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);
223 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
227 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
228 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
229 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
231 if (get_nr_swap_pages() > 0)
232 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
233 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
234 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
240 * lruvec_lru_size - Returns the number of pages on the given LRU list.
241 * @lruvec: lru vector
243 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
245 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
247 unsigned long lru_size;
250 if (!mem_cgroup_disabled())
251 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
253 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
255 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
256 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
259 if (!managed_zone(zone))
262 if (!mem_cgroup_disabled())
263 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
265 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
266 NR_ZONE_LRU_BASE + lru);
267 lru_size -= min(size, lru_size);
275 * Add a shrinker callback to be called from the vm.
277 int register_shrinker(struct shrinker *shrinker)
279 size_t size = sizeof(*shrinker->nr_deferred);
281 if (shrinker->flags & SHRINKER_NUMA_AWARE)
284 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
285 if (!shrinker->nr_deferred)
288 down_write(&shrinker_rwsem);
289 list_add_tail(&shrinker->list, &shrinker_list);
290 up_write(&shrinker_rwsem);
293 EXPORT_SYMBOL(register_shrinker);
298 void unregister_shrinker(struct shrinker *shrinker)
300 if (!shrinker->nr_deferred)
302 down_write(&shrinker_rwsem);
303 list_del(&shrinker->list);
304 up_write(&shrinker_rwsem);
305 kfree(shrinker->nr_deferred);
306 shrinker->nr_deferred = NULL;
308 EXPORT_SYMBOL(unregister_shrinker);
310 #define SHRINK_BATCH 128
312 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
313 struct shrinker *shrinker, int priority)
315 unsigned long freed = 0;
316 unsigned long long delta;
321 int nid = shrinkctl->nid;
322 long batch_size = shrinker->batch ? shrinker->batch
324 long scanned = 0, next_deferred;
326 freeable = shrinker->count_objects(shrinker, shrinkctl);
331 * copy the current shrinker scan count into a local variable
332 * and zero it so that other concurrent shrinker invocations
333 * don't also do this scanning work.
335 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
338 delta = freeable >> priority;
340 do_div(delta, shrinker->seeks);
342 if (total_scan < 0) {
343 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
344 shrinker->scan_objects, total_scan);
345 total_scan = freeable;
348 next_deferred = total_scan;
351 * We need to avoid excessive windup on filesystem shrinkers
352 * due to large numbers of GFP_NOFS allocations causing the
353 * shrinkers to return -1 all the time. This results in a large
354 * nr being built up so when a shrink that can do some work
355 * comes along it empties the entire cache due to nr >>>
356 * freeable. This is bad for sustaining a working set in
359 * Hence only allow the shrinker to scan the entire cache when
360 * a large delta change is calculated directly.
362 if (delta < freeable / 4)
363 total_scan = min(total_scan, freeable / 2);
366 * Avoid risking looping forever due to too large nr value:
367 * never try to free more than twice the estimate number of
370 if (total_scan > freeable * 2)
371 total_scan = freeable * 2;
373 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
374 freeable, delta, total_scan, priority);
377 * Normally, we should not scan less than batch_size objects in one
378 * pass to avoid too frequent shrinker calls, but if the slab has less
379 * than batch_size objects in total and we are really tight on memory,
380 * we will try to reclaim all available objects, otherwise we can end
381 * up failing allocations although there are plenty of reclaimable
382 * objects spread over several slabs with usage less than the
385 * We detect the "tight on memory" situations by looking at the total
386 * number of objects we want to scan (total_scan). If it is greater
387 * than the total number of objects on slab (freeable), we must be
388 * scanning at high prio and therefore should try to reclaim as much as
391 while (total_scan >= batch_size ||
392 total_scan >= freeable) {
394 unsigned long nr_to_scan = min(batch_size, total_scan);
396 shrinkctl->nr_to_scan = nr_to_scan;
397 shrinkctl->nr_scanned = nr_to_scan;
398 ret = shrinker->scan_objects(shrinker, shrinkctl);
399 if (ret == SHRINK_STOP)
403 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
404 total_scan -= shrinkctl->nr_scanned;
405 scanned += shrinkctl->nr_scanned;
410 if (next_deferred >= scanned)
411 next_deferred -= scanned;
415 * move the unused scan count back into the shrinker in a
416 * manner that handles concurrent updates. If we exhausted the
417 * scan, there is no need to do an update.
419 if (next_deferred > 0)
420 new_nr = atomic_long_add_return(next_deferred,
421 &shrinker->nr_deferred[nid]);
423 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
425 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
430 * shrink_slab - shrink slab caches
431 * @gfp_mask: allocation context
432 * @nid: node whose slab caches to target
433 * @memcg: memory cgroup whose slab caches to target
434 * @priority: the reclaim priority
436 * Call the shrink functions to age shrinkable caches.
438 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
439 * unaware shrinkers will receive a node id of 0 instead.
441 * @memcg specifies the memory cgroup to target. If it is not NULL,
442 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
443 * objects from the memory cgroup specified. Otherwise, only unaware
444 * shrinkers are called.
446 * @priority is sc->priority, we take the number of objects and >> by priority
447 * in order to get the scan target.
449 * Returns the number of reclaimed slab objects.
451 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
452 struct mem_cgroup *memcg,
455 struct shrinker *shrinker;
456 unsigned long freed = 0;
458 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
461 if (!down_read_trylock(&shrinker_rwsem)) {
463 * If we would return 0, our callers would understand that we
464 * have nothing else to shrink and give up trying. By returning
465 * 1 we keep it going and assume we'll be able to shrink next
472 list_for_each_entry(shrinker, &shrinker_list, list) {
473 struct shrink_control sc = {
474 .gfp_mask = gfp_mask,
480 * If kernel memory accounting is disabled, we ignore
481 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
482 * passing NULL for memcg.
484 if (memcg_kmem_enabled() &&
485 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
488 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
491 freed += do_shrink_slab(&sc, shrinker, priority);
493 * Bail out if someone want to register a new shrinker to
494 * prevent the regsitration from being stalled for long periods
495 * by parallel ongoing shrinking.
497 if (rwsem_is_contended(&shrinker_rwsem)) {
503 up_read(&shrinker_rwsem);
509 void drop_slab_node(int nid)
514 struct mem_cgroup *memcg = NULL;
518 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
519 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
520 } while (freed > 10);
527 for_each_online_node(nid)
531 static inline int is_page_cache_freeable(struct page *page)
534 * A freeable page cache page is referenced only by the caller
535 * that isolated the page, the page cache radix tree and
536 * optional buffer heads at page->private.
538 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
540 return page_count(page) - page_has_private(page) == 1 + radix_pins;
543 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
545 if (current->flags & PF_SWAPWRITE)
547 if (!inode_write_congested(inode))
549 if (inode_to_bdi(inode) == current->backing_dev_info)
555 * We detected a synchronous write error writing a page out. Probably
556 * -ENOSPC. We need to propagate that into the address_space for a subsequent
557 * fsync(), msync() or close().
559 * The tricky part is that after writepage we cannot touch the mapping: nothing
560 * prevents it from being freed up. But we have a ref on the page and once
561 * that page is locked, the mapping is pinned.
563 * We're allowed to run sleeping lock_page() here because we know the caller has
566 static void handle_write_error(struct address_space *mapping,
567 struct page *page, int error)
570 if (page_mapping(page) == mapping)
571 mapping_set_error(mapping, error);
575 /* possible outcome of pageout() */
577 /* failed to write page out, page is locked */
579 /* move page to the active list, page is locked */
581 /* page has been sent to the disk successfully, page is unlocked */
583 /* page is clean and locked */
588 * pageout is called by shrink_page_list() for each dirty page.
589 * Calls ->writepage().
591 static pageout_t pageout(struct page *page, struct address_space *mapping,
592 struct scan_control *sc)
595 * If the page is dirty, only perform writeback if that write
596 * will be non-blocking. To prevent this allocation from being
597 * stalled by pagecache activity. But note that there may be
598 * stalls if we need to run get_block(). We could test
599 * PagePrivate for that.
601 * If this process is currently in __generic_file_write_iter() against
602 * this page's queue, we can perform writeback even if that
605 * If the page is swapcache, write it back even if that would
606 * block, for some throttling. This happens by accident, because
607 * swap_backing_dev_info is bust: it doesn't reflect the
608 * congestion state of the swapdevs. Easy to fix, if needed.
610 if (!is_page_cache_freeable(page))
614 * Some data journaling orphaned pages can have
615 * page->mapping == NULL while being dirty with clean buffers.
617 if (page_has_private(page)) {
618 if (try_to_free_buffers(page)) {
619 ClearPageDirty(page);
620 pr_info("%s: orphaned page\n", __func__);
626 if (mapping->a_ops->writepage == NULL)
627 return PAGE_ACTIVATE;
628 if (!may_write_to_inode(mapping->host, sc))
631 if (clear_page_dirty_for_io(page)) {
633 struct writeback_control wbc = {
634 .sync_mode = WB_SYNC_NONE,
635 .nr_to_write = SWAP_CLUSTER_MAX,
637 .range_end = LLONG_MAX,
641 SetPageReclaim(page);
642 res = mapping->a_ops->writepage(page, &wbc);
644 handle_write_error(mapping, page, res);
645 if (res == AOP_WRITEPAGE_ACTIVATE) {
646 ClearPageReclaim(page);
647 return PAGE_ACTIVATE;
650 if (!PageWriteback(page)) {
651 /* synchronous write or broken a_ops? */
652 ClearPageReclaim(page);
654 trace_mm_vmscan_writepage(page);
655 inc_node_page_state(page, NR_VMSCAN_WRITE);
663 * Same as remove_mapping, but if the page is removed from the mapping, it
664 * gets returned with a refcount of 0.
666 static int __remove_mapping(struct address_space *mapping, struct page *page,
672 BUG_ON(!PageLocked(page));
673 BUG_ON(mapping != page_mapping(page));
675 spin_lock_irqsave(&mapping->tree_lock, flags);
677 * The non racy check for a busy page.
679 * Must be careful with the order of the tests. When someone has
680 * a ref to the page, it may be possible that they dirty it then
681 * drop the reference. So if PageDirty is tested before page_count
682 * here, then the following race may occur:
684 * get_user_pages(&page);
685 * [user mapping goes away]
687 * !PageDirty(page) [good]
688 * SetPageDirty(page);
690 * !page_count(page) [good, discard it]
692 * [oops, our write_to data is lost]
694 * Reversing the order of the tests ensures such a situation cannot
695 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
696 * load is not satisfied before that of page->_refcount.
698 * Note that if SetPageDirty is always performed via set_page_dirty,
699 * and thus under tree_lock, then this ordering is not required.
701 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
702 refcount = 1 + HPAGE_PMD_NR;
705 if (!page_ref_freeze(page, refcount))
707 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
708 if (unlikely(PageDirty(page))) {
709 page_ref_unfreeze(page, refcount);
713 if (PageSwapCache(page)) {
714 swp_entry_t swap = { .val = page_private(page) };
715 mem_cgroup_swapout(page, swap);
716 __delete_from_swap_cache(page);
717 spin_unlock_irqrestore(&mapping->tree_lock, flags);
718 put_swap_page(page, swap);
720 void (*freepage)(struct page *);
723 freepage = mapping->a_ops->freepage;
725 * Remember a shadow entry for reclaimed file cache in
726 * order to detect refaults, thus thrashing, later on.
728 * But don't store shadows in an address space that is
729 * already exiting. This is not just an optizimation,
730 * inode reclaim needs to empty out the radix tree or
731 * the nodes are lost. Don't plant shadows behind its
734 * We also don't store shadows for DAX mappings because the
735 * only page cache pages found in these are zero pages
736 * covering holes, and because we don't want to mix DAX
737 * exceptional entries and shadow exceptional entries in the
740 if (reclaimed && page_is_file_cache(page) &&
741 !mapping_exiting(mapping) && !dax_mapping(mapping))
742 shadow = workingset_eviction(mapping, page);
743 __delete_from_page_cache(page, shadow);
744 spin_unlock_irqrestore(&mapping->tree_lock, flags);
746 if (freepage != NULL)
753 spin_unlock_irqrestore(&mapping->tree_lock, flags);
758 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
759 * someone else has a ref on the page, abort and return 0. If it was
760 * successfully detached, return 1. Assumes the caller has a single ref on
763 int remove_mapping(struct address_space *mapping, struct page *page)
765 if (__remove_mapping(mapping, page, false)) {
767 * Unfreezing the refcount with 1 rather than 2 effectively
768 * drops the pagecache ref for us without requiring another
771 page_ref_unfreeze(page, 1);
778 * putback_lru_page - put previously isolated page onto appropriate LRU list
779 * @page: page to be put back to appropriate lru list
781 * Add previously isolated @page to appropriate LRU list.
782 * Page may still be unevictable for other reasons.
784 * lru_lock must not be held, interrupts must be enabled.
786 void putback_lru_page(struct page *page)
789 int was_unevictable = PageUnevictable(page);
791 VM_BUG_ON_PAGE(PageLRU(page), page);
794 ClearPageUnevictable(page);
796 if (page_evictable(page)) {
798 * For evictable pages, we can use the cache.
799 * In event of a race, worst case is we end up with an
800 * unevictable page on [in]active list.
801 * We know how to handle that.
803 is_unevictable = false;
807 * Put unevictable pages directly on zone's unevictable
810 is_unevictable = true;
811 add_page_to_unevictable_list(page);
813 * When racing with an mlock or AS_UNEVICTABLE clearing
814 * (page is unlocked) make sure that if the other thread
815 * does not observe our setting of PG_lru and fails
816 * isolation/check_move_unevictable_pages,
817 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
818 * the page back to the evictable list.
820 * The other side is TestClearPageMlocked() or shmem_lock().
826 * page's status can change while we move it among lru. If an evictable
827 * page is on unevictable list, it never be freed. To avoid that,
828 * check after we added it to the list, again.
830 if (is_unevictable && page_evictable(page)) {
831 if (!isolate_lru_page(page)) {
835 /* This means someone else dropped this page from LRU
836 * So, it will be freed or putback to LRU again. There is
837 * nothing to do here.
841 if (was_unevictable && !is_unevictable)
842 count_vm_event(UNEVICTABLE_PGRESCUED);
843 else if (!was_unevictable && is_unevictable)
844 count_vm_event(UNEVICTABLE_PGCULLED);
846 put_page(page); /* drop ref from isolate */
849 enum page_references {
851 PAGEREF_RECLAIM_CLEAN,
856 static enum page_references page_check_references(struct page *page,
857 struct scan_control *sc)
859 int referenced_ptes, referenced_page;
860 unsigned long vm_flags;
862 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
864 referenced_page = TestClearPageReferenced(page);
867 * Mlock lost the isolation race with us. Let try_to_unmap()
868 * move the page to the unevictable list.
870 if (vm_flags & VM_LOCKED)
871 return PAGEREF_RECLAIM;
873 if (referenced_ptes) {
874 if (PageSwapBacked(page))
875 return PAGEREF_ACTIVATE;
877 * All mapped pages start out with page table
878 * references from the instantiating fault, so we need
879 * to look twice if a mapped file page is used more
882 * Mark it and spare it for another trip around the
883 * inactive list. Another page table reference will
884 * lead to its activation.
886 * Note: the mark is set for activated pages as well
887 * so that recently deactivated but used pages are
890 SetPageReferenced(page);
892 if (referenced_page || referenced_ptes > 1)
893 return PAGEREF_ACTIVATE;
896 * Activate file-backed executable pages after first usage.
898 if (vm_flags & VM_EXEC)
899 return PAGEREF_ACTIVATE;
904 /* Reclaim if clean, defer dirty pages to writeback */
905 if (referenced_page && !PageSwapBacked(page))
906 return PAGEREF_RECLAIM_CLEAN;
908 return PAGEREF_RECLAIM;
911 /* Check if a page is dirty or under writeback */
912 static void page_check_dirty_writeback(struct page *page,
913 bool *dirty, bool *writeback)
915 struct address_space *mapping;
918 * Anonymous pages are not handled by flushers and must be written
919 * from reclaim context. Do not stall reclaim based on them
921 if (!page_is_file_cache(page) ||
922 (PageAnon(page) && !PageSwapBacked(page))) {
928 /* By default assume that the page flags are accurate */
929 *dirty = PageDirty(page);
930 *writeback = PageWriteback(page);
932 /* Verify dirty/writeback state if the filesystem supports it */
933 if (!page_has_private(page))
936 mapping = page_mapping(page);
937 if (mapping && mapping->a_ops->is_dirty_writeback)
938 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
941 struct reclaim_stat {
943 unsigned nr_unqueued_dirty;
944 unsigned nr_congested;
945 unsigned nr_writeback;
946 unsigned nr_immediate;
947 unsigned nr_activate;
948 unsigned nr_ref_keep;
949 unsigned nr_unmap_fail;
953 * shrink_page_list() returns the number of reclaimed pages
955 static unsigned long shrink_page_list(struct list_head *page_list,
956 struct pglist_data *pgdat,
957 struct scan_control *sc,
958 enum ttu_flags ttu_flags,
959 struct reclaim_stat *stat,
962 LIST_HEAD(ret_pages);
963 LIST_HEAD(free_pages);
965 unsigned nr_unqueued_dirty = 0;
966 unsigned nr_dirty = 0;
967 unsigned nr_congested = 0;
968 unsigned nr_reclaimed = 0;
969 unsigned nr_writeback = 0;
970 unsigned nr_immediate = 0;
971 unsigned nr_ref_keep = 0;
972 unsigned nr_unmap_fail = 0;
976 while (!list_empty(page_list)) {
977 struct address_space *mapping;
980 enum page_references references = PAGEREF_RECLAIM_CLEAN;
981 bool dirty, writeback;
985 page = lru_to_page(page_list);
986 list_del(&page->lru);
988 if (!trylock_page(page))
991 VM_BUG_ON_PAGE(PageActive(page), page);
995 if (unlikely(!page_evictable(page)))
996 goto activate_locked;
998 if (!sc->may_unmap && page_mapped(page))
1001 /* Double the slab pressure for mapped and swapcache pages */
1002 if ((page_mapped(page) || PageSwapCache(page)) &&
1003 !(PageAnon(page) && !PageSwapBacked(page)))
1006 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1007 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1010 * The number of dirty pages determines if a zone is marked
1011 * reclaim_congested which affects wait_iff_congested. kswapd
1012 * will stall and start writing pages if the tail of the LRU
1013 * is all dirty unqueued pages.
1015 page_check_dirty_writeback(page, &dirty, &writeback);
1016 if (dirty || writeback)
1019 if (dirty && !writeback)
1020 nr_unqueued_dirty++;
1023 * Treat this page as congested if the underlying BDI is or if
1024 * pages are cycling through the LRU so quickly that the
1025 * pages marked for immediate reclaim are making it to the
1026 * end of the LRU a second time.
1028 mapping = page_mapping(page);
1029 if (((dirty || writeback) && mapping &&
1030 inode_write_congested(mapping->host)) ||
1031 (writeback && PageReclaim(page)))
1035 * If a page at the tail of the LRU is under writeback, there
1036 * are three cases to consider.
1038 * 1) If reclaim is encountering an excessive number of pages
1039 * under writeback and this page is both under writeback and
1040 * PageReclaim then it indicates that pages are being queued
1041 * for IO but are being recycled through the LRU before the
1042 * IO can complete. Waiting on the page itself risks an
1043 * indefinite stall if it is impossible to writeback the
1044 * page due to IO error or disconnected storage so instead
1045 * note that the LRU is being scanned too quickly and the
1046 * caller can stall after page list has been processed.
1048 * 2) Global or new memcg reclaim encounters a page that is
1049 * not marked for immediate reclaim, or the caller does not
1050 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1051 * not to fs). In this case mark the page for immediate
1052 * reclaim and continue scanning.
1054 * Require may_enter_fs because we would wait on fs, which
1055 * may not have submitted IO yet. And the loop driver might
1056 * enter reclaim, and deadlock if it waits on a page for
1057 * which it is needed to do the write (loop masks off
1058 * __GFP_IO|__GFP_FS for this reason); but more thought
1059 * would probably show more reasons.
1061 * 3) Legacy memcg encounters a page that is already marked
1062 * PageReclaim. memcg does not have any dirty pages
1063 * throttling so we could easily OOM just because too many
1064 * pages are in writeback and there is nothing else to
1065 * reclaim. Wait for the writeback to complete.
1067 * In cases 1) and 2) we activate the pages to get them out of
1068 * the way while we continue scanning for clean pages on the
1069 * inactive list and refilling from the active list. The
1070 * observation here is that waiting for disk writes is more
1071 * expensive than potentially causing reloads down the line.
1072 * Since they're marked for immediate reclaim, they won't put
1073 * memory pressure on the cache working set any longer than it
1074 * takes to write them to disk.
1076 if (PageWriteback(page)) {
1078 if (current_is_kswapd() &&
1079 PageReclaim(page) &&
1080 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1082 goto activate_locked;
1085 } else if (sane_reclaim(sc) ||
1086 !PageReclaim(page) || !may_enter_fs) {
1088 * This is slightly racy - end_page_writeback()
1089 * might have just cleared PageReclaim, then
1090 * setting PageReclaim here end up interpreted
1091 * as PageReadahead - but that does not matter
1092 * enough to care. What we do want is for this
1093 * page to have PageReclaim set next time memcg
1094 * reclaim reaches the tests above, so it will
1095 * then wait_on_page_writeback() to avoid OOM;
1096 * and it's also appropriate in global reclaim.
1098 SetPageReclaim(page);
1100 goto activate_locked;
1105 wait_on_page_writeback(page);
1106 /* then go back and try same page again */
1107 list_add_tail(&page->lru, page_list);
1113 references = page_check_references(page, sc);
1115 switch (references) {
1116 case PAGEREF_ACTIVATE:
1117 goto activate_locked;
1121 case PAGEREF_RECLAIM:
1122 case PAGEREF_RECLAIM_CLEAN:
1123 ; /* try to reclaim the page below */
1127 * Anonymous process memory has backing store?
1128 * Try to allocate it some swap space here.
1129 * Lazyfree page could be freed directly
1131 if (PageAnon(page) && PageSwapBacked(page)) {
1132 if (!PageSwapCache(page)) {
1133 if (!(sc->gfp_mask & __GFP_IO))
1135 if (PageTransHuge(page)) {
1136 /* cannot split THP, skip it */
1137 if (!can_split_huge_page(page, NULL))
1138 goto activate_locked;
1140 * Split pages without a PMD map right
1141 * away. Chances are some or all of the
1142 * tail pages can be freed without IO.
1144 if (!compound_mapcount(page) &&
1145 split_huge_page_to_list(page,
1147 goto activate_locked;
1149 if (!add_to_swap(page)) {
1150 if (!PageTransHuge(page))
1151 goto activate_locked;
1152 /* Fallback to swap normal pages */
1153 if (split_huge_page_to_list(page,
1155 goto activate_locked;
1156 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1157 count_vm_event(THP_SWPOUT_FALLBACK);
1159 if (!add_to_swap(page))
1160 goto activate_locked;
1165 /* Adding to swap updated mapping */
1166 mapping = page_mapping(page);
1168 } else if (unlikely(PageTransHuge(page))) {
1169 /* Split file THP */
1170 if (split_huge_page_to_list(page, page_list))
1175 * The page is mapped into the page tables of one or more
1176 * processes. Try to unmap it here.
1178 if (page_mapped(page)) {
1179 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1181 if (unlikely(PageTransHuge(page)))
1182 flags |= TTU_SPLIT_HUGE_PMD;
1183 if (!try_to_unmap(page, flags)) {
1185 goto activate_locked;
1189 if (PageDirty(page)) {
1191 * Only kswapd can writeback filesystem pages
1192 * to avoid risk of stack overflow. But avoid
1193 * injecting inefficient single-page IO into
1194 * flusher writeback as much as possible: only
1195 * write pages when we've encountered many
1196 * dirty pages, and when we've already scanned
1197 * the rest of the LRU for clean pages and see
1198 * the same dirty pages again (PageReclaim).
1200 if (page_is_file_cache(page) &&
1201 (!current_is_kswapd() || !PageReclaim(page) ||
1202 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1204 * Immediately reclaim when written back.
1205 * Similar in principal to deactivate_page()
1206 * except we already have the page isolated
1207 * and know it's dirty
1209 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1210 SetPageReclaim(page);
1212 goto activate_locked;
1215 if (references == PAGEREF_RECLAIM_CLEAN)
1219 if (!sc->may_writepage)
1223 * Page is dirty. Flush the TLB if a writable entry
1224 * potentially exists to avoid CPU writes after IO
1225 * starts and then write it out here.
1227 try_to_unmap_flush_dirty();
1228 switch (pageout(page, mapping, sc)) {
1232 goto activate_locked;
1234 if (PageWriteback(page))
1236 if (PageDirty(page))
1240 * A synchronous write - probably a ramdisk. Go
1241 * ahead and try to reclaim the page.
1243 if (!trylock_page(page))
1245 if (PageDirty(page) || PageWriteback(page))
1247 mapping = page_mapping(page);
1249 ; /* try to free the page below */
1254 * If the page has buffers, try to free the buffer mappings
1255 * associated with this page. If we succeed we try to free
1258 * We do this even if the page is PageDirty().
1259 * try_to_release_page() does not perform I/O, but it is
1260 * possible for a page to have PageDirty set, but it is actually
1261 * clean (all its buffers are clean). This happens if the
1262 * buffers were written out directly, with submit_bh(). ext3
1263 * will do this, as well as the blockdev mapping.
1264 * try_to_release_page() will discover that cleanness and will
1265 * drop the buffers and mark the page clean - it can be freed.
1267 * Rarely, pages can have buffers and no ->mapping. These are
1268 * the pages which were not successfully invalidated in
1269 * truncate_complete_page(). We try to drop those buffers here
1270 * and if that worked, and the page is no longer mapped into
1271 * process address space (page_count == 1) it can be freed.
1272 * Otherwise, leave the page on the LRU so it is swappable.
1274 if (page_has_private(page)) {
1275 if (!try_to_release_page(page, sc->gfp_mask))
1276 goto activate_locked;
1277 if (!mapping && page_count(page) == 1) {
1279 if (put_page_testzero(page))
1283 * rare race with speculative reference.
1284 * the speculative reference will free
1285 * this page shortly, so we may
1286 * increment nr_reclaimed here (and
1287 * leave it off the LRU).
1295 if (PageAnon(page) && !PageSwapBacked(page)) {
1296 /* follow __remove_mapping for reference */
1297 if (!page_ref_freeze(page, 1))
1299 if (PageDirty(page)) {
1300 page_ref_unfreeze(page, 1);
1304 count_vm_event(PGLAZYFREED);
1305 count_memcg_page_event(page, PGLAZYFREED);
1306 } else if (!mapping || !__remove_mapping(mapping, page, true))
1309 * At this point, we have no other references and there is
1310 * no way to pick any more up (removed from LRU, removed
1311 * from pagecache). Can use non-atomic bitops now (and
1312 * we obviously don't have to worry about waking up a process
1313 * waiting on the page lock, because there are no references.
1315 __ClearPageLocked(page);
1320 * Is there need to periodically free_page_list? It would
1321 * appear not as the counts should be low
1323 if (unlikely(PageTransHuge(page))) {
1324 mem_cgroup_uncharge(page);
1325 (*get_compound_page_dtor(page))(page);
1327 list_add(&page->lru, &free_pages);
1331 /* Not a candidate for swapping, so reclaim swap space. */
1332 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1334 try_to_free_swap(page);
1335 VM_BUG_ON_PAGE(PageActive(page), page);
1336 if (!PageMlocked(page)) {
1337 SetPageActive(page);
1339 count_memcg_page_event(page, PGACTIVATE);
1344 list_add(&page->lru, &ret_pages);
1345 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1348 mem_cgroup_uncharge_list(&free_pages);
1349 try_to_unmap_flush();
1350 free_unref_page_list(&free_pages);
1352 list_splice(&ret_pages, page_list);
1353 count_vm_events(PGACTIVATE, pgactivate);
1356 stat->nr_dirty = nr_dirty;
1357 stat->nr_congested = nr_congested;
1358 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1359 stat->nr_writeback = nr_writeback;
1360 stat->nr_immediate = nr_immediate;
1361 stat->nr_activate = pgactivate;
1362 stat->nr_ref_keep = nr_ref_keep;
1363 stat->nr_unmap_fail = nr_unmap_fail;
1365 return nr_reclaimed;
1368 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1369 struct list_head *page_list)
1371 struct scan_control sc = {
1372 .gfp_mask = GFP_KERNEL,
1373 .priority = DEF_PRIORITY,
1377 struct page *page, *next;
1378 LIST_HEAD(clean_pages);
1380 list_for_each_entry_safe(page, next, page_list, lru) {
1381 if (page_is_file_cache(page) && !PageDirty(page) &&
1382 !__PageMovable(page)) {
1383 ClearPageActive(page);
1384 list_move(&page->lru, &clean_pages);
1388 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1389 TTU_IGNORE_ACCESS, NULL, true);
1390 list_splice(&clean_pages, page_list);
1391 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1396 * Attempt to remove the specified page from its LRU. Only take this page
1397 * if it is of the appropriate PageActive status. Pages which are being
1398 * freed elsewhere are also ignored.
1400 * page: page to consider
1401 * mode: one of the LRU isolation modes defined above
1403 * returns 0 on success, -ve errno on failure.
1405 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1409 /* Only take pages on the LRU. */
1413 /* Compaction should not handle unevictable pages but CMA can do so */
1414 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1420 * To minimise LRU disruption, the caller can indicate that it only
1421 * wants to isolate pages it will be able to operate on without
1422 * blocking - clean pages for the most part.
1424 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1425 * that it is possible to migrate without blocking
1427 if (mode & ISOLATE_ASYNC_MIGRATE) {
1428 /* All the caller can do on PageWriteback is block */
1429 if (PageWriteback(page))
1432 if (PageDirty(page)) {
1433 struct address_space *mapping;
1436 * Only pages without mappings or that have a
1437 * ->migratepage callback are possible to migrate
1440 mapping = page_mapping(page);
1441 if (mapping && !mapping->a_ops->migratepage)
1446 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1449 if (likely(get_page_unless_zero(page))) {
1451 * Be careful not to clear PageLRU until after we're
1452 * sure the page is not being freed elsewhere -- the
1453 * page release code relies on it.
1464 * Update LRU sizes after isolating pages. The LRU size updates must
1465 * be complete before mem_cgroup_update_lru_size due to a santity check.
1467 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1468 enum lru_list lru, unsigned long *nr_zone_taken)
1472 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1473 if (!nr_zone_taken[zid])
1476 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1478 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1485 * zone_lru_lock is heavily contended. Some of the functions that
1486 * shrink the lists perform better by taking out a batch of pages
1487 * and working on them outside the LRU lock.
1489 * For pagecache intensive workloads, this function is the hottest
1490 * spot in the kernel (apart from copy_*_user functions).
1492 * Appropriate locks must be held before calling this function.
1494 * @nr_to_scan: The number of eligible pages to look through on the list.
1495 * @lruvec: The LRU vector to pull pages from.
1496 * @dst: The temp list to put pages on to.
1497 * @nr_scanned: The number of pages that were scanned.
1498 * @sc: The scan_control struct for this reclaim session
1499 * @mode: One of the LRU isolation modes
1500 * @lru: LRU list id for isolating
1502 * returns how many pages were moved onto *@dst.
1504 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1505 struct lruvec *lruvec, struct list_head *dst,
1506 unsigned long *nr_scanned, struct scan_control *sc,
1507 isolate_mode_t mode, enum lru_list lru)
1509 struct list_head *src = &lruvec->lists[lru];
1510 unsigned long nr_taken = 0;
1511 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1512 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1513 unsigned long skipped = 0;
1514 unsigned long scan, total_scan, nr_pages;
1515 LIST_HEAD(pages_skipped);
1518 for (total_scan = 0;
1519 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1523 page = lru_to_page(src);
1524 prefetchw_prev_lru_page(page, src, flags);
1526 VM_BUG_ON_PAGE(!PageLRU(page), page);
1528 if (page_zonenum(page) > sc->reclaim_idx) {
1529 list_move(&page->lru, &pages_skipped);
1530 nr_skipped[page_zonenum(page)]++;
1535 * Do not count skipped pages because that makes the function
1536 * return with no isolated pages if the LRU mostly contains
1537 * ineligible pages. This causes the VM to not reclaim any
1538 * pages, triggering a premature OOM.
1541 switch (__isolate_lru_page(page, mode)) {
1543 nr_pages = hpage_nr_pages(page);
1544 nr_taken += nr_pages;
1545 nr_zone_taken[page_zonenum(page)] += nr_pages;
1546 list_move(&page->lru, dst);
1550 /* else it is being freed elsewhere */
1551 list_move(&page->lru, src);
1560 * Splice any skipped pages to the start of the LRU list. Note that
1561 * this disrupts the LRU order when reclaiming for lower zones but
1562 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1563 * scanning would soon rescan the same pages to skip and put the
1564 * system at risk of premature OOM.
1566 if (!list_empty(&pages_skipped)) {
1569 list_splice(&pages_skipped, src);
1570 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1571 if (!nr_skipped[zid])
1574 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1575 skipped += nr_skipped[zid];
1578 *nr_scanned = total_scan;
1579 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1580 total_scan, skipped, nr_taken, mode, lru);
1581 update_lru_sizes(lruvec, lru, nr_zone_taken);
1586 * isolate_lru_page - tries to isolate a page from its LRU list
1587 * @page: page to isolate from its LRU list
1589 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1590 * vmstat statistic corresponding to whatever LRU list the page was on.
1592 * Returns 0 if the page was removed from an LRU list.
1593 * Returns -EBUSY if the page was not on an LRU list.
1595 * The returned page will have PageLRU() cleared. If it was found on
1596 * the active list, it will have PageActive set. If it was found on
1597 * the unevictable list, it will have the PageUnevictable bit set. That flag
1598 * may need to be cleared by the caller before letting the page go.
1600 * The vmstat statistic corresponding to the list on which the page was
1601 * found will be decremented.
1604 * (1) Must be called with an elevated refcount on the page. This is a
1605 * fundamentnal difference from isolate_lru_pages (which is called
1606 * without a stable reference).
1607 * (2) the lru_lock must not be held.
1608 * (3) interrupts must be enabled.
1610 int isolate_lru_page(struct page *page)
1614 VM_BUG_ON_PAGE(!page_count(page), page);
1615 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1617 if (PageLRU(page)) {
1618 struct zone *zone = page_zone(page);
1619 struct lruvec *lruvec;
1621 spin_lock_irq(zone_lru_lock(zone));
1622 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1623 if (PageLRU(page)) {
1624 int lru = page_lru(page);
1627 del_page_from_lru_list(page, lruvec, lru);
1630 spin_unlock_irq(zone_lru_lock(zone));
1636 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1637 * then get resheduled. When there are massive number of tasks doing page
1638 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1639 * the LRU list will go small and be scanned faster than necessary, leading to
1640 * unnecessary swapping, thrashing and OOM.
1642 static int too_many_isolated(struct pglist_data *pgdat, int file,
1643 struct scan_control *sc)
1645 unsigned long inactive, isolated;
1647 if (current_is_kswapd())
1650 if (!sane_reclaim(sc))
1654 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1655 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1657 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1658 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1662 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1663 * won't get blocked by normal direct-reclaimers, forming a circular
1666 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1669 return isolated > inactive;
1672 static noinline_for_stack void
1673 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1675 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1676 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1677 LIST_HEAD(pages_to_free);
1680 * Put back any unfreeable pages.
1682 while (!list_empty(page_list)) {
1683 struct page *page = lru_to_page(page_list);
1686 VM_BUG_ON_PAGE(PageLRU(page), page);
1687 list_del(&page->lru);
1688 if (unlikely(!page_evictable(page))) {
1689 spin_unlock_irq(&pgdat->lru_lock);
1690 putback_lru_page(page);
1691 spin_lock_irq(&pgdat->lru_lock);
1695 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1698 lru = page_lru(page);
1699 add_page_to_lru_list(page, lruvec, lru);
1701 if (is_active_lru(lru)) {
1702 int file = is_file_lru(lru);
1703 int numpages = hpage_nr_pages(page);
1704 reclaim_stat->recent_rotated[file] += numpages;
1706 if (put_page_testzero(page)) {
1707 __ClearPageLRU(page);
1708 __ClearPageActive(page);
1709 del_page_from_lru_list(page, lruvec, lru);
1711 if (unlikely(PageCompound(page))) {
1712 spin_unlock_irq(&pgdat->lru_lock);
1713 mem_cgroup_uncharge(page);
1714 (*get_compound_page_dtor(page))(page);
1715 spin_lock_irq(&pgdat->lru_lock);
1717 list_add(&page->lru, &pages_to_free);
1722 * To save our caller's stack, now use input list for pages to free.
1724 list_splice(&pages_to_free, page_list);
1728 * If a kernel thread (such as nfsd for loop-back mounts) services
1729 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1730 * In that case we should only throttle if the backing device it is
1731 * writing to is congested. In other cases it is safe to throttle.
1733 static int current_may_throttle(void)
1735 return !(current->flags & PF_LESS_THROTTLE) ||
1736 current->backing_dev_info == NULL ||
1737 bdi_write_congested(current->backing_dev_info);
1741 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1742 * of reclaimed pages
1744 static noinline_for_stack unsigned long
1745 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1746 struct scan_control *sc, enum lru_list lru)
1748 LIST_HEAD(page_list);
1749 unsigned long nr_scanned;
1750 unsigned long nr_reclaimed = 0;
1751 unsigned long nr_taken;
1752 struct reclaim_stat stat = {};
1753 isolate_mode_t isolate_mode = 0;
1754 int file = is_file_lru(lru);
1755 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1756 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1757 bool stalled = false;
1759 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1763 /* wait a bit for the reclaimer. */
1767 /* We are about to die and free our memory. Return now. */
1768 if (fatal_signal_pending(current))
1769 return SWAP_CLUSTER_MAX;
1775 isolate_mode |= ISOLATE_UNMAPPED;
1777 spin_lock_irq(&pgdat->lru_lock);
1779 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1780 &nr_scanned, sc, isolate_mode, lru);
1782 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1783 reclaim_stat->recent_scanned[file] += nr_taken;
1785 if (current_is_kswapd()) {
1786 if (global_reclaim(sc))
1787 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1788 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1791 if (global_reclaim(sc))
1792 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1793 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1796 spin_unlock_irq(&pgdat->lru_lock);
1801 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1804 spin_lock_irq(&pgdat->lru_lock);
1806 if (current_is_kswapd()) {
1807 if (global_reclaim(sc))
1808 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1809 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1812 if (global_reclaim(sc))
1813 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1814 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1818 putback_inactive_pages(lruvec, &page_list);
1820 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1822 spin_unlock_irq(&pgdat->lru_lock);
1824 mem_cgroup_uncharge_list(&page_list);
1825 free_unref_page_list(&page_list);
1828 * If reclaim is isolating dirty pages under writeback, it implies
1829 * that the long-lived page allocation rate is exceeding the page
1830 * laundering rate. Either the global limits are not being effective
1831 * at throttling processes due to the page distribution throughout
1832 * zones or there is heavy usage of a slow backing device. The
1833 * only option is to throttle from reclaim context which is not ideal
1834 * as there is no guarantee the dirtying process is throttled in the
1835 * same way balance_dirty_pages() manages.
1837 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1838 * of pages under pages flagged for immediate reclaim and stall if any
1839 * are encountered in the nr_immediate check below.
1841 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1842 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1845 * Legacy memcg will stall in page writeback so avoid forcibly
1848 if (sane_reclaim(sc)) {
1850 * Tag a zone as congested if all the dirty pages scanned were
1851 * backed by a congested BDI and wait_iff_congested will stall.
1853 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1854 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1857 * If dirty pages are scanned that are not queued for IO, it
1858 * implies that flushers are not doing their job. This can
1859 * happen when memory pressure pushes dirty pages to the end of
1860 * the LRU before the dirty limits are breached and the dirty
1861 * data has expired. It can also happen when the proportion of
1862 * dirty pages grows not through writes but through memory
1863 * pressure reclaiming all the clean cache. And in some cases,
1864 * the flushers simply cannot keep up with the allocation
1865 * rate. Nudge the flusher threads in case they are asleep, but
1866 * also allow kswapd to start writing pages during reclaim.
1868 if (stat.nr_unqueued_dirty == nr_taken) {
1869 wakeup_flusher_threads(WB_REASON_VMSCAN);
1870 set_bit(PGDAT_DIRTY, &pgdat->flags);
1874 * If kswapd scans pages marked marked for immediate
1875 * reclaim and under writeback (nr_immediate), it implies
1876 * that pages are cycling through the LRU faster than
1877 * they are written so also forcibly stall.
1879 if (stat.nr_immediate && current_may_throttle())
1880 congestion_wait(BLK_RW_ASYNC, HZ/10);
1884 * Stall direct reclaim for IO completions if underlying BDIs or zone
1885 * is congested. Allow kswapd to continue until it starts encountering
1886 * unqueued dirty pages or cycling through the LRU too quickly.
1888 if (!sc->hibernation_mode && !current_is_kswapd() &&
1889 current_may_throttle())
1890 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1892 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1893 nr_scanned, nr_reclaimed,
1894 stat.nr_dirty, stat.nr_writeback,
1895 stat.nr_congested, stat.nr_immediate,
1896 stat.nr_activate, stat.nr_ref_keep,
1898 sc->priority, file);
1899 return nr_reclaimed;
1903 * This moves pages from the active list to the inactive list.
1905 * We move them the other way if the page is referenced by one or more
1906 * processes, from rmap.
1908 * If the pages are mostly unmapped, the processing is fast and it is
1909 * appropriate to hold zone_lru_lock across the whole operation. But if
1910 * the pages are mapped, the processing is slow (page_referenced()) so we
1911 * should drop zone_lru_lock around each page. It's impossible to balance
1912 * this, so instead we remove the pages from the LRU while processing them.
1913 * It is safe to rely on PG_active against the non-LRU pages in here because
1914 * nobody will play with that bit on a non-LRU page.
1916 * The downside is that we have to touch page->_refcount against each page.
1917 * But we had to alter page->flags anyway.
1919 * Returns the number of pages moved to the given lru.
1922 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1923 struct list_head *list,
1924 struct list_head *pages_to_free,
1927 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1932 while (!list_empty(list)) {
1933 page = lru_to_page(list);
1934 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1936 VM_BUG_ON_PAGE(PageLRU(page), page);
1939 nr_pages = hpage_nr_pages(page);
1940 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1941 list_move(&page->lru, &lruvec->lists[lru]);
1943 if (put_page_testzero(page)) {
1944 __ClearPageLRU(page);
1945 __ClearPageActive(page);
1946 del_page_from_lru_list(page, lruvec, lru);
1948 if (unlikely(PageCompound(page))) {
1949 spin_unlock_irq(&pgdat->lru_lock);
1950 mem_cgroup_uncharge(page);
1951 (*get_compound_page_dtor(page))(page);
1952 spin_lock_irq(&pgdat->lru_lock);
1954 list_add(&page->lru, pages_to_free);
1956 nr_moved += nr_pages;
1960 if (!is_active_lru(lru)) {
1961 __count_vm_events(PGDEACTIVATE, nr_moved);
1962 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1969 static void shrink_active_list(unsigned long nr_to_scan,
1970 struct lruvec *lruvec,
1971 struct scan_control *sc,
1974 unsigned long nr_taken;
1975 unsigned long nr_scanned;
1976 unsigned long vm_flags;
1977 LIST_HEAD(l_hold); /* The pages which were snipped off */
1978 LIST_HEAD(l_active);
1979 LIST_HEAD(l_inactive);
1981 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1982 unsigned nr_deactivate, nr_activate;
1983 unsigned nr_rotated = 0;
1984 isolate_mode_t isolate_mode = 0;
1985 int file = is_file_lru(lru);
1986 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1991 isolate_mode |= ISOLATE_UNMAPPED;
1993 spin_lock_irq(&pgdat->lru_lock);
1995 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1996 &nr_scanned, sc, isolate_mode, lru);
1998 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1999 reclaim_stat->recent_scanned[file] += nr_taken;
2001 __count_vm_events(PGREFILL, nr_scanned);
2002 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2004 spin_unlock_irq(&pgdat->lru_lock);
2006 while (!list_empty(&l_hold)) {
2008 page = lru_to_page(&l_hold);
2009 list_del(&page->lru);
2011 if (unlikely(!page_evictable(page))) {
2012 putback_lru_page(page);
2016 if (unlikely(buffer_heads_over_limit)) {
2017 if (page_has_private(page) && trylock_page(page)) {
2018 if (page_has_private(page))
2019 try_to_release_page(page, 0);
2024 if (page_referenced(page, 0, sc->target_mem_cgroup,
2026 nr_rotated += hpage_nr_pages(page);
2028 * Identify referenced, file-backed active pages and
2029 * give them one more trip around the active list. So
2030 * that executable code get better chances to stay in
2031 * memory under moderate memory pressure. Anon pages
2032 * are not likely to be evicted by use-once streaming
2033 * IO, plus JVM can create lots of anon VM_EXEC pages,
2034 * so we ignore them here.
2036 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2037 list_add(&page->lru, &l_active);
2042 ClearPageActive(page); /* we are de-activating */
2043 list_add(&page->lru, &l_inactive);
2047 * Move pages back to the lru list.
2049 spin_lock_irq(&pgdat->lru_lock);
2051 * Count referenced pages from currently used mappings as rotated,
2052 * even though only some of them are actually re-activated. This
2053 * helps balance scan pressure between file and anonymous pages in
2056 reclaim_stat->recent_rotated[file] += nr_rotated;
2058 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2059 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2060 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2061 spin_unlock_irq(&pgdat->lru_lock);
2063 mem_cgroup_uncharge_list(&l_hold);
2064 free_unref_page_list(&l_hold);
2065 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2066 nr_deactivate, nr_rotated, sc->priority, file);
2070 * The inactive anon list should be small enough that the VM never has
2071 * to do too much work.
2073 * The inactive file list should be small enough to leave most memory
2074 * to the established workingset on the scan-resistant active list,
2075 * but large enough to avoid thrashing the aggregate readahead window.
2077 * Both inactive lists should also be large enough that each inactive
2078 * page has a chance to be referenced again before it is reclaimed.
2080 * If that fails and refaulting is observed, the inactive list grows.
2082 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2083 * on this LRU, maintained by the pageout code. An inactive_ratio
2084 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2087 * memory ratio inactive
2088 * -------------------------------------
2097 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2098 struct mem_cgroup *memcg,
2099 struct scan_control *sc, bool actual_reclaim)
2101 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2102 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2103 enum lru_list inactive_lru = file * LRU_FILE;
2104 unsigned long inactive, active;
2105 unsigned long inactive_ratio;
2106 unsigned long refaults;
2110 * If we don't have swap space, anonymous page deactivation
2113 if (!file && !total_swap_pages)
2116 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2117 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2120 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2122 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2125 * When refaults are being observed, it means a new workingset
2126 * is being established. Disable active list protection to get
2127 * rid of the stale workingset quickly.
2129 if (file && actual_reclaim && lruvec->refaults != refaults) {
2132 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2134 inactive_ratio = int_sqrt(10 * gb);
2140 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2141 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2142 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2143 inactive_ratio, file);
2145 return inactive * inactive_ratio < active;
2148 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2149 struct lruvec *lruvec, struct mem_cgroup *memcg,
2150 struct scan_control *sc)
2152 if (is_active_lru(lru)) {
2153 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2155 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2159 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2170 * Determine how aggressively the anon and file LRU lists should be
2171 * scanned. The relative value of each set of LRU lists is determined
2172 * by looking at the fraction of the pages scanned we did rotate back
2173 * onto the active list instead of evict.
2175 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2176 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2178 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2179 struct scan_control *sc, unsigned long *nr,
2180 unsigned long *lru_pages)
2182 int swappiness = mem_cgroup_swappiness(memcg);
2183 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2185 u64 denominator = 0; /* gcc */
2186 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2187 unsigned long anon_prio, file_prio;
2188 enum scan_balance scan_balance;
2189 unsigned long anon, file;
2190 unsigned long ap, fp;
2193 /* If we have no swap space, do not bother scanning anon pages. */
2194 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2195 scan_balance = SCAN_FILE;
2200 * Global reclaim will swap to prevent OOM even with no
2201 * swappiness, but memcg users want to use this knob to
2202 * disable swapping for individual groups completely when
2203 * using the memory controller's swap limit feature would be
2206 if (!global_reclaim(sc) && !swappiness) {
2207 scan_balance = SCAN_FILE;
2212 * Do not apply any pressure balancing cleverness when the
2213 * system is close to OOM, scan both anon and file equally
2214 * (unless the swappiness setting disagrees with swapping).
2216 if (!sc->priority && swappiness) {
2217 scan_balance = SCAN_EQUAL;
2222 * Prevent the reclaimer from falling into the cache trap: as
2223 * cache pages start out inactive, every cache fault will tip
2224 * the scan balance towards the file LRU. And as the file LRU
2225 * shrinks, so does the window for rotation from references.
2226 * This means we have a runaway feedback loop where a tiny
2227 * thrashing file LRU becomes infinitely more attractive than
2228 * anon pages. Try to detect this based on file LRU size.
2230 if (global_reclaim(sc)) {
2231 unsigned long pgdatfile;
2232 unsigned long pgdatfree;
2234 unsigned long total_high_wmark = 0;
2236 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2237 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2238 node_page_state(pgdat, NR_INACTIVE_FILE);
2240 for (z = 0; z < MAX_NR_ZONES; z++) {
2241 struct zone *zone = &pgdat->node_zones[z];
2242 if (!managed_zone(zone))
2245 total_high_wmark += high_wmark_pages(zone);
2248 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2250 * Force SCAN_ANON if there are enough inactive
2251 * anonymous pages on the LRU in eligible zones.
2252 * Otherwise, the small LRU gets thrashed.
2254 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2255 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2257 scan_balance = SCAN_ANON;
2264 * If there is enough inactive page cache, i.e. if the size of the
2265 * inactive list is greater than that of the active list *and* the
2266 * inactive list actually has some pages to scan on this priority, we
2267 * do not reclaim anything from the anonymous working set right now.
2268 * Without the second condition we could end up never scanning an
2269 * lruvec even if it has plenty of old anonymous pages unless the
2270 * system is under heavy pressure.
2272 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2273 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2274 scan_balance = SCAN_FILE;
2278 scan_balance = SCAN_FRACT;
2281 * With swappiness at 100, anonymous and file have the same priority.
2282 * This scanning priority is essentially the inverse of IO cost.
2284 anon_prio = swappiness;
2285 file_prio = 200 - anon_prio;
2288 * OK, so we have swap space and a fair amount of page cache
2289 * pages. We use the recently rotated / recently scanned
2290 * ratios to determine how valuable each cache is.
2292 * Because workloads change over time (and to avoid overflow)
2293 * we keep these statistics as a floating average, which ends
2294 * up weighing recent references more than old ones.
2296 * anon in [0], file in [1]
2299 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2300 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2301 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2302 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2304 spin_lock_irq(&pgdat->lru_lock);
2305 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2306 reclaim_stat->recent_scanned[0] /= 2;
2307 reclaim_stat->recent_rotated[0] /= 2;
2310 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2311 reclaim_stat->recent_scanned[1] /= 2;
2312 reclaim_stat->recent_rotated[1] /= 2;
2316 * The amount of pressure on anon vs file pages is inversely
2317 * proportional to the fraction of recently scanned pages on
2318 * each list that were recently referenced and in active use.
2320 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2321 ap /= reclaim_stat->recent_rotated[0] + 1;
2323 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2324 fp /= reclaim_stat->recent_rotated[1] + 1;
2325 spin_unlock_irq(&pgdat->lru_lock);
2329 denominator = ap + fp + 1;
2332 for_each_evictable_lru(lru) {
2333 int file = is_file_lru(lru);
2337 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2338 scan = size >> sc->priority;
2340 * If the cgroup's already been deleted, make sure to
2341 * scrape out the remaining cache.
2343 if (!scan && !mem_cgroup_online(memcg))
2344 scan = min(size, SWAP_CLUSTER_MAX);
2346 switch (scan_balance) {
2348 /* Scan lists relative to size */
2352 * Scan types proportional to swappiness and
2353 * their relative recent reclaim efficiency.
2355 scan = div64_u64(scan * fraction[file],
2360 /* Scan one type exclusively */
2361 if ((scan_balance == SCAN_FILE) != file) {
2367 /* Look ma, no brain */
2377 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2379 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2380 struct scan_control *sc, unsigned long *lru_pages)
2382 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2383 unsigned long nr[NR_LRU_LISTS];
2384 unsigned long targets[NR_LRU_LISTS];
2385 unsigned long nr_to_scan;
2387 unsigned long nr_reclaimed = 0;
2388 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2389 struct blk_plug plug;
2392 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2394 /* Record the original scan target for proportional adjustments later */
2395 memcpy(targets, nr, sizeof(nr));
2398 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2399 * event that can occur when there is little memory pressure e.g.
2400 * multiple streaming readers/writers. Hence, we do not abort scanning
2401 * when the requested number of pages are reclaimed when scanning at
2402 * DEF_PRIORITY on the assumption that the fact we are direct
2403 * reclaiming implies that kswapd is not keeping up and it is best to
2404 * do a batch of work at once. For memcg reclaim one check is made to
2405 * abort proportional reclaim if either the file or anon lru has already
2406 * dropped to zero at the first pass.
2408 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2409 sc->priority == DEF_PRIORITY);
2411 blk_start_plug(&plug);
2412 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2413 nr[LRU_INACTIVE_FILE]) {
2414 unsigned long nr_anon, nr_file, percentage;
2415 unsigned long nr_scanned;
2417 for_each_evictable_lru(lru) {
2419 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2420 nr[lru] -= nr_to_scan;
2422 nr_reclaimed += shrink_list(lru, nr_to_scan,
2429 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2433 * For kswapd and memcg, reclaim at least the number of pages
2434 * requested. Ensure that the anon and file LRUs are scanned
2435 * proportionally what was requested by get_scan_count(). We
2436 * stop reclaiming one LRU and reduce the amount scanning
2437 * proportional to the original scan target.
2439 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2440 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2443 * It's just vindictive to attack the larger once the smaller
2444 * has gone to zero. And given the way we stop scanning the
2445 * smaller below, this makes sure that we only make one nudge
2446 * towards proportionality once we've got nr_to_reclaim.
2448 if (!nr_file || !nr_anon)
2451 if (nr_file > nr_anon) {
2452 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2453 targets[LRU_ACTIVE_ANON] + 1;
2455 percentage = nr_anon * 100 / scan_target;
2457 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2458 targets[LRU_ACTIVE_FILE] + 1;
2460 percentage = nr_file * 100 / scan_target;
2463 /* Stop scanning the smaller of the LRU */
2465 nr[lru + LRU_ACTIVE] = 0;
2468 * Recalculate the other LRU scan count based on its original
2469 * scan target and the percentage scanning already complete
2471 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2472 nr_scanned = targets[lru] - nr[lru];
2473 nr[lru] = targets[lru] * (100 - percentage) / 100;
2474 nr[lru] -= min(nr[lru], nr_scanned);
2477 nr_scanned = targets[lru] - nr[lru];
2478 nr[lru] = targets[lru] * (100 - percentage) / 100;
2479 nr[lru] -= min(nr[lru], nr_scanned);
2481 scan_adjusted = true;
2483 blk_finish_plug(&plug);
2484 sc->nr_reclaimed += nr_reclaimed;
2487 * Even if we did not try to evict anon pages at all, we want to
2488 * rebalance the anon lru active/inactive ratio.
2490 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2491 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2492 sc, LRU_ACTIVE_ANON);
2495 /* Use reclaim/compaction for costly allocs or under memory pressure */
2496 static bool in_reclaim_compaction(struct scan_control *sc)
2498 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2499 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2500 sc->priority < DEF_PRIORITY - 2))
2507 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2508 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2509 * true if more pages should be reclaimed such that when the page allocator
2510 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2511 * It will give up earlier than that if there is difficulty reclaiming pages.
2513 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2514 unsigned long nr_reclaimed,
2515 unsigned long nr_scanned,
2516 struct scan_control *sc)
2518 unsigned long pages_for_compaction;
2519 unsigned long inactive_lru_pages;
2522 /* If not in reclaim/compaction mode, stop */
2523 if (!in_reclaim_compaction(sc))
2526 /* Consider stopping depending on scan and reclaim activity */
2527 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2529 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2530 * full LRU list has been scanned and we are still failing
2531 * to reclaim pages. This full LRU scan is potentially
2532 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2534 if (!nr_reclaimed && !nr_scanned)
2538 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2539 * fail without consequence, stop if we failed to reclaim
2540 * any pages from the last SWAP_CLUSTER_MAX number of
2541 * pages that were scanned. This will return to the
2542 * caller faster at the risk reclaim/compaction and
2543 * the resulting allocation attempt fails
2550 * If we have not reclaimed enough pages for compaction and the
2551 * inactive lists are large enough, continue reclaiming
2553 pages_for_compaction = compact_gap(sc->order);
2554 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2555 if (get_nr_swap_pages() > 0)
2556 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2557 if (sc->nr_reclaimed < pages_for_compaction &&
2558 inactive_lru_pages > pages_for_compaction)
2561 /* If compaction would go ahead or the allocation would succeed, stop */
2562 for (z = 0; z <= sc->reclaim_idx; z++) {
2563 struct zone *zone = &pgdat->node_zones[z];
2564 if (!managed_zone(zone))
2567 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2568 case COMPACT_SUCCESS:
2569 case COMPACT_CONTINUE:
2572 /* check next zone */
2579 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2581 struct reclaim_state *reclaim_state = current->reclaim_state;
2582 unsigned long nr_reclaimed, nr_scanned;
2583 bool reclaimable = false;
2586 struct mem_cgroup *root = sc->target_mem_cgroup;
2587 struct mem_cgroup_reclaim_cookie reclaim = {
2589 .priority = sc->priority,
2591 unsigned long node_lru_pages = 0;
2592 struct mem_cgroup *memcg;
2594 nr_reclaimed = sc->nr_reclaimed;
2595 nr_scanned = sc->nr_scanned;
2597 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2599 unsigned long lru_pages;
2600 unsigned long reclaimed;
2601 unsigned long scanned;
2603 if (mem_cgroup_low(root, memcg)) {
2604 if (!sc->memcg_low_reclaim) {
2605 sc->memcg_low_skipped = 1;
2608 mem_cgroup_event(memcg, MEMCG_LOW);
2611 reclaimed = sc->nr_reclaimed;
2612 scanned = sc->nr_scanned;
2613 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2614 node_lru_pages += lru_pages;
2617 shrink_slab(sc->gfp_mask, pgdat->node_id,
2618 memcg, sc->priority);
2620 /* Record the group's reclaim efficiency */
2621 vmpressure(sc->gfp_mask, memcg, false,
2622 sc->nr_scanned - scanned,
2623 sc->nr_reclaimed - reclaimed);
2626 * Direct reclaim and kswapd have to scan all memory
2627 * cgroups to fulfill the overall scan target for the
2630 * Limit reclaim, on the other hand, only cares about
2631 * nr_to_reclaim pages to be reclaimed and it will
2632 * retry with decreasing priority if one round over the
2633 * whole hierarchy is not sufficient.
2635 if (!global_reclaim(sc) &&
2636 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2637 mem_cgroup_iter_break(root, memcg);
2640 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2642 if (global_reclaim(sc))
2643 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2646 if (reclaim_state) {
2647 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2648 reclaim_state->reclaimed_slab = 0;
2651 /* Record the subtree's reclaim efficiency */
2652 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2653 sc->nr_scanned - nr_scanned,
2654 sc->nr_reclaimed - nr_reclaimed);
2656 if (sc->nr_reclaimed - nr_reclaimed)
2659 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2660 sc->nr_scanned - nr_scanned, sc));
2663 * Kswapd gives up on balancing particular nodes after too
2664 * many failures to reclaim anything from them and goes to
2665 * sleep. On reclaim progress, reset the failure counter. A
2666 * successful direct reclaim run will revive a dormant kswapd.
2669 pgdat->kswapd_failures = 0;
2675 * Returns true if compaction should go ahead for a costly-order request, or
2676 * the allocation would already succeed without compaction. Return false if we
2677 * should reclaim first.
2679 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2681 unsigned long watermark;
2682 enum compact_result suitable;
2684 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2685 if (suitable == COMPACT_SUCCESS)
2686 /* Allocation should succeed already. Don't reclaim. */
2688 if (suitable == COMPACT_SKIPPED)
2689 /* Compaction cannot yet proceed. Do reclaim. */
2693 * Compaction is already possible, but it takes time to run and there
2694 * are potentially other callers using the pages just freed. So proceed
2695 * with reclaim to make a buffer of free pages available to give
2696 * compaction a reasonable chance of completing and allocating the page.
2697 * Note that we won't actually reclaim the whole buffer in one attempt
2698 * as the target watermark in should_continue_reclaim() is lower. But if
2699 * we are already above the high+gap watermark, don't reclaim at all.
2701 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2703 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2707 * This is the direct reclaim path, for page-allocating processes. We only
2708 * try to reclaim pages from zones which will satisfy the caller's allocation
2711 * If a zone is deemed to be full of pinned pages then just give it a light
2712 * scan then give up on it.
2714 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2718 unsigned long nr_soft_reclaimed;
2719 unsigned long nr_soft_scanned;
2721 pg_data_t *last_pgdat = NULL;
2724 * If the number of buffer_heads in the machine exceeds the maximum
2725 * allowed level, force direct reclaim to scan the highmem zone as
2726 * highmem pages could be pinning lowmem pages storing buffer_heads
2728 orig_mask = sc->gfp_mask;
2729 if (buffer_heads_over_limit) {
2730 sc->gfp_mask |= __GFP_HIGHMEM;
2731 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2734 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2735 sc->reclaim_idx, sc->nodemask) {
2737 * Take care memory controller reclaiming has small influence
2740 if (global_reclaim(sc)) {
2741 if (!cpuset_zone_allowed(zone,
2742 GFP_KERNEL | __GFP_HARDWALL))
2746 * If we already have plenty of memory free for
2747 * compaction in this zone, don't free any more.
2748 * Even though compaction is invoked for any
2749 * non-zero order, only frequent costly order
2750 * reclamation is disruptive enough to become a
2751 * noticeable problem, like transparent huge
2754 if (IS_ENABLED(CONFIG_COMPACTION) &&
2755 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2756 compaction_ready(zone, sc)) {
2757 sc->compaction_ready = true;
2762 * Shrink each node in the zonelist once. If the
2763 * zonelist is ordered by zone (not the default) then a
2764 * node may be shrunk multiple times but in that case
2765 * the user prefers lower zones being preserved.
2767 if (zone->zone_pgdat == last_pgdat)
2771 * This steals pages from memory cgroups over softlimit
2772 * and returns the number of reclaimed pages and
2773 * scanned pages. This works for global memory pressure
2774 * and balancing, not for a memcg's limit.
2776 nr_soft_scanned = 0;
2777 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2778 sc->order, sc->gfp_mask,
2780 sc->nr_reclaimed += nr_soft_reclaimed;
2781 sc->nr_scanned += nr_soft_scanned;
2782 /* need some check for avoid more shrink_zone() */
2785 /* See comment about same check for global reclaim above */
2786 if (zone->zone_pgdat == last_pgdat)
2788 last_pgdat = zone->zone_pgdat;
2789 shrink_node(zone->zone_pgdat, sc);
2793 * Restore to original mask to avoid the impact on the caller if we
2794 * promoted it to __GFP_HIGHMEM.
2796 sc->gfp_mask = orig_mask;
2799 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2801 struct mem_cgroup *memcg;
2803 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2805 unsigned long refaults;
2806 struct lruvec *lruvec;
2809 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2811 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2813 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2814 lruvec->refaults = refaults;
2815 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2819 * This is the main entry point to direct page reclaim.
2821 * If a full scan of the inactive list fails to free enough memory then we
2822 * are "out of memory" and something needs to be killed.
2824 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2825 * high - the zone may be full of dirty or under-writeback pages, which this
2826 * caller can't do much about. We kick the writeback threads and take explicit
2827 * naps in the hope that some of these pages can be written. But if the
2828 * allocating task holds filesystem locks which prevent writeout this might not
2829 * work, and the allocation attempt will fail.
2831 * returns: 0, if no pages reclaimed
2832 * else, the number of pages reclaimed
2834 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2835 struct scan_control *sc)
2837 int initial_priority = sc->priority;
2838 pg_data_t *last_pgdat;
2842 delayacct_freepages_start();
2844 if (global_reclaim(sc))
2845 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2848 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2851 shrink_zones(zonelist, sc);
2853 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2856 if (sc->compaction_ready)
2860 * If we're getting trouble reclaiming, start doing
2861 * writepage even in laptop mode.
2863 if (sc->priority < DEF_PRIORITY - 2)
2864 sc->may_writepage = 1;
2865 } while (--sc->priority >= 0);
2868 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2870 if (zone->zone_pgdat == last_pgdat)
2872 last_pgdat = zone->zone_pgdat;
2873 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2876 delayacct_freepages_end();
2878 if (sc->nr_reclaimed)
2879 return sc->nr_reclaimed;
2881 /* Aborted reclaim to try compaction? don't OOM, then */
2882 if (sc->compaction_ready)
2885 /* Untapped cgroup reserves? Don't OOM, retry. */
2886 if (sc->memcg_low_skipped) {
2887 sc->priority = initial_priority;
2888 sc->memcg_low_reclaim = 1;
2889 sc->memcg_low_skipped = 0;
2896 static bool allow_direct_reclaim(pg_data_t *pgdat)
2899 unsigned long pfmemalloc_reserve = 0;
2900 unsigned long free_pages = 0;
2904 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2907 for (i = 0; i <= ZONE_NORMAL; i++) {
2908 zone = &pgdat->node_zones[i];
2909 if (!managed_zone(zone))
2912 if (!zone_reclaimable_pages(zone))
2915 pfmemalloc_reserve += min_wmark_pages(zone);
2916 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2919 /* If there are no reserves (unexpected config) then do not throttle */
2920 if (!pfmemalloc_reserve)
2923 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2925 /* kswapd must be awake if processes are being throttled */
2926 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2927 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2928 (enum zone_type)ZONE_NORMAL);
2929 wake_up_interruptible(&pgdat->kswapd_wait);
2936 * Throttle direct reclaimers if backing storage is backed by the network
2937 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2938 * depleted. kswapd will continue to make progress and wake the processes
2939 * when the low watermark is reached.
2941 * Returns true if a fatal signal was delivered during throttling. If this
2942 * happens, the page allocator should not consider triggering the OOM killer.
2944 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2945 nodemask_t *nodemask)
2949 pg_data_t *pgdat = NULL;
2952 * Kernel threads should not be throttled as they may be indirectly
2953 * responsible for cleaning pages necessary for reclaim to make forward
2954 * progress. kjournald for example may enter direct reclaim while
2955 * committing a transaction where throttling it could forcing other
2956 * processes to block on log_wait_commit().
2958 if (current->flags & PF_KTHREAD)
2962 * If a fatal signal is pending, this process should not throttle.
2963 * It should return quickly so it can exit and free its memory
2965 if (fatal_signal_pending(current))
2969 * Check if the pfmemalloc reserves are ok by finding the first node
2970 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2971 * GFP_KERNEL will be required for allocating network buffers when
2972 * swapping over the network so ZONE_HIGHMEM is unusable.
2974 * Throttling is based on the first usable node and throttled processes
2975 * wait on a queue until kswapd makes progress and wakes them. There
2976 * is an affinity then between processes waking up and where reclaim
2977 * progress has been made assuming the process wakes on the same node.
2978 * More importantly, processes running on remote nodes will not compete
2979 * for remote pfmemalloc reserves and processes on different nodes
2980 * should make reasonable progress.
2982 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2983 gfp_zone(gfp_mask), nodemask) {
2984 if (zone_idx(zone) > ZONE_NORMAL)
2987 /* Throttle based on the first usable node */
2988 pgdat = zone->zone_pgdat;
2989 if (allow_direct_reclaim(pgdat))
2994 /* If no zone was usable by the allocation flags then do not throttle */
2998 /* Account for the throttling */
2999 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3002 * If the caller cannot enter the filesystem, it's possible that it
3003 * is due to the caller holding an FS lock or performing a journal
3004 * transaction in the case of a filesystem like ext[3|4]. In this case,
3005 * it is not safe to block on pfmemalloc_wait as kswapd could be
3006 * blocked waiting on the same lock. Instead, throttle for up to a
3007 * second before continuing.
3009 if (!(gfp_mask & __GFP_FS)) {
3010 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3011 allow_direct_reclaim(pgdat), HZ);
3016 /* Throttle until kswapd wakes the process */
3017 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3018 allow_direct_reclaim(pgdat));
3021 if (fatal_signal_pending(current))
3028 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3029 gfp_t gfp_mask, nodemask_t *nodemask)
3031 unsigned long nr_reclaimed;
3032 struct scan_control sc = {
3033 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3034 .gfp_mask = current_gfp_context(gfp_mask),
3035 .reclaim_idx = gfp_zone(gfp_mask),
3037 .nodemask = nodemask,
3038 .priority = DEF_PRIORITY,
3039 .may_writepage = !laptop_mode,
3045 * Do not enter reclaim if fatal signal was delivered while throttled.
3046 * 1 is returned so that the page allocator does not OOM kill at this
3049 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3052 trace_mm_vmscan_direct_reclaim_begin(order,
3057 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3059 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3061 return nr_reclaimed;
3066 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3067 gfp_t gfp_mask, bool noswap,
3069 unsigned long *nr_scanned)
3071 struct scan_control sc = {
3072 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3073 .target_mem_cgroup = memcg,
3074 .may_writepage = !laptop_mode,
3076 .reclaim_idx = MAX_NR_ZONES - 1,
3077 .may_swap = !noswap,
3079 unsigned long lru_pages;
3081 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3082 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3084 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3090 * NOTE: Although we can get the priority field, using it
3091 * here is not a good idea, since it limits the pages we can scan.
3092 * if we don't reclaim here, the shrink_node from balance_pgdat
3093 * will pick up pages from other mem cgroup's as well. We hack
3094 * the priority and make it zero.
3096 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3098 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3100 *nr_scanned = sc.nr_scanned;
3101 return sc.nr_reclaimed;
3104 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3105 unsigned long nr_pages,
3109 struct zonelist *zonelist;
3110 unsigned long nr_reclaimed;
3112 unsigned int noreclaim_flag;
3113 struct scan_control sc = {
3114 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3115 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3116 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3117 .reclaim_idx = MAX_NR_ZONES - 1,
3118 .target_mem_cgroup = memcg,
3119 .priority = DEF_PRIORITY,
3120 .may_writepage = !laptop_mode,
3122 .may_swap = may_swap,
3126 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3127 * take care of from where we get pages. So the node where we start the
3128 * scan does not need to be the current node.
3130 nid = mem_cgroup_select_victim_node(memcg);
3132 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3134 trace_mm_vmscan_memcg_reclaim_begin(0,
3139 noreclaim_flag = memalloc_noreclaim_save();
3140 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3141 memalloc_noreclaim_restore(noreclaim_flag);
3143 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3145 return nr_reclaimed;
3149 static void age_active_anon(struct pglist_data *pgdat,
3150 struct scan_control *sc)
3152 struct mem_cgroup *memcg;
3154 if (!total_swap_pages)
3157 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3159 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3161 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3162 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3163 sc, LRU_ACTIVE_ANON);
3165 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3170 * Returns true if there is an eligible zone balanced for the request order
3173 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3176 unsigned long mark = -1;
3179 for (i = 0; i <= classzone_idx; i++) {
3180 zone = pgdat->node_zones + i;
3182 if (!managed_zone(zone))
3185 mark = high_wmark_pages(zone);
3186 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3191 * If a node has no populated zone within classzone_idx, it does not
3192 * need balancing by definition. This can happen if a zone-restricted
3193 * allocation tries to wake a remote kswapd.
3201 /* Clear pgdat state for congested, dirty or under writeback. */
3202 static void clear_pgdat_congested(pg_data_t *pgdat)
3204 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3205 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3206 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3210 * Prepare kswapd for sleeping. This verifies that there are no processes
3211 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3213 * Returns true if kswapd is ready to sleep
3215 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3218 * The throttled processes are normally woken up in balance_pgdat() as
3219 * soon as allow_direct_reclaim() is true. But there is a potential
3220 * race between when kswapd checks the watermarks and a process gets
3221 * throttled. There is also a potential race if processes get
3222 * throttled, kswapd wakes, a large process exits thereby balancing the
3223 * zones, which causes kswapd to exit balance_pgdat() before reaching
3224 * the wake up checks. If kswapd is going to sleep, no process should
3225 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3226 * the wake up is premature, processes will wake kswapd and get
3227 * throttled again. The difference from wake ups in balance_pgdat() is
3228 * that here we are under prepare_to_wait().
3230 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3231 wake_up_all(&pgdat->pfmemalloc_wait);
3233 /* Hopeless node, leave it to direct reclaim */
3234 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3237 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3238 clear_pgdat_congested(pgdat);
3246 * kswapd shrinks a node of pages that are at or below the highest usable
3247 * zone that is currently unbalanced.
3249 * Returns true if kswapd scanned at least the requested number of pages to
3250 * reclaim or if the lack of progress was due to pages under writeback.
3251 * This is used to determine if the scanning priority needs to be raised.
3253 static bool kswapd_shrink_node(pg_data_t *pgdat,
3254 struct scan_control *sc)
3259 /* Reclaim a number of pages proportional to the number of zones */
3260 sc->nr_to_reclaim = 0;
3261 for (z = 0; z <= sc->reclaim_idx; z++) {
3262 zone = pgdat->node_zones + z;
3263 if (!managed_zone(zone))
3266 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3270 * Historically care was taken to put equal pressure on all zones but
3271 * now pressure is applied based on node LRU order.
3273 shrink_node(pgdat, sc);
3276 * Fragmentation may mean that the system cannot be rebalanced for
3277 * high-order allocations. If twice the allocation size has been
3278 * reclaimed then recheck watermarks only at order-0 to prevent
3279 * excessive reclaim. Assume that a process requested a high-order
3280 * can direct reclaim/compact.
3282 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3285 return sc->nr_scanned >= sc->nr_to_reclaim;
3289 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3290 * that are eligible for use by the caller until at least one zone is
3293 * Returns the order kswapd finished reclaiming at.
3295 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3296 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3297 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3298 * or lower is eligible for reclaim until at least one usable zone is
3301 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3304 unsigned long nr_soft_reclaimed;
3305 unsigned long nr_soft_scanned;
3307 struct scan_control sc = {
3308 .gfp_mask = GFP_KERNEL,
3310 .priority = DEF_PRIORITY,
3311 .may_writepage = !laptop_mode,
3315 count_vm_event(PAGEOUTRUN);
3318 unsigned long nr_reclaimed = sc.nr_reclaimed;
3319 bool raise_priority = true;
3321 sc.reclaim_idx = classzone_idx;
3324 * If the number of buffer_heads exceeds the maximum allowed
3325 * then consider reclaiming from all zones. This has a dual
3326 * purpose -- on 64-bit systems it is expected that
3327 * buffer_heads are stripped during active rotation. On 32-bit
3328 * systems, highmem pages can pin lowmem memory and shrinking
3329 * buffers can relieve lowmem pressure. Reclaim may still not
3330 * go ahead if all eligible zones for the original allocation
3331 * request are balanced to avoid excessive reclaim from kswapd.
3333 if (buffer_heads_over_limit) {
3334 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3335 zone = pgdat->node_zones + i;
3336 if (!managed_zone(zone))
3345 * Only reclaim if there are no eligible zones. Note that
3346 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3349 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3353 * Do some background aging of the anon list, to give
3354 * pages a chance to be referenced before reclaiming. All
3355 * pages are rotated regardless of classzone as this is
3356 * about consistent aging.
3358 age_active_anon(pgdat, &sc);
3361 * If we're getting trouble reclaiming, start doing writepage
3362 * even in laptop mode.
3364 if (sc.priority < DEF_PRIORITY - 2)
3365 sc.may_writepage = 1;
3367 /* Call soft limit reclaim before calling shrink_node. */
3369 nr_soft_scanned = 0;
3370 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3371 sc.gfp_mask, &nr_soft_scanned);
3372 sc.nr_reclaimed += nr_soft_reclaimed;
3375 * There should be no need to raise the scanning priority if
3376 * enough pages are already being scanned that that high
3377 * watermark would be met at 100% efficiency.
3379 if (kswapd_shrink_node(pgdat, &sc))
3380 raise_priority = false;
3383 * If the low watermark is met there is no need for processes
3384 * to be throttled on pfmemalloc_wait as they should not be
3385 * able to safely make forward progress. Wake them
3387 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3388 allow_direct_reclaim(pgdat))
3389 wake_up_all(&pgdat->pfmemalloc_wait);
3391 /* Check if kswapd should be suspending */
3392 if (try_to_freeze() || kthread_should_stop())
3396 * Raise priority if scanning rate is too low or there was no
3397 * progress in reclaiming pages
3399 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3400 if (raise_priority || !nr_reclaimed)
3402 } while (sc.priority >= 1);
3404 if (!sc.nr_reclaimed)
3405 pgdat->kswapd_failures++;
3408 snapshot_refaults(NULL, pgdat);
3410 * Return the order kswapd stopped reclaiming at as
3411 * prepare_kswapd_sleep() takes it into account. If another caller
3412 * entered the allocator slow path while kswapd was awake, order will
3413 * remain at the higher level.
3419 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3420 * allocation request woke kswapd for. When kswapd has not woken recently,
3421 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3422 * given classzone and returns it or the highest classzone index kswapd
3423 * was recently woke for.
3425 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3426 enum zone_type classzone_idx)
3428 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3429 return classzone_idx;
3431 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3434 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3435 unsigned int classzone_idx)
3440 if (freezing(current) || kthread_should_stop())
3443 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3446 * Try to sleep for a short interval. Note that kcompactd will only be
3447 * woken if it is possible to sleep for a short interval. This is
3448 * deliberate on the assumption that if reclaim cannot keep an
3449 * eligible zone balanced that it's also unlikely that compaction will
3452 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3454 * Compaction records what page blocks it recently failed to
3455 * isolate pages from and skips them in the future scanning.
3456 * When kswapd is going to sleep, it is reasonable to assume
3457 * that pages and compaction may succeed so reset the cache.
3459 reset_isolation_suitable(pgdat);
3462 * We have freed the memory, now we should compact it to make
3463 * allocation of the requested order possible.
3465 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3467 remaining = schedule_timeout(HZ/10);
3470 * If woken prematurely then reset kswapd_classzone_idx and
3471 * order. The values will either be from a wakeup request or
3472 * the previous request that slept prematurely.
3475 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3476 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3479 finish_wait(&pgdat->kswapd_wait, &wait);
3480 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3484 * After a short sleep, check if it was a premature sleep. If not, then
3485 * go fully to sleep until explicitly woken up.
3488 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3489 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3492 * vmstat counters are not perfectly accurate and the estimated
3493 * value for counters such as NR_FREE_PAGES can deviate from the
3494 * true value by nr_online_cpus * threshold. To avoid the zone
3495 * watermarks being breached while under pressure, we reduce the
3496 * per-cpu vmstat threshold while kswapd is awake and restore
3497 * them before going back to sleep.
3499 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3501 if (!kthread_should_stop())
3504 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3507 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3509 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3511 finish_wait(&pgdat->kswapd_wait, &wait);
3515 * The background pageout daemon, started as a kernel thread
3516 * from the init process.
3518 * This basically trickles out pages so that we have _some_
3519 * free memory available even if there is no other activity
3520 * that frees anything up. This is needed for things like routing
3521 * etc, where we otherwise might have all activity going on in
3522 * asynchronous contexts that cannot page things out.
3524 * If there are applications that are active memory-allocators
3525 * (most normal use), this basically shouldn't matter.
3527 static int kswapd(void *p)
3529 unsigned int alloc_order, reclaim_order;
3530 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3531 pg_data_t *pgdat = (pg_data_t*)p;
3532 struct task_struct *tsk = current;
3534 struct reclaim_state reclaim_state = {
3535 .reclaimed_slab = 0,
3537 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3539 if (!cpumask_empty(cpumask))
3540 set_cpus_allowed_ptr(tsk, cpumask);
3541 current->reclaim_state = &reclaim_state;
3544 * Tell the memory management that we're a "memory allocator",
3545 * and that if we need more memory we should get access to it
3546 * regardless (see "__alloc_pages()"). "kswapd" should
3547 * never get caught in the normal page freeing logic.
3549 * (Kswapd normally doesn't need memory anyway, but sometimes
3550 * you need a small amount of memory in order to be able to
3551 * page out something else, and this flag essentially protects
3552 * us from recursively trying to free more memory as we're
3553 * trying to free the first piece of memory in the first place).
3555 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3558 pgdat->kswapd_order = 0;
3559 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3563 alloc_order = reclaim_order = pgdat->kswapd_order;
3564 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3567 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3570 /* Read the new order and classzone_idx */
3571 alloc_order = reclaim_order = pgdat->kswapd_order;
3572 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3573 pgdat->kswapd_order = 0;
3574 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3576 ret = try_to_freeze();
3577 if (kthread_should_stop())
3581 * We can speed up thawing tasks if we don't call balance_pgdat
3582 * after returning from the refrigerator
3588 * Reclaim begins at the requested order but if a high-order
3589 * reclaim fails then kswapd falls back to reclaiming for
3590 * order-0. If that happens, kswapd will consider sleeping
3591 * for the order it finished reclaiming at (reclaim_order)
3592 * but kcompactd is woken to compact for the original
3593 * request (alloc_order).
3595 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3597 fs_reclaim_acquire(GFP_KERNEL);
3598 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3599 fs_reclaim_release(GFP_KERNEL);
3600 if (reclaim_order < alloc_order)
3601 goto kswapd_try_sleep;
3604 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3605 current->reclaim_state = NULL;
3611 * A zone is low on free memory, so wake its kswapd task to service it.
3613 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3617 if (!managed_zone(zone))
3620 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3622 pgdat = zone->zone_pgdat;
3623 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3625 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3626 if (!waitqueue_active(&pgdat->kswapd_wait))
3629 /* Hopeless node, leave it to direct reclaim */
3630 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3633 if (pgdat_balanced(pgdat, order, classzone_idx))
3636 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
3637 wake_up_interruptible(&pgdat->kswapd_wait);
3640 #ifdef CONFIG_HIBERNATION
3642 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3645 * Rather than trying to age LRUs the aim is to preserve the overall
3646 * LRU order by reclaiming preferentially
3647 * inactive > active > active referenced > active mapped
3649 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3651 struct reclaim_state reclaim_state;
3652 struct scan_control sc = {
3653 .nr_to_reclaim = nr_to_reclaim,
3654 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3655 .reclaim_idx = MAX_NR_ZONES - 1,
3656 .priority = DEF_PRIORITY,
3660 .hibernation_mode = 1,
3662 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3663 struct task_struct *p = current;
3664 unsigned long nr_reclaimed;
3665 unsigned int noreclaim_flag;
3667 noreclaim_flag = memalloc_noreclaim_save();
3668 fs_reclaim_acquire(sc.gfp_mask);
3669 reclaim_state.reclaimed_slab = 0;
3670 p->reclaim_state = &reclaim_state;
3672 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3674 p->reclaim_state = NULL;
3675 fs_reclaim_release(sc.gfp_mask);
3676 memalloc_noreclaim_restore(noreclaim_flag);
3678 return nr_reclaimed;
3680 #endif /* CONFIG_HIBERNATION */
3682 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3683 not required for correctness. So if the last cpu in a node goes
3684 away, we get changed to run anywhere: as the first one comes back,
3685 restore their cpu bindings. */
3686 static int kswapd_cpu_online(unsigned int cpu)
3690 for_each_node_state(nid, N_MEMORY) {
3691 pg_data_t *pgdat = NODE_DATA(nid);
3692 const struct cpumask *mask;
3694 mask = cpumask_of_node(pgdat->node_id);
3696 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3697 /* One of our CPUs online: restore mask */
3698 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3704 * This kswapd start function will be called by init and node-hot-add.
3705 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3707 int kswapd_run(int nid)
3709 pg_data_t *pgdat = NODE_DATA(nid);
3715 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3716 if (IS_ERR(pgdat->kswapd)) {
3717 /* failure at boot is fatal */
3718 BUG_ON(system_state < SYSTEM_RUNNING);
3719 pr_err("Failed to start kswapd on node %d\n", nid);
3720 ret = PTR_ERR(pgdat->kswapd);
3721 pgdat->kswapd = NULL;
3727 * Called by memory hotplug when all memory in a node is offlined. Caller must
3728 * hold mem_hotplug_begin/end().
3730 void kswapd_stop(int nid)
3732 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3735 kthread_stop(kswapd);
3736 NODE_DATA(nid)->kswapd = NULL;
3740 static int __init kswapd_init(void)
3745 for_each_node_state(nid, N_MEMORY)
3747 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3748 "mm/vmscan:online", kswapd_cpu_online,
3754 module_init(kswapd_init)
3760 * If non-zero call node_reclaim when the number of free pages falls below
3763 int node_reclaim_mode __read_mostly;
3765 #define RECLAIM_OFF 0
3766 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3767 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3768 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3771 * Priority for NODE_RECLAIM. This determines the fraction of pages
3772 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3775 #define NODE_RECLAIM_PRIORITY 4
3778 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3781 int sysctl_min_unmapped_ratio = 1;
3784 * If the number of slab pages in a zone grows beyond this percentage then
3785 * slab reclaim needs to occur.
3787 int sysctl_min_slab_ratio = 5;
3789 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3791 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3792 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3793 node_page_state(pgdat, NR_ACTIVE_FILE);
3796 * It's possible for there to be more file mapped pages than
3797 * accounted for by the pages on the file LRU lists because
3798 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3800 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3803 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3804 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3806 unsigned long nr_pagecache_reclaimable;
3807 unsigned long delta = 0;
3810 * If RECLAIM_UNMAP is set, then all file pages are considered
3811 * potentially reclaimable. Otherwise, we have to worry about
3812 * pages like swapcache and node_unmapped_file_pages() provides
3815 if (node_reclaim_mode & RECLAIM_UNMAP)
3816 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3818 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3820 /* If we can't clean pages, remove dirty pages from consideration */
3821 if (!(node_reclaim_mode & RECLAIM_WRITE))
3822 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3824 /* Watch for any possible underflows due to delta */
3825 if (unlikely(delta > nr_pagecache_reclaimable))
3826 delta = nr_pagecache_reclaimable;
3828 return nr_pagecache_reclaimable - delta;
3832 * Try to free up some pages from this node through reclaim.
3834 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3836 /* Minimum pages needed in order to stay on node */
3837 const unsigned long nr_pages = 1 << order;
3838 struct task_struct *p = current;
3839 struct reclaim_state reclaim_state;
3840 unsigned int noreclaim_flag;
3841 struct scan_control sc = {
3842 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3843 .gfp_mask = current_gfp_context(gfp_mask),
3845 .priority = NODE_RECLAIM_PRIORITY,
3846 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3847 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3849 .reclaim_idx = gfp_zone(gfp_mask),
3854 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3855 * and we also need to be able to write out pages for RECLAIM_WRITE
3856 * and RECLAIM_UNMAP.
3858 noreclaim_flag = memalloc_noreclaim_save();
3859 p->flags |= PF_SWAPWRITE;
3860 fs_reclaim_acquire(sc.gfp_mask);
3861 reclaim_state.reclaimed_slab = 0;
3862 p->reclaim_state = &reclaim_state;
3864 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3866 * Free memory by calling shrink zone with increasing
3867 * priorities until we have enough memory freed.
3870 shrink_node(pgdat, &sc);
3871 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3874 p->reclaim_state = NULL;
3875 fs_reclaim_release(gfp_mask);
3876 current->flags &= ~PF_SWAPWRITE;
3877 memalloc_noreclaim_restore(noreclaim_flag);
3878 return sc.nr_reclaimed >= nr_pages;
3881 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3886 * Node reclaim reclaims unmapped file backed pages and
3887 * slab pages if we are over the defined limits.
3889 * A small portion of unmapped file backed pages is needed for
3890 * file I/O otherwise pages read by file I/O will be immediately
3891 * thrown out if the node is overallocated. So we do not reclaim
3892 * if less than a specified percentage of the node is used by
3893 * unmapped file backed pages.
3895 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3896 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3897 return NODE_RECLAIM_FULL;
3900 * Do not scan if the allocation should not be delayed.
3902 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3903 return NODE_RECLAIM_NOSCAN;
3906 * Only run node reclaim on the local node or on nodes that do not
3907 * have associated processors. This will favor the local processor
3908 * over remote processors and spread off node memory allocations
3909 * as wide as possible.
3911 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3912 return NODE_RECLAIM_NOSCAN;
3914 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3915 return NODE_RECLAIM_NOSCAN;
3917 ret = __node_reclaim(pgdat, gfp_mask, order);
3918 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3921 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3928 * page_evictable - test whether a page is evictable
3929 * @page: the page to test
3931 * Test whether page is evictable--i.e., should be placed on active/inactive
3932 * lists vs unevictable list.
3934 * Reasons page might not be evictable:
3935 * (1) page's mapping marked unevictable
3936 * (2) page is part of an mlocked VMA
3939 int page_evictable(struct page *page)
3941 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3946 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3947 * @pages: array of pages to check
3948 * @nr_pages: number of pages to check
3950 * Checks pages for evictability and moves them to the appropriate lru list.
3952 * This function is only used for SysV IPC SHM_UNLOCK.
3954 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3956 struct lruvec *lruvec;
3957 struct pglist_data *pgdat = NULL;
3962 for (i = 0; i < nr_pages; i++) {
3963 struct page *page = pages[i];
3964 struct pglist_data *pagepgdat = page_pgdat(page);
3967 if (pagepgdat != pgdat) {
3969 spin_unlock_irq(&pgdat->lru_lock);
3971 spin_lock_irq(&pgdat->lru_lock);
3973 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3975 if (!PageLRU(page) || !PageUnevictable(page))
3978 if (page_evictable(page)) {
3979 enum lru_list lru = page_lru_base_type(page);
3981 VM_BUG_ON_PAGE(PageActive(page), page);
3982 ClearPageUnevictable(page);
3983 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3984 add_page_to_lru_list(page, lruvec, lru);
3990 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3991 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3992 spin_unlock_irq(&pgdat->lru_lock);
3995 #endif /* CONFIG_SHMEM */