an FS lock or performing a journal
2998 * transaction in the case of a filesystem like ext[3|4]. In this case,
2999 * it is not safe to block on pfmemalloc_wait as kswapd could be
3000 * blocked waiting on the same lock. Instead, throttle for up to a
3001 * second before continuing.
3003 if (!(gfp_mask & __GFP_FS)) {
3004 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3005 allow_direct_reclaim(pgdat), HZ);
3010 /* Throttle until kswapd wakes the process */
3011 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3012 allow_direct_reclaim(pgdat));
3015 if (fatal_signal_pending(current))
3022 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3023 gfp_t gfp_mask, nodemask_t *nodemask)
3025 unsigned long nr_reclaimed;
3026 struct scan_control sc = {
3027 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3028 .gfp_mask = current_gfp_context(gfp_mask),
3029 .reclaim_idx = gfp_zone(gfp_mask),
3031 .nodemask = nodemask,
3032 .priority = DEF_PRIORITY,
3033 .may_writepage = !laptop_mode,
3039 * Do not enter reclaim if fatal signal was delivered while throttled.
3040 * 1 is returned so that the page allocator does not OOM kill at this
3043 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3046 trace_mm_vmscan_direct_reclaim_begin(order,
3051 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3053 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3055 return nr_reclaimed;
3060 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3061 gfp_t gfp_mask, bool noswap,
3063 unsigned long *nr_scanned)
3065 struct scan_control sc = {
3066 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3067 .target_mem_cgroup = memcg,
3068 .may_writepage = !laptop_mode,
3070 .reclaim_idx = MAX_NR_ZONES - 1,
3071 .may_swap = !noswap,
3073 unsigned long lru_pages;
3075 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3076 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3078 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3084 * NOTE: Although we can get the priority field, using it
3085 * here is not a good idea, since it limits the pages we can scan.
3086 * if we don't reclaim here, the shrink_node from balance_pgdat
3087 * will pick up pages from other mem cgroup's as well. We hack
3088 * the priority and make it zero.
3090 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3092 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3094 *nr_scanned = sc.nr_scanned;
3095 return sc.nr_reclaimed;
3098 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3099 unsigned long nr_pages,
3103 struct zonelist *zonelist;
3104 unsigned long nr_reclaimed;
3106 unsigned int noreclaim_flag;
3107 struct scan_control sc = {
3108 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3109 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3110 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3111 .reclaim_idx = MAX_NR_ZONES - 1,
3112 .target_mem_cgroup = memcg,
3113 .priority = DEF_PRIORITY,
3114 .may_writepage = !laptop_mode,
3116 .may_swap = may_swap,
3120 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3121 * take care of from where we get pages. So the node where we start the
3122 * scan does not need to be the current node.
3124 nid = mem_cgroup_select_victim_node(memcg);
3126 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3128 trace_mm_vmscan_memcg_reclaim_begin(0,
3133 noreclaim_flag = memalloc_noreclaim_save();
3134 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3135 memalloc_noreclaim_restore(noreclaim_flag);
3137 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3139 return nr_reclaimed;
3143 static void age_active_anon(struct pglist_data *pgdat,
3144 struct scan_control *sc)
3146 struct mem_cgroup *memcg;
3148 if (!total_swap_pages)
3151 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3153 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3155 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3156 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3157 sc, LRU_ACTIVE_ANON);
3159 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3164 * Returns true if there is an eligible zone balanced for the request order
3167 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3170 unsigned long mark = -1;
3173 for (i = 0; i <= classzone_idx; i++) {
3174 zone = pgdat->node_zones + i;
3176 if (!managed_zone(zone))
3179 mark = high_wmark_pages(zone);
3180 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3185 * If a node has no populated zone within classzone_idx, it does not
3186 * need balancing by definition. This can happen if a zone-restricted
3187 * allocation tries to wake a remote kswapd.
3195 /* Clear pgdat state for congested, dirty or under writeback. */
3196 static void clear_pgdat_congested(pg_data_t *pgdat)
3198 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3199 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3200 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3204 * Prepare kswapd for sleeping. This verifies that there are no processes
3205 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3207 * Returns true if kswapd is ready to sleep
3209 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3212 * The throttled processes are normally woken up in balance_pgdat() as
3213 * soon as allow_direct_reclaim() is true. But there is a potential
3214 * race between when kswapd checks the watermarks and a process gets
3215 * throttled. There is also a potential race if processes get
3216 * throttled, kswapd wakes, a large process exits thereby balancing the
3217 * zones, which causes kswapd to exit balance_pgdat() before reaching
3218 * the wake up checks. If kswapd is going to sleep, no process should
3219 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3220 * the wake up is premature, processes will wake kswapd and get
3221 * throttled again. The difference from wake ups in balance_pgdat() is
3222 * that here we are under prepare_to_wait().
3224 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3225 wake_up_all(&pgdat->pfmemalloc_wait);
3227 /* Hopeless node, leave it to direct reclaim */
3228 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3231 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3232 clear_pgdat_congested(pgdat);
3240 * kswapd shrinks a node of pages that are at or below the highest usable
3241 * zone that is currently unbalanced.
3243 * Returns true if kswapd scanned at least the requested number of pages to
3244 * reclaim or if the lack of progress was due to pages under writeback.
3245 * This is used to determine if the scanning priority needs to be raised.
3247 static bool kswapd_shrink_node(pg_data_t *pgdat,
3248 struct scan_control *sc)
3253 /* Reclaim a number of pages proportional to the number of zones */
3254 sc->nr_to_reclaim = 0;
3255 for (z = 0; z <= sc->reclaim_idx; z++) {
3256 zone = pgdat->node_zones + z;
3257 if (!managed_zone(zone))
3260 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3264 * Historically care was taken to put equal pressure on all zones but
3265 * now pressure is applied based on node LRU order.
3267 shrink_node(pgdat, sc);
3270 * Fragmentation may mean that the system cannot be rebalanced for
3271 * high-order allocations. If twice the allocation size has been
3272 * reclaimed then recheck watermarks only at order-0 to prevent
3273 * excessive reclaim. Assume that a process requested a high-order
3274 * can direct reclaim/compact.
3276 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3279 return sc->nr_scanned >= sc->nr_to_reclaim;
3283 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3284 * that are eligible for use by the caller until at least one zone is
3287 * Returns the order kswapd finished reclaiming at.
3289 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3290 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3291 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3292 * or lower is eligible for reclaim until at least one usable zone is
3295 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3298 unsigned long nr_soft_reclaimed;
3299 unsigned long nr_soft_scanned;
3301 struct scan_control sc = {
3302 .gfp_mask = GFP_KERNEL,
3304 .priority = DEF_PRIORITY,
3305 .may_writepage = !laptop_mode,
3309 count_vm_event(PAGEOUTRUN);
3312 unsigned long nr_reclaimed = sc.nr_reclaimed;
3313 bool raise_priority = true;
3315 sc.reclaim_idx = classzone_idx;
3318 * If the number of buffer_heads exceeds the maximum allowed
3319 * then consider reclaiming from all zones. This has a dual
3320 * purpose -- on 64-bit systems it is expected that
3321 * buffer_heads are stripped during active rotation. On 32-bit
3322 * systems, highmem pages can pin lowmem memory and shrinking
3323 * buffers can relieve lowmem pressure. Reclaim may still not
3324 * go ahead if all eligible zones for the original allocation
3325 * request are balanced to avoid excessive reclaim from kswapd.
3327 if (buffer_heads_over_limit) {
3328 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3329 zone = pgdat->node_zones + i;
3330 if (!managed_zone(zone))
3339 * Only reclaim if there are no eligible zones. Note that
3340 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3343 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3347 * Do some background aging of the anon list, to give
3348 * pages a chance to be referenced before reclaiming. All
3349 * pages are rotated regardless of classzone as this is
3350 * about consistent aging.
3352 age_active_anon(pgdat, &sc);
3355 * If we're getting trouble reclaiming, start doing writepage
3356 * even in laptop mode.
3358 if (sc.priority < DEF_PRIORITY - 2)
3359 sc.may_writepage = 1;
3361 /* Call soft limit reclaim before calling shrink_node. */
3363 nr_soft_scanned = 0;
3364 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3365 sc.gfp_mask, &nr_soft_scanned);
3366 sc.nr_reclaimed += nr_soft_reclaimed;
3369 * There should be no need to raise the scanning priority if
3370 * enough pages are already being scanned that that high
3371 * watermark would be met at 100% efficiency.
3373 if (kswapd_shrink_node(pgdat, &sc))
3374 raise_priority = false;
3377 * If the low watermark is met there is no need for processes
3378 * to be throttled on pfmemalloc_wait as they should not be
3379 * able to safely make forward progress. Wake them
3381 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3382 allow_direct_reclaim(pgdat))
3383 wake_up_all(&pgdat->pfmemalloc_wait);
3385 /* Check if kswapd should be suspending */
3386 if (try_to_freeze() || kthread_should_stop())
3390 * Raise priority if scanning rate is too low or there was no
3391 * progress in reclaiming pages
3393 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3394 if (raise_priority || !nr_reclaimed)
3396 } while (sc.priority >= 1);
3398 if (!sc.nr_reclaimed)
3399 pgdat->kswapd_failures++;
3402 snapshot_refaults(NULL, pgdat);
3404 * Return the order kswapd stopped reclaiming at as
3405 * prepare_kswapd_sleep() takes it into account. If another caller
3406 * entered the allocator slow path while kswapd was awake, order will
3407 * remain at the higher level.
3413 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3414 * allocation request woke kswapd for. When kswapd has not woken recently,
3415 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3416 * given classzone and returns it or the highest classzone index kswapd
3417 * was recently woke for.
3419 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3420 enum zone_type classzone_idx)
3422 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3423 return classzone_idx;
3425 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3428 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3429 unsigned int classzone_idx)
3434 if (freezing(current) || kthread_should_stop())
3437 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3440 * Try to sleep for a short interval. Note that kcompactd will only be
3441 * woken if it is possible to sleep for a short interval. This is
3442 * deliberate on the assumption that if reclaim cannot keep an
3443 * eligible zone balanced that it's also unlikely that compaction will
3446 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3448 * Compaction records what page blocks it recently failed to
3449 * isolate pages from and skips them in the future scanning.
3450 * When kswapd is going to sleep, it is reasonable to assume
3451 * that pages and compaction may succeed so reset the cache.
3453 reset_isolation_suitable(pgdat);
3456 * We have freed the memory, now we should compact it to make
3457 * allocation of the requested order possible.
3459 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3461 remaining = schedule_timeout(HZ/10);
3464 * If woken prematurely then reset kswapd_classzone_idx and
3465 * order. The values will either be from a wakeup request or
3466 * the previous request that slept prematurely.
3469 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3470 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3473 finish_wait(&pgdat->kswapd_wait, &wait);
3474 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3478 * After a short sleep, check if it was a premature sleep. If not, then
3479 * go fully to sleep until explicitly woken up.
3482 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3483 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3486 * vmstat counters are not perfectly accurate and the estimated
3487 * value for counters such as NR_FREE_PAGES can deviate from the
3488 * true value by nr_online_cpus * threshold. To avoid the zone
3489 * watermarks being breached while under pressure, we reduce the
3490 * per-cpu vmstat threshold while kswapd is awake and restore
3491 * them before going back to sleep.
3493 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3495 if (!kthread_should_stop())
3498 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3501 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3503 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3505 finish_wait(&pgdat->kswapd_wait, &wait);
3509 * The background pageout daemon, started as a kernel thread
3510 * from the init process.
3512 * This basically trickles out pages so that we have _some_
3513 * free memory available even if there is no other activity
3514 * that frees anything up. This is needed for things like routing
3515 * etc, where we otherwise might have all activity going on in
3516 * asynchronous contexts that cannot page things out.
3518 * If there are applications that are active memory-allocators
3519 * (most normal use), this basically shouldn't matter.
3521 static int kswapd(void *p)
3523 unsigned int alloc_order, reclaim_order;
3524 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3525 pg_data_t *pgdat = (pg_data_t*)p;
3526 struct task_struct *tsk = current;
3528 struct reclaim_state reclaim_state = {
3529 .reclaimed_slab = 0,
3531 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3533 if (!cpumask_empty(cpumask))
3534 set_cpus_allowed_ptr(tsk, cpumask);
3535 current->reclaim_state = &reclaim_state;
3538 * Tell the memory management that we're a "memory allocator",
3539 * and that if we need more memory we should get access to it
3540 * regardless (see "__alloc_pages()"). "kswapd" should
3541 * never get caught in the normal page freeing logic.
3543 * (Kswapd normally doesn't need memory anyway, but sometimes
3544 * you need a small amount of memory in order to be able to
3545 * page out something else, and this flag essentially protects
3546 * us from recursively trying to free more memory as we're
3547 * trying to free the first piece of memory in the first place).
3549 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3552 pgdat->kswapd_order = 0;
3553 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3557 alloc_order = reclaim_order = pgdat->kswapd_order;
3558 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3561 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3564 /* Read the new order and classzone_idx */
3565 alloc_order = reclaim_order = pgdat->kswapd_order;
3566 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3567 pgdat->kswapd_order = 0;
3568 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3570 ret = try_to_freeze();
3571 if (kthread_should_stop())
3575 * We can speed up thawing tasks if we don't call balance_pgdat
3576 * after returning from the refrigerator
3582 * Reclaim begins at the requested order but if a high-order
3583 * reclaim fails then kswapd falls back to reclaiming for
3584 * order-0. If that happens, kswapd will consider sleeping
3585 * for the order it finished reclaiming at (reclaim_order)
3586 * but kcompactd is woken to compact for the original
3587 * request (alloc_order).
3589 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3591 fs_reclaim_acquire(GFP_KERNEL);
3592 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3593 fs_reclaim_release(GFP_KERNEL);
3594 if (reclaim_order < alloc_order)
3595 goto kswapd_try_sleep;
3598 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3599 current->reclaim_state = NULL;
3605 * A zone is low on free memory, so wake its kswapd task to service it.
3607 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3611 if (!managed_zone(zone))
3614 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3616 pgdat = zone->zone_pgdat;
3617 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3619 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3620 if (!waitqueue_active(&pgdat->kswapd_wait))
3623 /* Hopeless node, leave it to direct reclaim */
3624 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3627 if (pgdat_balanced(pgdat, order, classzone_idx))
3630 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
3631 wake_up_interruptible(&pgdat->kswapd_wait);
3634 #ifdef CONFIG_HIBERNATION
3636 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3639 * Rather than trying to age LRUs the aim is to preserve the overall
3640 * LRU order by reclaiming preferentially
3641 * inactive > active > active referenced > active mapped
3643 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3645 struct reclaim_state reclaim_state;
3646 struct scan_control sc = {
3647 .nr_to_reclaim = nr_to_reclaim,
3648 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3649 .reclaim_idx = MAX_NR_ZONES - 1,
3650 .priority = DEF_PRIORITY,
3654 .hibernation_mode = 1,
3656 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3657 struct task_struct *p = current;
3658 unsigned long nr_reclaimed;
3659 unsigned int noreclaim_flag;
3661 noreclaim_flag = memalloc_noreclaim_save();
3662 fs_reclaim_acquire(sc.gfp_mask);
3663 reclaim_state.reclaimed_slab = 0;
3664 p->reclaim_state = &reclaim_state;
3666 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3668 p->reclaim_state = NULL;
3669 fs_reclaim_release(sc.gfp_mask);
3670 memalloc_noreclaim_restore(noreclaim_flag);
3672 return nr_reclaimed;
3674 #endif /* CONFIG_HIBERNATION */
3676 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3677 not required for correctness. So if the last cpu in a node goes
3678 away, we get changed to run anywhere: as the first one comes back,
3679 restore their cpu bindings. */
3680 static int kswapd_cpu_online(unsigned int cpu)
3684 for_each_node_state(nid, N_MEMORY) {
3685 pg_data_t *pgdat = NODE_DATA(nid);
3686 const struct cpumask *mask;
3688 mask = cpumask_of_node(pgdat->node_id);
3690 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3691 /* One of our CPUs online: restore mask */
3692 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3698 * This kswapd start function will be called by init and node-hot-add.
3699 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3701 int kswapd_run(int nid)
3703 pg_data_t *pgdat = NODE_DATA(nid);
3709 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3710 if (IS_ERR(pgdat->kswapd)) {
3711 /* failure at boot is fatal */
3712 BUG_ON(system_state < SYSTEM_RUNNING);
3713 pr_err("Failed to start kswapd on node %d\n", nid);
3714 ret = PTR_ERR(pgdat->kswapd);
3715 pgdat->kswapd = NULL;
3721 * Called by memory hotplug when all memory in a node is offlined. Caller must
3722 * hold mem_hotplug_begin/end().
3724 void kswapd_stop(int nid)
3726 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3729 kthread_stop(kswapd);
3730 NODE_DATA(nid)->kswapd = NULL;
3734 static int __init kswapd_init(void)
3739 for_each_node_state(nid, N_MEMORY)
3741 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3742 "mm/vmscan:online", kswapd_cpu_online,
3748 module_init(kswapd_init)
3754 * If non-zero call node_reclaim when the number of free pages falls below
3757 int node_reclaim_mode __read_mostly;
3759 #define RECLAIM_OFF 0
3760 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3761 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3762 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3765 * Priority for NODE_RECLAIM. This determines the fraction of pages
3766 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3769 #define NODE_RECLAIM_PRIORITY 4
3772 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3775 int sysctl_min_unmapped_ratio = 1;
3778 * If the number of slab pages in a zone grows beyond this percentage then
3779 * slab reclaim needs to occur.
3781 int sysctl_min_slab_ratio = 5;
3783 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3785 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3786 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3787 node_page_state(pgdat, NR_ACTIVE_FILE);
3790 * It's possible for there to be more file mapped pages than
3791 * accounted for by the pages on the file LRU lists because
3792 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3794 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3797 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3798 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3800 unsigned long nr_pagecache_reclaimable;
3801 unsigned long delta = 0;
3804 * If RECLAIM_UNMAP is set, then all file pages are considered
3805 * potentially reclaimable. Otherwise, we have to worry about
3806 * pages like swapcache and node_unmapped_file_pages() provides
3809 if (node_reclaim_mode & RECLAIM_UNMAP)
3810 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3812 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3814 /* If we can't clean pages, remove dirty pages from consideration */
3815 if (!(node_reclaim_mode & RECLAIM_WRITE))
3816 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3818 /* Watch for any possible underflows due to delta */
3819 if (unlikely(delta > nr_pagecache_reclaimable))
3820 delta = nr_pagecache_reclaimable;
3822 return nr_pagecache_reclaimable - delta;
3826 * Try to free up some pages from this node through reclaim.
3828 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3830 /* Minimum pages needed in order to stay on node */
3831 const unsigned long nr_pages = 1 << order;
3832 struct task_struct *p = current;
3833 struct reclaim_state reclaim_state;
3834 unsigned int noreclaim_flag;
3835 struct scan_control sc = {
3836 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3837 .gfp_mask = current_gfp_context(gfp_mask),
3839 .priority = NODE_RECLAIM_PRIORITY,
3840 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3841 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3843 .reclaim_idx = gfp_zone(gfp_mask),
3848 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3849 * and we also need to be able to write out pages for RECLAIM_WRITE
3850 * and RECLAIM_UNMAP.
3852 noreclaim_flag = memalloc_noreclaim_save();
3853 p->flags |= PF_SWAPWRITE;
3854 fs_reclaim_acquire(sc.gfp_mask);
3855 reclaim_state.reclaimed_slab = 0;
3856 p->reclaim_state = &reclaim_state;
3858 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3860 * Free memory by calling shrink zone with increasing
3861 * priorities until we have enough memory freed.
3864 shrink_node(pgdat, &sc);
3865 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3868 p->reclaim_state = NULL;
3869 fs_reclaim_release(gfp_mask);
3870 current->flags &= ~PF_SWAPWRITE;
3871 memalloc_noreclaim_restore(noreclaim_flag);
3872 return sc.nr_reclaimed >= nr_pages;
3875 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3880 * Node reclaim reclaims unmapped file backed pages and
3881 * slab pages if we are over the defined limits.
3883 * A small portion of unmapped file backed pages is needed for
3884 * file I/O otherwise pages read by file I/O will be immediately
3885 * thrown out if the node is overallocated. So we do not reclaim
3886 * if less than a specified percentage of the node is used by
3887 * unmapped file backed pages.
3889 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3890 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3891 return NODE_RECLAIM_FULL;
3894 * Do not scan if the allocation should not be delayed.
3896 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3897 return NODE_RECLAIM_NOSCAN;
3900 * Only run node reclaim on the local node or on nodes that do not
3901 * have associated processors. This will favor the local processor
3902 * over remote processors and spread off node memory allocations
3903 * as wide as possible.
3905 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3906 return NODE_RECLAIM_NOSCAN;
3908 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3909 return NODE_RECLAIM_NOSCAN;
3911 ret = __node_reclaim(pgdat, gfp_mask, order);
3912 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3915 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3922 * page_evictable - test whether a page is evictable
3923 * @page: the page to test
3925 * Test whether page is evictable--i.e., should be placed on active/inactive
3926 * lists vs unevictable list.
3928 * Reasons page might not be evictable:
3929 * (1) page's mapping marked unevictable
3930 * (2) page is part of an mlocked VMA
3933 int page_evictable(struct page *page)
3935 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3940 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3941 * @pages: array of pages to check
3942 * @nr_pages: number of pages to check
3944 * Checks pages for evictability and moves them to the appropriate lru list.
3946 * This function is only used for SysV IPC SHM_UNLOCK.
3948 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3950 struct lruvec *lruvec;
3951 struct pglist_data *pgdat = NULL;
3956 for (i = 0; i < nr_pages; i++) {
3957 struct page *page = pages[i];
3958 struct pglist_data *pagepgdat = page_pgdat(page);
3961 if (pagepgdat != pgdat) {
3963 spin_unlock_irq(&pgdat->lru_lock);
3965 spin_lock_irq(&pgdat->lru_lock);
3967 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3969 if (!PageLRU(page) || !PageUnevictable(page))
3972 if (page_evictable(page)) {
3973 enum lru_list lru = page_lru_base_type(page);
3975 VM_BUG_ON_PAGE(PageActive(page), page);
3976 ClearPageUnevictable(page);
3977 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3978 add_page_to_lru_list(page, lruvec, lru);
3984 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3985 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3986 spin_unlock_irq(&pgdat->lru_lock);
3989 #endif /* CONFIG_SHMEM */