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);
506 * Bail out if someone want to register a new shrinker to
507 * prevent the regsitration from being stalled for long periods
508 * by parallel ongoing shrinking.
510 if (rwsem_is_contended(&shrinker_rwsem)) {
516 up_read(&shrinker_rwsem);
522 void drop_slab_node(int nid)
527 struct mem_cgroup *memcg = NULL;
531 freed += shrink_slab(GFP_KERNEL, nid, memcg,
533 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
534 } while (freed > 10);
541 for_each_online_node(nid)
545 static inline int is_page_cache_freeable(struct page *page)
548 * A freeable page cache page is referenced only by the caller
549 * that isolated the page, the page cache radix tree and
550 * optional buffer heads at page->private.
552 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
554 return page_count(page) - page_has_private(page) == 1 + radix_pins;
557 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
559 if (current->flags & PF_SWAPWRITE)
561 if (!inode_write_congested(inode))
563 if (inode_to_bdi(inode) == current->backing_dev_info)
569 * We detected a synchronous write error writing a page out. Probably
570 * -ENOSPC. We need to propagate that into the address_space for a subsequent
571 * fsync(), msync() or close().
573 * The tricky part is that after writepage we cannot touch the mapping: nothing
574 * prevents it from being freed up. But we have a ref on the page and once
575 * that page is locked, the mapping is pinned.
577 * We're allowed to run sleeping lock_page() here because we know the caller has
580 static void handle_write_error(struct address_space *mapping,
581 struct page *page, int error)
584 if (page_mapping(page) == mapping)
585 mapping_set_error(mapping, error);
589 /* possible outcome of pageout() */
591 /* failed to write page out, page is locked */
593 /* move page to the active list, page is locked */
595 /* page has been sent to the disk successfully, page is unlocked */
597 /* page is clean and locked */
602 * pageout is called by shrink_page_list() for each dirty page.
603 * Calls ->writepage().
605 static pageout_t pageout(struct page *page, struct address_space *mapping,
606 struct scan_control *sc)
609 * If the page is dirty, only perform writeback if that write
610 * will be non-blocking. To prevent this allocation from being
611 * stalled by pagecache activity. But note that there may be
612 * stalls if we need to run get_block(). We could test
613 * PagePrivate for that.
615 * If this process is currently in __generic_file_write_iter() against
616 * this page's queue, we can perform writeback even if that
619 * If the page is swapcache, write it back even if that would
620 * block, for some throttling. This happens by accident, because
621 * swap_backing_dev_info is bust: it doesn't reflect the
622 * congestion state of the swapdevs. Easy to fix, if needed.
624 if (!is_page_cache_freeable(page))
628 * Some data journaling orphaned pages can have
629 * page->mapping == NULL while being dirty with clean buffers.
631 if (page_has_private(page)) {
632 if (try_to_free_buffers(page)) {
633 ClearPageDirty(page);
634 pr_info("%s: orphaned page\n", __func__);
640 if (mapping->a_ops->writepage == NULL)
641 return PAGE_ACTIVATE;
642 if (!may_write_to_inode(mapping->host, sc))
645 if (clear_page_dirty_for_io(page)) {
647 struct writeback_control wbc = {
648 .sync_mode = WB_SYNC_NONE,
649 .nr_to_write = SWAP_CLUSTER_MAX,
651 .range_end = LLONG_MAX,
655 SetPageReclaim(page);
656 res = mapping->a_ops->writepage(page, &wbc);
658 handle_write_error(mapping, page, res);
659 if (res == AOP_WRITEPAGE_ACTIVATE) {
660 ClearPageReclaim(page);
661 return PAGE_ACTIVATE;
664 if (!PageWriteback(page)) {
665 /* synchronous write or broken a_ops? */
666 ClearPageReclaim(page);
668 trace_mm_vmscan_writepage(page);
669 inc_node_page_state(page, NR_VMSCAN_WRITE);
677 * Same as remove_mapping, but if the page is removed from the mapping, it
678 * gets returned with a refcount of 0.
680 static int __remove_mapping(struct address_space *mapping, struct page *page,
686 BUG_ON(!PageLocked(page));
687 BUG_ON(mapping != page_mapping(page));
689 spin_lock_irqsave(&mapping->tree_lock, flags);
691 * The non racy check for a busy page.
693 * Must be careful with the order of the tests. When someone has
694 * a ref to the page, it may be possible that they dirty it then
695 * drop the reference. So if PageDirty is tested before page_count
696 * here, then the following race may occur:
698 * get_user_pages(&page);
699 * [user mapping goes away]
701 * !PageDirty(page) [good]
702 * SetPageDirty(page);
704 * !page_count(page) [good, discard it]
706 * [oops, our write_to data is lost]
708 * Reversing the order of the tests ensures such a situation cannot
709 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
710 * load is not satisfied before that of page->_refcount.
712 * Note that if SetPageDirty is always performed via set_page_dirty,
713 * and thus under tree_lock, then this ordering is not required.
715 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
716 refcount = 1 + HPAGE_PMD_NR;
719 if (!page_ref_freeze(page, refcount))
721 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
722 if (unlikely(PageDirty(page))) {
723 page_ref_unfreeze(page, refcount);
727 if (PageSwapCache(page)) {
728 swp_entry_t swap = { .val = page_private(page) };
729 mem_cgroup_swapout(page, swap);
730 __delete_from_swap_cache(page);
731 spin_unlock_irqrestore(&mapping->tree_lock, flags);
732 put_swap_page(page, swap);
734 void (*freepage)(struct page *);
737 freepage = mapping->a_ops->freepage;
739 * Remember a shadow entry for reclaimed file cache in
740 * order to detect refaults, thus thrashing, later on.
742 * But don't store shadows in an address space that is
743 * already exiting. This is not just an optizimation,
744 * inode reclaim needs to empty out the radix tree or
745 * the nodes are lost. Don't plant shadows behind its
748 * We also don't store shadows for DAX mappings because the
749 * only page cache pages found in these are zero pages
750 * covering holes, and because we don't want to mix DAX
751 * exceptional entries and shadow exceptional entries in the
754 if (reclaimed && page_is_file_cache(page) &&
755 !mapping_exiting(mapping) && !dax_mapping(mapping))
756 shadow = workingset_eviction(mapping, page);
757 __delete_from_page_cache(page, shadow);
758 spin_unlock_irqrestore(&mapping->tree_lock, flags);
760 if (freepage != NULL)
767 spin_unlock_irqrestore(&mapping->tree_lock, flags);
772 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
773 * someone else has a ref on the page, abort and return 0. If it was
774 * successfully detached, return 1. Assumes the caller has a single ref on
777 int remove_mapping(struct address_space *mapping, struct page *page)
779 if (__remove_mapping(mapping, page, false)) {
781 * Unfreezing the refcount with 1 rather than 2 effectively
782 * drops the pagecache ref for us without requiring another
785 page_ref_unfreeze(page, 1);
792 * putback_lru_page - put previously isolated page onto appropriate LRU list
793 * @page: page to be put back to appropriate lru list
795 * Add previously isolated @page to appropriate LRU list.
796 * Page may still be unevictable for other reasons.
798 * lru_lock must not be held, interrupts must be enabled.
800 void putback_lru_page(struct page *page)
803 int was_unevictable = PageUnevictable(page);
805 VM_BUG_ON_PAGE(PageLRU(page), page);
808 ClearPageUnevictable(page);
810 if (page_evictable(page)) {
812 * For evictable pages, we can use the cache.
813 * In event of a race, worst case is we end up with an
814 * unevictable page on [in]active list.
815 * We know how to handle that.
817 is_unevictable = false;
821 * Put unevictable pages directly on zone's unevictable
824 is_unevictable = true;
825 add_page_to_unevictable_list(page);
827 * When racing with an mlock or AS_UNEVICTABLE clearing
828 * (page is unlocked) make sure that if the other thread
829 * does not observe our setting of PG_lru and fails
830 * isolation/check_move_unevictable_pages,
831 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
832 * the page back to the evictable list.
834 * The other side is TestClearPageMlocked() or shmem_lock().
840 * page's status can change while we move it among lru. If an evictable
841 * page is on unevictable list, it never be freed. To avoid that,
842 * check after we added it to the list, again.
844 if (is_unevictable && page_evictable(page)) {
845 if (!isolate_lru_page(page)) {
849 /* This means someone else dropped this page from LRU
850 * So, it will be freed or putback to LRU again. There is
851 * nothing to do here.
855 if (was_unevictable && !is_unevictable)
856 count_vm_event(UNEVICTABLE_PGRESCUED);
857 else if (!was_unevictable && is_unevictable)
858 count_vm_event(UNEVICTABLE_PGCULLED);
860 put_page(page); /* drop ref from isolate */
863 enum page_references {
865 PAGEREF_RECLAIM_CLEAN,
870 static enum page_references page_check_references(struct page *page,
871 struct scan_control *sc)
873 int referenced_ptes, referenced_page;
874 unsigned long vm_flags;
876 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
878 referenced_page = TestClearPageReferenced(page);
881 * Mlock lost the isolation race with us. Let try_to_unmap()
882 * move the page to the unevictable list.
884 if (vm_flags & VM_LOCKED)
885 return PAGEREF_RECLAIM;
887 if (referenced_ptes) {
888 if (PageSwapBacked(page))
889 return PAGEREF_ACTIVATE;
891 * All mapped pages start out with page table
892 * references from the instantiating fault, so we need
893 * to look twice if a mapped file page is used more
896 * Mark it and spare it for another trip around the
897 * inactive list. Another page table reference will
898 * lead to its activation.
900 * Note: the mark is set for activated pages as well
901 * so that recently deactivated but used pages are
904 SetPageReferenced(page);
906 if (referenced_page || referenced_ptes > 1)
907 return PAGEREF_ACTIVATE;
910 * Activate file-backed executable pages after first usage.
912 if (vm_flags & VM_EXEC)
913 return PAGEREF_ACTIVATE;
918 /* Reclaim if clean, defer dirty pages to writeback */
919 if (referenced_page && !PageSwapBacked(page))
920 return PAGEREF_RECLAIM_CLEAN;
922 return PAGEREF_RECLAIM;
925 /* Check if a page is dirty or under writeback */
926 static void page_check_dirty_writeback(struct page *page,
927 bool *dirty, bool *writeback)
929 struct address_space *mapping;
932 * Anonymous pages are not handled by flushers and must be written
933 * from reclaim context. Do not stall reclaim based on them
935 if (!page_is_file_cache(page) ||
936 (PageAnon(page) && !PageSwapBacked(page))) {
942 /* By default assume that the page flags are accurate */
943 *dirty = PageDirty(page);
944 *writeback = PageWriteback(page);
946 /* Verify dirty/writeback state if the filesystem supports it */
947 if (!page_has_private(page))
950 mapping = page_mapping(page);
951 if (mapping && mapping->a_ops->is_dirty_writeback)
952 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
955 struct reclaim_stat {
957 unsigned nr_unqueued_dirty;
958 unsigned nr_congested;
959 unsigned nr_writeback;
960 unsigned nr_immediate;
961 unsigned nr_activate;
962 unsigned nr_ref_keep;
963 unsigned nr_unmap_fail;
967 * shrink_page_list() returns the number of reclaimed pages
969 static unsigned long shrink_page_list(struct list_head *page_list,
970 struct pglist_data *pgdat,
971 struct scan_control *sc,
972 enum ttu_flags ttu_flags,
973 struct reclaim_stat *stat,
976 LIST_HEAD(ret_pages);
977 LIST_HEAD(free_pages);
979 unsigned nr_unqueued_dirty = 0;
980 unsigned nr_dirty = 0;
981 unsigned nr_congested = 0;
982 unsigned nr_reclaimed = 0;
983 unsigned nr_writeback = 0;
984 unsigned nr_immediate = 0;
985 unsigned nr_ref_keep = 0;
986 unsigned nr_unmap_fail = 0;
990 while (!list_empty(page_list)) {
991 struct address_space *mapping;
994 enum page_references references = PAGEREF_RECLAIM_CLEAN;
995 bool dirty, writeback;
999 page = lru_to_page(page_list);
1000 list_del(&page->lru);
1002 if (!trylock_page(page))
1005 VM_BUG_ON_PAGE(PageActive(page), page);
1009 if (unlikely(!page_evictable(page)))
1010 goto activate_locked;
1012 if (!sc->may_unmap && page_mapped(page))
1015 /* Double the slab pressure for mapped and swapcache pages */
1016 if ((page_mapped(page) || PageSwapCache(page)) &&
1017 !(PageAnon(page) && !PageSwapBacked(page)))
1020 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1021 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1024 * The number of dirty pages determines if a zone is marked
1025 * reclaim_congested which affects wait_iff_congested. kswapd
1026 * will stall and start writing pages if the tail of the LRU
1027 * is all dirty unqueued pages.
1029 page_check_dirty_writeback(page, &dirty, &writeback);
1030 if (dirty || writeback)
1033 if (dirty && !writeback)
1034 nr_unqueued_dirty++;
1037 * Treat this page as congested if the underlying BDI is or if
1038 * pages are cycling through the LRU so quickly that the
1039 * pages marked for immediate reclaim are making it to the
1040 * end of the LRU a second time.
1042 mapping = page_mapping(page);
1043 if (((dirty || writeback) && mapping &&
1044 inode_write_congested(mapping->host)) ||
1045 (writeback && PageReclaim(page)))
1049 * If a page at the tail of the LRU is under writeback, there
1050 * are three cases to consider.
1052 * 1) If reclaim is encountering an excessive number of pages
1053 * under writeback and this page is both under writeback and
1054 * PageReclaim then it indicates that pages are being queued
1055 * for IO but are being recycled through the LRU before the
1056 * IO can complete. Waiting on the page itself risks an
1057 * indefinite stall if it is impossible to writeback the
1058 * page due to IO error or disconnected storage so instead
1059 * note that the LRU is being scanned too quickly and the
1060 * caller can stall after page list has been processed.
1062 * 2) Global or new memcg reclaim encounters a page that is
1063 * not marked for immediate reclaim, or the caller does not
1064 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1065 * not to fs). In this case mark the page for immediate
1066 * reclaim and continue scanning.
1068 * Require may_enter_fs because we would wait on fs, which
1069 * may not have submitted IO yet. And the loop driver might
1070 * enter reclaim, and deadlock if it waits on a page for
1071 * which it is needed to do the write (loop masks off
1072 * __GFP_IO|__GFP_FS for this reason); but more thought
1073 * would probably show more reasons.
1075 * 3) Legacy memcg encounters a page that is already marked
1076 * PageReclaim. memcg does not have any dirty pages
1077 * throttling so we could easily OOM just because too many
1078 * pages are in writeback and there is nothing else to
1079 * reclaim. Wait for the writeback to complete.
1081 * In cases 1) and 2) we activate the pages to get them out of
1082 * the way while we continue scanning for clean pages on the
1083 * inactive list and refilling from the active list. The
1084 * observation here is that waiting for disk writes is more
1085 * expensive than potentially causing reloads down the line.
1086 * Since they're marked for immediate reclaim, they won't put
1087 * memory pressure on the cache working set any longer than it
1088 * takes to write them to disk.
1090 if (PageWriteback(page)) {
1092 if (current_is_kswapd() &&
1093 PageReclaim(page) &&
1094 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1096 goto activate_locked;
1099 } else if (sane_reclaim(sc) ||
1100 !PageReclaim(page) || !may_enter_fs) {
1102 * This is slightly racy - end_page_writeback()
1103 * might have just cleared PageReclaim, then
1104 * setting PageReclaim here end up interpreted
1105 * as PageReadahead - but that does not matter
1106 * enough to care. What we do want is for this
1107 * page to have PageReclaim set next time memcg
1108 * reclaim reaches the tests above, so it will
1109 * then wait_on_page_writeback() to avoid OOM;
1110 * and it's also appropriate in global reclaim.
1112 SetPageReclaim(page);
1114 goto activate_locked;
1119 wait_on_page_writeback(page);
1120 /* then go back and try same page again */
1121 list_add_tail(&page->lru, page_list);
1127 references = page_check_references(page, sc);
1129 switch (references) {
1130 case PAGEREF_ACTIVATE:
1131 goto activate_locked;
1135 case PAGEREF_RECLAIM:
1136 case PAGEREF_RECLAIM_CLEAN:
1137 ; /* try to reclaim the page below */
1141 * Anonymous process memory has backing store?
1142 * Try to allocate it some swap space here.
1143 * Lazyfree page could be freed directly
1145 if (PageAnon(page) && PageSwapBacked(page)) {
1146 if (!PageSwapCache(page)) {
1147 if (!(sc->gfp_mask & __GFP_IO))
1149 if (PageTransHuge(page)) {
1150 /* cannot split THP, skip it */
1151 if (!can_split_huge_page(page, NULL))
1152 goto activate_locked;
1154 * Split pages without a PMD map right
1155 * away. Chances are some or all of the
1156 * tail pages can be freed without IO.
1158 if (!compound_mapcount(page) &&
1159 split_huge_page_to_list(page,
1161 goto activate_locked;
1163 if (!add_to_swap(page)) {
1164 if (!PageTransHuge(page))
1165 goto activate_locked;
1166 /* Fallback to swap normal pages */
1167 if (split_huge_page_to_list(page,
1169 goto activate_locked;
1170 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1171 count_vm_event(THP_SWPOUT_FALLBACK);
1173 if (!add_to_swap(page))
1174 goto activate_locked;
1179 /* Adding to swap updated mapping */
1180 mapping = page_mapping(page);
1182 } else if (unlikely(PageTransHuge(page))) {
1183 /* Split file THP */
1184 if (split_huge_page_to_list(page, page_list))
1189 * The page is mapped into the page tables of one or more
1190 * processes. Try to unmap it here.
1192 if (page_mapped(page)) {
1193 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1195 if (unlikely(PageTransHuge(page)))
1196 flags |= TTU_SPLIT_HUGE_PMD;
1197 if (!try_to_unmap(page, flags)) {
1199 goto activate_locked;
1203 if (PageDirty(page)) {
1205 * Only kswapd can writeback filesystem pages
1206 * to avoid risk of stack overflow. But avoid
1207 * injecting inefficient single-page IO into
1208 * flusher writeback as much as possible: only
1209 * write pages when we've encountered many
1210 * dirty pages, and when we've already scanned
1211 * the rest of the LRU for clean pages and see
1212 * the same dirty pages again (PageReclaim).
1214 if (page_is_file_cache(page) &&
1215 (!current_is_kswapd() || !PageReclaim(page) ||
1216 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1218 * Immediately reclaim when written back.
1219 * Similar in principal to deactivate_page()
1220 * except we already have the page isolated
1221 * and know it's dirty
1223 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1224 SetPageReclaim(page);
1226 goto activate_locked;
1229 if (references == PAGEREF_RECLAIM_CLEAN)
1233 if (!sc->may_writepage)
1237 * Page is dirty. Flush the TLB if a writable entry
1238 * potentially exists to avoid CPU writes after IO
1239 * starts and then write it out here.
1241 try_to_unmap_flush_dirty();
1242 switch (pageout(page, mapping, sc)) {
1246 goto activate_locked;
1248 if (PageWriteback(page))
1250 if (PageDirty(page))
1254 * A synchronous write - probably a ramdisk. Go
1255 * ahead and try to reclaim the page.
1257 if (!trylock_page(page))
1259 if (PageDirty(page) || PageWriteback(page))
1261 mapping = page_mapping(page);
1263 ; /* try to free the page below */
1268 * If the page has buffers, try to free the buffer mappings
1269 * associated with this page. If we succeed we try to free
1272 * We do this even if the page is PageDirty().
1273 * try_to_release_page() does not perform I/O, but it is
1274 * possible for a page to have PageDirty set, but it is actually
1275 * clean (all its buffers are clean). This happens if the
1276 * buffers were written out directly, with submit_bh(). ext3
1277 * will do this, as well as the blockdev mapping.
1278 * try_to_release_page() will discover that cleanness and will
1279 * drop the buffers and mark the page clean - it can be freed.
1281 * Rarely, pages can have buffers and no ->mapping. These are
1282 * the pages which were not successfully invalidated in
1283 * truncate_complete_page(). We try to drop those buffers here
1284 * and if that worked, and the page is no longer mapped into
1285 * process address space (page_count == 1) it can be freed.
1286 * Otherwise, leave the page on the LRU so it is swappable.
1288 if (page_has_private(page)) {
1289 if (!try_to_release_page(page, sc->gfp_mask))
1290 goto activate_locked;
1291 if (!mapping && page_count(page) == 1) {
1293 if (put_page_testzero(page))
1297 * rare race with speculative reference.
1298 * the speculative reference will free
1299 * this page shortly, so we may
1300 * increment nr_reclaimed here (and
1301 * leave it off the LRU).
1309 if (PageAnon(page) && !PageSwapBacked(page)) {
1310 /* follow __remove_mapping for reference */
1311 if (!page_ref_freeze(page, 1))
1313 if (PageDirty(page)) {
1314 page_ref_unfreeze(page, 1);
1318 count_vm_event(PGLAZYFREED);
1319 count_memcg_page_event(page, PGLAZYFREED);
1320 } else if (!mapping || !__remove_mapping(mapping, page, true))
1323 * At this point, we have no other references and there is
1324 * no way to pick any more up (removed from LRU, removed
1325 * from pagecache). Can use non-atomic bitops now (and
1326 * we obviously don't have to worry about waking up a process
1327 * waiting on the page lock, because there are no references.
1329 __ClearPageLocked(page);
1334 * Is there need to periodically free_page_list? It would
1335 * appear not as the counts should be low
1337 if (unlikely(PageTransHuge(page))) {
1338 mem_cgroup_uncharge(page);
1339 (*get_compound_page_dtor(page))(page);
1341 list_add(&page->lru, &free_pages);
1345 /* Not a candidate for swapping, so reclaim swap space. */
1346 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1348 try_to_free_swap(page);
1349 VM_BUG_ON_PAGE(PageActive(page), page);
1350 if (!PageMlocked(page)) {
1351 SetPageActive(page);
1353 count_memcg_page_event(page, PGACTIVATE);
1358 list_add(&page->lru, &ret_pages);
1359 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1362 mem_cgroup_uncharge_list(&free_pages);
1363 try_to_unmap_flush();
1364 free_hot_cold_page_list(&free_pages, true);
1366 list_splice(&ret_pages, page_list);
1367 count_vm_events(PGACTIVATE, pgactivate);
1370 stat->nr_dirty = nr_dirty;
1371 stat->nr_congested = nr_congested;
1372 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1373 stat->nr_writeback = nr_writeback;
1374 stat->nr_immediate = nr_immediate;
1375 stat->nr_activate = pgactivate;
1376 stat->nr_ref_keep = nr_ref_keep;
1377 stat->nr_unmap_fail = nr_unmap_fail;
1379 return nr_reclaimed;
1382 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1383 struct list_head *page_list)
1385 struct scan_control sc = {
1386 .gfp_mask = GFP_KERNEL,
1387 .priority = DEF_PRIORITY,
1391 struct page *page, *next;
1392 LIST_HEAD(clean_pages);
1394 list_for_each_entry_safe(page, next, page_list, lru) {
1395 if (page_is_file_cache(page) && !PageDirty(page) &&
1396 !__PageMovable(page) && !PageUnevictable(page)) {
1397 ClearPageActive(page);
1398 list_move(&page->lru, &clean_pages);
1402 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1403 TTU_IGNORE_ACCESS, NULL, true);
1404 list_splice(&clean_pages, page_list);
1405 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1410 * Attempt to remove the specified page from its LRU. Only take this page
1411 * if it is of the appropriate PageActive status. Pages which are being
1412 * freed elsewhere are also ignored.
1414 * page: page to consider
1415 * mode: one of the LRU isolation modes defined above
1417 * returns 0 on success, -ve errno on failure.
1419 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1423 /* Only take pages on the LRU. */
1427 /* Compaction should not handle unevictable pages but CMA can do so */
1428 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1434 * To minimise LRU disruption, the caller can indicate that it only
1435 * wants to isolate pages it will be able to operate on without
1436 * blocking - clean pages for the most part.
1438 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1439 * that it is possible to migrate without blocking
1441 if (mode & ISOLATE_ASYNC_MIGRATE) {
1442 /* All the caller can do on PageWriteback is block */
1443 if (PageWriteback(page))
1446 if (PageDirty(page)) {
1447 struct address_space *mapping;
1451 * Only pages without mappings or that have a
1452 * ->migratepage callback are possible to migrate
1453 * without blocking. However, we can be racing with
1454 * truncation so it's necessary to lock the page
1455 * to stabilise the mapping as truncation holds
1456 * the page lock until after the page is removed
1457 * from the page cache.
1459 if (!trylock_page(page))
1462 mapping = page_mapping(page);
1463 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1470 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1473 if (likely(get_page_unless_zero(page))) {
1475 * Be careful not to clear PageLRU until after we're
1476 * sure the page is not being freed elsewhere -- the
1477 * page release code relies on it.
1488 * Update LRU sizes after isolating pages. The LRU size updates must
1489 * be complete before mem_cgroup_update_lru_size due to a santity check.
1491 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1492 enum lru_list lru, unsigned long *nr_zone_taken)
1496 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1497 if (!nr_zone_taken[zid])
1500 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1502 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1509 * zone_lru_lock is heavily contended. Some of the functions that
1510 * shrink the lists perform better by taking out a batch of pages
1511 * and working on them outside the LRU lock.
1513 * For pagecache intensive workloads, this function is the hottest
1514 * spot in the kernel (apart from copy_*_user functions).
1516 * Appropriate locks must be held before calling this function.
1518 * @nr_to_scan: The number of eligible pages to look through on the list.
1519 * @lruvec: The LRU vector to pull pages from.
1520 * @dst: The temp list to put pages on to.
1521 * @nr_scanned: The number of pages that were scanned.
1522 * @sc: The scan_control struct for this reclaim session
1523 * @mode: One of the LRU isolation modes
1524 * @lru: LRU list id for isolating
1526 * returns how many pages were moved onto *@dst.
1528 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1529 struct lruvec *lruvec, struct list_head *dst,
1530 unsigned long *nr_scanned, struct scan_control *sc,
1531 isolate_mode_t mode, enum lru_list lru)
1533 struct list_head *src = &lruvec->lists[lru];
1534 unsigned long nr_taken = 0;
1535 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1536 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1537 unsigned long skipped = 0;
1538 unsigned long scan, total_scan, nr_pages;
1539 LIST_HEAD(pages_skipped);
1542 for (total_scan = 0;
1543 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1547 page = lru_to_page(src);
1548 prefetchw_prev_lru_page(page, src, flags);
1550 VM_BUG_ON_PAGE(!PageLRU(page), page);
1552 if (page_zonenum(page) > sc->reclaim_idx) {
1553 list_move(&page->lru, &pages_skipped);
1554 nr_skipped[page_zonenum(page)]++;
1559 * Do not count skipped pages because that makes the function
1560 * return with no isolated pages if the LRU mostly contains
1561 * ineligible pages. This causes the VM to not reclaim any
1562 * pages, triggering a premature OOM.
1565 switch (__isolate_lru_page(page, mode)) {
1567 nr_pages = hpage_nr_pages(page);
1568 nr_taken += nr_pages;
1569 nr_zone_taken[page_zonenum(page)] += nr_pages;
1570 list_move(&page->lru, dst);
1574 /* else it is being freed elsewhere */
1575 list_move(&page->lru, src);
1584 * Splice any skipped pages to the start of the LRU list. Note that
1585 * this disrupts the LRU order when reclaiming for lower zones but
1586 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1587 * scanning would soon rescan the same pages to skip and put the
1588 * system at risk of premature OOM.
1590 if (!list_empty(&pages_skipped)) {
1593 list_splice(&pages_skipped, src);
1594 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1595 if (!nr_skipped[zid])
1598 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1599 skipped += nr_skipped[zid];
1602 *nr_scanned = total_scan;
1603 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1604 total_scan, skipped, nr_taken, mode, lru);
1605 update_lru_sizes(lruvec, lru, nr_zone_taken);
1610 * isolate_lru_page - tries to isolate a page from its LRU list
1611 * @page: page to isolate from its LRU list
1613 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1614 * vmstat statistic corresponding to whatever LRU list the page was on.
1616 * Returns 0 if the page was removed from an LRU list.
1617 * Returns -EBUSY if the page was not on an LRU list.
1619 * The returned page will have PageLRU() cleared. If it was found on
1620 * the active list, it will have PageActive set. If it was found on
1621 * the unevictable list, it will have the PageUnevictable bit set. That flag
1622 * may need to be cleared by the caller before letting the page go.
1624 * The vmstat statistic corresponding to the list on which the page was
1625 * found will be decremented.
1628 * (1) Must be called with an elevated refcount on the page. This is a
1629 * fundamentnal difference from isolate_lru_pages (which is called
1630 * without a stable reference).
1631 * (2) the lru_lock must not be held.
1632 * (3) interrupts must be enabled.
1634 int isolate_lru_page(struct page *page)
1638 VM_BUG_ON_PAGE(!page_count(page), page);
1639 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1641 if (PageLRU(page)) {
1642 struct zone *zone = page_zone(page);
1643 struct lruvec *lruvec;
1645 spin_lock_irq(zone_lru_lock(zone));
1646 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1647 if (PageLRU(page)) {
1648 int lru = page_lru(page);
1651 del_page_from_lru_list(page, lruvec, lru);
1654 spin_unlock_irq(zone_lru_lock(zone));
1660 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1661 * then get resheduled. When there are massive number of tasks doing page
1662 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1663 * the LRU list will go small and be scanned faster than necessary, leading to
1664 * unnecessary swapping, thrashing and OOM.
1666 static int too_many_isolated(struct pglist_data *pgdat, int file,
1667 struct scan_control *sc)
1669 unsigned long inactive, isolated;
1671 if (current_is_kswapd())
1674 if (!sane_reclaim(sc))
1678 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1679 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1681 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1682 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1686 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1687 * won't get blocked by normal direct-reclaimers, forming a circular
1690 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1693 return isolated > inactive;
1696 static noinline_for_stack void
1697 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1699 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1700 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1701 LIST_HEAD(pages_to_free);
1704 * Put back any unfreeable pages.
1706 while (!list_empty(page_list)) {
1707 struct page *page = lru_to_page(page_list);
1710 VM_BUG_ON_PAGE(PageLRU(page), page);
1711 list_del(&page->lru);
1712 if (unlikely(!page_evictable(page))) {
1713 spin_unlock_irq(&pgdat->lru_lock);
1714 putback_lru_page(page);
1715 spin_lock_irq(&pgdat->lru_lock);
1719 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1722 lru = page_lru(page);
1723 add_page_to_lru_list(page, lruvec, lru);
1725 if (is_active_lru(lru)) {
1726 int file = is_file_lru(lru);
1727 int numpages = hpage_nr_pages(page);
1728 reclaim_stat->recent_rotated[file] += numpages;
1730 if (put_page_testzero(page)) {
1731 __ClearPageLRU(page);
1732 __ClearPageActive(page);
1733 del_page_from_lru_list(page, lruvec, lru);
1735 if (unlikely(PageCompound(page))) {
1736 spin_unlock_irq(&pgdat->lru_lock);
1737 mem_cgroup_uncharge(page);
1738 (*get_compound_page_dtor(page))(page);
1739 spin_lock_irq(&pgdat->lru_lock);
1741 list_add(&page->lru, &pages_to_free);
1746 * To save our caller's stack, now use input list for pages to free.
1748 list_splice(&pages_to_free, page_list);
1752 * If a kernel thread (such as nfsd for loop-back mounts) services
1753 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1754 * In that case we should only throttle if the backing device it is
1755 * writing to is congested. In other cases it is safe to throttle.
1757 static int current_may_throttle(void)
1759 return !(current->flags & PF_LESS_THROTTLE) ||
1760 current->backing_dev_info == NULL ||
1761 bdi_write_congested(current->backing_dev_info);
1765 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1766 * of reclaimed pages
1768 static noinline_for_stack unsigned long
1769 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1770 struct scan_control *sc, enum lru_list lru)
1772 LIST_HEAD(page_list);
1773 unsigned long nr_scanned;
1774 unsigned long nr_reclaimed = 0;
1775 unsigned long nr_taken;
1776 struct reclaim_stat stat = {};
1777 isolate_mode_t isolate_mode = 0;
1778 int file = is_file_lru(lru);
1779 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1780 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1781 bool stalled = false;
1783 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1787 /* wait a bit for the reclaimer. */
1791 /* We are about to die and free our memory. Return now. */
1792 if (fatal_signal_pending(current))
1793 return SWAP_CLUSTER_MAX;
1799 isolate_mode |= ISOLATE_UNMAPPED;
1801 spin_lock_irq(&pgdat->lru_lock);
1803 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1804 &nr_scanned, sc, isolate_mode, lru);
1806 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1807 reclaim_stat->recent_scanned[file] += nr_taken;
1809 if (current_is_kswapd()) {
1810 if (global_reclaim(sc))
1811 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1812 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1815 if (global_reclaim(sc))
1816 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1817 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1820 spin_unlock_irq(&pgdat->lru_lock);
1825 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1828 spin_lock_irq(&pgdat->lru_lock);
1830 if (current_is_kswapd()) {
1831 if (global_reclaim(sc))
1832 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1833 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1836 if (global_reclaim(sc))
1837 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1838 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1842 putback_inactive_pages(lruvec, &page_list);
1844 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1846 spin_unlock_irq(&pgdat->lru_lock);
1848 mem_cgroup_uncharge_list(&page_list);
1849 free_hot_cold_page_list(&page_list, true);
1852 * If reclaim is isolating dirty pages under writeback, it implies
1853 * that the long-lived page allocation rate is exceeding the page
1854 * laundering rate. Either the global limits are not being effective
1855 * at throttling processes due to the page distribution throughout
1856 * zones or there is heavy usage of a slow backing device. The
1857 * only option is to throttle from reclaim context which is not ideal
1858 * as there is no guarantee the dirtying process is throttled in the
1859 * same way balance_dirty_pages() manages.
1861 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1862 * of pages under pages flagged for immediate reclaim and stall if any
1863 * are encountered in the nr_immediate check below.
1865 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1866 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1869 * If dirty pages are scanned that are not queued for IO, it
1870 * implies that flushers are not doing their job. This can
1871 * happen when memory pressure pushes dirty pages to the end of
1872 * the LRU before the dirty limits are breached and the dirty
1873 * data has expired. It can also happen when the proportion of
1874 * dirty pages grows not through writes but through memory
1875 * pressure reclaiming all the clean cache. And in some cases,
1876 * the flushers simply cannot keep up with the allocation
1877 * rate. Nudge the flusher threads in case they are asleep.
1879 if (stat.nr_unqueued_dirty == nr_taken)
1880 wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1883 * Legacy memcg will stall in page writeback so avoid forcibly
1886 if (sane_reclaim(sc)) {
1888 * Tag a zone as congested if all the dirty pages scanned were
1889 * backed by a congested BDI and wait_iff_congested will stall.
1891 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1892 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1894 /* Allow kswapd to start writing pages during reclaim. */
1895 if (stat.nr_unqueued_dirty == nr_taken)
1896 set_bit(PGDAT_DIRTY, &pgdat->flags);
1899 * If kswapd scans pages marked marked for immediate
1900 * reclaim and under writeback (nr_immediate), it implies
1901 * that pages are cycling through the LRU faster than
1902 * they are written so also forcibly stall.
1904 if (stat.nr_immediate && current_may_throttle())
1905 congestion_wait(BLK_RW_ASYNC, HZ/10);
1909 * Stall direct reclaim for IO completions if underlying BDIs or zone
1910 * is congested. Allow kswapd to continue until it starts encountering
1911 * unqueued dirty pages or cycling through the LRU too quickly.
1913 if (!sc->hibernation_mode && !current_is_kswapd() &&
1914 current_may_throttle())
1915 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1917 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1918 nr_scanned, nr_reclaimed,
1919 stat.nr_dirty, stat.nr_writeback,
1920 stat.nr_congested, stat.nr_immediate,
1921 stat.nr_activate, stat.nr_ref_keep,
1923 sc->priority, file);
1924 return nr_reclaimed;
1928 * This moves pages from the active list to the inactive list.
1930 * We move them the other way if the page is referenced by one or more
1931 * processes, from rmap.
1933 * If the pages are mostly unmapped, the processing is fast and it is
1934 * appropriate to hold zone_lru_lock across the whole operation. But if
1935 * the pages are mapped, the processing is slow (page_referenced()) so we
1936 * should drop zone_lru_lock around each page. It's impossible to balance
1937 * this, so instead we remove the pages from the LRU while processing them.
1938 * It is safe to rely on PG_active against the non-LRU pages in here because
1939 * nobody will play with that bit on a non-LRU page.
1941 * The downside is that we have to touch page->_refcount against each page.
1942 * But we had to alter page->flags anyway.
1944 * Returns the number of pages moved to the given lru.
1947 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1948 struct list_head *list,
1949 struct list_head *pages_to_free,
1952 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1957 while (!list_empty(list)) {
1958 page = lru_to_page(list);
1959 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1961 VM_BUG_ON_PAGE(PageLRU(page), page);
1964 nr_pages = hpage_nr_pages(page);
1965 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1966 list_move(&page->lru, &lruvec->lists[lru]);
1968 if (put_page_testzero(page)) {
1969 __ClearPageLRU(page);
1970 __ClearPageActive(page);
1971 del_page_from_lru_list(page, lruvec, lru);
1973 if (unlikely(PageCompound(page))) {
1974 spin_unlock_irq(&pgdat->lru_lock);
1975 mem_cgroup_uncharge(page);
1976 (*get_compound_page_dtor(page))(page);
1977 spin_lock_irq(&pgdat->lru_lock);
1979 list_add(&page->lru, pages_to_free);
1981 nr_moved += nr_pages;
1985 if (!is_active_lru(lru)) {
1986 __count_vm_events(PGDEACTIVATE, nr_moved);
1987 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1994 static void shrink_active_list(unsigned long nr_to_scan,
1995 struct lruvec *lruvec,
1996 struct scan_control *sc,
1999 unsigned long nr_taken;
2000 unsigned long nr_scanned;
2001 unsigned long vm_flags;
2002 LIST_HEAD(l_hold); /* The pages which were snipped off */
2003 LIST_HEAD(l_active);
2004 LIST_HEAD(l_inactive);
2006 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2007 unsigned nr_deactivate, nr_activate;
2008 unsigned nr_rotated = 0;
2009 isolate_mode_t isolate_mode = 0;
2010 int file = is_file_lru(lru);
2011 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2016 isolate_mode |= ISOLATE_UNMAPPED;
2018 spin_lock_irq(&pgdat->lru_lock);
2020 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2021 &nr_scanned, sc, isolate_mode, lru);
2023 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2024 reclaim_stat->recent_scanned[file] += nr_taken;
2026 __count_vm_events(PGREFILL, nr_scanned);
2027 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2029 spin_unlock_irq(&pgdat->lru_lock);
2031 while (!list_empty(&l_hold)) {
2033 page = lru_to_page(&l_hold);
2034 list_del(&page->lru);
2036 if (unlikely(!page_evictable(page))) {
2037 putback_lru_page(page);
2041 if (unlikely(buffer_heads_over_limit)) {
2042 if (page_has_private(page) && trylock_page(page)) {
2043 if (page_has_private(page))
2044 try_to_release_page(page, 0);
2049 if (page_referenced(page, 0, sc->target_mem_cgroup,
2051 nr_rotated += hpage_nr_pages(page);
2053 * Identify referenced, file-backed active pages and
2054 * give them one more trip around the active list. So
2055 * that executable code get better chances to stay in
2056 * memory under moderate memory pressure. Anon pages
2057 * are not likely to be evicted by use-once streaming
2058 * IO, plus JVM can create lots of anon VM_EXEC pages,
2059 * so we ignore them here.
2061 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2062 list_add(&page->lru, &l_active);
2067 ClearPageActive(page); /* we are de-activating */
2068 list_add(&page->lru, &l_inactive);
2072 * Move pages back to the lru list.
2074 spin_lock_irq(&pgdat->lru_lock);
2076 * Count referenced pages from currently used mappings as rotated,
2077 * even though only some of them are actually re-activated. This
2078 * helps balance scan pressure between file and anonymous pages in
2081 reclaim_stat->recent_rotated[file] += nr_rotated;
2083 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2084 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2085 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2086 spin_unlock_irq(&pgdat->lru_lock);
2088 mem_cgroup_uncharge_list(&l_hold);
2089 free_hot_cold_page_list(&l_hold, true);
2090 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2091 nr_deactivate, nr_rotated, sc->priority, file);
2095 * The inactive anon list should be small enough that the VM never has
2096 * to do too much work.
2098 * The inactive file list should be small enough to leave most memory
2099 * to the established workingset on the scan-resistant active list,
2100 * but large enough to avoid thrashing the aggregate readahead window.
2102 * Both inactive lists should also be large enough that each inactive
2103 * page has a chance to be referenced again before it is reclaimed.
2105 * If that fails and refaulting is observed, the inactive list grows.
2107 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2108 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2109 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2112 * memory ratio inactive
2113 * -------------------------------------
2122 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2123 struct scan_control *sc, bool trace)
2125 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2126 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2127 enum lru_list inactive_lru = file * LRU_FILE;
2128 unsigned long inactive, active;
2129 unsigned long inactive_ratio;
2130 unsigned long refaults;
2134 * If we don't have swap space, anonymous page deactivation
2137 if (!file && !total_swap_pages)
2140 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2141 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2144 * When refaults are being observed, it means a new workingset
2145 * is being established. Disable active list protection to get
2146 * rid of the stale workingset quickly.
2148 refaults = lruvec_page_state(lruvec, WORKINGSET_ACTIVATE);
2149 if (file && lruvec->refaults != refaults) {
2152 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2154 inactive_ratio = int_sqrt(10 * gb);
2160 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2161 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2162 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2163 inactive_ratio, file);
2165 return inactive * inactive_ratio < active;
2168 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2169 struct lruvec *lruvec, struct scan_control *sc)
2171 if (is_active_lru(lru)) {
2172 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2173 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2177 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2188 * Determine how aggressively the anon and file LRU lists should be
2189 * scanned. The relative value of each set of LRU lists is determined
2190 * by looking at the fraction of the pages scanned we did rotate back
2191 * onto the active list instead of evict.
2193 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2194 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2196 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2197 struct scan_control *sc, unsigned long *nr,
2198 unsigned long *lru_pages)
2200 int swappiness = mem_cgroup_swappiness(memcg);
2201 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2203 u64 denominator = 0; /* gcc */
2204 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2205 unsigned long anon_prio, file_prio;
2206 enum scan_balance scan_balance;
2207 unsigned long anon, file;
2208 unsigned long ap, fp;
2211 /* If we have no swap space, do not bother scanning anon pages. */
2212 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2213 scan_balance = SCAN_FILE;
2218 * Global reclaim will swap to prevent OOM even with no
2219 * swappiness, but memcg users want to use this knob to
2220 * disable swapping for individual groups completely when
2221 * using the memory controller's swap limit feature would be
2224 if (!global_reclaim(sc) && !swappiness) {
2225 scan_balance = SCAN_FILE;
2230 * Do not apply any pressure balancing cleverness when the
2231 * system is close to OOM, scan both anon and file equally
2232 * (unless the swappiness setting disagrees with swapping).
2234 if (!sc->priority && swappiness) {
2235 scan_balance = SCAN_EQUAL;
2240 * Prevent the reclaimer from falling into the cache trap: as
2241 * cache pages start out inactive, every cache fault will tip
2242 * the scan balance towards the file LRU. And as the file LRU
2243 * shrinks, so does the window for rotation from references.
2244 * This means we have a runaway feedback loop where a tiny
2245 * thrashing file LRU becomes infinitely more attractive than
2246 * anon pages. Try to detect this based on file LRU size.
2248 if (global_reclaim(sc)) {
2249 unsigned long pgdatfile;
2250 unsigned long pgdatfree;
2252 unsigned long total_high_wmark = 0;
2254 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2255 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2256 node_page_state(pgdat, NR_INACTIVE_FILE);
2258 for (z = 0; z < MAX_NR_ZONES; z++) {
2259 struct zone *zone = &pgdat->node_zones[z];
2260 if (!managed_zone(zone))
2263 total_high_wmark += high_wmark_pages(zone);
2266 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2268 * Force SCAN_ANON if there are enough inactive
2269 * anonymous pages on the LRU in eligible zones.
2270 * Otherwise, the small LRU gets thrashed.
2272 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2273 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2275 scan_balance = SCAN_ANON;
2282 * If there is enough inactive page cache, i.e. if the size of the
2283 * inactive list is greater than that of the active list *and* the
2284 * inactive list actually has some pages to scan on this priority, we
2285 * do not reclaim anything from the anonymous working set right now.
2286 * Without the second condition we could end up never scanning an
2287 * lruvec even if it has plenty of old anonymous pages unless the
2288 * system is under heavy pressure.
2290 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2291 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2292 scan_balance = SCAN_FILE;
2296 scan_balance = SCAN_FRACT;
2299 * With swappiness at 100, anonymous and file have the same priority.
2300 * This scanning priority is essentially the inverse of IO cost.
2302 anon_prio = swappiness;
2303 file_prio = 200 - anon_prio;
2306 * OK, so we have swap space and a fair amount of page cache
2307 * pages. We use the recently rotated / recently scanned
2308 * ratios to determine how valuable each cache is.
2310 * Because workloads change over time (and to avoid overflow)
2311 * we keep these statistics as a floating average, which ends
2312 * up weighing recent references more than old ones.
2314 * anon in [0], file in [1]
2317 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2318 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2319 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2320 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2322 spin_lock_irq(&pgdat->lru_lock);
2323 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2324 reclaim_stat->recent_scanned[0] /= 2;
2325 reclaim_stat->recent_rotated[0] /= 2;
2328 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2329 reclaim_stat->recent_scanned[1] /= 2;
2330 reclaim_stat->recent_rotated[1] /= 2;
2334 * The amount of pressure on anon vs file pages is inversely
2335 * proportional to the fraction of recently scanned pages on
2336 * each list that were recently referenced and in active use.
2338 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2339 ap /= reclaim_stat->recent_rotated[0] + 1;
2341 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2342 fp /= reclaim_stat->recent_rotated[1] + 1;
2343 spin_unlock_irq(&pgdat->lru_lock);
2347 denominator = ap + fp + 1;
2350 for_each_evictable_lru(lru) {
2351 int file = is_file_lru(lru);
2355 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2356 scan = size >> sc->priority;
2358 * If the cgroup's already been deleted, make sure to
2359 * scrape out the remaining cache.
2361 if (!scan && !mem_cgroup_online(memcg))
2362 scan = min(size, SWAP_CLUSTER_MAX);
2364 switch (scan_balance) {
2366 /* Scan lists relative to size */
2370 * Scan types proportional to swappiness and
2371 * their relative recent reclaim efficiency.
2372 * Make sure we don't miss the last page on
2373 * the offlined memory cgroups because of a
2376 scan = mem_cgroup_online(memcg) ?
2377 div64_u64(scan * fraction[file], denominator) :
2378 DIV64_U64_ROUND_UP(scan * fraction[file],
2383 /* Scan one type exclusively */
2384 if ((scan_balance == SCAN_FILE) != file) {
2390 /* Look ma, no brain */
2400 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2402 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2403 struct scan_control *sc, unsigned long *lru_pages)
2405 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2406 unsigned long nr[NR_LRU_LISTS];
2407 unsigned long targets[NR_LRU_LISTS];
2408 unsigned long nr_to_scan;
2410 unsigned long nr_reclaimed = 0;
2411 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2412 struct blk_plug plug;
2415 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2417 /* Record the original scan target for proportional adjustments later */
2418 memcpy(targets, nr, sizeof(nr));
2421 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2422 * event that can occur when there is little memory pressure e.g.
2423 * multiple streaming readers/writers. Hence, we do not abort scanning
2424 * when the requested number of pages are reclaimed when scanning at
2425 * DEF_PRIORITY on the assumption that the fact we are direct
2426 * reclaiming implies that kswapd is not keeping up and it is best to
2427 * do a batch of work at once. For memcg reclaim one check is made to
2428 * abort proportional reclaim if either the file or anon lru has already
2429 * dropped to zero at the first pass.
2431 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2432 sc->priority == DEF_PRIORITY);
2434 blk_start_plug(&plug);
2435 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2436 nr[LRU_INACTIVE_FILE]) {
2437 unsigned long nr_anon, nr_file, percentage;
2438 unsigned long nr_scanned;
2440 for_each_evictable_lru(lru) {
2442 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2443 nr[lru] -= nr_to_scan;
2445 nr_reclaimed += shrink_list(lru, nr_to_scan,
2452 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2456 * For kswapd and memcg, reclaim at least the number of pages
2457 * requested. Ensure that the anon and file LRUs are scanned
2458 * proportionally what was requested by get_scan_count(). We
2459 * stop reclaiming one LRU and reduce the amount scanning
2460 * proportional to the original scan target.
2462 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2463 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2466 * It's just vindictive to attack the larger once the smaller
2467 * has gone to zero. And given the way we stop scanning the
2468 * smaller below, this makes sure that we only make one nudge
2469 * towards proportionality once we've got nr_to_reclaim.
2471 if (!nr_file || !nr_anon)
2474 if (nr_file > nr_anon) {
2475 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2476 targets[LRU_ACTIVE_ANON] + 1;
2478 percentage = nr_anon * 100 / scan_target;
2480 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2481 targets[LRU_ACTIVE_FILE] + 1;
2483 percentage = nr_file * 100 / scan_target;
2486 /* Stop scanning the smaller of the LRU */
2488 nr[lru + LRU_ACTIVE] = 0;
2491 * Recalculate the other LRU scan count based on its original
2492 * scan target and the percentage scanning already complete
2494 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2495 nr_scanned = targets[lru] - nr[lru];
2496 nr[lru] = targets[lru] * (100 - percentage) / 100;
2497 nr[lru] -= min(nr[lru], nr_scanned);
2500 nr_scanned = targets[lru] - nr[lru];
2501 nr[lru] = targets[lru] * (100 - percentage) / 100;
2502 nr[lru] -= min(nr[lru], nr_scanned);
2504 scan_adjusted = true;
2506 blk_finish_plug(&plug);
2507 sc->nr_reclaimed += nr_reclaimed;
2510 * Even if we did not try to evict anon pages at all, we want to
2511 * rebalance the anon lru active/inactive ratio.
2513 if (inactive_list_is_low(lruvec, false, sc, true))
2514 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2515 sc, LRU_ACTIVE_ANON);
2518 /* Use reclaim/compaction for costly allocs or under memory pressure */
2519 static bool in_reclaim_compaction(struct scan_control *sc)
2521 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2522 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2523 sc->priority < DEF_PRIORITY - 2))
2530 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2531 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2532 * true if more pages should be reclaimed such that when the page allocator
2533 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2534 * It will give up earlier than that if there is difficulty reclaiming pages.
2536 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2537 unsigned long nr_reclaimed,
2538 unsigned long nr_scanned,
2539 struct scan_control *sc)
2541 unsigned long pages_for_compaction;
2542 unsigned long inactive_lru_pages;
2545 /* If not in reclaim/compaction mode, stop */
2546 if (!in_reclaim_compaction(sc))
2549 /* Consider stopping depending on scan and reclaim activity */
2550 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2552 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2553 * full LRU list has been scanned and we are still failing
2554 * to reclaim pages. This full LRU scan is potentially
2555 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2557 if (!nr_reclaimed && !nr_scanned)
2561 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2562 * fail without consequence, stop if we failed to reclaim
2563 * any pages from the last SWAP_CLUSTER_MAX number of
2564 * pages that were scanned. This will return to the
2565 * caller faster at the risk reclaim/compaction and
2566 * the resulting allocation attempt fails
2573 * If we have not reclaimed enough pages for compaction and the
2574 * inactive lists are large enough, continue reclaiming
2576 pages_for_compaction = compact_gap(sc->order);
2577 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2578 if (get_nr_swap_pages() > 0)
2579 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2580 if (sc->nr_reclaimed < pages_for_compaction &&
2581 inactive_lru_pages > pages_for_compaction)
2584 /* If compaction would go ahead or the allocation would succeed, stop */
2585 for (z = 0; z <= sc->reclaim_idx; z++) {
2586 struct zone *zone = &pgdat->node_zones[z];
2587 if (!managed_zone(zone))
2590 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2591 case COMPACT_SUCCESS:
2592 case COMPACT_CONTINUE:
2595 /* check next zone */
2602 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2604 struct reclaim_state *reclaim_state = current->reclaim_state;
2605 unsigned long nr_reclaimed, nr_scanned;
2606 bool reclaimable = false;
2609 struct mem_cgroup *root = sc->target_mem_cgroup;
2610 struct mem_cgroup_reclaim_cookie reclaim = {
2612 .priority = sc->priority,
2614 unsigned long node_lru_pages = 0;
2615 struct mem_cgroup *memcg;
2617 nr_reclaimed = sc->nr_reclaimed;
2618 nr_scanned = sc->nr_scanned;
2620 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2622 unsigned long lru_pages;
2623 unsigned long reclaimed;
2624 unsigned long scanned;
2626 if (mem_cgroup_low(root, memcg)) {
2627 if (!sc->memcg_low_reclaim) {
2628 sc->memcg_low_skipped = 1;
2631 memcg_memory_event(memcg, MEMCG_LOW);
2634 reclaimed = sc->nr_reclaimed;
2635 scanned = sc->nr_scanned;
2637 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2638 node_lru_pages += lru_pages;
2641 shrink_slab(sc->gfp_mask, pgdat->node_id,
2642 memcg, sc->nr_scanned - scanned,
2645 /* Record the group's reclaim efficiency */
2646 vmpressure(sc->gfp_mask, memcg, false,
2647 sc->nr_scanned - scanned,
2648 sc->nr_reclaimed - reclaimed);
2651 * Direct reclaim and kswapd have to scan all memory
2652 * cgroups to fulfill the overall scan target for the
2655 * Limit reclaim, on the other hand, only cares about
2656 * nr_to_reclaim pages to be reclaimed and it will
2657 * retry with decreasing priority if one round over the
2658 * whole hierarchy is not sufficient.
2660 if (!global_reclaim(sc) &&
2661 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2662 mem_cgroup_iter_break(root, memcg);
2665 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2668 * Shrink the slab caches in the same proportion that
2669 * the eligible LRU pages were scanned.
2671 if (global_reclaim(sc))
2672 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2673 sc->nr_scanned - nr_scanned,
2676 if (reclaim_state) {
2677 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2678 reclaim_state->reclaimed_slab = 0;
2681 /* Record the subtree's reclaim efficiency */
2682 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2683 sc->nr_scanned - nr_scanned,
2684 sc->nr_reclaimed - nr_reclaimed);
2686 if (sc->nr_reclaimed - nr_reclaimed)
2689 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2690 sc->nr_scanned - nr_scanned, sc));
2693 * Kswapd gives up on balancing particular nodes after too
2694 * many failures to reclaim anything from them and goes to
2695 * sleep. On reclaim progress, reset the failure counter. A
2696 * successful direct reclaim run will revive a dormant kswapd.
2699 pgdat->kswapd_failures = 0;
2705 * Returns true if compaction should go ahead for a costly-order request, or
2706 * the allocation would already succeed without compaction. Return false if we
2707 * should reclaim first.
2709 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2711 unsigned long watermark;
2712 enum compact_result suitable;
2714 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2715 if (suitable == COMPACT_SUCCESS)
2716 /* Allocation should succeed already. Don't reclaim. */
2718 if (suitable == COMPACT_SKIPPED)
2719 /* Compaction cannot yet proceed. Do reclaim. */
2723 * Compaction is already possible, but it takes time to run and there
2724 * are potentially other callers using the pages just freed. So proceed
2725 * with reclaim to make a buffer of free pages available to give
2726 * compaction a reasonable chance of completing and allocating the page.
2727 * Note that we won't actually reclaim the whole buffer in one attempt
2728 * as the target watermark in should_continue_reclaim() is lower. But if
2729 * we are already above the high+gap watermark, don't reclaim at all.
2731 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2733 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2737 * This is the direct reclaim path, for page-allocating processes. We only
2738 * try to reclaim pages from zones which will satisfy the caller's allocation
2741 * If a zone is deemed to be full of pinned pages then just give it a light
2742 * scan then give up on it.
2744 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2748 unsigned long nr_soft_reclaimed;
2749 unsigned long nr_soft_scanned;
2751 pg_data_t *last_pgdat = NULL;
2754 * If the number of buffer_heads in the machine exceeds the maximum
2755 * allowed level, force direct reclaim to scan the highmem zone as
2756 * highmem pages could be pinning lowmem pages storing buffer_heads
2758 orig_mask = sc->gfp_mask;
2759 if (buffer_heads_over_limit) {
2760 sc->gfp_mask |= __GFP_HIGHMEM;
2761 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2764 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2765 sc->reclaim_idx, sc->nodemask) {
2767 * Take care memory controller reclaiming has small influence
2770 if (global_reclaim(sc)) {
2771 if (!cpuset_zone_allowed(zone,
2772 GFP_KERNEL | __GFP_HARDWALL))
2776 * If we already have plenty of memory free for
2777 * compaction in this zone, don't free any more.
2778 * Even though compaction is invoked for any
2779 * non-zero order, only frequent costly order
2780 * reclamation is disruptive enough to become a
2781 * noticeable problem, like transparent huge
2784 if (IS_ENABLED(CONFIG_COMPACTION) &&
2785 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2786 compaction_ready(zone, sc)) {
2787 sc->compaction_ready = true;
2792 * Shrink each node in the zonelist once. If the
2793 * zonelist is ordered by zone (not the default) then a
2794 * node may be shrunk multiple times but in that case
2795 * the user prefers lower zones being preserved.
2797 if (zone->zone_pgdat == last_pgdat)
2801 * This steals pages from memory cgroups over softlimit
2802 * and returns the number of reclaimed pages and
2803 * scanned pages. This works for global memory pressure
2804 * and balancing, not for a memcg's limit.
2806 nr_soft_scanned = 0;
2807 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2808 sc->order, sc->gfp_mask,
2810 sc->nr_reclaimed += nr_soft_reclaimed;
2811 sc->nr_scanned += nr_soft_scanned;
2812 /* need some check for avoid more shrink_zone() */
2815 /* See comment about same check for global reclaim above */
2816 if (zone->zone_pgdat == last_pgdat)
2818 last_pgdat = zone->zone_pgdat;
2819 shrink_node(zone->zone_pgdat, sc);
2823 * Restore to original mask to avoid the impact on the caller if we
2824 * promoted it to __GFP_HIGHMEM.
2826 sc->gfp_mask = orig_mask;
2829 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2831 struct mem_cgroup *memcg;
2833 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2835 unsigned long refaults;
2836 struct lruvec *lruvec;
2838 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2839 refaults = lruvec_page_state(lruvec, WORKINGSET_ACTIVATE);
2840 lruvec->refaults = refaults;
2841 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2845 * This is the main entry point to direct page reclaim.
2847 * If a full scan of the inactive list fails to free enough memory then we
2848 * are "out of memory" and something needs to be killed.
2850 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2851 * high - the zone may be full of dirty or under-writeback pages, which this
2852 * caller can't do much about. We kick the writeback threads and take explicit
2853 * naps in the hope that some of these pages can be written. But if the
2854 * allocating task holds filesystem locks which prevent writeout this might not
2855 * work, and the allocation attempt will fail.
2857 * returns: 0, if no pages reclaimed
2858 * else, the number of pages reclaimed
2860 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2861 struct scan_control *sc)
2863 int initial_priority = sc->priority;
2864 pg_data_t *last_pgdat;
2868 delayacct_freepages_start();
2870 if (global_reclaim(sc))
2871 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2874 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2877 shrink_zones(zonelist, sc);
2879 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2882 if (sc->compaction_ready)
2886 * If we're getting trouble reclaiming, start doing
2887 * writepage even in laptop mode.
2889 if (sc->priority < DEF_PRIORITY - 2)
2890 sc->may_writepage = 1;
2891 } while (--sc->priority >= 0);
2894 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2896 if (zone->zone_pgdat == last_pgdat)
2898 last_pgdat = zone->zone_pgdat;
2899 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2902 delayacct_freepages_end();
2904 if (sc->nr_reclaimed)
2905 return sc->nr_reclaimed;
2907 /* Aborted reclaim to try compaction? don't OOM, then */
2908 if (sc->compaction_ready)
2911 /* Untapped cgroup reserves? Don't OOM, retry. */
2912 if (sc->memcg_low_skipped) {
2913 sc->priority = initial_priority;
2914 sc->memcg_low_reclaim = 1;
2915 sc->memcg_low_skipped = 0;
2922 static bool allow_direct_reclaim(pg_data_t *pgdat)
2925 unsigned long pfmemalloc_reserve = 0;
2926 unsigned long free_pages = 0;
2930 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2933 for (i = 0; i <= ZONE_NORMAL; i++) {
2934 zone = &pgdat->node_zones[i];
2935 if (!managed_zone(zone))
2938 if (!zone_reclaimable_pages(zone))
2941 pfmemalloc_reserve += min_wmark_pages(zone);
2942 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2945 /* If there are no reserves (unexpected config) then do not throttle */
2946 if (!pfmemalloc_reserve)
2949 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2951 /* kswapd must be awake if processes are being throttled */
2952 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2953 if (READ_ONCE(pgdat->kswapd_classzone_idx) > ZONE_NORMAL)
2954 WRITE_ONCE(pgdat->kswapd_classzone_idx, ZONE_NORMAL);
2956 wake_up_interruptible(&pgdat->kswapd_wait);
2963 * Throttle direct reclaimers if backing storage is backed by the network
2964 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2965 * depleted. kswapd will continue to make progress and wake the processes
2966 * when the low watermark is reached.
2968 * Returns true if a fatal signal was delivered during throttling. If this
2969 * happens, the page allocator should not consider triggering the OOM killer.
2971 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2972 nodemask_t *nodemask)
2976 pg_data_t *pgdat = NULL;
2979 * Kernel threads should not be throttled as they may be indirectly
2980 * responsible for cleaning pages necessary for reclaim to make forward
2981 * progress. kjournald for example may enter direct reclaim while
2982 * committing a transaction where throttling it could forcing other
2983 * processes to block on log_wait_commit().
2985 if (current->flags & PF_KTHREAD)
2989 * If a fatal signal is pending, this process should not throttle.
2990 * It should return quickly so it can exit and free its memory
2992 if (fatal_signal_pending(current))
2996 * Check if the pfmemalloc reserves are ok by finding the first node
2997 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2998 * GFP_KERNEL will be required for allocating network buffers when
2999 * swapping over the network so ZONE_HIGHMEM is unusable.
3001 * Throttling is based on the first usable node and throttled processes
3002 * wait on a queue until kswapd makes progress and wakes them. There
3003 * is an affinity then between processes waking up and where reclaim
3004 * progress has been made assuming the process wakes on the same node.
3005 * More importantly, processes running on remote nodes will not compete
3006 * for remote pfmemalloc reserves and processes on different nodes
3007 * should make reasonable progress.
3009 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3010 gfp_zone(gfp_mask), nodemask) {
3011 if (zone_idx(zone) > ZONE_NORMAL)
3014 /* Throttle based on the first usable node */
3015 pgdat = zone->zone_pgdat;
3016 if (allow_direct_reclaim(pgdat))
3021 /* If no zone was usable by the allocation flags then do not throttle */
3025 /* Account for the throttling */
3026 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3029 * If the caller cannot enter the filesystem, it's possible that it
3030 * is due to the caller holding an FS lock or performing a journal
3031 * transaction in the case of a filesystem like ext[3|4]. In this case,
3032 * it is not safe to block on pfmemalloc_wait as kswapd could be
3033 * blocked waiting on the same lock. Instead, throttle for up to a
3034 * second before continuing.
3036 if (!(gfp_mask & __GFP_FS)) {
3037 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3038 allow_direct_reclaim(pgdat), HZ);
3043 /* Throttle until kswapd wakes the process */
3044 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3045 allow_direct_reclaim(pgdat));
3048 if (fatal_signal_pending(current))
3055 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3056 gfp_t gfp_mask, nodemask_t *nodemask)
3058 unsigned long nr_reclaimed;
3059 struct scan_control sc = {
3060 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3061 .gfp_mask = current_gfp_context(gfp_mask),
3062 .reclaim_idx = gfp_zone(gfp_mask),
3064 .nodemask = nodemask,
3065 .priority = DEF_PRIORITY,
3066 .may_writepage = !laptop_mode,
3072 * Do not enter reclaim if fatal signal was delivered while throttled.
3073 * 1 is returned so that the page allocator does not OOM kill at this
3076 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3079 trace_mm_vmscan_direct_reclaim_begin(order,
3084 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3086 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3088 return nr_reclaimed;
3093 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3094 gfp_t gfp_mask, bool noswap,
3096 unsigned long *nr_scanned)
3098 struct scan_control sc = {
3099 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3100 .target_mem_cgroup = memcg,
3101 .may_writepage = !laptop_mode,
3103 .reclaim_idx = MAX_NR_ZONES - 1,
3104 .may_swap = !noswap,
3106 unsigned long lru_pages;
3108 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3109 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3111 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3117 * NOTE: Although we can get the priority field, using it
3118 * here is not a good idea, since it limits the pages we can scan.
3119 * if we don't reclaim here, the shrink_node from balance_pgdat
3120 * will pick up pages from other mem cgroup's as well. We hack
3121 * the priority and make it zero.
3123 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3125 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3127 *nr_scanned = sc.nr_scanned;
3128 return sc.nr_reclaimed;
3131 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3132 unsigned long nr_pages,
3136 struct zonelist *zonelist;
3137 unsigned long nr_reclaimed;
3139 unsigned int noreclaim_flag;
3140 struct scan_control sc = {
3141 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3142 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3143 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3144 .reclaim_idx = MAX_NR_ZONES - 1,
3145 .target_mem_cgroup = memcg,
3146 .priority = DEF_PRIORITY,
3147 .may_writepage = !laptop_mode,
3149 .may_swap = may_swap,
3153 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3154 * take care of from where we get pages. So the node where we start the
3155 * scan does not need to be the current node.
3157 nid = mem_cgroup_select_victim_node(memcg);
3159 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3161 trace_mm_vmscan_memcg_reclaim_begin(0,
3166 noreclaim_flag = memalloc_noreclaim_save();
3167 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3168 memalloc_noreclaim_restore(noreclaim_flag);
3170 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3172 return nr_reclaimed;
3176 static void age_active_anon(struct pglist_data *pgdat,
3177 struct scan_control *sc)
3179 struct mem_cgroup *memcg;
3181 if (!total_swap_pages)
3184 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3186 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3188 if (inactive_list_is_low(lruvec, false, sc, true))
3189 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3190 sc, LRU_ACTIVE_ANON);
3192 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3197 * Returns true if there is an eligible zone balanced for the request order
3200 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3203 unsigned long mark = -1;
3206 for (i = 0; i <= classzone_idx; i++) {
3207 zone = pgdat->node_zones + i;
3209 if (!managed_zone(zone))
3212 mark = high_wmark_pages(zone);
3213 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3218 * If a node has no populated zone within classzone_idx, it does not
3219 * need balancing by definition. This can happen if a zone-restricted
3220 * allocation tries to wake a remote kswapd.
3228 /* Clear pgdat state for congested, dirty or under writeback. */
3229 static void clear_pgdat_congested(pg_data_t *pgdat)
3231 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3232 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3233 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3237 * Prepare kswapd for sleeping. This verifies that there are no processes
3238 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3240 * Returns true if kswapd is ready to sleep
3242 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3245 * The throttled processes are normally woken up in balance_pgdat() as
3246 * soon as allow_direct_reclaim() is true. But there is a potential
3247 * race between when kswapd checks the watermarks and a process gets
3248 * throttled. There is also a potential race if processes get
3249 * throttled, kswapd wakes, a large process exits thereby balancing the
3250 * zones, which causes kswapd to exit balance_pgdat() before reaching
3251 * the wake up checks. If kswapd is going to sleep, no process should
3252 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3253 * the wake up is premature, processes will wake kswapd and get
3254 * throttled again. The difference from wake ups in balance_pgdat() is
3255 * that here we are under prepare_to_wait().
3257 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3258 wake_up_all(&pgdat->pfmemalloc_wait);
3260 /* Hopeless node, leave it to direct reclaim */
3261 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3264 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3265 clear_pgdat_congested(pgdat);
3273 * kswapd shrinks a node of pages that are at or below the highest usable
3274 * zone that is currently unbalanced.
3276 * Returns true if kswapd scanned at least the requested number of pages to
3277 * reclaim or if the lack of progress was due to pages under writeback.
3278 * This is used to determine if the scanning priority needs to be raised.
3280 static bool kswapd_shrink_node(pg_data_t *pgdat,
3281 struct scan_control *sc)
3286 /* Reclaim a number of pages proportional to the number of zones */
3287 sc->nr_to_reclaim = 0;
3288 for (z = 0; z <= sc->reclaim_idx; z++) {
3289 zone = pgdat->node_zones + z;
3290 if (!managed_zone(zone))
3293 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3297 * Historically care was taken to put equal pressure on all zones but
3298 * now pressure is applied based on node LRU order.
3300 shrink_node(pgdat, sc);
3303 * Fragmentation may mean that the system cannot be rebalanced for
3304 * high-order allocations. If twice the allocation size has been
3305 * reclaimed then recheck watermarks only at order-0 to prevent
3306 * excessive reclaim. Assume that a process requested a high-order
3307 * can direct reclaim/compact.
3309 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3312 return sc->nr_scanned >= sc->nr_to_reclaim;
3316 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3317 * that are eligible for use by the caller until at least one zone is
3320 * Returns the order kswapd finished reclaiming at.
3322 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3323 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3324 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3325 * or lower is eligible for reclaim until at least one usable zone is
3328 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3331 unsigned long nr_soft_reclaimed;
3332 unsigned long nr_soft_scanned;
3334 struct scan_control sc = {
3335 .gfp_mask = GFP_KERNEL,
3337 .priority = DEF_PRIORITY,
3338 .may_writepage = !laptop_mode,
3342 count_vm_event(PAGEOUTRUN);
3345 unsigned long nr_reclaimed = sc.nr_reclaimed;
3346 bool raise_priority = true;
3348 sc.reclaim_idx = classzone_idx;
3351 * If the number of buffer_heads exceeds the maximum allowed
3352 * then consider reclaiming from all zones. This has a dual
3353 * purpose -- on 64-bit systems it is expected that
3354 * buffer_heads are stripped during active rotation. On 32-bit
3355 * systems, highmem pages can pin lowmem memory and shrinking
3356 * buffers can relieve lowmem pressure. Reclaim may still not
3357 * go ahead if all eligible zones for the original allocation
3358 * request are balanced to avoid excessive reclaim from kswapd.
3360 if (buffer_heads_over_limit) {
3361 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3362 zone = pgdat->node_zones + i;
3363 if (!managed_zone(zone))
3372 * Only reclaim if there are no eligible zones. Note that
3373 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3376 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3380 * Do some background aging of the anon list, to give
3381 * pages a chance to be referenced before reclaiming. All
3382 * pages are rotated regardless of classzone as this is
3383 * about consistent aging.
3385 age_active_anon(pgdat, &sc);
3388 * If we're getting trouble reclaiming, start doing writepage
3389 * even in laptop mode.
3391 if (sc.priority < DEF_PRIORITY - 2)
3392 sc.may_writepage = 1;
3394 /* Call soft limit reclaim before calling shrink_node. */
3396 nr_soft_scanned = 0;
3397 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3398 sc.gfp_mask, &nr_soft_scanned);
3399 sc.nr_reclaimed += nr_soft_reclaimed;
3402 * There should be no need to raise the scanning priority if
3403 * enough pages are already being scanned that that high
3404 * watermark would be met at 100% efficiency.
3406 if (kswapd_shrink_node(pgdat, &sc))
3407 raise_priority = false;
3410 * If the low watermark is met there is no need for processes
3411 * to be throttled on pfmemalloc_wait as they should not be
3412 * able to safely make forward progress. Wake them
3414 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3415 allow_direct_reclaim(pgdat))
3416 wake_up_all(&pgdat->pfmemalloc_wait);
3418 /* Check if kswapd should be suspending */
3419 if (try_to_freeze() || kthread_should_stop())
3423 * Raise priority if scanning rate is too low or there was no
3424 * progress in reclaiming pages
3426 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3427 if (raise_priority || !nr_reclaimed)
3429 } while (sc.priority >= 1);
3431 if (!sc.nr_reclaimed)
3432 pgdat->kswapd_failures++;
3435 snapshot_refaults(NULL, pgdat);
3437 * Return the order kswapd stopped reclaiming at as
3438 * prepare_kswapd_sleep() takes it into account. If another caller
3439 * entered the allocator slow path while kswapd was awake, order will
3440 * remain at the higher level.
3446 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3447 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3448 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3449 * after previous reclaim attempt (node is still unbalanced). In that case
3450 * return the zone index of the previous kswapd reclaim cycle.
3452 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3453 enum zone_type prev_classzone_idx)
3455 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_classzone_idx);
3457 return curr_idx == MAX_NR_ZONES ? prev_classzone_idx : curr_idx;
3460 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3461 unsigned int classzone_idx)
3466 if (freezing(current) || kthread_should_stop())
3469 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3472 * Try to sleep for a short interval. Note that kcompactd will only be
3473 * woken if it is possible to sleep for a short interval. This is
3474 * deliberate on the assumption that if reclaim cannot keep an
3475 * eligible zone balanced that it's also unlikely that compaction will
3478 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3480 * Compaction records what page blocks it recently failed to
3481 * isolate pages from and skips them in the future scanning.
3482 * When kswapd is going to sleep, it is reasonable to assume
3483 * that pages and compaction may succeed so reset the cache.
3485 reset_isolation_suitable(pgdat);
3488 * We have freed the memory, now we should compact it to make
3489 * allocation of the requested order possible.
3491 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3493 remaining = schedule_timeout(HZ/10);
3496 * If woken prematurely then reset kswapd_classzone_idx and
3497 * order. The values will either be from a wakeup request or
3498 * the previous request that slept prematurely.
3501 WRITE_ONCE(pgdat->kswapd_classzone_idx,
3502 kswapd_classzone_idx(pgdat, classzone_idx));
3504 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3505 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3508 finish_wait(&pgdat->kswapd_wait, &wait);
3509 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3513 * After a short sleep, check if it was a premature sleep. If not, then
3514 * go fully to sleep until explicitly woken up.
3517 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3518 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3521 * vmstat counters are not perfectly accurate and the estimated
3522 * value for counters such as NR_FREE_PAGES can deviate from the
3523 * true value by nr_online_cpus * threshold. To avoid the zone
3524 * watermarks being breached while under pressure, we reduce the
3525 * per-cpu vmstat threshold while kswapd is awake and restore
3526 * them before going back to sleep.
3528 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3530 if (!kthread_should_stop())
3533 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3536 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3538 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3540 finish_wait(&pgdat->kswapd_wait, &wait);
3544 * The background pageout daemon, started as a kernel thread
3545 * from the init process.
3547 * This basically trickles out pages so that we have _some_
3548 * free memory available even if there is no other activity
3549 * that frees anything up. This is needed for things like routing
3550 * etc, where we otherwise might have all activity going on in
3551 * asynchronous contexts that cannot page things out.
3553 * If there are applications that are active memory-allocators
3554 * (most normal use), this basically shouldn't matter.
3556 static int kswapd(void *p)
3558 unsigned int alloc_order, reclaim_order;
3559 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3560 pg_data_t *pgdat = (pg_data_t*)p;
3561 struct task_struct *tsk = current;
3563 struct reclaim_state reclaim_state = {
3564 .reclaimed_slab = 0,
3566 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3568 if (!cpumask_empty(cpumask))
3569 set_cpus_allowed_ptr(tsk, cpumask);
3570 current->reclaim_state = &reclaim_state;
3573 * Tell the memory management that we're a "memory allocator",
3574 * and that if we need more memory we should get access to it
3575 * regardless (see "__alloc_pages()"). "kswapd" should
3576 * never get caught in the normal page freeing logic.
3578 * (Kswapd normally doesn't need memory anyway, but sometimes
3579 * you need a small amount of memory in order to be able to
3580 * page out something else, and this flag essentially protects
3581 * us from recursively trying to free more memory as we're
3582 * trying to free the first piece of memory in the first place).
3584 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3587 WRITE_ONCE(pgdat->kswapd_order, 0);
3588 WRITE_ONCE(pgdat->kswapd_classzone_idx, MAX_NR_ZONES);
3592 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3593 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3596 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3599 /* Read the new order and classzone_idx */
3600 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3601 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3602 WRITE_ONCE(pgdat->kswapd_order, 0);
3603 WRITE_ONCE(pgdat->kswapd_classzone_idx, MAX_NR_ZONES);
3605 ret = try_to_freeze();
3606 if (kthread_should_stop())
3610 * We can speed up thawing tasks if we don't call balance_pgdat
3611 * after returning from the refrigerator
3617 * Reclaim begins at the requested order but if a high-order
3618 * reclaim fails then kswapd falls back to reclaiming for
3619 * order-0. If that happens, kswapd will consider sleeping
3620 * for the order it finished reclaiming at (reclaim_order)
3621 * but kcompactd is woken to compact for the original
3622 * request (alloc_order).
3624 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3626 fs_reclaim_acquire(GFP_KERNEL);
3627 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3628 fs_reclaim_release(GFP_KERNEL);
3629 if (reclaim_order < alloc_order)
3630 goto kswapd_try_sleep;
3633 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3634 current->reclaim_state = NULL;
3640 * A zone is low on free memory, so wake its kswapd task to service it.
3642 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3645 enum zone_type curr_idx;
3647 if (!managed_zone(zone))
3650 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3653 pgdat = zone->zone_pgdat;
3654 curr_idx = READ_ONCE(pgdat->kswapd_classzone_idx);
3656 if (curr_idx == MAX_NR_ZONES || curr_idx < classzone_idx)
3657 WRITE_ONCE(pgdat->kswapd_classzone_idx, classzone_idx);
3659 if (READ_ONCE(pgdat->kswapd_order) < order)
3660 WRITE_ONCE(pgdat->kswapd_order, order);
3662 if (!waitqueue_active(&pgdat->kswapd_wait))
3665 /* Hopeless node, leave it to direct reclaim */
3666 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3669 if (pgdat_balanced(pgdat, order, classzone_idx))
3672 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
3673 wake_up_interruptible(&pgdat->kswapd_wait);
3676 #ifdef CONFIG_HIBERNATION
3678 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3681 * Rather than trying to age LRUs the aim is to preserve the overall
3682 * LRU order by reclaiming preferentially
3683 * inactive > active > active referenced > active mapped
3685 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3687 struct reclaim_state reclaim_state;
3688 struct scan_control sc = {
3689 .nr_to_reclaim = nr_to_reclaim,
3690 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3691 .reclaim_idx = MAX_NR_ZONES - 1,
3692 .priority = DEF_PRIORITY,
3696 .hibernation_mode = 1,
3698 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3699 struct task_struct *p = current;
3700 unsigned long nr_reclaimed;
3701 unsigned int noreclaim_flag;
3703 noreclaim_flag = memalloc_noreclaim_save();
3704 fs_reclaim_acquire(sc.gfp_mask);
3705 reclaim_state.reclaimed_slab = 0;
3706 p->reclaim_state = &reclaim_state;
3708 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3710 p->reclaim_state = NULL;
3711 fs_reclaim_release(sc.gfp_mask);
3712 memalloc_noreclaim_restore(noreclaim_flag);
3714 return nr_reclaimed;
3716 #endif /* CONFIG_HIBERNATION */
3718 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3719 not required for correctness. So if the last cpu in a node goes
3720 away, we get changed to run anywhere: as the first one comes back,
3721 restore their cpu bindings. */
3722 static int kswapd_cpu_online(unsigned int cpu)
3726 for_each_node_state(nid, N_MEMORY) {
3727 pg_data_t *pgdat = NODE_DATA(nid);
3728 const struct cpumask *mask;
3730 mask = cpumask_of_node(pgdat->node_id);
3732 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3733 /* One of our CPUs online: restore mask */
3734 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3740 * This kswapd start function will be called by init and node-hot-add.
3741 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3743 int kswapd_run(int nid)
3745 pg_data_t *pgdat = NODE_DATA(nid);
3751 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3752 if (IS_ERR(pgdat->kswapd)) {
3753 /* failure at boot is fatal */
3754 BUG_ON(system_state < SYSTEM_RUNNING);
3755 pr_err("Failed to start kswapd on node %d\n", nid);
3756 ret = PTR_ERR(pgdat->kswapd);
3757 pgdat->kswapd = NULL;
3763 * Called by memory hotplug when all memory in a node is offlined. Caller must
3764 * hold mem_hotplug_begin/end().
3766 void kswapd_stop(int nid)
3768 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3771 kthread_stop(kswapd);
3772 NODE_DATA(nid)->kswapd = NULL;
3776 static int __init kswapd_init(void)
3781 for_each_node_state(nid, N_MEMORY)
3783 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3784 "mm/vmscan:online", kswapd_cpu_online,
3790 module_init(kswapd_init)
3796 * If non-zero call node_reclaim when the number of free pages falls below
3799 int node_reclaim_mode __read_mostly;
3801 #define RECLAIM_OFF 0
3802 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3803 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3804 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3807 * Priority for NODE_RECLAIM. This determines the fraction of pages
3808 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3811 #define NODE_RECLAIM_PRIORITY 4
3814 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3817 int sysctl_min_unmapped_ratio = 1;
3820 * If the number of slab pages in a zone grows beyond this percentage then
3821 * slab reclaim needs to occur.
3823 int sysctl_min_slab_ratio = 5;
3825 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3827 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3828 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3829 node_page_state(pgdat, NR_ACTIVE_FILE);
3832 * It's possible for there to be more file mapped pages than
3833 * accounted for by the pages on the file LRU lists because
3834 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3836 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3839 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3840 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3842 unsigned long nr_pagecache_reclaimable;
3843 unsigned long delta = 0;
3846 * If RECLAIM_UNMAP is set, then all file pages are considered
3847 * potentially reclaimable. Otherwise, we have to worry about
3848 * pages like swapcache and node_unmapped_file_pages() provides
3851 if (node_reclaim_mode & RECLAIM_UNMAP)
3852 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3854 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3856 /* If we can't clean pages, remove dirty pages from consideration */
3857 if (!(node_reclaim_mode & RECLAIM_WRITE))
3858 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3860 /* Watch for any possible underflows due to delta */
3861 if (unlikely(delta > nr_pagecache_reclaimable))
3862 delta = nr_pagecache_reclaimable;
3864 return nr_pagecache_reclaimable - delta;
3868 * Try to free up some pages from this node through reclaim.
3870 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3872 /* Minimum pages needed in order to stay on node */
3873 const unsigned long nr_pages = 1 << order;
3874 struct task_struct *p = current;
3875 struct reclaim_state reclaim_state;
3876 unsigned int noreclaim_flag;
3877 struct scan_control sc = {
3878 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3879 .gfp_mask = current_gfp_context(gfp_mask),
3881 .priority = NODE_RECLAIM_PRIORITY,
3882 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3883 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3885 .reclaim_idx = gfp_zone(gfp_mask),
3890 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3891 * and we also need to be able to write out pages for RECLAIM_WRITE
3892 * and RECLAIM_UNMAP.
3894 noreclaim_flag = memalloc_noreclaim_save();
3895 p->flags |= PF_SWAPWRITE;
3896 fs_reclaim_acquire(sc.gfp_mask);
3897 reclaim_state.reclaimed_slab = 0;
3898 p->reclaim_state = &reclaim_state;
3900 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3902 * Free memory by calling shrink zone with increasing
3903 * priorities until we have enough memory freed.
3906 shrink_node(pgdat, &sc);
3907 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3910 p->reclaim_state = NULL;
3911 fs_reclaim_release(gfp_mask);
3912 current->flags &= ~PF_SWAPWRITE;
3913 memalloc_noreclaim_restore(noreclaim_flag);
3914 return sc.nr_reclaimed >= nr_pages;
3917 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3922 * Node reclaim reclaims unmapped file backed pages and
3923 * slab pages if we are over the defined limits.
3925 * A small portion of unmapped file backed pages is needed for
3926 * file I/O otherwise pages read by file I/O will be immediately
3927 * thrown out if the node is overallocated. So we do not reclaim
3928 * if less than a specified percentage of the node is used by
3929 * unmapped file backed pages.
3931 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3932 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3933 return NODE_RECLAIM_FULL;
3936 * Do not scan if the allocation should not be delayed.
3938 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3939 return NODE_RECLAIM_NOSCAN;
3942 * Only run node reclaim on the local node or on nodes that do not
3943 * have associated processors. This will favor the local processor
3944 * over remote processors and spread off node memory allocations
3945 * as wide as possible.
3947 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3948 return NODE_RECLAIM_NOSCAN;
3950 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3951 return NODE_RECLAIM_NOSCAN;
3953 ret = __node_reclaim(pgdat, gfp_mask, order);
3954 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3957 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3964 * page_evictable - test whether a page is evictable
3965 * @page: the page to test
3967 * Test whether page is evictable--i.e., should be placed on active/inactive
3968 * lists vs unevictable list.
3970 * Reasons page might not be evictable:
3971 * (1) page's mapping marked unevictable
3972 * (2) page is part of an mlocked VMA
3975 int page_evictable(struct page *page)
3979 /* Prevent address_space of inode and swap cache from being freed */
3981 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3988 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3989 * @pages: array of pages to check
3990 * @nr_pages: number of pages to check
3992 * Checks pages for evictability and moves them to the appropriate lru list.
3994 * This function is only used for SysV IPC SHM_UNLOCK.
3996 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3998 struct lruvec *lruvec;
3999 struct pglist_data *pgdat = NULL;
4004 for (i = 0; i < nr_pages; i++) {
4005 struct page *page = pages[i];
4006 struct pglist_data *pagepgdat = page_pgdat(page);
4009 if (pagepgdat != pgdat) {
4011 spin_unlock_irq(&pgdat->lru_lock);
4013 spin_lock_irq(&pgdat->lru_lock);
4015 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4017 if (!PageLRU(page) || !PageUnevictable(page))
4020 if (page_evictable(page)) {
4021 enum lru_list lru = page_lru_base_type(page);
4023 VM_BUG_ON_PAGE(PageActive(page), page);
4024 ClearPageUnevictable(page);
4025 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4026 add_page_to_lru_list(page, lruvec, lru);
4032 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4033 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4034 spin_unlock_irq(&pgdat->lru_lock);
4037 #endif /* CONFIG_SHMEM */