4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup *target_mem_cgroup;
84 /* Scan (total_size >> priority) pages at once */
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
90 unsigned int may_writepage:1;
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
101 unsigned int hibernation_mode:1;
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
113 #ifdef ARCH_HAS_PREFETCH
114 #define prefetch_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
124 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
127 #ifdef ARCH_HAS_PREFETCHW
128 #define prefetchw_prev_lru_page(_page, _base, _field) \
130 if ((_page)->lru.prev != _base) { \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
138 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
142 * From 0 .. 100. Higher means more swappy.
144 int vm_swappiness = 60;
146 * The total number of pages which are beyond the high watermark within all
149 unsigned long vm_total_pages;
151 static LIST_HEAD(shrinker_list);
152 static DECLARE_RWSEM(shrinker_rwsem);
155 static bool global_reclaim(struct scan_control *sc)
157 return !sc->target_mem_cgroup;
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
173 static bool sane_reclaim(struct scan_control *sc)
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
179 #ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
186 static bool global_reclaim(struct scan_control *sc)
191 static bool sane_reclaim(struct scan_control *sc)
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
202 unsigned long zone_reclaimable_pages(struct zone *zone)
206 nr = zone_page_state_snapshot(zone, NR_ZONE_LRU_FILE);
207 if (get_nr_swap_pages() > 0)
208 nr += zone_page_state_snapshot(zone, NR_ZONE_LRU_ANON);
213 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
217 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
218 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
219 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
221 if (get_nr_swap_pages() > 0)
222 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
223 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
224 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
229 bool pgdat_reclaimable(struct pglist_data *pgdat)
231 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
232 pgdat_reclaimable_pages(pgdat) * 6;
235 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
237 if (!mem_cgroup_disabled())
238 return mem_cgroup_get_lru_size(lruvec, lru);
240 return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
244 * Add a shrinker callback to be called from the vm.
246 int register_shrinker(struct shrinker *shrinker)
248 size_t size = sizeof(*shrinker->nr_deferred);
250 if (shrinker->flags & SHRINKER_NUMA_AWARE)
253 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
254 if (!shrinker->nr_deferred)
257 down_write(&shrinker_rwsem);
258 list_add_tail(&shrinker->list, &shrinker_list);
259 up_write(&shrinker_rwsem);
262 EXPORT_SYMBOL(register_shrinker);
267 void unregister_shrinker(struct shrinker *shrinker)
269 down_write(&shrinker_rwsem);
270 list_del(&shrinker->list);
271 up_write(&shrinker_rwsem);
272 kfree(shrinker->nr_deferred);
274 EXPORT_SYMBOL(unregister_shrinker);
276 #define SHRINK_BATCH 128
278 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
279 struct shrinker *shrinker,
280 unsigned long nr_scanned,
281 unsigned long nr_eligible)
283 unsigned long freed = 0;
284 unsigned long long delta;
289 int nid = shrinkctl->nid;
290 long batch_size = shrinker->batch ? shrinker->batch
293 freeable = shrinker->count_objects(shrinker, shrinkctl);
298 * copy the current shrinker scan count into a local variable
299 * and zero it so that other concurrent shrinker invocations
300 * don't also do this scanning work.
302 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
305 delta = (4 * nr_scanned) / shrinker->seeks;
307 do_div(delta, nr_eligible + 1);
309 if (total_scan < 0) {
310 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
311 shrinker->scan_objects, total_scan);
312 total_scan = freeable;
316 * We need to avoid excessive windup on filesystem shrinkers
317 * due to large numbers of GFP_NOFS allocations causing the
318 * shrinkers to return -1 all the time. This results in a large
319 * nr being built up so when a shrink that can do some work
320 * comes along it empties the entire cache due to nr >>>
321 * freeable. This is bad for sustaining a working set in
324 * Hence only allow the shrinker to scan the entire cache when
325 * a large delta change is calculated directly.
327 if (delta < freeable / 4)
328 total_scan = min(total_scan, freeable / 2);
331 * Avoid risking looping forever due to too large nr value:
332 * never try to free more than twice the estimate number of
335 if (total_scan > freeable * 2)
336 total_scan = freeable * 2;
338 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
339 nr_scanned, nr_eligible,
340 freeable, delta, total_scan);
343 * Normally, we should not scan less than batch_size objects in one
344 * pass to avoid too frequent shrinker calls, but if the slab has less
345 * than batch_size objects in total and we are really tight on memory,
346 * we will try to reclaim all available objects, otherwise we can end
347 * up failing allocations although there are plenty of reclaimable
348 * objects spread over several slabs with usage less than the
351 * We detect the "tight on memory" situations by looking at the total
352 * number of objects we want to scan (total_scan). If it is greater
353 * than the total number of objects on slab (freeable), we must be
354 * scanning at high prio and therefore should try to reclaim as much as
357 while (total_scan >= batch_size ||
358 total_scan >= freeable) {
360 unsigned long nr_to_scan = min(batch_size, total_scan);
362 shrinkctl->nr_to_scan = nr_to_scan;
363 ret = shrinker->scan_objects(shrinker, shrinkctl);
364 if (ret == SHRINK_STOP)
368 count_vm_events(SLABS_SCANNED, nr_to_scan);
369 total_scan -= nr_to_scan;
375 * move the unused scan count back into the shrinker in a
376 * manner that handles concurrent updates. If we exhausted the
377 * scan, there is no need to do an update.
380 new_nr = atomic_long_add_return(total_scan,
381 &shrinker->nr_deferred[nid]);
383 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
385 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
390 * shrink_slab - shrink slab caches
391 * @gfp_mask: allocation context
392 * @nid: node whose slab caches to target
393 * @memcg: memory cgroup whose slab caches to target
394 * @nr_scanned: pressure numerator
395 * @nr_eligible: pressure denominator
397 * Call the shrink functions to age shrinkable caches.
399 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
400 * unaware shrinkers will receive a node id of 0 instead.
402 * @memcg specifies the memory cgroup to target. If it is not NULL,
403 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
404 * objects from the memory cgroup specified. Otherwise, only unaware
405 * shrinkers are called.
407 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
408 * the available objects should be scanned. Page reclaim for example
409 * passes the number of pages scanned and the number of pages on the
410 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
411 * when it encountered mapped pages. The ratio is further biased by
412 * the ->seeks setting of the shrink function, which indicates the
413 * cost to recreate an object relative to that of an LRU page.
415 * Returns the number of reclaimed slab objects.
417 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
418 struct mem_cgroup *memcg,
419 unsigned long nr_scanned,
420 unsigned long nr_eligible)
422 struct shrinker *shrinker;
423 unsigned long freed = 0;
425 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
429 nr_scanned = SWAP_CLUSTER_MAX;
431 if (!down_read_trylock(&shrinker_rwsem)) {
433 * If we would return 0, our callers would understand that we
434 * have nothing else to shrink and give up trying. By returning
435 * 1 we keep it going and assume we'll be able to shrink next
442 list_for_each_entry(shrinker, &shrinker_list, list) {
443 struct shrink_control sc = {
444 .gfp_mask = gfp_mask,
450 * If kernel memory accounting is disabled, we ignore
451 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
452 * passing NULL for memcg.
454 if (memcg_kmem_enabled() &&
455 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
458 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
461 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
464 up_read(&shrinker_rwsem);
470 void drop_slab_node(int nid)
475 struct mem_cgroup *memcg = NULL;
479 freed += shrink_slab(GFP_KERNEL, nid, memcg,
481 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
482 } while (freed > 10);
489 for_each_online_node(nid)
493 static inline int is_page_cache_freeable(struct page *page)
496 * A freeable page cache page is referenced only by the caller
497 * that isolated the page, the page cache radix tree and
498 * optional buffer heads at page->private.
500 return page_count(page) - page_has_private(page) == 2;
503 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
505 if (current->flags & PF_SWAPWRITE)
507 if (!inode_write_congested(inode))
509 if (inode_to_bdi(inode) == current->backing_dev_info)
515 * We detected a synchronous write error writing a page out. Probably
516 * -ENOSPC. We need to propagate that into the address_space for a subsequent
517 * fsync(), msync() or close().
519 * The tricky part is that after writepage we cannot touch the mapping: nothing
520 * prevents it from being freed up. But we have a ref on the page and once
521 * that page is locked, the mapping is pinned.
523 * We're allowed to run sleeping lock_page() here because we know the caller has
526 static void handle_write_error(struct address_space *mapping,
527 struct page *page, int error)
530 if (page_mapping(page) == mapping)
531 mapping_set_error(mapping, error);
535 /* possible outcome of pageout() */
537 /* failed to write page out, page is locked */
539 /* move page to the active list, page is locked */
541 /* page has been sent to the disk successfully, page is unlocked */
543 /* page is clean and locked */
548 * pageout is called by shrink_page_list() for each dirty page.
549 * Calls ->writepage().
551 static pageout_t pageout(struct page *page, struct address_space *mapping,
552 struct scan_control *sc)
555 * If the page is dirty, only perform writeback if that write
556 * will be non-blocking. To prevent this allocation from being
557 * stalled by pagecache activity. But note that there may be
558 * stalls if we need to run get_block(). We could test
559 * PagePrivate for that.
561 * If this process is currently in __generic_file_write_iter() against
562 * this page's queue, we can perform writeback even if that
565 * If the page is swapcache, write it back even if that would
566 * block, for some throttling. This happens by accident, because
567 * swap_backing_dev_info is bust: it doesn't reflect the
568 * congestion state of the swapdevs. Easy to fix, if needed.
570 if (!is_page_cache_freeable(page))
574 * Some data journaling orphaned pages can have
575 * page->mapping == NULL while being dirty with clean buffers.
577 if (page_has_private(page)) {
578 if (try_to_free_buffers(page)) {
579 ClearPageDirty(page);
580 pr_info("%s: orphaned page\n", __func__);
586 if (mapping->a_ops->writepage == NULL)
587 return PAGE_ACTIVATE;
588 if (!may_write_to_inode(mapping->host, sc))
591 if (clear_page_dirty_for_io(page)) {
593 struct writeback_control wbc = {
594 .sync_mode = WB_SYNC_NONE,
595 .nr_to_write = SWAP_CLUSTER_MAX,
597 .range_end = LLONG_MAX,
601 SetPageReclaim(page);
602 res = mapping->a_ops->writepage(page, &wbc);
604 handle_write_error(mapping, page, res);
605 if (res == AOP_WRITEPAGE_ACTIVATE) {
606 ClearPageReclaim(page);
607 return PAGE_ACTIVATE;
610 if (!PageWriteback(page)) {
611 /* synchronous write or broken a_ops? */
612 ClearPageReclaim(page);
614 trace_mm_vmscan_writepage(page);
615 inc_zone_page_state(page, NR_VMSCAN_WRITE);
623 * Same as remove_mapping, but if the page is removed from the mapping, it
624 * gets returned with a refcount of 0.
626 static int __remove_mapping(struct address_space *mapping, struct page *page,
631 BUG_ON(!PageLocked(page));
632 BUG_ON(mapping != page_mapping(page));
634 spin_lock_irqsave(&mapping->tree_lock, flags);
636 * The non racy check for a busy page.
638 * Must be careful with the order of the tests. When someone has
639 * a ref to the page, it may be possible that they dirty it then
640 * drop the reference. So if PageDirty is tested before page_count
641 * here, then the following race may occur:
643 * get_user_pages(&page);
644 * [user mapping goes away]
646 * !PageDirty(page) [good]
647 * SetPageDirty(page);
649 * !page_count(page) [good, discard it]
651 * [oops, our write_to data is lost]
653 * Reversing the order of the tests ensures such a situation cannot
654 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
655 * load is not satisfied before that of page->_refcount.
657 * Note that if SetPageDirty is always performed via set_page_dirty,
658 * and thus under tree_lock, then this ordering is not required.
660 if (!page_ref_freeze(page, 2))
662 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
663 if (unlikely(PageDirty(page))) {
664 page_ref_unfreeze(page, 2);
668 if (PageSwapCache(page)) {
669 swp_entry_t swap = { .val = page_private(page) };
670 mem_cgroup_swapout(page, swap);
671 __delete_from_swap_cache(page);
672 spin_unlock_irqrestore(&mapping->tree_lock, flags);
673 swapcache_free(swap);
675 void (*freepage)(struct page *);
678 freepage = mapping->a_ops->freepage;
680 * Remember a shadow entry for reclaimed file cache in
681 * order to detect refaults, thus thrashing, later on.
683 * But don't store shadows in an address space that is
684 * already exiting. This is not just an optizimation,
685 * inode reclaim needs to empty out the radix tree or
686 * the nodes are lost. Don't plant shadows behind its
689 * We also don't store shadows for DAX mappings because the
690 * only page cache pages found in these are zero pages
691 * covering holes, and because we don't want to mix DAX
692 * exceptional entries and shadow exceptional entries in the
695 if (reclaimed && page_is_file_cache(page) &&
696 !mapping_exiting(mapping) && !dax_mapping(mapping))
697 shadow = workingset_eviction(mapping, page);
698 __delete_from_page_cache(page, shadow);
699 spin_unlock_irqrestore(&mapping->tree_lock, flags);
701 if (freepage != NULL)
708 spin_unlock_irqrestore(&mapping->tree_lock, flags);
713 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
714 * someone else has a ref on the page, abort and return 0. If it was
715 * successfully detached, return 1. Assumes the caller has a single ref on
718 int remove_mapping(struct address_space *mapping, struct page *page)
720 if (__remove_mapping(mapping, page, false)) {
722 * Unfreezing the refcount with 1 rather than 2 effectively
723 * drops the pagecache ref for us without requiring another
726 page_ref_unfreeze(page, 1);
733 * putback_lru_page - put previously isolated page onto appropriate LRU list
734 * @page: page to be put back to appropriate lru list
736 * Add previously isolated @page to appropriate LRU list.
737 * Page may still be unevictable for other reasons.
739 * lru_lock must not be held, interrupts must be enabled.
741 void putback_lru_page(struct page *page)
744 int was_unevictable = PageUnevictable(page);
746 VM_BUG_ON_PAGE(PageLRU(page), page);
749 ClearPageUnevictable(page);
751 if (page_evictable(page)) {
753 * For evictable pages, we can use the cache.
754 * In event of a race, worst case is we end up with an
755 * unevictable page on [in]active list.
756 * We know how to handle that.
758 is_unevictable = false;
762 * Put unevictable pages directly on zone's unevictable
765 is_unevictable = true;
766 add_page_to_unevictable_list(page);
768 * When racing with an mlock or AS_UNEVICTABLE clearing
769 * (page is unlocked) make sure that if the other thread
770 * does not observe our setting of PG_lru and fails
771 * isolation/check_move_unevictable_pages,
772 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
773 * the page back to the evictable list.
775 * The other side is TestClearPageMlocked() or shmem_lock().
781 * page's status can change while we move it among lru. If an evictable
782 * page is on unevictable list, it never be freed. To avoid that,
783 * check after we added it to the list, again.
785 if (is_unevictable && page_evictable(page)) {
786 if (!isolate_lru_page(page)) {
790 /* This means someone else dropped this page from LRU
791 * So, it will be freed or putback to LRU again. There is
792 * nothing to do here.
796 if (was_unevictable && !is_unevictable)
797 count_vm_event(UNEVICTABLE_PGRESCUED);
798 else if (!was_unevictable && is_unevictable)
799 count_vm_event(UNEVICTABLE_PGCULLED);
801 put_page(page); /* drop ref from isolate */
804 enum page_references {
806 PAGEREF_RECLAIM_CLEAN,
811 static enum page_references page_check_references(struct page *page,
812 struct scan_control *sc)
814 int referenced_ptes, referenced_page;
815 unsigned long vm_flags;
817 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
819 referenced_page = TestClearPageReferenced(page);
822 * Mlock lost the isolation race with us. Let try_to_unmap()
823 * move the page to the unevictable list.
825 if (vm_flags & VM_LOCKED)
826 return PAGEREF_RECLAIM;
828 if (referenced_ptes) {
829 if (PageSwapBacked(page))
830 return PAGEREF_ACTIVATE;
832 * All mapped pages start out with page table
833 * references from the instantiating fault, so we need
834 * to look twice if a mapped file page is used more
837 * Mark it and spare it for another trip around the
838 * inactive list. Another page table reference will
839 * lead to its activation.
841 * Note: the mark is set for activated pages as well
842 * so that recently deactivated but used pages are
845 SetPageReferenced(page);
847 if (referenced_page || referenced_ptes > 1)
848 return PAGEREF_ACTIVATE;
851 * Activate file-backed executable pages after first usage.
853 if (vm_flags & VM_EXEC)
854 return PAGEREF_ACTIVATE;
859 /* Reclaim if clean, defer dirty pages to writeback */
860 if (referenced_page && !PageSwapBacked(page))
861 return PAGEREF_RECLAIM_CLEAN;
863 return PAGEREF_RECLAIM;
866 /* Check if a page is dirty or under writeback */
867 static void page_check_dirty_writeback(struct page *page,
868 bool *dirty, bool *writeback)
870 struct address_space *mapping;
873 * Anonymous pages are not handled by flushers and must be written
874 * from reclaim context. Do not stall reclaim based on them
876 if (!page_is_file_cache(page)) {
882 /* By default assume that the page flags are accurate */
883 *dirty = PageDirty(page);
884 *writeback = PageWriteback(page);
886 /* Verify dirty/writeback state if the filesystem supports it */
887 if (!page_has_private(page))
890 mapping = page_mapping(page);
891 if (mapping && mapping->a_ops->is_dirty_writeback)
892 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
896 * shrink_page_list() returns the number of reclaimed pages
898 static unsigned long shrink_page_list(struct list_head *page_list,
899 struct pglist_data *pgdat,
900 struct scan_control *sc,
901 enum ttu_flags ttu_flags,
902 unsigned long *ret_nr_dirty,
903 unsigned long *ret_nr_unqueued_dirty,
904 unsigned long *ret_nr_congested,
905 unsigned long *ret_nr_writeback,
906 unsigned long *ret_nr_immediate,
909 LIST_HEAD(ret_pages);
910 LIST_HEAD(free_pages);
912 unsigned long nr_unqueued_dirty = 0;
913 unsigned long nr_dirty = 0;
914 unsigned long nr_congested = 0;
915 unsigned long nr_reclaimed = 0;
916 unsigned long nr_writeback = 0;
917 unsigned long nr_immediate = 0;
921 while (!list_empty(page_list)) {
922 struct address_space *mapping;
925 enum page_references references = PAGEREF_RECLAIM_CLEAN;
926 bool dirty, writeback;
927 bool lazyfree = false;
928 int ret = SWAP_SUCCESS;
932 page = lru_to_page(page_list);
933 list_del(&page->lru);
935 if (!trylock_page(page))
938 VM_BUG_ON_PAGE(PageActive(page), page);
942 if (unlikely(!page_evictable(page)))
945 if (!sc->may_unmap && page_mapped(page))
948 /* Double the slab pressure for mapped and swapcache pages */
949 if (page_mapped(page) || PageSwapCache(page))
952 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
953 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
956 * The number of dirty pages determines if a zone is marked
957 * reclaim_congested which affects wait_iff_congested. kswapd
958 * will stall and start writing pages if the tail of the LRU
959 * is all dirty unqueued pages.
961 page_check_dirty_writeback(page, &dirty, &writeback);
962 if (dirty || writeback)
965 if (dirty && !writeback)
969 * Treat this page as congested if the underlying BDI is or if
970 * pages are cycling through the LRU so quickly that the
971 * pages marked for immediate reclaim are making it to the
972 * end of the LRU a second time.
974 mapping = page_mapping(page);
975 if (((dirty || writeback) && mapping &&
976 inode_write_congested(mapping->host)) ||
977 (writeback && PageReclaim(page)))
981 * If a page at the tail of the LRU is under writeback, there
982 * are three cases to consider.
984 * 1) If reclaim is encountering an excessive number of pages
985 * under writeback and this page is both under writeback and
986 * PageReclaim then it indicates that pages are being queued
987 * for IO but are being recycled through the LRU before the
988 * IO can complete. Waiting on the page itself risks an
989 * indefinite stall if it is impossible to writeback the
990 * page due to IO error or disconnected storage so instead
991 * note that the LRU is being scanned too quickly and the
992 * caller can stall after page list has been processed.
994 * 2) Global or new memcg reclaim encounters a page that is
995 * not marked for immediate reclaim, or the caller does not
996 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
997 * not to fs). In this case mark the page for immediate
998 * reclaim and continue scanning.
1000 * Require may_enter_fs because we would wait on fs, which
1001 * may not have submitted IO yet. And the loop driver might
1002 * enter reclaim, and deadlock if it waits on a page for
1003 * which it is needed to do the write (loop masks off
1004 * __GFP_IO|__GFP_FS for this reason); but more thought
1005 * would probably show more reasons.
1007 * 3) Legacy memcg encounters a page that is already marked
1008 * PageReclaim. memcg does not have any dirty pages
1009 * throttling so we could easily OOM just because too many
1010 * pages are in writeback and there is nothing else to
1011 * reclaim. Wait for the writeback to complete.
1013 if (PageWriteback(page)) {
1015 if (current_is_kswapd() &&
1016 PageReclaim(page) &&
1017 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1022 } else if (sane_reclaim(sc) ||
1023 !PageReclaim(page) || !may_enter_fs) {
1025 * This is slightly racy - end_page_writeback()
1026 * might have just cleared PageReclaim, then
1027 * setting PageReclaim here end up interpreted
1028 * as PageReadahead - but that does not matter
1029 * enough to care. What we do want is for this
1030 * page to have PageReclaim set next time memcg
1031 * reclaim reaches the tests above, so it will
1032 * then wait_on_page_writeback() to avoid OOM;
1033 * and it's also appropriate in global reclaim.
1035 SetPageReclaim(page);
1042 wait_on_page_writeback(page);
1043 /* then go back and try same page again */
1044 list_add_tail(&page->lru, page_list);
1050 references = page_check_references(page, sc);
1052 switch (references) {
1053 case PAGEREF_ACTIVATE:
1054 goto activate_locked;
1057 case PAGEREF_RECLAIM:
1058 case PAGEREF_RECLAIM_CLEAN:
1059 ; /* try to reclaim the page below */
1063 * Anonymous process memory has backing store?
1064 * Try to allocate it some swap space here.
1066 if (PageAnon(page) && !PageSwapCache(page)) {
1067 if (!(sc->gfp_mask & __GFP_IO))
1069 if (!add_to_swap(page, page_list))
1070 goto activate_locked;
1074 /* Adding to swap updated mapping */
1075 mapping = page_mapping(page);
1076 } else if (unlikely(PageTransHuge(page))) {
1077 /* Split file THP */
1078 if (split_huge_page_to_list(page, page_list))
1082 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1085 * The page is mapped into the page tables of one or more
1086 * processes. Try to unmap it here.
1088 if (page_mapped(page) && mapping) {
1089 switch (ret = try_to_unmap(page, lazyfree ?
1090 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1091 (ttu_flags | TTU_BATCH_FLUSH))) {
1093 goto activate_locked;
1101 ; /* try to free the page below */
1105 if (PageDirty(page)) {
1107 * Only kswapd can writeback filesystem pages to
1108 * avoid risk of stack overflow but only writeback
1109 * if many dirty pages have been encountered.
1111 if (page_is_file_cache(page) &&
1112 (!current_is_kswapd() ||
1113 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1115 * Immediately reclaim when written back.
1116 * Similar in principal to deactivate_page()
1117 * except we already have the page isolated
1118 * and know it's dirty
1120 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1121 SetPageReclaim(page);
1126 if (references == PAGEREF_RECLAIM_CLEAN)
1130 if (!sc->may_writepage)
1134 * Page is dirty. Flush the TLB if a writable entry
1135 * potentially exists to avoid CPU writes after IO
1136 * starts and then write it out here.
1138 try_to_unmap_flush_dirty();
1139 switch (pageout(page, mapping, sc)) {
1143 goto activate_locked;
1145 if (PageWriteback(page))
1147 if (PageDirty(page))
1151 * A synchronous write - probably a ramdisk. Go
1152 * ahead and try to reclaim the page.
1154 if (!trylock_page(page))
1156 if (PageDirty(page) || PageWriteback(page))
1158 mapping = page_mapping(page);
1160 ; /* try to free the page below */
1165 * If the page has buffers, try to free the buffer mappings
1166 * associated with this page. If we succeed we try to free
1169 * We do this even if the page is PageDirty().
1170 * try_to_release_page() does not perform I/O, but it is
1171 * possible for a page to have PageDirty set, but it is actually
1172 * clean (all its buffers are clean). This happens if the
1173 * buffers were written out directly, with submit_bh(). ext3
1174 * will do this, as well as the blockdev mapping.
1175 * try_to_release_page() will discover that cleanness and will
1176 * drop the buffers and mark the page clean - it can be freed.
1178 * Rarely, pages can have buffers and no ->mapping. These are
1179 * the pages which were not successfully invalidated in
1180 * truncate_complete_page(). We try to drop those buffers here
1181 * and if that worked, and the page is no longer mapped into
1182 * process address space (page_count == 1) it can be freed.
1183 * Otherwise, leave the page on the LRU so it is swappable.
1185 if (page_has_private(page)) {
1186 if (!try_to_release_page(page, sc->gfp_mask))
1187 goto activate_locked;
1188 if (!mapping && page_count(page) == 1) {
1190 if (put_page_testzero(page))
1194 * rare race with speculative reference.
1195 * the speculative reference will free
1196 * this page shortly, so we may
1197 * increment nr_reclaimed here (and
1198 * leave it off the LRU).
1207 if (!mapping || !__remove_mapping(mapping, page, true))
1211 * At this point, we have no other references and there is
1212 * no way to pick any more up (removed from LRU, removed
1213 * from pagecache). Can use non-atomic bitops now (and
1214 * we obviously don't have to worry about waking up a process
1215 * waiting on the page lock, because there are no references.
1217 __ClearPageLocked(page);
1219 if (ret == SWAP_LZFREE)
1220 count_vm_event(PGLAZYFREED);
1225 * Is there need to periodically free_page_list? It would
1226 * appear not as the counts should be low
1228 list_add(&page->lru, &free_pages);
1232 if (PageSwapCache(page))
1233 try_to_free_swap(page);
1235 list_add(&page->lru, &ret_pages);
1239 /* Not a candidate for swapping, so reclaim swap space. */
1240 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1241 try_to_free_swap(page);
1242 VM_BUG_ON_PAGE(PageActive(page), page);
1243 SetPageActive(page);
1248 list_add(&page->lru, &ret_pages);
1249 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1252 mem_cgroup_uncharge_list(&free_pages);
1253 try_to_unmap_flush();
1254 free_hot_cold_page_list(&free_pages, true);
1256 list_splice(&ret_pages, page_list);
1257 count_vm_events(PGACTIVATE, pgactivate);
1259 *ret_nr_dirty += nr_dirty;
1260 *ret_nr_congested += nr_congested;
1261 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1262 *ret_nr_writeback += nr_writeback;
1263 *ret_nr_immediate += nr_immediate;
1264 return nr_reclaimed;
1267 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1268 struct list_head *page_list)
1270 struct scan_control sc = {
1271 .gfp_mask = GFP_KERNEL,
1272 .priority = DEF_PRIORITY,
1275 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1276 struct page *page, *next;
1277 LIST_HEAD(clean_pages);
1279 list_for_each_entry_safe(page, next, page_list, lru) {
1280 if (page_is_file_cache(page) && !PageDirty(page) &&
1281 !__PageMovable(page)) {
1282 ClearPageActive(page);
1283 list_move(&page->lru, &clean_pages);
1287 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1288 TTU_UNMAP|TTU_IGNORE_ACCESS,
1289 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1290 list_splice(&clean_pages, page_list);
1291 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1296 * Attempt to remove the specified page from its LRU. Only take this page
1297 * if it is of the appropriate PageActive status. Pages which are being
1298 * freed elsewhere are also ignored.
1300 * page: page to consider
1301 * mode: one of the LRU isolation modes defined above
1303 * returns 0 on success, -ve errno on failure.
1305 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1309 /* Only take pages on the LRU. */
1313 /* Compaction should not handle unevictable pages but CMA can do so */
1314 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1320 * To minimise LRU disruption, the caller can indicate that it only
1321 * wants to isolate pages it will be able to operate on without
1322 * blocking - clean pages for the most part.
1324 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1325 * is used by reclaim when it is cannot write to backing storage
1327 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1328 * that it is possible to migrate without blocking
1330 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1331 /* All the caller can do on PageWriteback is block */
1332 if (PageWriteback(page))
1335 if (PageDirty(page)) {
1336 struct address_space *mapping;
1338 /* ISOLATE_CLEAN means only clean pages */
1339 if (mode & ISOLATE_CLEAN)
1343 * Only pages without mappings or that have a
1344 * ->migratepage callback are possible to migrate
1347 mapping = page_mapping(page);
1348 if (mapping && !mapping->a_ops->migratepage)
1353 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1356 if (likely(get_page_unless_zero(page))) {
1358 * Be careful not to clear PageLRU until after we're
1359 * sure the page is not being freed elsewhere -- the
1360 * page release code relies on it.
1370 * zone_lru_lock is heavily contended. Some of the functions that
1371 * shrink the lists perform better by taking out a batch of pages
1372 * and working on them outside the LRU lock.
1374 * For pagecache intensive workloads, this function is the hottest
1375 * spot in the kernel (apart from copy_*_user functions).
1377 * Appropriate locks must be held before calling this function.
1379 * @nr_to_scan: The number of pages to look through on the list.
1380 * @lruvec: The LRU vector to pull pages from.
1381 * @dst: The temp list to put pages on to.
1382 * @nr_scanned: The number of pages that were scanned.
1383 * @sc: The scan_control struct for this reclaim session
1384 * @mode: One of the LRU isolation modes
1385 * @lru: LRU list id for isolating
1387 * returns how many pages were moved onto *@dst.
1389 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1390 struct lruvec *lruvec, struct list_head *dst,
1391 unsigned long *nr_scanned, struct scan_control *sc,
1392 isolate_mode_t mode, enum lru_list lru)
1394 struct list_head *src = &lruvec->lists[lru];
1395 unsigned long nr_taken = 0;
1396 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1397 unsigned long scan, nr_pages;
1398 LIST_HEAD(pages_skipped);
1400 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1401 !list_empty(src); scan++) {
1404 page = lru_to_page(src);
1405 prefetchw_prev_lru_page(page, src, flags);
1407 VM_BUG_ON_PAGE(!PageLRU(page), page);
1409 if (page_zonenum(page) > sc->reclaim_idx) {
1410 list_move(&page->lru, &pages_skipped);
1414 switch (__isolate_lru_page(page, mode)) {
1416 nr_pages = hpage_nr_pages(page);
1417 nr_taken += nr_pages;
1418 nr_zone_taken[page_zonenum(page)] += nr_pages;
1419 list_move(&page->lru, dst);
1423 /* else it is being freed elsewhere */
1424 list_move(&page->lru, src);
1433 * Splice any skipped pages to the start of the LRU list. Note that
1434 * this disrupts the LRU order when reclaiming for lower zones but
1435 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1436 * scanning would soon rescan the same pages to skip and put the
1437 * system at risk of premature OOM.
1439 if (!list_empty(&pages_skipped))
1440 list_splice(&pages_skipped, src);
1442 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1443 nr_taken, mode, is_file_lru(lru));
1444 for (scan = 0; scan < MAX_NR_ZONES; scan++) {
1445 nr_pages = nr_zone_taken[scan];
1449 update_lru_size(lruvec, lru, scan, -nr_pages);
1455 * isolate_lru_page - tries to isolate a page from its LRU list
1456 * @page: page to isolate from its LRU list
1458 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1459 * vmstat statistic corresponding to whatever LRU list the page was on.
1461 * Returns 0 if the page was removed from an LRU list.
1462 * Returns -EBUSY if the page was not on an LRU list.
1464 * The returned page will have PageLRU() cleared. If it was found on
1465 * the active list, it will have PageActive set. If it was found on
1466 * the unevictable list, it will have the PageUnevictable bit set. That flag
1467 * may need to be cleared by the caller before letting the page go.
1469 * The vmstat statistic corresponding to the list on which the page was
1470 * found will be decremented.
1473 * (1) Must be called with an elevated refcount on the page. This is a
1474 * fundamentnal difference from isolate_lru_pages (which is called
1475 * without a stable reference).
1476 * (2) the lru_lock must not be held.
1477 * (3) interrupts must be enabled.
1479 int isolate_lru_page(struct page *page)
1483 VM_BUG_ON_PAGE(!page_count(page), page);
1484 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1486 if (PageLRU(page)) {
1487 struct zone *zone = page_zone(page);
1488 struct lruvec *lruvec;
1490 spin_lock_irq(zone_lru_lock(zone));
1491 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1492 if (PageLRU(page)) {
1493 int lru = page_lru(page);
1496 del_page_from_lru_list(page, lruvec, lru);
1499 spin_unlock_irq(zone_lru_lock(zone));
1505 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1506 * then get resheduled. When there are massive number of tasks doing page
1507 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1508 * the LRU list will go small and be scanned faster than necessary, leading to
1509 * unnecessary swapping, thrashing and OOM.
1511 static int too_many_isolated(struct pglist_data *pgdat, int file,
1512 struct scan_control *sc)
1514 unsigned long inactive, isolated;
1516 if (current_is_kswapd())
1519 if (!sane_reclaim(sc))
1523 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1524 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1526 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1527 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1531 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1532 * won't get blocked by normal direct-reclaimers, forming a circular
1535 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1538 return isolated > inactive;
1541 static noinline_for_stack void
1542 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1544 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1545 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1546 LIST_HEAD(pages_to_free);
1549 * Put back any unfreeable pages.
1551 while (!list_empty(page_list)) {
1552 struct page *page = lru_to_page(page_list);
1555 VM_BUG_ON_PAGE(PageLRU(page), page);
1556 list_del(&page->lru);
1557 if (unlikely(!page_evictable(page))) {
1558 spin_unlock_irq(&pgdat->lru_lock);
1559 putback_lru_page(page);
1560 spin_lock_irq(&pgdat->lru_lock);
1564 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1567 lru = page_lru(page);
1568 add_page_to_lru_list(page, lruvec, lru);
1570 if (is_active_lru(lru)) {
1571 int file = is_file_lru(lru);
1572 int numpages = hpage_nr_pages(page);
1573 reclaim_stat->recent_rotated[file] += numpages;
1575 if (put_page_testzero(page)) {
1576 __ClearPageLRU(page);
1577 __ClearPageActive(page);
1578 del_page_from_lru_list(page, lruvec, lru);
1580 if (unlikely(PageCompound(page))) {
1581 spin_unlock_irq(&pgdat->lru_lock);
1582 mem_cgroup_uncharge(page);
1583 (*get_compound_page_dtor(page))(page);
1584 spin_lock_irq(&pgdat->lru_lock);
1586 list_add(&page->lru, &pages_to_free);
1591 * To save our caller's stack, now use input list for pages to free.
1593 list_splice(&pages_to_free, page_list);
1597 * If a kernel thread (such as nfsd for loop-back mounts) services
1598 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1599 * In that case we should only throttle if the backing device it is
1600 * writing to is congested. In other cases it is safe to throttle.
1602 static int current_may_throttle(void)
1604 return !(current->flags & PF_LESS_THROTTLE) ||
1605 current->backing_dev_info == NULL ||
1606 bdi_write_congested(current->backing_dev_info);
1610 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1611 * of reclaimed pages
1613 static noinline_for_stack unsigned long
1614 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1615 struct scan_control *sc, enum lru_list lru)
1617 LIST_HEAD(page_list);
1618 unsigned long nr_scanned;
1619 unsigned long nr_reclaimed = 0;
1620 unsigned long nr_taken;
1621 unsigned long nr_dirty = 0;
1622 unsigned long nr_congested = 0;
1623 unsigned long nr_unqueued_dirty = 0;
1624 unsigned long nr_writeback = 0;
1625 unsigned long nr_immediate = 0;
1626 isolate_mode_t isolate_mode = 0;
1627 int file = is_file_lru(lru);
1628 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1629 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1631 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1632 congestion_wait(BLK_RW_ASYNC, HZ/10);
1634 /* We are about to die and free our memory. Return now. */
1635 if (fatal_signal_pending(current))
1636 return SWAP_CLUSTER_MAX;
1642 isolate_mode |= ISOLATE_UNMAPPED;
1643 if (!sc->may_writepage)
1644 isolate_mode |= ISOLATE_CLEAN;
1646 spin_lock_irq(&pgdat->lru_lock);
1648 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1649 &nr_scanned, sc, isolate_mode, lru);
1651 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1652 reclaim_stat->recent_scanned[file] += nr_taken;
1654 if (global_reclaim(sc)) {
1655 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1656 if (current_is_kswapd())
1657 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1659 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1661 spin_unlock_irq(&pgdat->lru_lock);
1666 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1667 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1668 &nr_writeback, &nr_immediate,
1671 spin_lock_irq(&pgdat->lru_lock);
1673 if (global_reclaim(sc)) {
1674 if (current_is_kswapd())
1675 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1677 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1680 putback_inactive_pages(lruvec, &page_list);
1682 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1684 spin_unlock_irq(&pgdat->lru_lock);
1686 mem_cgroup_uncharge_list(&page_list);
1687 free_hot_cold_page_list(&page_list, true);
1690 * If reclaim is isolating dirty pages under writeback, it implies
1691 * that the long-lived page allocation rate is exceeding the page
1692 * laundering rate. Either the global limits are not being effective
1693 * at throttling processes due to the page distribution throughout
1694 * zones or there is heavy usage of a slow backing device. The
1695 * only option is to throttle from reclaim context which is not ideal
1696 * as there is no guarantee the dirtying process is throttled in the
1697 * same way balance_dirty_pages() manages.
1699 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1700 * of pages under pages flagged for immediate reclaim and stall if any
1701 * are encountered in the nr_immediate check below.
1703 if (nr_writeback && nr_writeback == nr_taken)
1704 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1707 * Legacy memcg will stall in page writeback so avoid forcibly
1710 if (sane_reclaim(sc)) {
1712 * Tag a zone as congested if all the dirty pages scanned were
1713 * backed by a congested BDI and wait_iff_congested will stall.
1715 if (nr_dirty && nr_dirty == nr_congested)
1716 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1719 * If dirty pages are scanned that are not queued for IO, it
1720 * implies that flushers are not keeping up. In this case, flag
1721 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1724 if (nr_unqueued_dirty == nr_taken)
1725 set_bit(PGDAT_DIRTY, &pgdat->flags);
1728 * If kswapd scans pages marked marked for immediate
1729 * reclaim and under writeback (nr_immediate), it implies
1730 * that pages are cycling through the LRU faster than
1731 * they are written so also forcibly stall.
1733 if (nr_immediate && current_may_throttle())
1734 congestion_wait(BLK_RW_ASYNC, HZ/10);
1738 * Stall direct reclaim for IO completions if underlying BDIs or zone
1739 * is congested. Allow kswapd to continue until it starts encountering
1740 * unqueued dirty pages or cycling through the LRU too quickly.
1742 if (!sc->hibernation_mode && !current_is_kswapd() &&
1743 current_may_throttle())
1744 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1746 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1747 nr_scanned, nr_reclaimed,
1748 sc->priority, file);
1749 return nr_reclaimed;
1753 * This moves pages from the active list to the inactive list.
1755 * We move them the other way if the page is referenced by one or more
1756 * processes, from rmap.
1758 * If the pages are mostly unmapped, the processing is fast and it is
1759 * appropriate to hold zone_lru_lock across the whole operation. But if
1760 * the pages are mapped, the processing is slow (page_referenced()) so we
1761 * should drop zone_lru_lock around each page. It's impossible to balance
1762 * this, so instead we remove the pages from the LRU while processing them.
1763 * It is safe to rely on PG_active against the non-LRU pages in here because
1764 * nobody will play with that bit on a non-LRU page.
1766 * The downside is that we have to touch page->_refcount against each page.
1767 * But we had to alter page->flags anyway.
1770 static void move_active_pages_to_lru(struct lruvec *lruvec,
1771 struct list_head *list,
1772 struct list_head *pages_to_free,
1775 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1776 unsigned long pgmoved = 0;
1780 while (!list_empty(list)) {
1781 page = lru_to_page(list);
1782 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1784 VM_BUG_ON_PAGE(PageLRU(page), page);
1787 nr_pages = hpage_nr_pages(page);
1788 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1789 list_move(&page->lru, &lruvec->lists[lru]);
1790 pgmoved += nr_pages;
1792 if (put_page_testzero(page)) {
1793 __ClearPageLRU(page);
1794 __ClearPageActive(page);
1795 del_page_from_lru_list(page, lruvec, lru);
1797 if (unlikely(PageCompound(page))) {
1798 spin_unlock_irq(&pgdat->lru_lock);
1799 mem_cgroup_uncharge(page);
1800 (*get_compound_page_dtor(page))(page);
1801 spin_lock_irq(&pgdat->lru_lock);
1803 list_add(&page->lru, pages_to_free);
1807 if (!is_active_lru(lru))
1808 __count_vm_events(PGDEACTIVATE, pgmoved);
1811 static void shrink_active_list(unsigned long nr_to_scan,
1812 struct lruvec *lruvec,
1813 struct scan_control *sc,
1816 unsigned long nr_taken;
1817 unsigned long nr_scanned;
1818 unsigned long vm_flags;
1819 LIST_HEAD(l_hold); /* The pages which were snipped off */
1820 LIST_HEAD(l_active);
1821 LIST_HEAD(l_inactive);
1823 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1824 unsigned long nr_rotated = 0;
1825 isolate_mode_t isolate_mode = 0;
1826 int file = is_file_lru(lru);
1827 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1832 isolate_mode |= ISOLATE_UNMAPPED;
1833 if (!sc->may_writepage)
1834 isolate_mode |= ISOLATE_CLEAN;
1836 spin_lock_irq(&pgdat->lru_lock);
1838 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1839 &nr_scanned, sc, isolate_mode, lru);
1841 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1842 reclaim_stat->recent_scanned[file] += nr_taken;
1844 if (global_reclaim(sc))
1845 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1846 __count_vm_events(PGREFILL, nr_scanned);
1848 spin_unlock_irq(&pgdat->lru_lock);
1850 while (!list_empty(&l_hold)) {
1852 page = lru_to_page(&l_hold);
1853 list_del(&page->lru);
1855 if (unlikely(!page_evictable(page))) {
1856 putback_lru_page(page);
1860 if (unlikely(buffer_heads_over_limit)) {
1861 if (page_has_private(page) && trylock_page(page)) {
1862 if (page_has_private(page))
1863 try_to_release_page(page, 0);
1868 if (page_referenced(page, 0, sc->target_mem_cgroup,
1870 nr_rotated += hpage_nr_pages(page);
1872 * Identify referenced, file-backed active pages and
1873 * give them one more trip around the active list. So
1874 * that executable code get better chances to stay in
1875 * memory under moderate memory pressure. Anon pages
1876 * are not likely to be evicted by use-once streaming
1877 * IO, plus JVM can create lots of anon VM_EXEC pages,
1878 * so we ignore them here.
1880 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1881 list_add(&page->lru, &l_active);
1886 ClearPageActive(page); /* we are de-activating */
1887 list_add(&page->lru, &l_inactive);
1891 * Move pages back to the lru list.
1893 spin_lock_irq(&pgdat->lru_lock);
1895 * Count referenced pages from currently used mappings as rotated,
1896 * even though only some of them are actually re-activated. This
1897 * helps balance scan pressure between file and anonymous pages in
1900 reclaim_stat->recent_rotated[file] += nr_rotated;
1902 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1903 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1904 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1905 spin_unlock_irq(&pgdat->lru_lock);
1907 mem_cgroup_uncharge_list(&l_hold);
1908 free_hot_cold_page_list(&l_hold, true);
1912 * The inactive anon list should be small enough that the VM never has
1913 * to do too much work.
1915 * The inactive file list should be small enough to leave most memory
1916 * to the established workingset on the scan-resistant active list,
1917 * but large enough to avoid thrashing the aggregate readahead window.
1919 * Both inactive lists should also be large enough that each inactive
1920 * page has a chance to be referenced again before it is reclaimed.
1922 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1923 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1924 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1927 * memory ratio inactive
1928 * -------------------------------------
1937 static bool inactive_list_is_low(struct lruvec *lruvec, bool file)
1939 unsigned long inactive_ratio;
1940 unsigned long inactive;
1941 unsigned long active;
1945 * If we don't have swap space, anonymous page deactivation
1948 if (!file && !total_swap_pages)
1951 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
1952 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
1954 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1956 inactive_ratio = int_sqrt(10 * gb);
1960 return inactive * inactive_ratio < active;
1963 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1964 struct lruvec *lruvec, struct scan_control *sc)
1966 if (is_active_lru(lru)) {
1967 if (inactive_list_is_low(lruvec, is_file_lru(lru)))
1968 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1972 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1983 * Determine how aggressively the anon and file LRU lists should be
1984 * scanned. The relative value of each set of LRU lists is determined
1985 * by looking at the fraction of the pages scanned we did rotate back
1986 * onto the active list instead of evict.
1988 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1989 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1991 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
1992 struct scan_control *sc, unsigned long *nr,
1993 unsigned long *lru_pages)
1995 int swappiness = mem_cgroup_swappiness(memcg);
1996 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1998 u64 denominator = 0; /* gcc */
1999 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2000 unsigned long anon_prio, file_prio;
2001 enum scan_balance scan_balance;
2002 unsigned long anon, file;
2003 bool force_scan = false;
2004 unsigned long ap, fp;
2010 * If the zone or memcg is small, nr[l] can be 0. This
2011 * results in no scanning on this priority and a potential
2012 * priority drop. Global direct reclaim can go to the next
2013 * zone and tends to have no problems. Global kswapd is for
2014 * zone balancing and it needs to scan a minimum amount. When
2015 * reclaiming for a memcg, a priority drop can cause high
2016 * latencies, so it's better to scan a minimum amount there as
2019 if (current_is_kswapd()) {
2020 if (!pgdat_reclaimable(pgdat))
2022 if (!mem_cgroup_online(memcg))
2025 if (!global_reclaim(sc))
2028 /* If we have no swap space, do not bother scanning anon pages. */
2029 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2030 scan_balance = SCAN_FILE;
2035 * Global reclaim will swap to prevent OOM even with no
2036 * swappiness, but memcg users want to use this knob to
2037 * disable swapping for individual groups completely when
2038 * using the memory controller's swap limit feature would be
2041 if (!global_reclaim(sc) && !swappiness) {
2042 scan_balance = SCAN_FILE;
2047 * Do not apply any pressure balancing cleverness when the
2048 * system is close to OOM, scan both anon and file equally
2049 * (unless the swappiness setting disagrees with swapping).
2051 if (!sc->priority && swappiness) {
2052 scan_balance = SCAN_EQUAL;
2057 * Prevent the reclaimer from falling into the cache trap: as
2058 * cache pages start out inactive, every cache fault will tip
2059 * the scan balance towards the file LRU. And as the file LRU
2060 * shrinks, so does the window for rotation from references.
2061 * This means we have a runaway feedback loop where a tiny
2062 * thrashing file LRU becomes infinitely more attractive than
2063 * anon pages. Try to detect this based on file LRU size.
2065 if (global_reclaim(sc)) {
2066 unsigned long pgdatfile;
2067 unsigned long pgdatfree;
2069 unsigned long total_high_wmark = 0;
2071 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2072 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2073 node_page_state(pgdat, NR_INACTIVE_FILE);
2075 for (z = 0; z < MAX_NR_ZONES; z++) {
2076 struct zone *zone = &pgdat->node_zones[z];
2077 if (!populated_zone(zone))
2080 total_high_wmark += high_wmark_pages(zone);
2083 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2084 scan_balance = SCAN_ANON;
2090 * If there is enough inactive page cache, i.e. if the size of the
2091 * inactive list is greater than that of the active list *and* the
2092 * inactive list actually has some pages to scan on this priority, we
2093 * do not reclaim anything from the anonymous working set right now.
2094 * Without the second condition we could end up never scanning an
2095 * lruvec even if it has plenty of old anonymous pages unless the
2096 * system is under heavy pressure.
2098 if (!inactive_list_is_low(lruvec, true) &&
2099 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2100 scan_balance = SCAN_FILE;
2104 scan_balance = SCAN_FRACT;
2107 * With swappiness at 100, anonymous and file have the same priority.
2108 * This scanning priority is essentially the inverse of IO cost.
2110 anon_prio = swappiness;
2111 file_prio = 200 - anon_prio;
2114 * OK, so we have swap space and a fair amount of page cache
2115 * pages. We use the recently rotated / recently scanned
2116 * ratios to determine how valuable each cache is.
2118 * Because workloads change over time (and to avoid overflow)
2119 * we keep these statistics as a floating average, which ends
2120 * up weighing recent references more than old ones.
2122 * anon in [0], file in [1]
2125 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2126 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2127 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2128 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2130 spin_lock_irq(&pgdat->lru_lock);
2131 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2132 reclaim_stat->recent_scanned[0] /= 2;
2133 reclaim_stat->recent_rotated[0] /= 2;
2136 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2137 reclaim_stat->recent_scanned[1] /= 2;
2138 reclaim_stat->recent_rotated[1] /= 2;
2142 * The amount of pressure on anon vs file pages is inversely
2143 * proportional to the fraction of recently scanned pages on
2144 * each list that were recently referenced and in active use.
2146 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2147 ap /= reclaim_stat->recent_rotated[0] + 1;
2149 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2150 fp /= reclaim_stat->recent_rotated[1] + 1;
2151 spin_unlock_irq(&pgdat->lru_lock);
2155 denominator = ap + fp + 1;
2157 some_scanned = false;
2158 /* Only use force_scan on second pass. */
2159 for (pass = 0; !some_scanned && pass < 2; pass++) {
2161 for_each_evictable_lru(lru) {
2162 int file = is_file_lru(lru);
2166 size = lruvec_lru_size(lruvec, lru);
2167 scan = size >> sc->priority;
2169 if (!scan && pass && force_scan)
2170 scan = min(size, SWAP_CLUSTER_MAX);
2172 switch (scan_balance) {
2174 /* Scan lists relative to size */
2178 * Scan types proportional to swappiness and
2179 * their relative recent reclaim efficiency.
2181 scan = div64_u64(scan * fraction[file],
2186 /* Scan one type exclusively */
2187 if ((scan_balance == SCAN_FILE) != file) {
2193 /* Look ma, no brain */
2201 * Skip the second pass and don't force_scan,
2202 * if we found something to scan.
2204 some_scanned |= !!scan;
2209 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2210 static void init_tlb_ubc(void)
2213 * This deliberately does not clear the cpumask as it's expensive
2214 * and unnecessary. If there happens to be data in there then the
2215 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2216 * then will be cleared.
2218 current->tlb_ubc.flush_required = false;
2221 static inline void init_tlb_ubc(void)
2224 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2227 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2229 static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2230 struct scan_control *sc, unsigned long *lru_pages)
2232 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2233 unsigned long nr[NR_LRU_LISTS];
2234 unsigned long targets[NR_LRU_LISTS];
2235 unsigned long nr_to_scan;
2237 unsigned long nr_reclaimed = 0;
2238 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2239 struct blk_plug plug;
2242 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2244 /* Record the original scan target for proportional adjustments later */
2245 memcpy(targets, nr, sizeof(nr));
2248 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2249 * event that can occur when there is little memory pressure e.g.
2250 * multiple streaming readers/writers. Hence, we do not abort scanning
2251 * when the requested number of pages are reclaimed when scanning at
2252 * DEF_PRIORITY on the assumption that the fact we are direct
2253 * reclaiming implies that kswapd is not keeping up and it is best to
2254 * do a batch of work at once. For memcg reclaim one check is made to
2255 * abort proportional reclaim if either the file or anon lru has already
2256 * dropped to zero at the first pass.
2258 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2259 sc->priority == DEF_PRIORITY);
2263 blk_start_plug(&plug);
2264 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2265 nr[LRU_INACTIVE_FILE]) {
2266 unsigned long nr_anon, nr_file, percentage;
2267 unsigned long nr_scanned;
2269 for_each_evictable_lru(lru) {
2271 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2272 nr[lru] -= nr_to_scan;
2274 nr_reclaimed += shrink_list(lru, nr_to_scan,
2279 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2283 * For kswapd and memcg, reclaim at least the number of pages
2284 * requested. Ensure that the anon and file LRUs are scanned
2285 * proportionally what was requested by get_scan_count(). We
2286 * stop reclaiming one LRU and reduce the amount scanning
2287 * proportional to the original scan target.
2289 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2290 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2293 * It's just vindictive to attack the larger once the smaller
2294 * has gone to zero. And given the way we stop scanning the
2295 * smaller below, this makes sure that we only make one nudge
2296 * towards proportionality once we've got nr_to_reclaim.
2298 if (!nr_file || !nr_anon)
2301 if (nr_file > nr_anon) {
2302 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2303 targets[LRU_ACTIVE_ANON] + 1;
2305 percentage = nr_anon * 100 / scan_target;
2307 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2308 targets[LRU_ACTIVE_FILE] + 1;
2310 percentage = nr_file * 100 / scan_target;
2313 /* Stop scanning the smaller of the LRU */
2315 nr[lru + LRU_ACTIVE] = 0;
2318 * Recalculate the other LRU scan count based on its original
2319 * scan target and the percentage scanning already complete
2321 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2322 nr_scanned = targets[lru] - nr[lru];
2323 nr[lru] = targets[lru] * (100 - percentage) / 100;
2324 nr[lru] -= min(nr[lru], nr_scanned);
2327 nr_scanned = targets[lru] - nr[lru];
2328 nr[lru] = targets[lru] * (100 - percentage) / 100;
2329 nr[lru] -= min(nr[lru], nr_scanned);
2331 scan_adjusted = true;
2333 blk_finish_plug(&plug);
2334 sc->nr_reclaimed += nr_reclaimed;
2337 * Even if we did not try to evict anon pages at all, we want to
2338 * rebalance the anon lru active/inactive ratio.
2340 if (inactive_list_is_low(lruvec, false))
2341 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2342 sc, LRU_ACTIVE_ANON);
2344 throttle_vm_writeout(sc->gfp_mask);
2347 /* Use reclaim/compaction for costly allocs or under memory pressure */
2348 static bool in_reclaim_compaction(struct scan_control *sc)
2350 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2351 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2352 sc->priority < DEF_PRIORITY - 2))
2359 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2360 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2361 * true if more pages should be reclaimed such that when the page allocator
2362 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2363 * It will give up earlier than that if there is difficulty reclaiming pages.
2365 static inline bool should_continue_reclaim(struct zone *zone,
2366 unsigned long nr_reclaimed,
2367 unsigned long nr_scanned,
2368 struct scan_control *sc)
2370 unsigned long pages_for_compaction;
2371 unsigned long inactive_lru_pages;
2373 /* If not in reclaim/compaction mode, stop */
2374 if (!in_reclaim_compaction(sc))
2377 /* Consider stopping depending on scan and reclaim activity */
2378 if (sc->gfp_mask & __GFP_REPEAT) {
2380 * For __GFP_REPEAT allocations, stop reclaiming if the
2381 * full LRU list has been scanned and we are still failing
2382 * to reclaim pages. This full LRU scan is potentially
2383 * expensive but a __GFP_REPEAT caller really wants to succeed
2385 if (!nr_reclaimed && !nr_scanned)
2389 * For non-__GFP_REPEAT allocations which can presumably
2390 * fail without consequence, stop if we failed to reclaim
2391 * any pages from the last SWAP_CLUSTER_MAX number of
2392 * pages that were scanned. This will return to the
2393 * caller faster at the risk reclaim/compaction and
2394 * the resulting allocation attempt fails
2401 * If we have not reclaimed enough pages for compaction and the
2402 * inactive lists are large enough, continue reclaiming
2404 pages_for_compaction = (2UL << sc->order);
2405 inactive_lru_pages = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE);
2406 if (get_nr_swap_pages() > 0)
2407 inactive_lru_pages += node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
2408 if (sc->nr_reclaimed < pages_for_compaction &&
2409 inactive_lru_pages > pages_for_compaction)
2412 /* If compaction would go ahead or the allocation would succeed, stop */
2413 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2414 case COMPACT_PARTIAL:
2415 case COMPACT_CONTINUE:
2422 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc,
2423 enum zone_type classzone_idx)
2425 struct reclaim_state *reclaim_state = current->reclaim_state;
2426 unsigned long nr_reclaimed, nr_scanned;
2427 bool reclaimable = false;
2428 struct zone *zone = &pgdat->node_zones[classzone_idx];
2431 struct mem_cgroup *root = sc->target_mem_cgroup;
2432 struct mem_cgroup_reclaim_cookie reclaim = {
2434 .priority = sc->priority,
2436 unsigned long zone_lru_pages = 0;
2437 struct mem_cgroup *memcg;
2439 nr_reclaimed = sc->nr_reclaimed;
2440 nr_scanned = sc->nr_scanned;
2442 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2444 unsigned long lru_pages;
2445 unsigned long reclaimed;
2446 unsigned long scanned;
2448 if (mem_cgroup_low(root, memcg)) {
2449 if (!sc->may_thrash)
2451 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2454 reclaimed = sc->nr_reclaimed;
2455 scanned = sc->nr_scanned;
2457 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
2458 zone_lru_pages += lru_pages;
2460 if (!global_reclaim(sc))
2461 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2462 memcg, sc->nr_scanned - scanned,
2465 /* Record the group's reclaim efficiency */
2466 vmpressure(sc->gfp_mask, memcg, false,
2467 sc->nr_scanned - scanned,
2468 sc->nr_reclaimed - reclaimed);
2471 * Direct reclaim and kswapd have to scan all memory
2472 * cgroups to fulfill the overall scan target for the
2475 * Limit reclaim, on the other hand, only cares about
2476 * nr_to_reclaim pages to be reclaimed and it will
2477 * retry with decreasing priority if one round over the
2478 * whole hierarchy is not sufficient.
2480 if (!global_reclaim(sc) &&
2481 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2482 mem_cgroup_iter_break(root, memcg);
2485 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2488 * Shrink the slab caches in the same proportion that
2489 * the eligible LRU pages were scanned.
2491 if (global_reclaim(sc))
2492 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2493 sc->nr_scanned - nr_scanned,
2496 if (reclaim_state) {
2497 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2498 reclaim_state->reclaimed_slab = 0;
2501 /* Record the subtree's reclaim efficiency */
2502 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2503 sc->nr_scanned - nr_scanned,
2504 sc->nr_reclaimed - nr_reclaimed);
2506 if (sc->nr_reclaimed - nr_reclaimed)
2509 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2510 sc->nr_scanned - nr_scanned, sc));
2516 * Returns true if compaction should go ahead for a high-order request, or
2517 * the high-order allocation would succeed without compaction.
2519 static inline bool compaction_ready(struct zone *zone, int order, int classzone_idx)
2521 unsigned long watermark;
2525 * Compaction takes time to run and there are potentially other
2526 * callers using the pages just freed. Continue reclaiming until
2527 * there is a buffer of free pages available to give compaction
2528 * a reasonable chance of completing and allocating the page
2530 watermark = high_wmark_pages(zone) + (2UL << order);
2531 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, classzone_idx);
2534 * If compaction is deferred, reclaim up to a point where
2535 * compaction will have a chance of success when re-enabled
2537 if (compaction_deferred(zone, order))
2538 return watermark_ok;
2541 * If compaction is not ready to start and allocation is not likely
2542 * to succeed without it, then keep reclaiming.
2544 if (compaction_suitable(zone, order, 0, classzone_idx) == COMPACT_SKIPPED)
2547 return watermark_ok;
2551 * This is the direct reclaim path, for page-allocating processes. We only
2552 * try to reclaim pages from zones which will satisfy the caller's allocation
2555 * If a zone is deemed to be full of pinned pages then just give it a light
2556 * scan then give up on it.
2558 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2562 unsigned long nr_soft_reclaimed;
2563 unsigned long nr_soft_scanned;
2565 enum zone_type classzone_idx;
2566 pg_data_t *last_pgdat = NULL;
2569 * If the number of buffer_heads in the machine exceeds the maximum
2570 * allowed level, force direct reclaim to scan the highmem zone as
2571 * highmem pages could be pinning lowmem pages storing buffer_heads
2573 orig_mask = sc->gfp_mask;
2574 if (buffer_heads_over_limit) {
2575 sc->gfp_mask |= __GFP_HIGHMEM;
2576 sc->reclaim_idx = classzone_idx = gfp_zone(sc->gfp_mask);
2579 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2580 sc->reclaim_idx, sc->nodemask) {
2581 if (!populated_zone(zone))
2585 * Note that reclaim_idx does not change as it is the highest
2586 * zone reclaimed from which for empty zones is a no-op but
2587 * classzone_idx is used by shrink_node to test if the slabs
2588 * should be shrunk on a given node.
2590 classzone_idx = sc->reclaim_idx;
2591 while (!populated_zone(zone->zone_pgdat->node_zones +
2596 * Take care memory controller reclaiming has small influence
2599 if (global_reclaim(sc)) {
2600 if (!cpuset_zone_allowed(zone,
2601 GFP_KERNEL | __GFP_HARDWALL))
2604 if (sc->priority != DEF_PRIORITY &&
2605 !pgdat_reclaimable(zone->zone_pgdat))
2606 continue; /* Let kswapd poll it */
2609 * If we already have plenty of memory free for
2610 * compaction in this zone, don't free any more.
2611 * Even though compaction is invoked for any
2612 * non-zero order, only frequent costly order
2613 * reclamation is disruptive enough to become a
2614 * noticeable problem, like transparent huge
2617 if (IS_ENABLED(CONFIG_COMPACTION) &&
2618 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2619 zonelist_zone_idx(z) <= classzone_idx &&
2620 compaction_ready(zone, sc->order, classzone_idx)) {
2621 sc->compaction_ready = true;
2626 * Shrink each node in the zonelist once. If the
2627 * zonelist is ordered by zone (not the default) then a
2628 * node may be shrunk multiple times but in that case
2629 * the user prefers lower zones being preserved.
2631 if (zone->zone_pgdat == last_pgdat)
2635 * This steals pages from memory cgroups over softlimit
2636 * and returns the number of reclaimed pages and
2637 * scanned pages. This works for global memory pressure
2638 * and balancing, not for a memcg's limit.
2640 nr_soft_scanned = 0;
2641 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2642 sc->order, sc->gfp_mask,
2644 sc->nr_reclaimed += nr_soft_reclaimed;
2645 sc->nr_scanned += nr_soft_scanned;
2646 /* need some check for avoid more shrink_zone() */
2649 /* See comment about same check for global reclaim above */
2650 if (zone->zone_pgdat == last_pgdat)
2652 last_pgdat = zone->zone_pgdat;
2653 shrink_node(zone->zone_pgdat, sc, classzone_idx);
2657 * Restore to original mask to avoid the impact on the caller if we
2658 * promoted it to __GFP_HIGHMEM.
2660 sc->gfp_mask = orig_mask;
2664 * This is the main entry point to direct page reclaim.
2666 * If a full scan of the inactive list fails to free enough memory then we
2667 * are "out of memory" and something needs to be killed.
2669 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2670 * high - the zone may be full of dirty or under-writeback pages, which this
2671 * caller can't do much about. We kick the writeback threads and take explicit
2672 * naps in the hope that some of these pages can be written. But if the
2673 * allocating task holds filesystem locks which prevent writeout this might not
2674 * work, and the allocation attempt will fail.
2676 * returns: 0, if no pages reclaimed
2677 * else, the number of pages reclaimed
2679 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2680 struct scan_control *sc)
2682 int initial_priority = sc->priority;
2683 unsigned long total_scanned = 0;
2684 unsigned long writeback_threshold;
2686 delayacct_freepages_start();
2688 if (global_reclaim(sc))
2689 count_vm_event(ALLOCSTALL);
2692 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2695 shrink_zones(zonelist, sc);
2697 total_scanned += sc->nr_scanned;
2698 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2701 if (sc->compaction_ready)
2705 * If we're getting trouble reclaiming, start doing
2706 * writepage even in laptop mode.
2708 if (sc->priority < DEF_PRIORITY - 2)
2709 sc->may_writepage = 1;
2712 * Try to write back as many pages as we just scanned. This
2713 * tends to cause slow streaming writers to write data to the
2714 * disk smoothly, at the dirtying rate, which is nice. But
2715 * that's undesirable in laptop mode, where we *want* lumpy
2716 * writeout. So in laptop mode, write out the whole world.
2718 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2719 if (total_scanned > writeback_threshold) {
2720 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2721 WB_REASON_TRY_TO_FREE_PAGES);
2722 sc->may_writepage = 1;
2724 } while (--sc->priority >= 0);
2726 delayacct_freepages_end();
2728 if (sc->nr_reclaimed)
2729 return sc->nr_reclaimed;
2731 /* Aborted reclaim to try compaction? don't OOM, then */
2732 if (sc->compaction_ready)
2735 /* Untapped cgroup reserves? Don't OOM, retry. */
2736 if (!sc->may_thrash) {
2737 sc->priority = initial_priority;
2745 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2748 unsigned long pfmemalloc_reserve = 0;
2749 unsigned long free_pages = 0;
2753 for (i = 0; i <= ZONE_NORMAL; i++) {
2754 zone = &pgdat->node_zones[i];
2755 if (!populated_zone(zone) ||
2756 pgdat_reclaimable_pages(pgdat) == 0)
2759 pfmemalloc_reserve += min_wmark_pages(zone);
2760 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2763 /* If there are no reserves (unexpected config) then do not throttle */
2764 if (!pfmemalloc_reserve)
2767 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2769 /* kswapd must be awake if processes are being throttled */
2770 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2771 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2772 (enum zone_type)ZONE_NORMAL);
2773 wake_up_interruptible(&pgdat->kswapd_wait);
2780 * Throttle direct reclaimers if backing storage is backed by the network
2781 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2782 * depleted. kswapd will continue to make progress and wake the processes
2783 * when the low watermark is reached.
2785 * Returns true if a fatal signal was delivered during throttling. If this
2786 * happens, the page allocator should not consider triggering the OOM killer.
2788 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2789 nodemask_t *nodemask)
2793 pg_data_t *pgdat = NULL;
2796 * Kernel threads should not be throttled as they may be indirectly
2797 * responsible for cleaning pages necessary for reclaim to make forward
2798 * progress. kjournald for example may enter direct reclaim while
2799 * committing a transaction where throttling it could forcing other
2800 * processes to block on log_wait_commit().
2802 if (current->flags & PF_KTHREAD)
2806 * If a fatal signal is pending, this process should not throttle.
2807 * It should return quickly so it can exit and free its memory
2809 if (fatal_signal_pending(current))
2813 * Check if the pfmemalloc reserves are ok by finding the first node
2814 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2815 * GFP_KERNEL will be required for allocating network buffers when
2816 * swapping over the network so ZONE_HIGHMEM is unusable.
2818 * Throttling is based on the first usable node and throttled processes
2819 * wait on a queue until kswapd makes progress and wakes them. There
2820 * is an affinity then between processes waking up and where reclaim
2821 * progress has been made assuming the process wakes on the same node.
2822 * More importantly, processes running on remote nodes will not compete
2823 * for remote pfmemalloc reserves and processes on different nodes
2824 * should make reasonable progress.
2826 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2827 gfp_zone(gfp_mask), nodemask) {
2828 if (zone_idx(zone) > ZONE_NORMAL)
2831 /* Throttle based on the first usable node */
2832 pgdat = zone->zone_pgdat;
2833 if (pfmemalloc_watermark_ok(pgdat))
2838 /* If no zone was usable by the allocation flags then do not throttle */
2842 /* Account for the throttling */
2843 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2846 * If the caller cannot enter the filesystem, it's possible that it
2847 * is due to the caller holding an FS lock or performing a journal
2848 * transaction in the case of a filesystem like ext[3|4]. In this case,
2849 * it is not safe to block on pfmemalloc_wait as kswapd could be
2850 * blocked waiting on the same lock. Instead, throttle for up to a
2851 * second before continuing.
2853 if (!(gfp_mask & __GFP_FS)) {
2854 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2855 pfmemalloc_watermark_ok(pgdat), HZ);
2860 /* Throttle until kswapd wakes the process */
2861 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2862 pfmemalloc_watermark_ok(pgdat));
2865 if (fatal_signal_pending(current))
2872 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2873 gfp_t gfp_mask, nodemask_t *nodemask)
2875 unsigned long nr_reclaimed;
2876 struct scan_control sc = {
2877 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2878 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2879 .reclaim_idx = gfp_zone(gfp_mask),
2881 .nodemask = nodemask,
2882 .priority = DEF_PRIORITY,
2883 .may_writepage = !laptop_mode,
2889 * Do not enter reclaim if fatal signal was delivered while throttled.
2890 * 1 is returned so that the page allocator does not OOM kill at this
2893 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2896 trace_mm_vmscan_direct_reclaim_begin(order,
2900 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2902 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2904 return nr_reclaimed;
2909 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2910 gfp_t gfp_mask, bool noswap,
2912 unsigned long *nr_scanned)
2914 struct scan_control sc = {
2915 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2916 .target_mem_cgroup = memcg,
2917 .may_writepage = !laptop_mode,
2919 .reclaim_idx = MAX_NR_ZONES - 1,
2920 .may_swap = !noswap,
2922 unsigned long lru_pages;
2924 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2925 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2927 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2932 * NOTE: Although we can get the priority field, using it
2933 * here is not a good idea, since it limits the pages we can scan.
2934 * if we don't reclaim here, the shrink_zone from balance_pgdat
2935 * will pick up pages from other mem cgroup's as well. We hack
2936 * the priority and make it zero.
2938 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
2940 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2942 *nr_scanned = sc.nr_scanned;
2943 return sc.nr_reclaimed;
2946 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2947 unsigned long nr_pages,
2951 struct zonelist *zonelist;
2952 unsigned long nr_reclaimed;
2954 struct scan_control sc = {
2955 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2956 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2957 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2958 .reclaim_idx = MAX_NR_ZONES - 1,
2959 .target_mem_cgroup = memcg,
2960 .priority = DEF_PRIORITY,
2961 .may_writepage = !laptop_mode,
2963 .may_swap = may_swap,
2967 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2968 * take care of from where we get pages. So the node where we start the
2969 * scan does not need to be the current node.
2971 nid = mem_cgroup_select_victim_node(memcg);
2973 zonelist = NODE_DATA(nid)->node_zonelists;
2975 trace_mm_vmscan_memcg_reclaim_begin(0,
2979 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2981 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2983 return nr_reclaimed;
2987 static void age_active_anon(struct pglist_data *pgdat,
2988 struct zone *zone, struct scan_control *sc)
2990 struct mem_cgroup *memcg;
2992 if (!total_swap_pages)
2995 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2997 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2999 if (inactive_list_is_low(lruvec, false))
3000 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3001 sc, LRU_ACTIVE_ANON);
3003 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3007 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3009 unsigned long mark = high_wmark_pages(zone);
3011 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3015 * If any eligible zone is balanced then the node is not considered
3016 * to be congested or dirty
3018 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3019 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3025 * Prepare kswapd for sleeping. This verifies that there are no processes
3026 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3028 * Returns true if kswapd is ready to sleep
3030 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3035 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3040 * The throttled processes are normally woken up in balance_pgdat() as
3041 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3042 * race between when kswapd checks the watermarks and a process gets
3043 * throttled. There is also a potential race if processes get
3044 * throttled, kswapd wakes, a large process exits thereby balancing the
3045 * zones, which causes kswapd to exit balance_pgdat() before reaching
3046 * the wake up checks. If kswapd is going to sleep, no process should
3047 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3048 * the wake up is premature, processes will wake kswapd and get
3049 * throttled again. The difference from wake ups in balance_pgdat() is
3050 * that here we are under prepare_to_wait().
3052 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3053 wake_up_all(&pgdat->pfmemalloc_wait);
3055 for (i = 0; i <= classzone_idx; i++) {
3056 struct zone *zone = pgdat->node_zones + i;
3058 if (!populated_zone(zone))
3061 if (!zone_balanced(zone, order, classzone_idx))
3069 * kswapd shrinks a node of pages that are at or below the highest usable
3070 * zone that is currently unbalanced.
3072 * Returns true if kswapd scanned at least the requested number of pages to
3073 * reclaim or if the lack of progress was due to pages under writeback.
3074 * This is used to determine if the scanning priority needs to be raised.
3076 static bool kswapd_shrink_node(pg_data_t *pgdat,
3078 struct scan_control *sc)
3083 /* Reclaim a number of pages proportional to the number of zones */
3084 sc->nr_to_reclaim = 0;
3085 for (z = 0; z <= classzone_idx; z++) {
3086 zone = pgdat->node_zones + z;
3087 if (!populated_zone(zone))
3090 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3094 * Historically care was taken to put equal pressure on all zones but
3095 * now pressure is applied based on node LRU order.
3097 shrink_node(pgdat, sc, classzone_idx);
3100 * Fragmentation may mean that the system cannot be rebalanced for
3101 * high-order allocations. If twice the allocation size has been
3102 * reclaimed then recheck watermarks only at order-0 to prevent
3103 * excessive reclaim. Assume that a process requested a high-order
3104 * can direct reclaim/compact.
3106 if (sc->order && sc->nr_reclaimed >= 2UL << sc->order)
3109 return sc->nr_scanned >= sc->nr_to_reclaim;
3113 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3114 * that are eligible for use by the caller until at least one zone is
3117 * Returns the order kswapd finished reclaiming at.
3119 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3120 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3121 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3122 * or lower is eligible for reclaim until at least one usable zone is
3125 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3128 unsigned long nr_soft_reclaimed;
3129 unsigned long nr_soft_scanned;
3131 struct scan_control sc = {
3132 .gfp_mask = GFP_KERNEL,
3134 .priority = DEF_PRIORITY,
3135 .may_writepage = !laptop_mode,
3138 .reclaim_idx = classzone_idx,
3140 count_vm_event(PAGEOUTRUN);
3143 bool raise_priority = true;
3145 sc.nr_reclaimed = 0;
3148 * If the number of buffer_heads in the machine exceeds the
3149 * maximum allowed level then reclaim from all zones. This is
3150 * not specific to highmem as highmem may not exist but it is
3151 * it is expected that buffer_heads are stripped in writeback.
3153 if (buffer_heads_over_limit) {
3154 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3155 zone = pgdat->node_zones + i;
3156 if (!populated_zone(zone))
3165 * Only reclaim if there are no eligible zones. Check from
3166 * high to low zone as allocations prefer higher zones.
3167 * Scanning from low to high zone would allow congestion to be
3168 * cleared during a very small window when a small low
3169 * zone was balanced even under extreme pressure when the
3170 * overall node may be congested.
3172 for (i = classzone_idx; i >= 0; i--) {
3173 zone = pgdat->node_zones + i;
3174 if (!populated_zone(zone))
3177 if (zone_balanced(zone, sc.order, classzone_idx))
3182 * Do some background aging of the anon list, to give
3183 * pages a chance to be referenced before reclaiming. All
3184 * pages are rotated regardless of classzone as this is
3185 * about consistent aging.
3187 age_active_anon(pgdat, &pgdat->node_zones[MAX_NR_ZONES - 1], &sc);
3190 * If we're getting trouble reclaiming, start doing writepage
3191 * even in laptop mode.
3193 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3194 sc.may_writepage = 1;
3196 /* Call soft limit reclaim before calling shrink_node. */
3198 nr_soft_scanned = 0;
3199 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, sc.order,
3200 sc.gfp_mask, &nr_soft_scanned);
3201 sc.nr_reclaimed += nr_soft_reclaimed;
3204 * There should be no need to raise the scanning priority if
3205 * enough pages are already being scanned that that high
3206 * watermark would be met at 100% efficiency.
3208 if (kswapd_shrink_node(pgdat, classzone_idx, &sc))
3209 raise_priority = false;
3212 * If the low watermark is met there is no need for processes
3213 * to be throttled on pfmemalloc_wait as they should not be
3214 * able to safely make forward progress. Wake them
3216 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3217 pfmemalloc_watermark_ok(pgdat))
3218 wake_up_all(&pgdat->pfmemalloc_wait);
3220 /* Check if kswapd should be suspending */
3221 if (try_to_freeze() || kthread_should_stop())
3225 * Raise priority if scanning rate is too low or there was no
3226 * progress in reclaiming pages
3228 if (raise_priority || !sc.nr_reclaimed)
3230 } while (sc.priority >= 1);
3234 * Return the order kswapd stopped reclaiming at as
3235 * prepare_kswapd_sleep() takes it into account. If another caller
3236 * entered the allocator slow path while kswapd was awake, order will
3237 * remain at the higher level.
3242 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3243 unsigned int classzone_idx)
3248 if (freezing(current) || kthread_should_stop())
3251 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3253 /* Try to sleep for a short interval */
3254 if (prepare_kswapd_sleep(pgdat, reclaim_order, remaining, classzone_idx)) {
3256 * Compaction records what page blocks it recently failed to
3257 * isolate pages from and skips them in the future scanning.
3258 * When kswapd is going to sleep, it is reasonable to assume
3259 * that pages and compaction may succeed so reset the cache.
3261 reset_isolation_suitable(pgdat);
3264 * We have freed the memory, now we should compact it to make
3265 * allocation of the requested order possible.
3267 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3269 remaining = schedule_timeout(HZ/10);
3272 * If woken prematurely then reset kswapd_classzone_idx and
3273 * order. The values will either be from a wakeup request or
3274 * the previous request that slept prematurely.
3277 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3278 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3281 finish_wait(&pgdat->kswapd_wait, &wait);
3282 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3286 * After a short sleep, check if it was a premature sleep. If not, then
3287 * go fully to sleep until explicitly woken up.
3289 if (prepare_kswapd_sleep(pgdat, reclaim_order, remaining, classzone_idx)) {
3290 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3293 * vmstat counters are not perfectly accurate and the estimated
3294 * value for counters such as NR_FREE_PAGES can deviate from the
3295 * true value by nr_online_cpus * threshold. To avoid the zone
3296 * watermarks being breached while under pressure, we reduce the
3297 * per-cpu vmstat threshold while kswapd is awake and restore
3298 * them before going back to sleep.
3300 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3302 if (!kthread_should_stop())
3305 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3308 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3310 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3312 finish_wait(&pgdat->kswapd_wait, &wait);
3316 * The background pageout daemon, started as a kernel thread
3317 * from the init process.
3319 * This basically trickles out pages so that we have _some_
3320 * free memory available even if there is no other activity
3321 * that frees anything up. This is needed for things like routing
3322 * etc, where we otherwise might have all activity going on in
3323 * asynchronous contexts that cannot page things out.
3325 * If there are applications that are active memory-allocators
3326 * (most normal use), this basically shouldn't matter.
3328 static int kswapd(void *p)
3330 unsigned int alloc_order, reclaim_order, classzone_idx;
3331 pg_data_t *pgdat = (pg_data_t*)p;
3332 struct task_struct *tsk = current;
3334 struct reclaim_state reclaim_state = {
3335 .reclaimed_slab = 0,
3337 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3339 lockdep_set_current_reclaim_state(GFP_KERNEL);
3341 if (!cpumask_empty(cpumask))
3342 set_cpus_allowed_ptr(tsk, cpumask);
3343 current->reclaim_state = &reclaim_state;
3346 * Tell the memory management that we're a "memory allocator",
3347 * and that if we need more memory we should get access to it
3348 * regardless (see "__alloc_pages()"). "kswapd" should
3349 * never get caught in the normal page freeing logic.
3351 * (Kswapd normally doesn't need memory anyway, but sometimes
3352 * you need a small amount of memory in order to be able to
3353 * page out something else, and this flag essentially protects
3354 * us from recursively trying to free more memory as we're
3355 * trying to free the first piece of memory in the first place).
3357 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3360 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3361 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3366 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3369 /* Read the new order and classzone_idx */
3370 alloc_order = reclaim_order = pgdat->kswapd_order;
3371 classzone_idx = pgdat->kswapd_classzone_idx;
3372 pgdat->kswapd_order = 0;
3373 pgdat->kswapd_classzone_idx = 0;
3375 ret = try_to_freeze();
3376 if (kthread_should_stop())
3380 * We can speed up thawing tasks if we don't call balance_pgdat
3381 * after returning from the refrigerator
3387 * Reclaim begins at the requested order but if a high-order
3388 * reclaim fails then kswapd falls back to reclaiming for
3389 * order-0. If that happens, kswapd will consider sleeping
3390 * for the order it finished reclaiming at (reclaim_order)
3391 * but kcompactd is woken to compact for the original
3392 * request (alloc_order).
3394 trace_mm_vmscan_kswapd_wake(pgdat->node_id, alloc_order);
3395 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3396 if (reclaim_order < alloc_order)
3397 goto kswapd_try_sleep;
3399 alloc_order = reclaim_order = pgdat->kswapd_order;
3400 classzone_idx = pgdat->kswapd_classzone_idx;
3403 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3404 current->reclaim_state = NULL;
3405 lockdep_clear_current_reclaim_state();
3411 * A zone is low on free memory, so wake its kswapd task to service it.
3413 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3417 if (!populated_zone(zone))
3420 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3422 pgdat = zone->zone_pgdat;
3423 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3424 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3425 if (!waitqueue_active(&pgdat->kswapd_wait))
3427 if (zone_balanced(zone, order, 0))
3430 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3431 wake_up_interruptible(&pgdat->kswapd_wait);
3434 #ifdef CONFIG_HIBERNATION
3436 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3439 * Rather than trying to age LRUs the aim is to preserve the overall
3440 * LRU order by reclaiming preferentially
3441 * inactive > active > active referenced > active mapped
3443 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3445 struct reclaim_state reclaim_state;
3446 struct scan_control sc = {
3447 .nr_to_reclaim = nr_to_reclaim,
3448 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3449 .reclaim_idx = MAX_NR_ZONES - 1,
3450 .priority = DEF_PRIORITY,
3454 .hibernation_mode = 1,
3456 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3457 struct task_struct *p = current;
3458 unsigned long nr_reclaimed;
3460 p->flags |= PF_MEMALLOC;
3461 lockdep_set_current_reclaim_state(sc.gfp_mask);
3462 reclaim_state.reclaimed_slab = 0;
3463 p->reclaim_state = &reclaim_state;
3465 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3467 p->reclaim_state = NULL;
3468 lockdep_clear_current_reclaim_state();
3469 p->flags &= ~PF_MEMALLOC;
3471 return nr_reclaimed;
3473 #endif /* CONFIG_HIBERNATION */
3475 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3476 not required for correctness. So if the last cpu in a node goes
3477 away, we get changed to run anywhere: as the first one comes back,
3478 restore their cpu bindings. */
3479 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3484 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3485 for_each_node_state(nid, N_MEMORY) {
3486 pg_data_t *pgdat = NODE_DATA(nid);
3487 const struct cpumask *mask;
3489 mask = cpumask_of_node(pgdat->node_id);
3491 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3492 /* One of our CPUs online: restore mask */
3493 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3500 * This kswapd start function will be called by init and node-hot-add.
3501 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3503 int kswapd_run(int nid)
3505 pg_data_t *pgdat = NODE_DATA(nid);
3511 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3512 if (IS_ERR(pgdat->kswapd)) {
3513 /* failure at boot is fatal */
3514 BUG_ON(system_state == SYSTEM_BOOTING);
3515 pr_err("Failed to start kswapd on node %d\n", nid);
3516 ret = PTR_ERR(pgdat->kswapd);
3517 pgdat->kswapd = NULL;
3523 * Called by memory hotplug when all memory in a node is offlined. Caller must
3524 * hold mem_hotplug_begin/end().
3526 void kswapd_stop(int nid)
3528 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3531 kthread_stop(kswapd);
3532 NODE_DATA(nid)->kswapd = NULL;
3536 static int __init kswapd_init(void)
3541 for_each_node_state(nid, N_MEMORY)
3543 hotcpu_notifier(cpu_callback, 0);
3547 module_init(kswapd_init)
3553 * If non-zero call zone_reclaim when the number of free pages falls below
3556 int zone_reclaim_mode __read_mostly;
3558 #define RECLAIM_OFF 0
3559 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3560 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3561 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3564 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3565 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3568 #define ZONE_RECLAIM_PRIORITY 4
3571 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3574 int sysctl_min_unmapped_ratio = 1;
3577 * If the number of slab pages in a zone grows beyond this percentage then
3578 * slab reclaim needs to occur.
3580 int sysctl_min_slab_ratio = 5;
3582 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3584 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3585 unsigned long file_lru = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
3586 node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE);
3589 * It's possible for there to be more file mapped pages than
3590 * accounted for by the pages on the file LRU lists because
3591 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3593 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3596 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3597 static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3599 unsigned long nr_pagecache_reclaimable;
3600 unsigned long delta = 0;
3603 * If RECLAIM_UNMAP is set, then all file pages are considered
3604 * potentially reclaimable. Otherwise, we have to worry about
3605 * pages like swapcache and zone_unmapped_file_pages() provides
3608 if (zone_reclaim_mode & RECLAIM_UNMAP)
3609 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3611 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3613 /* If we can't clean pages, remove dirty pages from consideration */
3614 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3615 delta += zone_page_state(zone, NR_FILE_DIRTY);
3617 /* Watch for any possible underflows due to delta */
3618 if (unlikely(delta > nr_pagecache_reclaimable))
3619 delta = nr_pagecache_reclaimable;
3621 return nr_pagecache_reclaimable - delta;
3625 * Try to free up some pages from this zone through reclaim.
3627 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3629 /* Minimum pages needed in order to stay on node */
3630 const unsigned long nr_pages = 1 << order;
3631 struct task_struct *p = current;
3632 struct reclaim_state reclaim_state;
3633 struct scan_control sc = {
3634 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3635 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3637 .priority = ZONE_RECLAIM_PRIORITY,
3638 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3639 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3641 .reclaim_idx = zone_idx(zone),
3646 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3647 * and we also need to be able to write out pages for RECLAIM_WRITE
3648 * and RECLAIM_UNMAP.
3650 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3651 lockdep_set_current_reclaim_state(gfp_mask);
3652 reclaim_state.reclaimed_slab = 0;
3653 p->reclaim_state = &reclaim_state;
3655 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3657 * Free memory by calling shrink zone with increasing
3658 * priorities until we have enough memory freed.
3661 shrink_node(zone->zone_pgdat, &sc, zone_idx(zone));
3662 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3665 p->reclaim_state = NULL;
3666 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3667 lockdep_clear_current_reclaim_state();
3668 return sc.nr_reclaimed >= nr_pages;
3671 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3677 * Zone reclaim reclaims unmapped file backed pages and
3678 * slab pages if we are over the defined limits.
3680 * A small portion of unmapped file backed pages is needed for
3681 * file I/O otherwise pages read by file I/O will be immediately
3682 * thrown out if the zone is overallocated. So we do not reclaim
3683 * if less than a specified percentage of the zone is used by
3684 * unmapped file backed pages.
3686 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3687 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3688 return ZONE_RECLAIM_FULL;
3690 if (!pgdat_reclaimable(zone->zone_pgdat))
3691 return ZONE_RECLAIM_FULL;
3694 * Do not scan if the allocation should not be delayed.
3696 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3697 return ZONE_RECLAIM_NOSCAN;
3700 * Only run zone reclaim on the local zone or on zones that do not
3701 * have associated processors. This will favor the local processor
3702 * over remote processors and spread off node memory allocations
3703 * as wide as possible.
3705 node_id = zone_to_nid(zone);
3706 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3707 return ZONE_RECLAIM_NOSCAN;
3709 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3710 return ZONE_RECLAIM_NOSCAN;
3712 ret = __zone_reclaim(zone, gfp_mask, order);
3713 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3716 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3723 * page_evictable - test whether a page is evictable
3724 * @page: the page to test
3726 * Test whether page is evictable--i.e., should be placed on active/inactive
3727 * lists vs unevictable list.
3729 * Reasons page might not be evictable:
3730 * (1) page's mapping marked unevictable
3731 * (2) page is part of an mlocked VMA
3734 int page_evictable(struct page *page)
3736 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3741 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3742 * @pages: array of pages to check
3743 * @nr_pages: number of pages to check
3745 * Checks pages for evictability and moves them to the appropriate lru list.
3747 * This function is only used for SysV IPC SHM_UNLOCK.
3749 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3751 struct lruvec *lruvec;
3752 struct zone *zone = NULL;
3757 for (i = 0; i < nr_pages; i++) {
3758 struct page *page = pages[i];
3759 struct zone *pagezone;
3762 pagezone = page_zone(page);
3763 if (pagezone != zone) {
3765 spin_unlock_irq(zone_lru_lock(zone));
3767 spin_lock_irq(zone_lru_lock(zone));
3769 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
3771 if (!PageLRU(page) || !PageUnevictable(page))
3774 if (page_evictable(page)) {
3775 enum lru_list lru = page_lru_base_type(page);
3777 VM_BUG_ON_PAGE(PageActive(page), page);
3778 ClearPageUnevictable(page);
3779 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3780 add_page_to_lru_list(page, lruvec, lru);
3786 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3787 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3788 spin_unlock_irq(zone_lru_lock(zone));
3791 #endif /* CONFIG_SHMEM */