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,
314 unsigned long nr_scanned,
315 unsigned long nr_eligible)
317 unsigned long freed = 0;
318 unsigned long long delta;
323 int nid = shrinkctl->nid;
324 long batch_size = shrinker->batch ? shrinker->batch
326 long scanned = 0, next_deferred;
328 freeable = shrinker->count_objects(shrinker, shrinkctl);
333 * copy the current shrinker scan count into a local variable
334 * and zero it so that other concurrent shrinker invocations
335 * don't also do this scanning work.
337 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
340 delta = (4 * nr_scanned) / shrinker->seeks;
342 do_div(delta, nr_eligible + 1);
344 if (total_scan < 0) {
345 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
346 shrinker->scan_objects, total_scan);
347 total_scan = freeable;
350 next_deferred = total_scan;
353 * We need to avoid excessive windup on filesystem shrinkers
354 * due to large numbers of GFP_NOFS allocations causing the
355 * shrinkers to return -1 all the time. This results in a large
356 * nr being built up so when a shrink that can do some work
357 * comes along it empties the entire cache due to nr >>>
358 * freeable. This is bad for sustaining a working set in
361 * Hence only allow the shrinker to scan the entire cache when
362 * a large delta change is calculated directly.
364 if (delta < freeable / 4)
365 total_scan = min(total_scan, freeable / 2);
368 * Avoid risking looping forever due to too large nr value:
369 * never try to free more than twice the estimate number of
372 if (total_scan > freeable * 2)
373 total_scan = freeable * 2;
375 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
376 nr_scanned, nr_eligible,
377 freeable, delta, total_scan);
380 * Normally, we should not scan less than batch_size objects in one
381 * pass to avoid too frequent shrinker calls, but if the slab has less
382 * than batch_size objects in total and we are really tight on memory,
383 * we will try to reclaim all available objects, otherwise we can end
384 * up failing allocations although there are plenty of reclaimable
385 * objects spread over several slabs with usage less than the
388 * We detect the "tight on memory" situations by looking at the total
389 * number of objects we want to scan (total_scan). If it is greater
390 * than the total number of objects on slab (freeable), we must be
391 * scanning at high prio and therefore should try to reclaim as much as
394 while (total_scan >= batch_size ||
395 total_scan >= freeable) {
397 unsigned long nr_to_scan = min(batch_size, total_scan);
399 shrinkctl->nr_to_scan = nr_to_scan;
400 shrinkctl->nr_scanned = nr_to_scan;
401 ret = shrinker->scan_objects(shrinker, shrinkctl);
402 if (ret == SHRINK_STOP)
406 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
407 total_scan -= shrinkctl->nr_scanned;
408 scanned += shrinkctl->nr_scanned;
413 if (next_deferred >= scanned)
414 next_deferred -= scanned;
418 * move the unused scan count back into the shrinker in a
419 * manner that handles concurrent updates. If we exhausted the
420 * scan, there is no need to do an update.
422 if (next_deferred > 0)
423 new_nr = atomic_long_add_return(next_deferred,
424 &shrinker->nr_deferred[nid]);
426 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
428 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
433 * shrink_slab - shrink slab caches
434 * @gfp_mask: allocation context
435 * @nid: node whose slab caches to target
436 * @memcg: memory cgroup whose slab caches to target
437 * @nr_scanned: pressure numerator
438 * @nr_eligible: pressure denominator
440 * Call the shrink functions to age shrinkable caches.
442 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
443 * unaware shrinkers will receive a node id of 0 instead.
445 * @memcg specifies the memory cgroup to target. If it is not NULL,
446 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
447 * objects from the memory cgroup specified. Otherwise, only unaware
448 * shrinkers are called.
450 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
451 * the available objects should be scanned. Page reclaim for example
452 * passes the number of pages scanned and the number of pages on the
453 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
454 * when it encountered mapped pages. The ratio is further biased by
455 * the ->seeks setting of the shrink function, which indicates the
456 * cost to recreate an object relative to that of an LRU page.
458 * Returns the number of reclaimed slab objects.
460 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
461 struct mem_cgroup *memcg,
462 unsigned long nr_scanned,
463 unsigned long nr_eligible)
465 struct shrinker *shrinker;
466 unsigned long freed = 0;
468 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
472 nr_scanned = SWAP_CLUSTER_MAX;
474 if (!down_read_trylock(&shrinker_rwsem)) {
476 * If we would return 0, our callers would understand that we
477 * have nothing else to shrink and give up trying. By returning
478 * 1 we keep it going and assume we'll be able to shrink next
485 list_for_each_entry(shrinker, &shrinker_list, list) {
486 struct shrink_control sc = {
487 .gfp_mask = gfp_mask,
493 * If kernel memory accounting is disabled, we ignore
494 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
495 * passing NULL for memcg.
497 if (memcg_kmem_enabled() &&
498 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
501 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
504 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
507 up_read(&shrinker_rwsem);
513 void drop_slab_node(int nid)
518 struct mem_cgroup *memcg = NULL;
522 freed += shrink_slab(GFP_KERNEL, nid, memcg,
524 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
525 } while (freed > 10);
532 for_each_online_node(nid)
536 static inline int is_page_cache_freeable(struct page *page)
539 * A freeable page cache page is referenced only by the caller
540 * that isolated the page, the page cache radix tree and
541 * optional buffer heads at page->private.
543 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
545 return page_count(page) - page_has_private(page) == 1 + radix_pins;
548 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
550 if (current->flags & PF_SWAPWRITE)
552 if (!inode_write_congested(inode))
554 if (inode_to_bdi(inode) == current->backing_dev_info)
560 * We detected a synchronous write error writing a page out. Probably
561 * -ENOSPC. We need to propagate that into the address_space for a subsequent
562 * fsync(), msync() or close().
564 * The tricky part is that after writepage we cannot touch the mapping: nothing
565 * prevents it from being freed up. But we have a ref on the page and once
566 * that page is locked, the mapping is pinned.
568 * We're allowed to run sleeping lock_page() here because we know the caller has
571 static void handle_write_error(struct address_space *mapping,
572 struct page *page, int error)
575 if (page_mapping(page) == mapping)
576 mapping_set_error(mapping, error);
580 /* possible outcome of pageout() */
582 /* failed to write page out, page is locked */
584 /* move page to the active list, page is locked */
586 /* page has been sent to the disk successfully, page is unlocked */
588 /* page is clean and locked */
593 * pageout is called by shrink_page_list() for each dirty page.
594 * Calls ->writepage().
596 static pageout_t pageout(struct page *page, struct address_space *mapping,
597 struct scan_control *sc)
600 * If the page is dirty, only perform writeback if that write
601 * will be non-blocking. To prevent this allocation from being
602 * stalled by pagecache activity. But note that there may be
603 * stalls if we need to run get_block(). We could test
604 * PagePrivate for that.
606 * If this process is currently in __generic_file_write_iter() against
607 * this page's queue, we can perform writeback even if that
610 * If the page is swapcache, write it back even if that would
611 * block, for some throttling. This happens by accident, because
612 * swap_backing_dev_info is bust: it doesn't reflect the
613 * congestion state of the swapdevs. Easy to fix, if needed.
615 if (!is_page_cache_freeable(page))
619 * Some data journaling orphaned pages can have
620 * page->mapping == NULL while being dirty with clean buffers.
622 if (page_has_private(page)) {
623 if (try_to_free_buffers(page)) {
624 ClearPageDirty(page);
625 pr_info("%s: orphaned page\n", __func__);
631 if (mapping->a_ops->writepage == NULL)
632 return PAGE_ACTIVATE;
633 if (!may_write_to_inode(mapping->host, sc))
636 if (clear_page_dirty_for_io(page)) {
638 struct writeback_control wbc = {
639 .sync_mode = WB_SYNC_NONE,
640 .nr_to_write = SWAP_CLUSTER_MAX,
642 .range_end = LLONG_MAX,
646 SetPageReclaim(page);
647 res = mapping->a_ops->writepage(page, &wbc);
649 handle_write_error(mapping, page, res);
650 if (res == AOP_WRITEPAGE_ACTIVATE) {
651 ClearPageReclaim(page);
652 return PAGE_ACTIVATE;
655 if (!PageWriteback(page)) {
656 /* synchronous write or broken a_ops? */
657 ClearPageReclaim(page);
659 trace_mm_vmscan_writepage(page);
660 inc_node_page_state(page, NR_VMSCAN_WRITE);
668 * Same as remove_mapping, but if the page is removed from the mapping, it
669 * gets returned with a refcount of 0.
671 static int __remove_mapping(struct address_space *mapping, struct page *page,
677 BUG_ON(!PageLocked(page));
678 BUG_ON(mapping != page_mapping(page));
680 spin_lock_irqsave(&mapping->tree_lock, flags);
682 * The non racy check for a busy page.
684 * Must be careful with the order of the tests. When someone has
685 * a ref to the page, it may be possible that they dirty it then
686 * drop the reference. So if PageDirty is tested before page_count
687 * here, then the following race may occur:
689 * get_user_pages(&page);
690 * [user mapping goes away]
692 * !PageDirty(page) [good]
693 * SetPageDirty(page);
695 * !page_count(page) [good, discard it]
697 * [oops, our write_to data is lost]
699 * Reversing the order of the tests ensures such a situation cannot
700 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
701 * load is not satisfied before that of page->_refcount.
703 * Note that if SetPageDirty is always performed via set_page_dirty,
704 * and thus under tree_lock, then this ordering is not required.
706 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
707 refcount = 1 + HPAGE_PMD_NR;
710 if (!page_ref_freeze(page, refcount))
712 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
713 if (unlikely(PageDirty(page))) {
714 page_ref_unfreeze(page, refcount);
718 if (PageSwapCache(page)) {
719 swp_entry_t swap = { .val = page_private(page) };
720 mem_cgroup_swapout(page, swap);
721 __delete_from_swap_cache(page);
722 spin_unlock_irqrestore(&mapping->tree_lock, flags);
723 put_swap_page(page, swap);
725 void (*freepage)(struct page *);
728 freepage = mapping->a_ops->freepage;
730 * Remember a shadow entry for reclaimed file cache in
731 * order to detect refaults, thus thrashing, later on.
733 * But don't store shadows in an address space that is
734 * already exiting. This is not just an optizimation,
735 * inode reclaim needs to empty out the radix tree or
736 * the nodes are lost. Don't plant shadows behind its
739 * We also don't store shadows for DAX mappings because the
740 * only page cache pages found in these are zero pages
741 * covering holes, and because we don't want to mix DAX
742 * exceptional entries and shadow exceptional entries in the
745 if (reclaimed && page_is_file_cache(page) &&
746 !mapping_exiting(mapping) && !dax_mapping(mapping))
747 shadow = workingset_eviction(mapping, page);
748 __delete_from_page_cache(page, shadow);
749 spin_unlock_irqrestore(&mapping->tree_lock, flags);
751 if (freepage != NULL)
758 spin_unlock_irqrestore(&mapping->tree_lock, flags);
763 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
764 * someone else has a ref on the page, abort and return 0. If it was
765 * successfully detached, return 1. Assumes the caller has a single ref on
768 int remove_mapping(struct address_space *mapping, struct page *page)
770 if (__remove_mapping(mapping, page, false)) {
772 * Unfreezing the refcount with 1 rather than 2 effectively
773 * drops the pagecache ref for us without requiring another
776 page_ref_unfreeze(page, 1);
783 * putback_lru_page - put previously isolated page onto appropriate LRU list
784 * @page: page to be put back to appropriate lru list
786 * Add previously isolated @page to appropriate LRU list.
787 * Page may still be unevictable for other reasons.
789 * lru_lock must not be held, interrupts must be enabled.
791 void putback_lru_page(struct page *page)
794 int was_unevictable = PageUnevictable(page);
796 VM_BUG_ON_PAGE(PageLRU(page), page);
799 ClearPageUnevictable(page);
801 if (page_evictable(page)) {
803 * For evictable pages, we can use the cache.
804 * In event of a race, worst case is we end up with an
805 * unevictable page on [in]active list.
806 * We know how to handle that.
808 is_unevictable = false;
812 * Put unevictable pages directly on zone's unevictable
815 is_unevictable = true;
816 add_page_to_unevictable_list(page);
818 * When racing with an mlock or AS_UNEVICTABLE clearing
819 * (page is unlocked) make sure that if the other thread
820 * does not observe our setting of PG_lru and fails
821 * isolation/check_move_unevictable_pages,
822 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
823 * the page back to the evictable list.
825 * The other side is TestClearPageMlocked() or shmem_lock().
831 * page's status can change while we move it among lru. If an evictable
832 * page is on unevictable list, it never be freed. To avoid that,
833 * check after we added it to the list, again.
835 if (is_unevictable && page_evictable(page)) {
836 if (!isolate_lru_page(page)) {
840 /* This means someone else dropped this page from LRU
841 * So, it will be freed or putback to LRU again. There is
842 * nothing to do here.
846 if (was_unevictable && !is_unevictable)
847 count_vm_event(UNEVICTABLE_PGRESCUED);
848 else if (!was_unevictable && is_unevictable)
849 count_vm_event(UNEVICTABLE_PGCULLED);
851 put_page(page); /* drop ref from isolate */
854 enum page_references {
856 PAGEREF_RECLAIM_CLEAN,
861 static enum page_references page_check_references(struct page *page,
862 struct scan_control *sc)
864 int referenced_ptes, referenced_page;
865 unsigned long vm_flags;
867 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
869 referenced_page = TestClearPageReferenced(page);
872 * Mlock lost the isolation race with us. Let try_to_unmap()
873 * move the page to the unevictable list.
875 if (vm_flags & VM_LOCKED)
876 return PAGEREF_RECLAIM;
878 if (referenced_ptes) {
879 if (PageSwapBacked(page))
880 return PAGEREF_ACTIVATE;
882 * All mapped pages start out with page table
883 * references from the instantiating fault, so we need
884 * to look twice if a mapped file page is used more
887 * Mark it and spare it for another trip around the
888 * inactive list. Another page table reference will
889 * lead to its activation.
891 * Note: the mark is set for activated pages as well
892 * so that recently deactivated but used pages are
895 SetPageReferenced(page);
897 if (referenced_page || referenced_ptes > 1)
898 return PAGEREF_ACTIVATE;
901 * Activate file-backed executable pages after first usage.
903 if (vm_flags & VM_EXEC)
904 return PAGEREF_ACTIVATE;
909 /* Reclaim if clean, defer dirty pages to writeback */
910 if (referenced_page && !PageSwapBacked(page))
911 return PAGEREF_RECLAIM_CLEAN;
913 return PAGEREF_RECLAIM;
916 /* Check if a page is dirty or under writeback */
917 static void page_check_dirty_writeback(struct page *page,
918 bool *dirty, bool *writeback)
920 struct address_space *mapping;
923 * Anonymous pages are not handled by flushers and must be written
924 * from reclaim context. Do not stall reclaim based on them
926 if (!page_is_file_cache(page) ||
927 (PageAnon(page) && !PageSwapBacked(page))) {
933 /* By default assume that the page flags are accurate */
934 *dirty = PageDirty(page);
935 *writeback = PageWriteback(page);
937 /* Verify dirty/writeback state if the filesystem supports it */
938 if (!page_has_private(page))
941 mapping = page_mapping(page);
942 if (mapping && mapping->a_ops->is_dirty_writeback)
943 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
946 struct reclaim_stat {
948 unsigned nr_unqueued_dirty;
949 unsigned nr_congested;
950 unsigned nr_writeback;
951 unsigned nr_immediate;
952 unsigned nr_activate;
953 unsigned nr_ref_keep;
954 unsigned nr_unmap_fail;
958 * shrink_page_list() returns the number of reclaimed pages
960 static unsigned long shrink_page_list(struct list_head *page_list,
961 struct pglist_data *pgdat,
962 struct scan_control *sc,
963 enum ttu_flags ttu_flags,
964 struct reclaim_stat *stat,
967 LIST_HEAD(ret_pages);
968 LIST_HEAD(free_pages);
970 unsigned nr_unqueued_dirty = 0;
971 unsigned nr_dirty = 0;
972 unsigned nr_congested = 0;
973 unsigned nr_reclaimed = 0;
974 unsigned nr_writeback = 0;
975 unsigned nr_immediate = 0;
976 unsigned nr_ref_keep = 0;
977 unsigned nr_unmap_fail = 0;
981 while (!list_empty(page_list)) {
982 struct address_space *mapping;
985 enum page_references references = PAGEREF_RECLAIM_CLEAN;
986 bool dirty, writeback;
990 page = lru_to_page(page_list);
991 list_del(&page->lru);
993 if (!trylock_page(page))
996 VM_BUG_ON_PAGE(PageActive(page), page);
1000 if (unlikely(!page_evictable(page)))
1001 goto activate_locked;
1003 if (!sc->may_unmap && page_mapped(page))
1006 /* Double the slab pressure for mapped and swapcache pages */
1007 if ((page_mapped(page) || PageSwapCache(page)) &&
1008 !(PageAnon(page) && !PageSwapBacked(page)))
1011 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1012 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1015 * The number of dirty pages determines if a zone is marked
1016 * reclaim_congested which affects wait_iff_congested. kswapd
1017 * will stall and start writing pages if the tail of the LRU
1018 * is all dirty unqueued pages.
1020 page_check_dirty_writeback(page, &dirty, &writeback);
1021 if (dirty || writeback)
1024 if (dirty && !writeback)
1025 nr_unqueued_dirty++;
1028 * Treat this page as congested if the underlying BDI is or if
1029 * pages are cycling through the LRU so quickly that the
1030 * pages marked for immediate reclaim are making it to the
1031 * end of the LRU a second time.
1033 mapping = page_mapping(page);
1034 if (((dirty || writeback) && mapping &&
1035 inode_write_congested(mapping->host)) ||
1036 (writeback && PageReclaim(page)))
1040 * If a page at the tail of the LRU is under writeback, there
1041 * are three cases to consider.
1043 * 1) If reclaim is encountering an excessive number of pages
1044 * under writeback and this page is both under writeback and
1045 * PageReclaim then it indicates that pages are being queued
1046 * for IO but are being recycled through the LRU before the
1047 * IO can complete. Waiting on the page itself risks an
1048 * indefinite stall if it is impossible to writeback the
1049 * page due to IO error or disconnected storage so instead
1050 * note that the LRU is being scanned too quickly and the
1051 * caller can stall after page list has been processed.
1053 * 2) Global or new memcg reclaim encounters a page that is
1054 * not marked for immediate reclaim, or the caller does not
1055 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1056 * not to fs). In this case mark the page for immediate
1057 * reclaim and continue scanning.
1059 * Require may_enter_fs because we would wait on fs, which
1060 * may not have submitted IO yet. And the loop driver might
1061 * enter reclaim, and deadlock if it waits on a page for
1062 * which it is needed to do the write (loop masks off
1063 * __GFP_IO|__GFP_FS for this reason); but more thought
1064 * would probably show more reasons.
1066 * 3) Legacy memcg encounters a page that is already marked
1067 * PageReclaim. memcg does not have any dirty pages
1068 * throttling so we could easily OOM just because too many
1069 * pages are in writeback and there is nothing else to
1070 * reclaim. Wait for the writeback to complete.
1072 * In cases 1) and 2) we activate the pages to get them out of
1073 * the way while we continue scanning for clean pages on the
1074 * inactive list and refilling from the active list. The
1075 * observation here is that waiting for disk writes is more
1076 * expensive than potentially causing reloads down the line.
1077 * Since they're marked for immediate reclaim, they won't put
1078 * memory pressure on the cache working set any longer than it
1079 * takes to write them to disk.
1081 if (PageWriteback(page)) {
1083 if (current_is_kswapd() &&
1084 PageReclaim(page) &&
1085 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1087 goto activate_locked;
1090 } else if (sane_reclaim(sc) ||
1091 !PageReclaim(page) || !may_enter_fs) {
1093 * This is slightly racy - end_page_writeback()
1094 * might have just cleared PageReclaim, then
1095 * setting PageReclaim here end up interpreted
1096 * as PageReadahead - but that does not matter
1097 * enough to care. What we do want is for this
1098 * page to have PageReclaim set next time memcg
1099 * reclaim reaches the tests above, so it will
1100 * then wait_on_page_writeback() to avoid OOM;
1101 * and it's also appropriate in global reclaim.
1103 SetPageReclaim(page);
1105 goto activate_locked;
1110 wait_on_page_writeback(page);
1111 /* then go back and try same page again */
1112 list_add_tail(&page->lru, page_list);
1118 references = page_check_references(page, sc);
1120 switch (references) {
1121 case PAGEREF_ACTIVATE:
1122 goto activate_locked;
1126 case PAGEREF_RECLAIM:
1127 case PAGEREF_RECLAIM_CLEAN:
1128 ; /* try to reclaim the page below */
1132 * Anonymous process memory has backing store?
1133 * Try to allocate it some swap space here.
1134 * Lazyfree page could be freed directly
1136 if (PageAnon(page) && PageSwapBacked(page)) {
1137 if (!PageSwapCache(page)) {
1138 if (!(sc->gfp_mask & __GFP_IO))
1140 if (PageTransHuge(page)) {
1141 /* cannot split THP, skip it */
1142 if (!can_split_huge_page(page, NULL))
1143 goto activate_locked;
1145 * Split pages without a PMD map right
1146 * away. Chances are some or all of the
1147 * tail pages can be freed without IO.
1149 if (!compound_mapcount(page) &&
1150 split_huge_page_to_list(page,
1152 goto activate_locked;
1154 if (!add_to_swap(page)) {
1155 if (!PageTransHuge(page))
1156 goto activate_locked;
1157 /* Fallback to swap normal pages */
1158 if (split_huge_page_to_list(page,
1160 goto activate_locked;
1161 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1162 count_vm_event(THP_SWPOUT_FALLBACK);
1164 if (!add_to_swap(page))
1165 goto activate_locked;
1170 /* Adding to swap updated mapping */
1171 mapping = page_mapping(page);
1173 } else if (unlikely(PageTransHuge(page))) {
1174 /* Split file THP */
1175 if (split_huge_page_to_list(page, page_list))
1180 * The page is mapped into the page tables of one or more
1181 * processes. Try to unmap it here.
1183 if (page_mapped(page)) {
1184 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1186 if (unlikely(PageTransHuge(page)))
1187 flags |= TTU_SPLIT_HUGE_PMD;
1188 if (!try_to_unmap(page, flags)) {
1190 goto activate_locked;
1194 if (PageDirty(page)) {
1196 * Only kswapd can writeback filesystem pages
1197 * to avoid risk of stack overflow. But avoid
1198 * injecting inefficient single-page IO into
1199 * flusher writeback as much as possible: only
1200 * write pages when we've encountered many
1201 * dirty pages, and when we've already scanned
1202 * the rest of the LRU for clean pages and see
1203 * the same dirty pages again (PageReclaim).
1205 if (page_is_file_cache(page) &&
1206 (!current_is_kswapd() || !PageReclaim(page) ||
1207 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1209 * Immediately reclaim when written back.
1210 * Similar in principal to deactivate_page()
1211 * except we already have the page isolated
1212 * and know it's dirty
1214 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1215 SetPageReclaim(page);
1217 goto activate_locked;
1220 if (references == PAGEREF_RECLAIM_CLEAN)
1224 if (!sc->may_writepage)
1228 * Page is dirty. Flush the TLB if a writable entry
1229 * potentially exists to avoid CPU writes after IO
1230 * starts and then write it out here.
1232 try_to_unmap_flush_dirty();
1233 switch (pageout(page, mapping, sc)) {
1237 goto activate_locked;
1239 if (PageWriteback(page))
1241 if (PageDirty(page))
1245 * A synchronous write - probably a ramdisk. Go
1246 * ahead and try to reclaim the page.
1248 if (!trylock_page(page))
1250 if (PageDirty(page) || PageWriteback(page))
1252 mapping = page_mapping(page);
1254 ; /* try to free the page below */
1259 * If the page has buffers, try to free the buffer mappings
1260 * associated with this page. If we succeed we try to free
1263 * We do this even if the page is PageDirty().
1264 * try_to_release_page() does not perform I/O, but it is
1265 * possible for a page to have PageDirty set, but it is actually
1266 * clean (all its buffers are clean). This happens if the
1267 * buffers were written out directly, with submit_bh(). ext3
1268 * will do this, as well as the blockdev mapping.
1269 * try_to_release_page() will discover that cleanness and will
1270 * drop the buffers and mark the page clean - it can be freed.
1272 * Rarely, pages can have buffers and no ->mapping. These are
1273 * the pages which were not successfully invalidated in
1274 * truncate_complete_page(). We try to drop those buffers here
1275 * and if that worked, and the page is no longer mapped into
1276 * process address space (page_count == 1) it can be freed.
1277 * Otherwise, leave the page on the LRU so it is swappable.
1279 if (page_has_private(page)) {
1280 if (!try_to_release_page(page, sc->gfp_mask))
1281 goto activate_locked;
1282 if (!mapping && page_count(page) == 1) {
1284 if (put_page_testzero(page))
1288 * rare race with speculative reference.
1289 * the speculative reference will free
1290 * this page shortly, so we may
1291 * increment nr_reclaimed here (and
1292 * leave it off the LRU).
1300 if (PageAnon(page) && !PageSwapBacked(page)) {
1301 /* follow __remove_mapping for reference */
1302 if (!page_ref_freeze(page, 1))
1304 if (PageDirty(page)) {
1305 page_ref_unfreeze(page, 1);
1309 count_vm_event(PGLAZYFREED);
1310 count_memcg_page_event(page, PGLAZYFREED);
1311 } else if (!mapping || !__remove_mapping(mapping, page, true))
1314 * At this point, we have no other references and there is
1315 * no way to pick any more up (removed from LRU, removed
1316 * from pagecache). Can use non-atomic bitops now (and
1317 * we obviously don't have to worry about waking up a process
1318 * waiting on the page lock, because there are no references.
1320 __ClearPageLocked(page);
1325 * Is there need to periodically free_page_list? It would
1326 * appear not as the counts should be low
1328 if (unlikely(PageTransHuge(page))) {
1329 mem_cgroup_uncharge(page);
1330 (*get_compound_page_dtor(page))(page);
1332 list_add(&page->lru, &free_pages);
1336 /* Not a candidate for swapping, so reclaim swap space. */
1337 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1339 try_to_free_swap(page);
1340 VM_BUG_ON_PAGE(PageActive(page), page);
1341 if (!PageMlocked(page)) {
1342 SetPageActive(page);
1344 count_memcg_page_event(page, PGACTIVATE);
1349 list_add(&page->lru, &ret_pages);
1350 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1353 mem_cgroup_uncharge_list(&free_pages);
1354 try_to_unmap_flush();
1355 free_hot_cold_page_list(&free_pages, true);
1357 list_splice(&ret_pages, page_list);
1358 count_vm_events(PGACTIVATE, pgactivate);
1361 stat->nr_dirty = nr_dirty;
1362 stat->nr_congested = nr_congested;
1363 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1364 stat->nr_writeback = nr_writeback;
1365 stat->nr_immediate = nr_immediate;
1366 stat->nr_activate = pgactivate;
1367 stat->nr_ref_keep = nr_ref_keep;
1368 stat->nr_unmap_fail = nr_unmap_fail;
1370 return nr_reclaimed;
1373 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1374 struct list_head *page_list)
1376 struct scan_control sc = {
1377 .gfp_mask = GFP_KERNEL,
1378 .priority = DEF_PRIORITY,
1382 struct page *page, *next;
1383 LIST_HEAD(clean_pages);
1385 list_for_each_entry_safe(page, next, page_list, lru) {
1386 if (page_is_file_cache(page) && !PageDirty(page) &&
1387 !__PageMovable(page)) {
1388 ClearPageActive(page);
1389 list_move(&page->lru, &clean_pages);
1393 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1394 TTU_IGNORE_ACCESS, NULL, true);
1395 list_splice(&clean_pages, page_list);
1396 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1401 * Attempt to remove the specified page from its LRU. Only take this page
1402 * if it is of the appropriate PageActive status. Pages which are being
1403 * freed elsewhere are also ignored.
1405 * page: page to consider
1406 * mode: one of the LRU isolation modes defined above
1408 * returns 0 on success, -ve errno on failure.
1410 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1414 /* Only take pages on the LRU. */
1418 /* Compaction should not handle unevictable pages but CMA can do so */
1419 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1425 * To minimise LRU disruption, the caller can indicate that it only
1426 * wants to isolate pages it will be able to operate on without
1427 * blocking - clean pages for the most part.
1429 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1430 * that it is possible to migrate without blocking
1432 if (mode & ISOLATE_ASYNC_MIGRATE) {
1433 /* All the caller can do on PageWriteback is block */
1434 if (PageWriteback(page))
1437 if (PageDirty(page)) {
1438 struct address_space *mapping;
1441 * Only pages without mappings or that have a
1442 * ->migratepage callback are possible to migrate
1445 mapping = page_mapping(page);
1446 if (mapping && !mapping->a_ops->migratepage)
1451 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1454 if (likely(get_page_unless_zero(page))) {
1456 * Be careful not to clear PageLRU until after we're
1457 * sure the page is not being freed elsewhere -- the
1458 * page release code relies on it.
1469 * Update LRU sizes after isolating pages. The LRU size updates must
1470 * be complete before mem_cgroup_update_lru_size due to a santity check.
1472 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1473 enum lru_list lru, unsigned long *nr_zone_taken)
1477 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1478 if (!nr_zone_taken[zid])
1481 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1483 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1490 * zone_lru_lock is heavily contended. Some of the functions that
1491 * shrink the lists perform better by taking out a batch of pages
1492 * and working on them outside the LRU lock.
1494 * For pagecache intensive workloads, this function is the hottest
1495 * spot in the kernel (apart from copy_*_user functions).
1497 * Appropriate locks must be held before calling this function.
1499 * @nr_to_scan: The number of eligible pages to look through on the list.
1500 * @lruvec: The LRU vector to pull pages from.
1501 * @dst: The temp list to put pages on to.
1502 * @nr_scanned: The number of pages that were scanned.
1503 * @sc: The scan_control struct for this reclaim session
1504 * @mode: One of the LRU isolation modes
1505 * @lru: LRU list id for isolating
1507 * returns how many pages were moved onto *@dst.
1509 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1510 struct lruvec *lruvec, struct list_head *dst,
1511 unsigned long *nr_scanned, struct scan_control *sc,
1512 isolate_mode_t mode, enum lru_list lru)
1514 struct list_head *src = &lruvec->lists[lru];
1515 unsigned long nr_taken = 0;
1516 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1517 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1518 unsigned long skipped = 0;
1519 unsigned long scan, total_scan, nr_pages;
1520 LIST_HEAD(pages_skipped);
1523 for (total_scan = 0;
1524 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1528 page = lru_to_page(src);
1529 prefetchw_prev_lru_page(page, src, flags);
1531 VM_BUG_ON_PAGE(!PageLRU(page), page);
1533 if (page_zonenum(page) > sc->reclaim_idx) {
1534 list_move(&page->lru, &pages_skipped);
1535 nr_skipped[page_zonenum(page)]++;
1540 * Do not count skipped pages because that makes the function
1541 * return with no isolated pages if the LRU mostly contains
1542 * ineligible pages. This causes the VM to not reclaim any
1543 * pages, triggering a premature OOM.
1546 switch (__isolate_lru_page(page, mode)) {
1548 nr_pages = hpage_nr_pages(page);
1549 nr_taken += nr_pages;
1550 nr_zone_taken[page_zonenum(page)] += nr_pages;
1551 list_move(&page->lru, dst);
1555 /* else it is being freed elsewhere */
1556 list_move(&page->lru, src);
1565 * Splice any skipped pages to the start of the LRU list. Note that
1566 * this disrupts the LRU order when reclaiming for lower zones but
1567 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1568 * scanning would soon rescan the same pages to skip and put the
1569 * system at risk of premature OOM.
1571 if (!list_empty(&pages_skipped)) {
1574 list_splice(&pages_skipped, src);
1575 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1576 if (!nr_skipped[zid])
1579 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1580 skipped += nr_skipped[zid];
1583 *nr_scanned = total_scan;
1584 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1585 total_scan, skipped, nr_taken, mode, lru);
1586 update_lru_sizes(lruvec, lru, nr_zone_taken);
1591 * isolate_lru_page - tries to isolate a page from its LRU list
1592 * @page: page to isolate from its LRU list
1594 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1595 * vmstat statistic corresponding to whatever LRU list the page was on.
1597 * Returns 0 if the page was removed from an LRU list.
1598 * Returns -EBUSY if the page was not on an LRU list.
1600 * The returned page will have PageLRU() cleared. If it was found on
1601 * the active list, it will have PageActive set. If it was found on
1602 * the unevictable list, it will have the PageUnevictable bit set. That flag
1603 * may need to be cleared by the caller before letting the page go.
1605 * The vmstat statistic corresponding to the list on which the page was
1606 * found will be decremented.
1609 * (1) Must be called with an elevated refcount on the page. This is a
1610 * fundamentnal difference from isolate_lru_pages (which is called
1611 * without a stable reference).
1612 * (2) the lru_lock must not be held.
1613 * (3) interrupts must be enabled.
1615 int isolate_lru_page(struct page *page)
1619 VM_BUG_ON_PAGE(!page_count(page), page);
1620 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1622 if (PageLRU(page)) {
1623 struct zone *zone = page_zone(page);
1624 struct lruvec *lruvec;
1626 spin_lock_irq(zone_lru_lock(zone));
1627 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1628 if (PageLRU(page)) {
1629 int lru = page_lru(page);
1632 del_page_from_lru_list(page, lruvec, lru);
1635 spin_unlock_irq(zone_lru_lock(zone));
1641 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1642 * then get resheduled. When there are massive number of tasks doing page
1643 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1644 * the LRU list will go small and be scanned faster than necessary, leading to
1645 * unnecessary swapping, thrashing and OOM.
1647 static int too_many_isolated(struct pglist_data *pgdat, int file,
1648 struct scan_control *sc)
1650 unsigned long inactive, isolated;
1652 if (current_is_kswapd())
1655 if (!sane_reclaim(sc))
1659 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1660 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1662 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1663 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1667 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1668 * won't get blocked by normal direct-reclaimers, forming a circular
1671 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1674 return isolated > inactive;
1677 static noinline_for_stack void
1678 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1680 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1681 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1682 LIST_HEAD(pages_to_free);
1685 * Put back any unfreeable pages.
1687 while (!list_empty(page_list)) {
1688 struct page *page = lru_to_page(page_list);
1691 VM_BUG_ON_PAGE(PageLRU(page), page);
1692 list_del(&page->lru);
1693 if (unlikely(!page_evictable(page))) {
1694 spin_unlock_irq(&pgdat->lru_lock);
1695 putback_lru_page(page);
1696 spin_lock_irq(&pgdat->lru_lock);
1700 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1703 lru = page_lru(page);
1704 add_page_to_lru_list(page, lruvec, lru);
1706 if (is_active_lru(lru)) {
1707 int file = is_file_lru(lru);
1708 int numpages = hpage_nr_pages(page);
1709 reclaim_stat->recent_rotated[file] += numpages;
1711 if (put_page_testzero(page)) {
1712 __ClearPageLRU(page);
1713 __ClearPageActive(page);
1714 del_page_from_lru_list(page, lruvec, lru);
1716 if (unlikely(PageCompound(page))) {
1717 spin_unlock_irq(&pgdat->lru_lock);
1718 mem_cgroup_uncharge(page);
1719 (*get_compound_page_dtor(page))(page);
1720 spin_lock_irq(&pgdat->lru_lock);
1722 list_add(&page->lru, &pages_to_free);
1727 * To save our caller's stack, now use input list for pages to free.
1729 list_splice(&pages_to_free, page_list);
1733 * If a kernel thread (such as nfsd for loop-back mounts) services
1734 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1735 * In that case we should only throttle if the backing device it is
1736 * writing to is congested. In other cases it is safe to throttle.
1738 static int current_may_throttle(void)
1740 return !(current->flags & PF_LESS_THROTTLE) ||
1741 current->backing_dev_info == NULL ||
1742 bdi_write_congested(current->backing_dev_info);
1746 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1747 * of reclaimed pages
1749 static noinline_for_stack unsigned long
1750 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1751 struct scan_control *sc, enum lru_list lru)
1753 LIST_HEAD(page_list);
1754 unsigned long nr_scanned;
1755 unsigned long nr_reclaimed = 0;
1756 unsigned long nr_taken;
1757 struct reclaim_stat stat = {};
1758 isolate_mode_t isolate_mode = 0;
1759 int file = is_file_lru(lru);
1760 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1761 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1762 bool stalled = false;
1764 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1768 /* wait a bit for the reclaimer. */
1772 /* We are about to die and free our memory. Return now. */
1773 if (fatal_signal_pending(current))
1774 return SWAP_CLUSTER_MAX;
1780 isolate_mode |= ISOLATE_UNMAPPED;
1782 spin_lock_irq(&pgdat->lru_lock);
1784 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1785 &nr_scanned, sc, isolate_mode, lru);
1787 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1788 reclaim_stat->recent_scanned[file] += nr_taken;
1790 if (current_is_kswapd()) {
1791 if (global_reclaim(sc))
1792 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1793 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1796 if (global_reclaim(sc))
1797 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1798 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1801 spin_unlock_irq(&pgdat->lru_lock);
1806 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1809 spin_lock_irq(&pgdat->lru_lock);
1811 if (current_is_kswapd()) {
1812 if (global_reclaim(sc))
1813 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1814 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1817 if (global_reclaim(sc))
1818 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1819 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1823 putback_inactive_pages(lruvec, &page_list);
1825 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1827 spin_unlock_irq(&pgdat->lru_lock);
1829 mem_cgroup_uncharge_list(&page_list);
1830 free_hot_cold_page_list(&page_list, true);
1833 * If reclaim is isolating dirty pages under writeback, it implies
1834 * that the long-lived page allocation rate is exceeding the page
1835 * laundering rate. Either the global limits are not being effective
1836 * at throttling processes due to the page distribution throughout
1837 * zones or there is heavy usage of a slow backing device. The
1838 * only option is to throttle from reclaim context which is not ideal
1839 * as there is no guarantee the dirtying process is throttled in the
1840 * same way balance_dirty_pages() manages.
1842 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1843 * of pages under pages flagged for immediate reclaim and stall if any
1844 * are encountered in the nr_immediate check below.
1846 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1847 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1850 * If dirty pages are scanned that are not queued for IO, it
1851 * implies that flushers are not doing their job. This can
1852 * happen when memory pressure pushes dirty pages to the end of
1853 * the LRU before the dirty limits are breached and the dirty
1854 * data has expired. It can also happen when the proportion of
1855 * dirty pages grows not through writes but through memory
1856 * pressure reclaiming all the clean cache. And in some cases,
1857 * the flushers simply cannot keep up with the allocation
1858 * rate. Nudge the flusher threads in case they are asleep.
1860 if (stat.nr_unqueued_dirty == nr_taken)
1861 wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1864 * Legacy memcg will stall in page writeback so avoid forcibly
1867 if (sane_reclaim(sc)) {
1869 * Tag a zone as congested if all the dirty pages scanned were
1870 * backed by a congested BDI and wait_iff_congested will stall.
1872 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1873 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1875 /* Allow kswapd to start writing pages during reclaim. */
1876 if (stat.nr_unqueued_dirty == nr_taken)
1877 set_bit(PGDAT_DIRTY, &pgdat->flags);
1880 * If kswapd scans pages marked marked for immediate
1881 * reclaim and under writeback (nr_immediate), it implies
1882 * that pages are cycling through the LRU faster than
1883 * they are written so also forcibly stall.
1885 if (stat.nr_immediate && current_may_throttle())
1886 congestion_wait(BLK_RW_ASYNC, HZ/10);
1890 * Stall direct reclaim for IO completions if underlying BDIs or zone
1891 * is congested. Allow kswapd to continue until it starts encountering
1892 * unqueued dirty pages or cycling through the LRU too quickly.
1894 if (!sc->hibernation_mode && !current_is_kswapd() &&
1895 current_may_throttle())
1896 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1898 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1899 nr_scanned, nr_reclaimed,
1900 stat.nr_dirty, stat.nr_writeback,
1901 stat.nr_congested, stat.nr_immediate,
1902 stat.nr_activate, stat.nr_ref_keep,
1904 sc->priority, file);
1905 return nr_reclaimed;
1909 * This moves pages from the active list to the inactive list.
1911 * We move them the other way if the page is referenced by one or more
1912 * processes, from rmap.
1914 * If the pages are mostly unmapped, the processing is fast and it is
1915 * appropriate to hold zone_lru_lock across the whole operation. But if
1916 * the pages are mapped, the processing is slow (page_referenced()) so we
1917 * should drop zone_lru_lock around each page. It's impossible to balance
1918 * this, so instead we remove the pages from the LRU while processing them.
1919 * It is safe to rely on PG_active against the non-LRU pages in here because
1920 * nobody will play with that bit on a non-LRU page.
1922 * The downside is that we have to touch page->_refcount against each page.
1923 * But we had to alter page->flags anyway.
1925 * Returns the number of pages moved to the given lru.
1928 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1929 struct list_head *list,
1930 struct list_head *pages_to_free,
1933 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1938 while (!list_empty(list)) {
1939 page = lru_to_page(list);
1940 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1942 VM_BUG_ON_PAGE(PageLRU(page), page);
1945 nr_pages = hpage_nr_pages(page);
1946 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1947 list_move(&page->lru, &lruvec->lists[lru]);
1949 if (put_page_testzero(page)) {
1950 __ClearPageLRU(page);
1951 __ClearPageActive(page);
1952 del_page_from_lru_list(page, lruvec, lru);
1954 if (unlikely(PageCompound(page))) {
1955 spin_unlock_irq(&pgdat->lru_lock);
1956 mem_cgroup_uncharge(page);
1957 (*get_compound_page_dtor(page))(page);
1958 spin_lock_irq(&pgdat->lru_lock);
1960 list_add(&page->lru, pages_to_free);
1962 nr_moved += nr_pages;
1966 if (!is_active_lru(lru)) {
1967 __count_vm_events(PGDEACTIVATE, nr_moved);
1968 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1975 static void shrink_active_list(unsigned long nr_to_scan,
1976 struct lruvec *lruvec,
1977 struct scan_control *sc,
1980 unsigned long nr_taken;
1981 unsigned long nr_scanned;
1982 unsigned long vm_flags;
1983 LIST_HEAD(l_hold); /* The pages which were snipped off */
1984 LIST_HEAD(l_active);
1985 LIST_HEAD(l_inactive);
1987 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1988 unsigned nr_deactivate, nr_activate;
1989 unsigned nr_rotated = 0;
1990 isolate_mode_t isolate_mode = 0;
1991 int file = is_file_lru(lru);
1992 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1997 isolate_mode |= ISOLATE_UNMAPPED;
1999 spin_lock_irq(&pgdat->lru_lock);
2001 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2002 &nr_scanned, sc, isolate_mode, lru);
2004 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2005 reclaim_stat->recent_scanned[file] += nr_taken;
2007 __count_vm_events(PGREFILL, nr_scanned);
2008 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2010 spin_unlock_irq(&pgdat->lru_lock);
2012 while (!list_empty(&l_hold)) {
2014 page = lru_to_page(&l_hold);
2015 list_del(&page->lru);
2017 if (unlikely(!page_evictable(page))) {
2018 putback_lru_page(page);
2022 if (unlikely(buffer_heads_over_limit)) {
2023 if (page_has_private(page) && trylock_page(page)) {
2024 if (page_has_private(page))
2025 try_to_release_page(page, 0);
2030 if (page_referenced(page, 0, sc->target_mem_cgroup,
2032 nr_rotated += hpage_nr_pages(page);
2034 * Identify referenced, file-backed active pages and
2035 * give them one more trip around the active list. So
2036 * that executable code get better chances to stay in
2037 * memory under moderate memory pressure. Anon pages
2038 * are not likely to be evicted by use-once streaming
2039 * IO, plus JVM can create lots of anon VM_EXEC pages,
2040 * so we ignore them here.
2042 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2043 list_add(&page->lru, &l_active);
2048 ClearPageActive(page); /* we are de-activating */
2049 list_add(&page->lru, &l_inactive);
2053 * Move pages back to the lru list.
2055 spin_lock_irq(&pgdat->lru_lock);
2057 * Count referenced pages from currently used mappings as rotated,
2058 * even though only some of them are actually re-activated. This
2059 * helps balance scan pressure between file and anonymous pages in
2062 reclaim_stat->recent_rotated[file] += nr_rotated;
2064 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2065 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2066 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2067 spin_unlock_irq(&pgdat->lru_lock);
2069 mem_cgroup_uncharge_list(&l_hold);
2070 free_hot_cold_page_list(&l_hold, true);
2071 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2072 nr_deactivate, nr_rotated, sc->priority, file);
2076 * The inactive anon list should be small enough that the VM never has
2077 * to do too much work.
2079 * The inactive file list should be small enough to leave most memory
2080 * to the established workingset on the scan-resistant active list,
2081 * but large enough to avoid thrashing the aggregate readahead window.
2083 * Both inactive lists should also be large enough that each inactive
2084 * page has a chance to be referenced again before it is reclaimed.
2086 * If that fails and refaulting is observed, the inactive list grows.
2088 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2089 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2090 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2093 * memory ratio inactive
2094 * -------------------------------------
2103 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2104 struct mem_cgroup *memcg,
2105 struct scan_control *sc, bool actual_reclaim)
2107 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2108 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2109 enum lru_list inactive_lru = file * LRU_FILE;
2110 unsigned long inactive, active;
2111 unsigned long inactive_ratio;
2112 unsigned long refaults;
2116 * If we don't have swap space, anonymous page deactivation
2119 if (!file && !total_swap_pages)
2122 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2123 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2126 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2128 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2131 * When refaults are being observed, it means a new workingset
2132 * is being established. Disable active list protection to get
2133 * rid of the stale workingset quickly.
2135 if (file && actual_reclaim && lruvec->refaults != refaults) {
2138 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2140 inactive_ratio = int_sqrt(10 * gb);
2146 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2147 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2148 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2149 inactive_ratio, file);
2151 return inactive * inactive_ratio < active;
2154 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2155 struct lruvec *lruvec, struct mem_cgroup *memcg,
2156 struct scan_control *sc)
2158 if (is_active_lru(lru)) {
2159 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2161 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2165 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2176 * Determine how aggressively the anon and file LRU lists should be
2177 * scanned. The relative value of each set of LRU lists is determined
2178 * by looking at the fraction of the pages scanned we did rotate back
2179 * onto the active list instead of evict.
2181 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2182 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2184 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2185 struct scan_control *sc, unsigned long *nr,
2186 unsigned long *lru_pages)
2188 int swappiness = mem_cgroup_swappiness(memcg);
2189 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2191 u64 denominator = 0; /* gcc */
2192 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2193 unsigned long anon_prio, file_prio;
2194 enum scan_balance scan_balance;
2195 unsigned long anon, file;
2196 unsigned long ap, fp;
2199 /* If we have no swap space, do not bother scanning anon pages. */
2200 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2201 scan_balance = SCAN_FILE;
2206 * Global reclaim will swap to prevent OOM even with no
2207 * swappiness, but memcg users want to use this knob to
2208 * disable swapping for individual groups completely when
2209 * using the memory controller's swap limit feature would be
2212 if (!global_reclaim(sc) && !swappiness) {
2213 scan_balance = SCAN_FILE;
2218 * Do not apply any pressure balancing cleverness when the
2219 * system is close to OOM, scan both anon and file equally
2220 * (unless the swappiness setting disagrees with swapping).
2222 if (!sc->priority && swappiness) {
2223 scan_balance = SCAN_EQUAL;
2228 * Prevent the reclaimer from falling into the cache trap: as
2229 * cache pages start out inactive, every cache fault will tip
2230 * the scan balance towards the file LRU. And as the file LRU
2231 * shrinks, so does the window for rotation from references.
2232 * This means we have a runaway feedback loop where a tiny
2233 * thrashing file LRU becomes infinitely more attractive than
2234 * anon pages. Try to detect this based on file LRU size.
2236 if (global_reclaim(sc)) {
2237 unsigned long pgdatfile;
2238 unsigned long pgdatfree;
2240 unsigned long total_high_wmark = 0;
2242 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2243 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2244 node_page_state(pgdat, NR_INACTIVE_FILE);
2246 for (z = 0; z < MAX_NR_ZONES; z++) {
2247 struct zone *zone = &pgdat->node_zones[z];
2248 if (!managed_zone(zone))
2251 total_high_wmark += high_wmark_pages(zone);
2254 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2256 * Force SCAN_ANON if there are enough inactive
2257 * anonymous pages on the LRU in eligible zones.
2258 * Otherwise, the small LRU gets thrashed.
2260 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2261 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2263 scan_balance = SCAN_ANON;
2270 * If there is enough inactive page cache, i.e. if the size of the
2271 * inactive list is greater than that of the active list *and* the
2272 * inactive list actually has some pages to scan on this priority, we
2273 * do not reclaim anything from the anonymous working set right now.
2274 * Without the second condition we could end up never scanning an
2275 * lruvec even if it has plenty of old anonymous pages unless the
2276 * system is under heavy pressure.
2278 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2279 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2280 scan_balance = SCAN_FILE;
2284 scan_balance = SCAN_FRACT;
2287 * With swappiness at 100, anonymous and file have the same priority.
2288 * This scanning priority is essentially the inverse of IO cost.
2290 anon_prio = swappiness;
2291 file_prio = 200 - anon_prio;
2294 * OK, so we have swap space and a fair amount of page cache
2295 * pages. We use the recently rotated / recently scanned
2296 * ratios to determine how valuable each cache is.
2298 * Because workloads change over time (and to avoid overflow)
2299 * we keep these statistics as a floating average, which ends
2300 * up weighing recent references more than old ones.
2302 * anon in [0], file in [1]
2305 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2306 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2307 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2308 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2310 spin_lock_irq(&pgdat->lru_lock);
2311 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2312 reclaim_stat->recent_scanned[0] /= 2;
2313 reclaim_stat->recent_rotated[0] /= 2;
2316 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2317 reclaim_stat->recent_scanned[1] /= 2;
2318 reclaim_stat->recent_rotated[1] /= 2;
2322 * The amount of pressure on anon vs file pages is inversely
2323 * proportional to the fraction of recently scanned pages on
2324 * each list that were recently referenced and in active use.
2326 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2327 ap /= reclaim_stat->recent_rotated[0] + 1;
2329 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2330 fp /= reclaim_stat->recent_rotated[1] + 1;
2331 spin_unlock_irq(&pgdat->lru_lock);
2335 denominator = ap + fp + 1;
2338 for_each_evictable_lru(lru) {
2339 int file = is_file_lru(lru);
2343 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2344 scan = size >> sc->priority;
2346 * If the cgroup's already been deleted, make sure to
2347 * scrape out the remaining cache.
2349 if (!scan && !mem_cgroup_online(memcg))
2350 scan = min(size, SWAP_CLUSTER_MAX);
2352 switch (scan_balance) {
2354 /* Scan lists relative to size */
2358 * Scan types proportional to swappiness and
2359 * their relative recent reclaim efficiency.
2361 scan = div64_u64(scan * fraction[file],
2366 /* Scan one type exclusively */
2367 if ((scan_balance == SCAN_FILE) != file) {
2373 /* Look ma, no brain */
2383 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2385 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2386 struct scan_control *sc, unsigned long *lru_pages)
2388 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2389 unsigned long nr[NR_LRU_LISTS];
2390 unsigned long targets[NR_LRU_LISTS];
2391 unsigned long nr_to_scan;
2393 unsigned long nr_reclaimed = 0;
2394 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2395 struct blk_plug plug;
2398 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2400 /* Record the original scan target for proportional adjustments later */
2401 memcpy(targets, nr, sizeof(nr));
2404 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2405 * event that can occur when there is little memory pressure e.g.
2406 * multiple streaming readers/writers. Hence, we do not abort scanning
2407 * when the requested number of pages are reclaimed when scanning at
2408 * DEF_PRIORITY on the assumption that the fact we are direct
2409 * reclaiming implies that kswapd is not keeping up and it is best to
2410 * do a batch of work at once. For memcg reclaim one check is made to
2411 * abort proportional reclaim if either the file or anon lru has already
2412 * dropped to zero at the first pass.
2414 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2415 sc->priority == DEF_PRIORITY);
2417 blk_start_plug(&plug);
2418 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2419 nr[LRU_INACTIVE_FILE]) {
2420 unsigned long nr_anon, nr_file, percentage;
2421 unsigned long nr_scanned;
2423 for_each_evictable_lru(lru) {
2425 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2426 nr[lru] -= nr_to_scan;
2428 nr_reclaimed += shrink_list(lru, nr_to_scan,
2435 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2439 * For kswapd and memcg, reclaim at least the number of pages
2440 * requested. Ensure that the anon and file LRUs are scanned
2441 * proportionally what was requested by get_scan_count(). We
2442 * stop reclaiming one LRU and reduce the amount scanning
2443 * proportional to the original scan target.
2445 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2446 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2449 * It's just vindictive to attack the larger once the smaller
2450 * has gone to zero. And given the way we stop scanning the
2451 * smaller below, this makes sure that we only make one nudge
2452 * towards proportionality once we've got nr_to_reclaim.
2454 if (!nr_file || !nr_anon)
2457 if (nr_file > nr_anon) {
2458 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2459 targets[LRU_ACTIVE_ANON] + 1;
2461 percentage = nr_anon * 100 / scan_target;
2463 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2464 targets[LRU_ACTIVE_FILE] + 1;
2466 percentage = nr_file * 100 / scan_target;
2469 /* Stop scanning the smaller of the LRU */
2471 nr[lru + LRU_ACTIVE] = 0;
2474 * Recalculate the other LRU scan count based on its original
2475 * scan target and the percentage scanning already complete
2477 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2478 nr_scanned = targets[lru] - nr[lru];
2479 nr[lru] = targets[lru] * (100 - percentage) / 100;
2480 nr[lru] -= min(nr[lru], nr_scanned);
2483 nr_scanned = targets[lru] - nr[lru];
2484 nr[lru] = targets[lru] * (100 - percentage) / 100;
2485 nr[lru] -= min(nr[lru], nr_scanned);
2487 scan_adjusted = true;
2489 blk_finish_plug(&plug);
2490 sc->nr_reclaimed += nr_reclaimed;
2493 * Even if we did not try to evict anon pages at all, we want to
2494 * rebalance the anon lru active/inactive ratio.
2496 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2497 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2498 sc, LRU_ACTIVE_ANON);
2501 /* Use reclaim/compaction for costly allocs or under memory pressure */
2502 static bool in_reclaim_compaction(struct scan_control *sc)
2504 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2505 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2506 sc->priority < DEF_PRIORITY - 2))
2513 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2514 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2515 * true if more pages should be reclaimed such that when the page allocator
2516 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2517 * It will give up earlier than that if there is difficulty reclaiming pages.
2519 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2520 unsigned long nr_reclaimed,
2521 unsigned long nr_scanned,
2522 struct scan_control *sc)
2524 unsigned long pages_for_compaction;
2525 unsigned long inactive_lru_pages;
2528 /* If not in reclaim/compaction mode, stop */
2529 if (!in_reclaim_compaction(sc))
2532 /* Consider stopping depending on scan and reclaim activity */
2533 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2535 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2536 * full LRU list has been scanned and we are still failing
2537 * to reclaim pages. This full LRU scan is potentially
2538 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2540 if (!nr_reclaimed && !nr_scanned)
2544 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2545 * fail without consequence, stop if we failed to reclaim
2546 * any pages from the last SWAP_CLUSTER_MAX number of
2547 * pages that were scanned. This will return to the
2548 * caller faster at the risk reclaim/compaction and
2549 * the resulting allocation attempt fails
2556 * If we have not reclaimed enough pages for compaction and the
2557 * inactive lists are large enough, continue reclaiming
2559 pages_for_compaction = compact_gap(sc->order);
2560 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2561 if (get_nr_swap_pages() > 0)
2562 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2563 if (sc->nr_reclaimed < pages_for_compaction &&
2564 inactive_lru_pages > pages_for_compaction)
2567 /* If compaction would go ahead or the allocation would succeed, stop */
2568 for (z = 0; z <= sc->reclaim_idx; z++) {
2569 struct zone *zone = &pgdat->node_zones[z];
2570 if (!managed_zone(zone))
2573 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2574 case COMPACT_SUCCESS:
2575 case COMPACT_CONTINUE:
2578 /* check next zone */
2585 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2587 struct reclaim_state *reclaim_state = current->reclaim_state;
2588 unsigned long nr_reclaimed, nr_scanned;
2589 bool reclaimable = false;
2592 struct mem_cgroup *root = sc->target_mem_cgroup;
2593 struct mem_cgroup_reclaim_cookie reclaim = {
2595 .priority = sc->priority,
2597 unsigned long node_lru_pages = 0;
2598 struct mem_cgroup *memcg;
2600 nr_reclaimed = sc->nr_reclaimed;
2601 nr_scanned = sc->nr_scanned;
2603 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2605 unsigned long lru_pages;
2606 unsigned long reclaimed;
2607 unsigned long scanned;
2609 if (mem_cgroup_low(root, memcg)) {
2610 if (!sc->memcg_low_reclaim) {
2611 sc->memcg_low_skipped = 1;
2614 mem_cgroup_event(memcg, MEMCG_LOW);
2617 reclaimed = sc->nr_reclaimed;
2618 scanned = sc->nr_scanned;
2620 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2621 node_lru_pages += lru_pages;
2624 shrink_slab(sc->gfp_mask, pgdat->node_id,
2625 memcg, sc->nr_scanned - scanned,
2628 /* Record the group's reclaim efficiency */
2629 vmpressure(sc->gfp_mask, memcg, false,
2630 sc->nr_scanned - scanned,
2631 sc->nr_reclaimed - reclaimed);
2634 * Direct reclaim and kswapd have to scan all memory
2635 * cgroups to fulfill the overall scan target for the
2638 * Limit reclaim, on the other hand, only cares about
2639 * nr_to_reclaim pages to be reclaimed and it will
2640 * retry with decreasing priority if one round over the
2641 * whole hierarchy is not sufficient.
2643 if (!global_reclaim(sc) &&
2644 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2645 mem_cgroup_iter_break(root, memcg);
2648 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2651 * Shrink the slab caches in the same proportion that
2652 * the eligible LRU pages were scanned.
2654 if (global_reclaim(sc))
2655 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2656 sc->nr_scanned - nr_scanned,
2659 if (reclaim_state) {
2660 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2661 reclaim_state->reclaimed_slab = 0;
2664 /* Record the subtree's reclaim efficiency */
2665 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2666 sc->nr_scanned - nr_scanned,
2667 sc->nr_reclaimed - nr_reclaimed);
2669 if (sc->nr_reclaimed - nr_reclaimed)
2672 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2673 sc->nr_scanned - nr_scanned, sc));
2676 * Kswapd gives up on balancing particular nodes after too
2677 * many failures to reclaim anything from them and goes to
2678 * sleep. On reclaim progress, reset the failure counter. A
2679 * successful direct reclaim run will revive a dormant kswapd.
2682 pgdat->kswapd_failures = 0;
2688 * Returns true if compaction should go ahead for a costly-order request, or
2689 * the allocation would already succeed without compaction. Return false if we
2690 * should reclaim first.
2692 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2694 unsigned long watermark;
2695 enum compact_result suitable;
2697 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2698 if (suitable == COMPACT_SUCCESS)
2699 /* Allocation should succeed already. Don't reclaim. */
2701 if (suitable == COMPACT_SKIPPED)
2702 /* Compaction cannot yet proceed. Do reclaim. */
2706 * Compaction is already possible, but it takes time to run and there
2707 * are potentially other callers using the pages just freed. So proceed
2708 * with reclaim to make a buffer of free pages available to give
2709 * compaction a reasonable chance of completing and allocating the page.
2710 * Note that we won't actually reclaim the whole buffer in one attempt
2711 * as the target watermark in should_continue_reclaim() is lower. But if
2712 * we are already above the high+gap watermark, don't reclaim at all.
2714 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2716 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2720 * This is the direct reclaim path, for page-allocating processes. We only
2721 * try to reclaim pages from zones which will satisfy the caller's allocation
2724 * If a zone is deemed to be full of pinned pages then just give it a light
2725 * scan then give up on it.
2727 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2731 unsigned long nr_soft_reclaimed;
2732 unsigned long nr_soft_scanned;
2734 pg_data_t *last_pgdat = NULL;
2737 * If the number of buffer_heads in the machine exceeds the maximum
2738 * allowed level, force direct reclaim to scan the highmem zone as
2739 * highmem pages could be pinning lowmem pages storing buffer_heads
2741 orig_mask = sc->gfp_mask;
2742 if (buffer_heads_over_limit) {
2743 sc->gfp_mask |= __GFP_HIGHMEM;
2744 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2747 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2748 sc->reclaim_idx, sc->nodemask) {
2750 * Take care memory controller reclaiming has small influence
2753 if (global_reclaim(sc)) {
2754 if (!cpuset_zone_allowed(zone,
2755 GFP_KERNEL | __GFP_HARDWALL))
2759 * If we already have plenty of memory free for
2760 * compaction in this zone, don't free any more.
2761 * Even though compaction is invoked for any
2762 * non-zero order, only frequent costly order
2763 * reclamation is disruptive enough to become a
2764 * noticeable problem, like transparent huge
2767 if (IS_ENABLED(CONFIG_COMPACTION) &&
2768 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2769 compaction_ready(zone, sc)) {
2770 sc->compaction_ready = true;
2775 * Shrink each node in the zonelist once. If the
2776 * zonelist is ordered by zone (not the default) then a
2777 * node may be shrunk multiple times but in that case
2778 * the user prefers lower zones being preserved.
2780 if (zone->zone_pgdat == last_pgdat)
2784 * This steals pages from memory cgroups over softlimit
2785 * and returns the number of reclaimed pages and
2786 * scanned pages. This works for global memory pressure
2787 * and balancing, not for a memcg's limit.
2789 nr_soft_scanned = 0;
2790 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2791 sc->order, sc->gfp_mask,
2793 sc->nr_reclaimed += nr_soft_reclaimed;
2794 sc->nr_scanned += nr_soft_scanned;
2795 /* need some check for avoid more shrink_zone() */
2798 /* See comment about same check for global reclaim above */
2799 if (zone->zone_pgdat == last_pgdat)
2801 last_pgdat = zone->zone_pgdat;
2802 shrink_node(zone->zone_pgdat, sc);
2806 * Restore to original mask to avoid the impact on the caller if we
2807 * promoted it to __GFP_HIGHMEM.
2809 sc->gfp_mask = orig_mask;
2812 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2814 struct mem_cgroup *memcg;
2816 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2818 unsigned long refaults;
2819 struct lruvec *lruvec;
2822 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2824 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2826 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2827 lruvec->refaults = refaults;
2828 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2832 * This is the main entry point to direct page reclaim.
2834 * If a full scan of the inactive list fails to free enough memory then we
2835 * are "out of memory" and something needs to be killed.
2837 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2838 * high - the zone may be full of dirty or under-writeback pages, which this
2839 * caller can't do much about. We kick the writeback threads and take explicit
2840 * naps in the hope that some of these pages can be written. But if the
2841 * allocating task holds filesystem locks which prevent writeout this might not
2842 * work, and the allocation attempt will fail.
2844 * returns: 0, if no pages reclaimed
2845 * else, the number of pages reclaimed
2847 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2848 struct scan_control *sc)
2850 int initial_priority = sc->priority;
2851 pg_data_t *last_pgdat;
2855 delayacct_freepages_start();
2857 if (global_reclaim(sc))
2858 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2861 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2864 shrink_zones(zonelist, sc);
2866 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2869 if (sc->compaction_ready)
2873 * If we're getting trouble reclaiming, start doing
2874 * writepage even in laptop mode.
2876 if (sc->priority < DEF_PRIORITY - 2)
2877 sc->may_writepage = 1;
2878 } while (--sc->priority >= 0);
2881 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2883 if (zone->zone_pgdat == last_pgdat)
2885 last_pgdat = zone->zone_pgdat;
2886 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2889 delayacct_freepages_end();
2891 if (sc->nr_reclaimed)
2892 return sc->nr_reclaimed;
2894 /* Aborted reclaim to try compaction? don't OOM, then */
2895 if (sc->compaction_ready)
2898 /* Untapped cgroup reserves? Don't OOM, retry. */
2899 if (sc->memcg_low_skipped) {
2900 sc->priority = initial_priority;
2901 sc->memcg_low_reclaim = 1;
2902 sc->memcg_low_skipped = 0;
2909 static bool allow_direct_reclaim(pg_data_t *pgdat)
2912 unsigned long pfmemalloc_reserve = 0;
2913 unsigned long free_pages = 0;
2917 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2920 for (i = 0; i <= ZONE_NORMAL; i++) {
2921 zone = &pgdat->node_zones[i];
2922 if (!managed_zone(zone))
2925 if (!zone_reclaimable_pages(zone))
2928 pfmemalloc_reserve += min_wmark_pages(zone);
2929 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2932 /* If there are no reserves (unexpected config) then do not throttle */
2933 if (!pfmemalloc_reserve)
2936 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2938 /* kswapd must be awake if processes are being throttled */
2939 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2940 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2941 (enum zone_type)ZONE_NORMAL);
2942 wake_up_interruptible(&pgdat->kswapd_wait);
2949 * Throttle direct reclaimers if backing storage is backed by the network
2950 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2951 * depleted. kswapd will continue to make progress and wake the processes
2952 * when the low watermark is reached.
2954 * Returns true if a fatal signal was delivered during throttling. If this
2955 * happens, the page allocator should not consider triggering the OOM killer.
2957 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2958 nodemask_t *nodemask)
2962 pg_data_t *pgdat = NULL;
2965 * Kernel threads should not be throttled as they may be indirectly
2966 * responsible for cleaning pages necessary for reclaim to make forward
2967 * progress. kjournald for example may enter direct reclaim while
2968 * committing a transaction where throttling it could forcing other
2969 * processes to block on log_wait_commit().
2971 if (current->flags & PF_KTHREAD)
2975 * If a fatal signal is pending, this process should not throttle.
2976 * It should return quickly so it can exit and free its memory
2978 if (fatal_signal_pending(current))
2982 * Check if the pfmemalloc reserves are ok by finding the first node
2983 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2984 * GFP_KERNEL will be required for allocating network buffers when
2985 * swapping over the network so ZONE_HIGHMEM is unusable.
2987 * Throttling is based on the first usable node and throttled processes
2988 * wait on a queue until kswapd makes progress and wakes them. There
2989 * is an affinity then between processes waking up and where reclaim
2990 * progress has been made assuming the process wakes on the same node.
2991 * More importantly, processes running on remote nodes will not compete
2992 * for remote pfmemalloc reserves and processes on different nodes
2993 * should make reasonable progress.
2995 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2996 gfp_zone(gfp_mask), nodemask) {
2997 if (zone_idx(zone) > ZONE_NORMAL)
3000 /* Throttle based on the first usable node */
3001 pgdat = zone->zone_pgdat;
3002 if (allow_direct_reclaim(pgdat))
3007 /* If no zone was usable by the allocation flags then do not throttle */
3011 /* Account for the throttling */
3012 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3015 * If the caller cannot enter the filesystem, it's possible that it
3016 * is due to the caller holding an FS lock or performing a journal
3017 * transaction in the case of a filesystem like ext[3|4]. In this case,
3018 * it is not safe to block on pfmemalloc_wait as kswapd could be
3019 * blocked waiting on the same lock. Instead, throttle for up to a
3020 * second before continuing.
3022 if (!(gfp_mask & __GFP_FS)) {
3023 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3024 allow_direct_reclaim(pgdat), HZ);
3029 /* Throttle until kswapd wakes the process */
3030 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3031 allow_direct_reclaim(pgdat));
3034 if (fatal_signal_pending(current))
3041 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3042 gfp_t gfp_mask, nodemask_t *nodemask)
3044 unsigned long nr_reclaimed;
3045 struct scan_control sc = {
3046 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3047 .gfp_mask = current_gfp_context(gfp_mask),
3048 .reclaim_idx = gfp_zone(gfp_mask),
3050 .nodemask = nodemask,
3051 .priority = DEF_PRIORITY,
3052 .may_writepage = !laptop_mode,
3058 * Do not enter reclaim if fatal signal was delivered while throttled.
3059 * 1 is returned so that the page allocator does not OOM kill at this
3062 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3065 trace_mm_vmscan_direct_reclaim_begin(order,
3070 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3072 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3074 return nr_reclaimed;
3079 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3080 gfp_t gfp_mask, bool noswap,
3082 unsigned long *nr_scanned)
3084 struct scan_control sc = {
3085 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3086 .target_mem_cgroup = memcg,
3087 .may_writepage = !laptop_mode,
3089 .reclaim_idx = MAX_NR_ZONES - 1,
3090 .may_swap = !noswap,
3092 unsigned long lru_pages;
3094 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3095 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3097 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3103 * NOTE: Although we can get the priority field, using it
3104 * here is not a good idea, since it limits the pages we can scan.
3105 * if we don't reclaim here, the shrink_node from balance_pgdat
3106 * will pick up pages from other mem cgroup's as well. We hack
3107 * the priority and make it zero.
3109 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3111 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3113 *nr_scanned = sc.nr_scanned;
3114 return sc.nr_reclaimed;
3117 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3118 unsigned long nr_pages,
3122 struct zonelist *zonelist;
3123 unsigned long nr_reclaimed;
3125 unsigned int noreclaim_flag;
3126 struct scan_control sc = {
3127 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3128 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3129 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3130 .reclaim_idx = MAX_NR_ZONES - 1,
3131 .target_mem_cgroup = memcg,
3132 .priority = DEF_PRIORITY,
3133 .may_writepage = !laptop_mode,
3135 .may_swap = may_swap,
3139 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3140 * take care of from where we get pages. So the node where we start the
3141 * scan does not need to be the current node.
3143 nid = mem_cgroup_select_victim_node(memcg);
3145 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3147 trace_mm_vmscan_memcg_reclaim_begin(0,
3152 noreclaim_flag = memalloc_noreclaim_save();
3153 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3154 memalloc_noreclaim_restore(noreclaim_flag);
3156 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3158 return nr_reclaimed;
3162 static void age_active_anon(struct pglist_data *pgdat,
3163 struct scan_control *sc)
3165 struct mem_cgroup *memcg;
3167 if (!total_swap_pages)
3170 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3172 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3174 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3175 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3176 sc, LRU_ACTIVE_ANON);
3178 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3183 * Returns true if there is an eligible zone balanced for the request order
3186 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3189 unsigned long mark = -1;
3192 for (i = 0; i <= classzone_idx; i++) {
3193 zone = pgdat->node_zones + i;
3195 if (!managed_zone(zone))
3198 mark = high_wmark_pages(zone);
3199 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3204 * If a node has no populated zone within classzone_idx, it does not
3205 * need balancing by definition. This can happen if a zone-restricted
3206 * allocation tries to wake a remote kswapd.
3214 /* Clear pgdat state for congested, dirty or under writeback. */
3215 static void clear_pgdat_congested(pg_data_t *pgdat)
3217 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3218 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3219 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3223 * Prepare kswapd for sleeping. This verifies that there are no processes
3224 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3226 * Returns true if kswapd is ready to sleep
3228 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3231 * The throttled processes are normally woken up in balance_pgdat() as
3232 * soon as allow_direct_reclaim() is true. But there is a potential
3233 * race between when kswapd checks the watermarks and a process gets
3234 * throttled. There is also a potential race if processes get
3235 * throttled, kswapd wakes, a large process exits thereby balancing the
3236 * zones, which causes kswapd to exit balance_pgdat() before reaching
3237 * the wake up checks. If kswapd is going to sleep, no process should
3238 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3239 * the wake up is premature, processes will wake kswapd and get
3240 * throttled again. The difference from wake ups in balance_pgdat() is
3241 * that here we are under prepare_to_wait().
3243 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3244 wake_up_all(&pgdat->pfmemalloc_wait);
3246 /* Hopeless node, leave it to direct reclaim */
3247 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3250 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3251 clear_pgdat_congested(pgdat);
3259 * kswapd shrinks a node of pages that are at or below the highest usable
3260 * zone that is currently unbalanced.
3262 * Returns true if kswapd scanned at least the requested number of pages to
3263 * reclaim or if the lack of progress was due to pages under writeback.
3264 * This is used to determine if the scanning priority needs to be raised.
3266 static bool kswapd_shrink_node(pg_data_t *pgdat,
3267 struct scan_control *sc)
3272 /* Reclaim a number of pages proportional to the number of zones */
3273 sc->nr_to_reclaim = 0;
3274 for (z = 0; z <= sc->reclaim_idx; z++) {
3275 zone = pgdat->node_zones + z;
3276 if (!managed_zone(zone))
3279 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3283 * Historically care was taken to put equal pressure on all zones but
3284 * now pressure is applied based on node LRU order.
3286 shrink_node(pgdat, sc);
3289 * Fragmentation may mean that the system cannot be rebalanced for
3290 * high-order allocations. If twice the allocation size has been
3291 * reclaimed then recheck watermarks only at order-0 to prevent
3292 * excessive reclaim. Assume that a process requested a high-order
3293 * can direct reclaim/compact.
3295 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3298 return sc->nr_scanned >= sc->nr_to_reclaim;
3302 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3303 * that are eligible for use by the caller until at least one zone is
3306 * Returns the order kswapd finished reclaiming at.
3308 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3309 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3310 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3311 * or lower is eligible for reclaim until at least one usable zone is
3314 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3317 unsigned long nr_soft_reclaimed;
3318 unsigned long nr_soft_scanned;
3320 struct scan_control sc = {
3321 .gfp_mask = GFP_KERNEL,
3323 .priority = DEF_PRIORITY,
3324 .may_writepage = !laptop_mode,
3328 count_vm_event(PAGEOUTRUN);
3331 unsigned long nr_reclaimed = sc.nr_reclaimed;
3332 bool raise_priority = true;
3334 sc.reclaim_idx = classzone_idx;
3337 * If the number of buffer_heads exceeds the maximum allowed
3338 * then consider reclaiming from all zones. This has a dual
3339 * purpose -- on 64-bit systems it is expected that
3340 * buffer_heads are stripped during active rotation. On 32-bit
3341 * systems, highmem pages can pin lowmem memory and shrinking
3342 * buffers can relieve lowmem pressure. Reclaim may still not
3343 * go ahead if all eligible zones for the original allocation
3344 * request are balanced to avoid excessive reclaim from kswapd.
3346 if (buffer_heads_over_limit) {
3347 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3348 zone = pgdat->node_zones + i;
3349 if (!managed_zone(zone))
3358 * Only reclaim if there are no eligible zones. Note that
3359 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3362 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3366 * Do some background aging of the anon list, to give
3367 * pages a chance to be referenced before reclaiming. All
3368 * pages are rotated regardless of classzone as this is
3369 * about consistent aging.
3371 age_active_anon(pgdat, &sc);
3374 * If we're getting trouble reclaiming, start doing writepage
3375 * even in laptop mode.
3377 if (sc.priority < DEF_PRIORITY - 2)
3378 sc.may_writepage = 1;
3380 /* Call soft limit reclaim before calling shrink_node. */
3382 nr_soft_scanned = 0;
3383 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3384 sc.gfp_mask, &nr_soft_scanned);
3385 sc.nr_reclaimed += nr_soft_reclaimed;
3388 * There should be no need to raise the scanning priority if
3389 * enough pages are already being scanned that that high
3390 * watermark would be met at 100% efficiency.
3392 if (kswapd_shrink_node(pgdat, &sc))
3393 raise_priority = false;
3396 * If the low watermark is met there is no need for processes
3397 * to be throttled on pfmemalloc_wait as they should not be
3398 * able to safely make forward progress. Wake them
3400 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3401 allow_direct_reclaim(pgdat))
3402 wake_up_all(&pgdat->pfmemalloc_wait);
3404 /* Check if kswapd should be suspending */
3405 if (try_to_freeze() || kthread_should_stop())
3409 * Raise priority if scanning rate is too low or there was no
3410 * progress in reclaiming pages
3412 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3413 if (raise_priority || !nr_reclaimed)
3415 } while (sc.priority >= 1);
3417 if (!sc.nr_reclaimed)
3418 pgdat->kswapd_failures++;
3421 snapshot_refaults(NULL, pgdat);
3423 * Return the order kswapd stopped reclaiming at as
3424 * prepare_kswapd_sleep() takes it into account. If another caller
3425 * entered the allocator slow path while kswapd was awake, order will
3426 * remain at the higher level.
3432 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3433 * allocation request woke kswapd for. When kswapd has not woken recently,
3434 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3435 * given classzone and returns it or the highest classzone index kswapd
3436 * was recently woke for.
3438 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3439 enum zone_type classzone_idx)
3441 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3442 return classzone_idx;
3444 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3447 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3448 unsigned int classzone_idx)
3453 if (freezing(current) || kthread_should_stop())
3456 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3459 * Try to sleep for a short interval. Note that kcompactd will only be
3460 * woken if it is possible to sleep for a short interval. This is
3461 * deliberate on the assumption that if reclaim cannot keep an
3462 * eligible zone balanced that it's also unlikely that compaction will
3465 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3467 * Compaction records what page blocks it recently failed to
3468 * isolate pages from and skips them in the future scanning.
3469 * When kswapd is going to sleep, it is reasonable to assume
3470 * that pages and compaction may succeed so reset the cache.
3472 reset_isolation_suitable(pgdat);
3475 * We have freed the memory, now we should compact it to make
3476 * allocation of the requested order possible.
3478 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3480 remaining = schedule_timeout(HZ/10);
3483 * If woken prematurely then reset kswapd_classzone_idx and
3484 * order. The values will either be from a wakeup request or
3485 * the previous request that slept prematurely.
3488 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3489 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3492 finish_wait(&pgdat->kswapd_wait, &wait);
3493 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3497 * After a short sleep, check if it was a premature sleep. If not, then
3498 * go fully to sleep until explicitly woken up.
3501 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3502 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3505 * vmstat counters are not perfectly accurate and the estimated
3506 * value for counters such as NR_FREE_PAGES can deviate from the
3507 * true value by nr_online_cpus * threshold. To avoid the zone
3508 * watermarks being breached while under pressure, we reduce the
3509 * per-cpu vmstat threshold while kswapd is awake and restore
3510 * them before going back to sleep.
3512 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3514 if (!kthread_should_stop())
3517 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3520 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3522 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3524 finish_wait(&pgdat->kswapd_wait, &wait);
3528 * The background pageout daemon, started as a kernel thread
3529 * from the init process.
3531 * This basically trickles out pages so that we have _some_
3532 * free memory available even if there is no other activity
3533 * that frees anything up. This is needed for things like routing
3534 * etc, where we otherwise might have all activity going on in
3535 * asynchronous contexts that cannot page things out.
3537 * If there are applications that are active memory-allocators
3538 * (most normal use), this basically shouldn't matter.
3540 static int kswapd(void *p)
3542 unsigned int alloc_order, reclaim_order;
3543 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3544 pg_data_t *pgdat = (pg_data_t*)p;
3545 struct task_struct *tsk = current;
3547 struct reclaim_state reclaim_state = {
3548 .reclaimed_slab = 0,
3550 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3552 if (!cpumask_empty(cpumask))
3553 set_cpus_allowed_ptr(tsk, cpumask);
3554 current->reclaim_state = &reclaim_state;
3557 * Tell the memory management that we're a "memory allocator",
3558 * and that if we need more memory we should get access to it
3559 * regardless (see "__alloc_pages()"). "kswapd" should
3560 * never get caught in the normal page freeing logic.
3562 * (Kswapd normally doesn't need memory anyway, but sometimes
3563 * you need a small amount of memory in order to be able to
3564 * page out something else, and this flag essentially protects
3565 * us from recursively trying to free more memory as we're
3566 * trying to free the first piece of memory in the first place).
3568 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3571 pgdat->kswapd_order = 0;
3572 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3576 alloc_order = reclaim_order = pgdat->kswapd_order;
3577 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3580 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3583 /* Read the new order and classzone_idx */
3584 alloc_order = reclaim_order = pgdat->kswapd_order;
3585 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3586 pgdat->kswapd_order = 0;
3587 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3589 ret = try_to_freeze();
3590 if (kthread_should_stop())
3594 * We can speed up thawing tasks if we don't call balance_pgdat
3595 * after returning from the refrigerator
3601 * Reclaim begins at the requested order but if a high-order
3602 * reclaim fails then kswapd falls back to reclaiming for
3603 * order-0. If that happens, kswapd will consider sleeping
3604 * for the order it finished reclaiming at (reclaim_order)
3605 * but kcompactd is woken to compact for the original
3606 * request (alloc_order).
3608 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3610 fs_reclaim_acquire(GFP_KERNEL);
3611 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3612 fs_reclaim_release(GFP_KERNEL);
3613 if (reclaim_order < alloc_order)
3614 goto kswapd_try_sleep;
3617 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3618 current->reclaim_state = NULL;
3624 * A zone is low on free memory, so wake its kswapd task to service it.
3626 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3630 if (!managed_zone(zone))
3633 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3635 pgdat = zone->zone_pgdat;
3636 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3638 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3639 if (!waitqueue_active(&pgdat->kswapd_wait))
3642 /* Hopeless node, leave it to direct reclaim */
3643 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3646 if (pgdat_balanced(pgdat, order, classzone_idx))
3649 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
3650 wake_up_interruptible(&pgdat->kswapd_wait);
3653 #ifdef CONFIG_HIBERNATION
3655 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3658 * Rather than trying to age LRUs the aim is to preserve the overall
3659 * LRU order by reclaiming preferentially
3660 * inactive > active > active referenced > active mapped
3662 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3664 struct reclaim_state reclaim_state;
3665 struct scan_control sc = {
3666 .nr_to_reclaim = nr_to_reclaim,
3667 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3668 .reclaim_idx = MAX_NR_ZONES - 1,
3669 .priority = DEF_PRIORITY,
3673 .hibernation_mode = 1,
3675 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3676 struct task_struct *p = current;
3677 unsigned long nr_reclaimed;
3678 unsigned int noreclaim_flag;
3680 noreclaim_flag = memalloc_noreclaim_save();
3681 fs_reclaim_acquire(sc.gfp_mask);
3682 reclaim_state.reclaimed_slab = 0;
3683 p->reclaim_state = &reclaim_state;
3685 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3687 p->reclaim_state = NULL;
3688 fs_reclaim_release(sc.gfp_mask);
3689 memalloc_noreclaim_restore(noreclaim_flag);
3691 return nr_reclaimed;
3693 #endif /* CONFIG_HIBERNATION */
3695 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3696 not required for correctness. So if the last cpu in a node goes
3697 away, we get changed to run anywhere: as the first one comes back,
3698 restore their cpu bindings. */
3699 static int kswapd_cpu_online(unsigned int cpu)
3703 for_each_node_state(nid, N_MEMORY) {
3704 pg_data_t *pgdat = NODE_DATA(nid);
3705 const struct cpumask *mask;
3707 mask = cpumask_of_node(pgdat->node_id);
3709 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3710 /* One of our CPUs online: restore mask */
3711 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3717 * This kswapd start function will be called by init and node-hot-add.
3718 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3720 int kswapd_run(int nid)
3722 pg_data_t *pgdat = NODE_DATA(nid);
3728 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3729 if (IS_ERR(pgdat->kswapd)) {
3730 /* failure at boot is fatal */
3731 BUG_ON(system_state < SYSTEM_RUNNING);
3732 pr_err("Failed to start kswapd on node %d\n", nid);
3733 ret = PTR_ERR(pgdat->kswapd);
3734 pgdat->kswapd = NULL;
3740 * Called by memory hotplug when all memory in a node is offlined. Caller must
3741 * hold mem_hotplug_begin/end().
3743 void kswapd_stop(int nid)
3745 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3748 kthread_stop(kswapd);
3749 NODE_DATA(nid)->kswapd = NULL;
3753 static int __init kswapd_init(void)
3758 for_each_node_state(nid, N_MEMORY)
3760 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3761 "mm/vmscan:online", kswapd_cpu_online,
3767 module_init(kswapd_init)
3773 * If non-zero call node_reclaim when the number of free pages falls below
3776 int node_reclaim_mode __read_mostly;
3778 #define RECLAIM_OFF 0
3779 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3780 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3781 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3784 * Priority for NODE_RECLAIM. This determines the fraction of pages
3785 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3788 #define NODE_RECLAIM_PRIORITY 4
3791 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3794 int sysctl_min_unmapped_ratio = 1;
3797 * If the number of slab pages in a zone grows beyond this percentage then
3798 * slab reclaim needs to occur.
3800 int sysctl_min_slab_ratio = 5;
3802 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3804 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3805 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3806 node_page_state(pgdat, NR_ACTIVE_FILE);
3809 * It's possible for there to be more file mapped pages than
3810 * accounted for by the pages on the file LRU lists because
3811 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3813 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3816 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3817 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3819 unsigned long nr_pagecache_reclaimable;
3820 unsigned long delta = 0;
3823 * If RECLAIM_UNMAP is set, then all file pages are considered
3824 * potentially reclaimable. Otherwise, we have to worry about
3825 * pages like swapcache and node_unmapped_file_pages() provides
3828 if (node_reclaim_mode & RECLAIM_UNMAP)
3829 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3831 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3833 /* If we can't clean pages, remove dirty pages from consideration */
3834 if (!(node_reclaim_mode & RECLAIM_WRITE))
3835 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3837 /* Watch for any possible underflows due to delta */
3838 if (unlikely(delta > nr_pagecache_reclaimable))
3839 delta = nr_pagecache_reclaimable;
3841 return nr_pagecache_reclaimable - delta;
3845 * Try to free up some pages from this node through reclaim.
3847 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3849 /* Minimum pages needed in order to stay on node */
3850 const unsigned long nr_pages = 1 << order;
3851 struct task_struct *p = current;
3852 struct reclaim_state reclaim_state;
3853 unsigned int noreclaim_flag;
3854 struct scan_control sc = {
3855 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3856 .gfp_mask = current_gfp_context(gfp_mask),
3858 .priority = NODE_RECLAIM_PRIORITY,
3859 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3860 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3862 .reclaim_idx = gfp_zone(gfp_mask),
3867 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3868 * and we also need to be able to write out pages for RECLAIM_WRITE
3869 * and RECLAIM_UNMAP.
3871 noreclaim_flag = memalloc_noreclaim_save();
3872 p->flags |= PF_SWAPWRITE;
3873 fs_reclaim_acquire(sc.gfp_mask);
3874 reclaim_state.reclaimed_slab = 0;
3875 p->reclaim_state = &reclaim_state;
3877 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3879 * Free memory by calling shrink zone with increasing
3880 * priorities until we have enough memory freed.
3883 shrink_node(pgdat, &sc);
3884 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3887 p->reclaim_state = NULL;
3888 fs_reclaim_release(gfp_mask);
3889 current->flags &= ~PF_SWAPWRITE;
3890 memalloc_noreclaim_restore(noreclaim_flag);
3891 return sc.nr_reclaimed >= nr_pages;
3894 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3899 * Node reclaim reclaims unmapped file backed pages and
3900 * slab pages if we are over the defined limits.
3902 * A small portion of unmapped file backed pages is needed for
3903 * file I/O otherwise pages read by file I/O will be immediately
3904 * thrown out if the node is overallocated. So we do not reclaim
3905 * if less than a specified percentage of the node is used by
3906 * unmapped file backed pages.
3908 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3909 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3910 return NODE_RECLAIM_FULL;
3913 * Do not scan if the allocation should not be delayed.
3915 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3916 return NODE_RECLAIM_NOSCAN;
3919 * Only run node reclaim on the local node or on nodes that do not
3920 * have associated processors. This will favor the local processor
3921 * over remote processors and spread off node memory allocations
3922 * as wide as possible.
3924 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3925 return NODE_RECLAIM_NOSCAN;
3927 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3928 return NODE_RECLAIM_NOSCAN;
3930 ret = __node_reclaim(pgdat, gfp_mask, order);
3931 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3934 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3941 * page_evictable - test whether a page is evictable
3942 * @page: the page to test
3944 * Test whether page is evictable--i.e., should be placed on active/inactive
3945 * lists vs unevictable list.
3947 * Reasons page might not be evictable:
3948 * (1) page's mapping marked unevictable
3949 * (2) page is part of an mlocked VMA
3952 int page_evictable(struct page *page)
3954 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3959 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3960 * @pages: array of pages to check
3961 * @nr_pages: number of pages to check
3963 * Checks pages for evictability and moves them to the appropriate lru list.
3965 * This function is only used for SysV IPC SHM_UNLOCK.
3967 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3969 struct lruvec *lruvec;
3970 struct pglist_data *pgdat = NULL;
3975 for (i = 0; i < nr_pages; i++) {
3976 struct page *page = pages[i];
3977 struct pglist_data *pagepgdat = page_pgdat(page);
3980 if (pagepgdat != pgdat) {
3982 spin_unlock_irq(&pgdat->lru_lock);
3984 spin_lock_irq(&pgdat->lru_lock);
3986 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3988 if (!PageLRU(page) || !PageUnevictable(page))
3991 if (page_evictable(page)) {
3992 enum lru_list lru = page_lru_base_type(page);
3994 VM_BUG_ON_PAGE(PageActive(page), page);
3995 ClearPageUnevictable(page);
3996 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3997 add_page_to_lru_list(page, lruvec, lru);
4003 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4004 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4005 spin_unlock_irq(&pgdat->lru_lock);
4008 #endif /* CONFIG_SHMEM */