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/debugfs.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 unsigned int may_writepage:1;
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
98 unsigned int hibernation_mode:1;
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
110 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
112 #ifdef ARCH_HAS_PREFETCH
113 #define prefetch_prev_lru_page(_page, _base, _field) \
115 if ((_page)->lru.prev != _base) { \
118 prev = lru_to_page(&(_page->lru)); \
119 prefetch(&prev->_field); \
123 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
126 #ifdef ARCH_HAS_PREFETCHW
127 #define prefetchw_prev_lru_page(_page, _base, _field) \
129 if ((_page)->lru.prev != _base) { \
132 prev = lru_to_page(&(_page->lru)); \
133 prefetchw(&prev->_field); \
137 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
141 * From 0 .. 100. Higher means more swappy.
143 int vm_swappiness = 60;
145 * The total number of pages which are beyond the high watermark within all
148 unsigned long vm_total_pages;
150 static LIST_HEAD(shrinker_list);
151 static DECLARE_RWSEM(shrinker_rwsem);
154 static bool global_reclaim(struct scan_control *sc)
156 return !sc->target_mem_cgroup;
160 * sane_reclaim - is the usual dirty throttling mechanism operational?
161 * @sc: scan_control in question
163 * The normal page dirty throttling mechanism in balance_dirty_pages() is
164 * completely broken with the legacy memcg and direct stalling in
165 * shrink_page_list() is used for throttling instead, which lacks all the
166 * niceties such as fairness, adaptive pausing, bandwidth proportional
167 * allocation and configurability.
169 * This function tests whether the vmscan currently in progress can assume
170 * that the normal dirty throttling mechanism is operational.
172 static bool sane_reclaim(struct scan_control *sc)
174 struct mem_cgroup *memcg = sc->target_mem_cgroup;
178 #ifdef CONFIG_CGROUP_WRITEBACK
179 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
185 static bool global_reclaim(struct scan_control *sc)
190 static bool sane_reclaim(struct scan_control *sc)
196 static unsigned long zone_reclaimable_pages(struct zone *zone)
200 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
201 zone_page_state(zone, NR_INACTIVE_FILE);
203 if (get_nr_swap_pages() > 0)
204 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
205 zone_page_state(zone, NR_INACTIVE_ANON);
210 bool zone_reclaimable(struct zone *zone)
212 return zone_page_state(zone, NR_PAGES_SCANNED) <
213 zone_reclaimable_pages(zone) * 6;
216 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
218 if (!mem_cgroup_disabled())
219 return mem_cgroup_get_lru_size(lruvec, lru);
221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
224 struct dentry *debug_file;
226 static int debug_shrinker_show(struct seq_file *s, void *unused)
228 struct shrinker *shrinker;
229 struct shrink_control sc;
234 down_read(&shrinker_rwsem);
235 list_for_each_entry(shrinker, &shrinker_list, list) {
238 num_objs = shrinker->count_objects(shrinker, &sc);
239 seq_printf(s, "%pf %d\n", shrinker->scan_objects, num_objs);
241 up_read(&shrinker_rwsem);
245 static int debug_shrinker_open(struct inode *inode, struct file *file)
247 return single_open(file, debug_shrinker_show, inode->i_private);
250 static const struct file_operations debug_shrinker_fops = {
251 .open = debug_shrinker_open,
254 .release = single_release,
258 * Add a shrinker callback to be called from the vm.
260 int register_shrinker(struct shrinker *shrinker)
262 size_t size = sizeof(*shrinker->nr_deferred);
265 * If we only have one possible node in the system anyway, save
266 * ourselves the trouble and disable NUMA aware behavior. This way we
267 * will save memory and some small loop time later.
269 if (nr_node_ids == 1)
270 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
272 if (shrinker->flags & SHRINKER_NUMA_AWARE)
275 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
276 if (!shrinker->nr_deferred)
279 down_write(&shrinker_rwsem);
280 list_add_tail(&shrinker->list, &shrinker_list);
281 up_write(&shrinker_rwsem);
284 EXPORT_SYMBOL(register_shrinker);
286 static int __init add_shrinker_debug(void)
288 debugfs_create_file("shrinker", 0644, NULL, NULL,
289 &debug_shrinker_fops);
293 late_initcall(add_shrinker_debug);
298 void unregister_shrinker(struct shrinker *shrinker)
300 down_write(&shrinker_rwsem);
301 list_del(&shrinker->list);
302 up_write(&shrinker_rwsem);
303 kfree(shrinker->nr_deferred);
305 EXPORT_SYMBOL(unregister_shrinker);
307 #define SHRINK_BATCH 128
309 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
310 struct shrinker *shrinker,
311 unsigned long nr_scanned,
312 unsigned long nr_eligible)
314 unsigned long freed = 0;
315 unsigned long long delta;
320 int nid = shrinkctl->nid;
321 long batch_size = shrinker->batch ? shrinker->batch
324 freeable = shrinker->count_objects(shrinker, shrinkctl);
329 * copy the current shrinker scan count into a local variable
330 * and zero it so that other concurrent shrinker invocations
331 * don't also do this scanning work.
333 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
336 delta = (4 * nr_scanned) / shrinker->seeks;
338 do_div(delta, nr_eligible + 1);
340 if (total_scan < 0) {
341 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
342 shrinker->scan_objects, total_scan);
343 total_scan = freeable;
347 * We need to avoid excessive windup on filesystem shrinkers
348 * due to large numbers of GFP_NOFS allocations causing the
349 * shrinkers to return -1 all the time. This results in a large
350 * nr being built up so when a shrink that can do some work
351 * comes along it empties the entire cache due to nr >>>
352 * freeable. This is bad for sustaining a working set in
355 * Hence only allow the shrinker to scan the entire cache when
356 * a large delta change is calculated directly.
358 if (delta < freeable / 4)
359 total_scan = min(total_scan, freeable / 2);
362 * Avoid risking looping forever due to too large nr value:
363 * never try to free more than twice the estimate number of
366 if (total_scan > freeable * 2)
367 total_scan = freeable * 2;
369 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
370 nr_scanned, nr_eligible,
371 freeable, delta, total_scan);
374 * Normally, we should not scan less than batch_size objects in one
375 * pass to avoid too frequent shrinker calls, but if the slab has less
376 * than batch_size objects in total and we are really tight on memory,
377 * we will try to reclaim all available objects, otherwise we can end
378 * up failing allocations although there are plenty of reclaimable
379 * objects spread over several slabs with usage less than the
382 * We detect the "tight on memory" situations by looking at the total
383 * number of objects we want to scan (total_scan). If it is greater
384 * than the total number of objects on slab (freeable), we must be
385 * scanning at high prio and therefore should try to reclaim as much as
388 while (total_scan >= batch_size ||
389 total_scan >= freeable) {
391 unsigned long nr_to_scan = min(batch_size, total_scan);
393 shrinkctl->nr_to_scan = nr_to_scan;
394 ret = shrinker->scan_objects(shrinker, shrinkctl);
395 if (ret == SHRINK_STOP)
399 count_vm_events(SLABS_SCANNED, nr_to_scan);
400 total_scan -= nr_to_scan;
406 * move the unused scan count back into the shrinker in a
407 * manner that handles concurrent updates. If we exhausted the
408 * scan, there is no need to do an update.
411 new_nr = atomic_long_add_return(total_scan,
412 &shrinker->nr_deferred[nid]);
414 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
416 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
421 * shrink_slab - shrink slab caches
422 * @gfp_mask: allocation context
423 * @nid: node whose slab caches to target
424 * @memcg: memory cgroup whose slab caches to target
425 * @nr_scanned: pressure numerator
426 * @nr_eligible: pressure denominator
428 * Call the shrink functions to age shrinkable caches.
430 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
431 * unaware shrinkers will receive a node id of 0 instead.
433 * @memcg specifies the memory cgroup to target. If it is not NULL,
434 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
435 * objects from the memory cgroup specified. Otherwise all shrinkers
436 * are called, and memcg aware shrinkers are supposed to scan the
439 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
440 * the available objects should be scanned. Page reclaim for example
441 * passes the number of pages scanned and the number of pages on the
442 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
443 * when it encountered mapped pages. The ratio is further biased by
444 * the ->seeks setting of the shrink function, which indicates the
445 * cost to recreate an object relative to that of an LRU page.
447 * Returns the number of reclaimed slab objects.
449 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
450 struct mem_cgroup *memcg,
451 unsigned long nr_scanned,
452 unsigned long nr_eligible)
454 struct shrinker *shrinker;
455 unsigned long freed = 0;
457 if (memcg && !memcg_kmem_is_active(memcg))
461 nr_scanned = SWAP_CLUSTER_MAX;
463 if (!down_read_trylock(&shrinker_rwsem)) {
465 * If we would return 0, our callers would understand that we
466 * have nothing else to shrink and give up trying. By returning
467 * 1 we keep it going and assume we'll be able to shrink next
474 list_for_each_entry(shrinker, &shrinker_list, list) {
475 struct shrink_control sc = {
476 .gfp_mask = gfp_mask,
481 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
484 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
487 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
490 up_read(&shrinker_rwsem);
496 void drop_slab_node(int nid)
501 struct mem_cgroup *memcg = NULL;
505 freed += shrink_slab(GFP_KERNEL, nid, memcg,
507 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
508 } while (freed > 10);
515 for_each_online_node(nid)
519 static inline int is_page_cache_freeable(struct page *page)
522 * A freeable page cache page is referenced only by the caller
523 * that isolated the page, the page cache radix tree and
524 * optional buffer heads at page->private.
526 return page_count(page) - page_has_private(page) == 2;
529 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
531 if (current->flags & PF_SWAPWRITE)
533 if (!inode_write_congested(inode))
535 if (inode_to_bdi(inode) == current->backing_dev_info)
541 * We detected a synchronous write error writing a page out. Probably
542 * -ENOSPC. We need to propagate that into the address_space for a subsequent
543 * fsync(), msync() or close().
545 * The tricky part is that after writepage we cannot touch the mapping: nothing
546 * prevents it from being freed up. But we have a ref on the page and once
547 * that page is locked, the mapping is pinned.
549 * We're allowed to run sleeping lock_page() here because we know the caller has
552 static void handle_write_error(struct address_space *mapping,
553 struct page *page, int error)
556 if (page_mapping(page) == mapping)
557 mapping_set_error(mapping, error);
561 /* possible outcome of pageout() */
563 /* failed to write page out, page is locked */
565 /* move page to the active list, page is locked */
567 /* page has been sent to the disk successfully, page is unlocked */
569 /* page is clean and locked */
574 * pageout is called by shrink_page_list() for each dirty page.
575 * Calls ->writepage().
577 static pageout_t pageout(struct page *page, struct address_space *mapping,
578 struct scan_control *sc)
581 * If the page is dirty, only perform writeback if that write
582 * will be non-blocking. To prevent this allocation from being
583 * stalled by pagecache activity. But note that there may be
584 * stalls if we need to run get_block(). We could test
585 * PagePrivate for that.
587 * If this process is currently in __generic_file_write_iter() against
588 * this page's queue, we can perform writeback even if that
591 * If the page is swapcache, write it back even if that would
592 * block, for some throttling. This happens by accident, because
593 * swap_backing_dev_info is bust: it doesn't reflect the
594 * congestion state of the swapdevs. Easy to fix, if needed.
596 if (!is_page_cache_freeable(page))
600 * Some data journaling orphaned pages can have
601 * page->mapping == NULL while being dirty with clean buffers.
603 if (page_has_private(page)) {
604 if (try_to_free_buffers(page)) {
605 ClearPageDirty(page);
606 pr_info("%s: orphaned page\n", __func__);
612 if (mapping->a_ops->writepage == NULL)
613 return PAGE_ACTIVATE;
614 if (!may_write_to_inode(mapping->host, sc))
617 if (clear_page_dirty_for_io(page)) {
619 struct writeback_control wbc = {
620 .sync_mode = WB_SYNC_NONE,
621 .nr_to_write = SWAP_CLUSTER_MAX,
623 .range_end = LLONG_MAX,
627 SetPageReclaim(page);
628 res = mapping->a_ops->writepage(page, &wbc);
630 handle_write_error(mapping, page, res);
631 if (res == AOP_WRITEPAGE_ACTIVATE) {
632 ClearPageReclaim(page);
633 return PAGE_ACTIVATE;
636 if (!PageWriteback(page)) {
637 /* synchronous write or broken a_ops? */
638 ClearPageReclaim(page);
640 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
641 inc_zone_page_state(page, NR_VMSCAN_WRITE);
649 * Same as remove_mapping, but if the page is removed from the mapping, it
650 * gets returned with a refcount of 0.
652 static int __remove_mapping(struct address_space *mapping, struct page *page,
656 struct mem_cgroup *memcg;
658 BUG_ON(!PageLocked(page));
659 BUG_ON(mapping != page_mapping(page));
661 memcg = mem_cgroup_begin_page_stat(page);
662 spin_lock_irqsave(&mapping->tree_lock, flags);
664 * The non racy check for a busy page.
666 * Must be careful with the order of the tests. When someone has
667 * a ref to the page, it may be possible that they dirty it then
668 * drop the reference. So if PageDirty is tested before page_count
669 * here, then the following race may occur:
671 * get_user_pages(&page);
672 * [user mapping goes away]
674 * !PageDirty(page) [good]
675 * SetPageDirty(page);
677 * !page_count(page) [good, discard it]
679 * [oops, our write_to data is lost]
681 * Reversing the order of the tests ensures such a situation cannot
682 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
683 * load is not satisfied before that of page->_count.
685 * Note that if SetPageDirty is always performed via set_page_dirty,
686 * and thus under tree_lock, then this ordering is not required.
688 if (!page_freeze_refs(page, 2))
690 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
691 if (unlikely(PageDirty(page))) {
692 page_unfreeze_refs(page, 2);
696 if (PageSwapCache(page)) {
697 swp_entry_t swap = { .val = page_private(page) };
698 mem_cgroup_swapout(page, swap);
699 __delete_from_swap_cache(page);
700 spin_unlock_irqrestore(&mapping->tree_lock, flags);
701 mem_cgroup_end_page_stat(memcg);
702 swapcache_free(swap);
704 void (*freepage)(struct page *);
707 freepage = mapping->a_ops->freepage;
709 * Remember a shadow entry for reclaimed file cache in
710 * order to detect refaults, thus thrashing, later on.
712 * But don't store shadows in an address space that is
713 * already exiting. This is not just an optizimation,
714 * inode reclaim needs to empty out the radix tree or
715 * the nodes are lost. Don't plant shadows behind its
718 if (reclaimed && page_is_file_cache(page) &&
719 !mapping_exiting(mapping))
720 shadow = workingset_eviction(mapping, page);
721 __delete_from_page_cache(page, shadow, memcg);
722 spin_unlock_irqrestore(&mapping->tree_lock, flags);
723 mem_cgroup_end_page_stat(memcg);
725 if (freepage != NULL)
732 spin_unlock_irqrestore(&mapping->tree_lock, flags);
733 mem_cgroup_end_page_stat(memcg);
738 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
739 * someone else has a ref on the page, abort and return 0. If it was
740 * successfully detached, return 1. Assumes the caller has a single ref on
743 int remove_mapping(struct address_space *mapping, struct page *page)
745 if (__remove_mapping(mapping, page, false)) {
747 * Unfreezing the refcount with 1 rather than 2 effectively
748 * drops the pagecache ref for us without requiring another
751 page_unfreeze_refs(page, 1);
758 * putback_lru_page - put previously isolated page onto appropriate LRU list
759 * @page: page to be put back to appropriate lru list
761 * Add previously isolated @page to appropriate LRU list.
762 * Page may still be unevictable for other reasons.
764 * lru_lock must not be held, interrupts must be enabled.
766 void putback_lru_page(struct page *page)
769 int was_unevictable = PageUnevictable(page);
771 VM_BUG_ON_PAGE(PageLRU(page), page);
774 ClearPageUnevictable(page);
776 if (page_evictable(page)) {
778 * For evictable pages, we can use the cache.
779 * In event of a race, worst case is we end up with an
780 * unevictable page on [in]active list.
781 * We know how to handle that.
783 is_unevictable = false;
787 * Put unevictable pages directly on zone's unevictable
790 is_unevictable = true;
791 add_page_to_unevictable_list(page);
793 * When racing with an mlock or AS_UNEVICTABLE clearing
794 * (page is unlocked) make sure that if the other thread
795 * does not observe our setting of PG_lru and fails
796 * isolation/check_move_unevictable_pages,
797 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
798 * the page back to the evictable list.
800 * The other side is TestClearPageMlocked() or shmem_lock().
806 * page's status can change while we move it among lru. If an evictable
807 * page is on unevictable list, it never be freed. To avoid that,
808 * check after we added it to the list, again.
810 if (is_unevictable && page_evictable(page)) {
811 if (!isolate_lru_page(page)) {
815 /* This means someone else dropped this page from LRU
816 * So, it will be freed or putback to LRU again. There is
817 * nothing to do here.
821 if (was_unevictable && !is_unevictable)
822 count_vm_event(UNEVICTABLE_PGRESCUED);
823 else if (!was_unevictable && is_unevictable)
824 count_vm_event(UNEVICTABLE_PGCULLED);
826 put_page(page); /* drop ref from isolate */
829 enum page_references {
831 PAGEREF_RECLAIM_CLEAN,
836 static enum page_references page_check_references(struct page *page,
837 struct scan_control *sc)
839 int referenced_ptes, referenced_page;
840 unsigned long vm_flags;
842 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
844 referenced_page = TestClearPageReferenced(page);
847 * Mlock lost the isolation race with us. Let try_to_unmap()
848 * move the page to the unevictable list.
850 if (vm_flags & VM_LOCKED)
851 return PAGEREF_RECLAIM;
853 if (referenced_ptes) {
854 if (PageSwapBacked(page))
855 return PAGEREF_ACTIVATE;
857 * All mapped pages start out with page table
858 * references from the instantiating fault, so we need
859 * to look twice if a mapped file page is used more
862 * Mark it and spare it for another trip around the
863 * inactive list. Another page table reference will
864 * lead to its activation.
866 * Note: the mark is set for activated pages as well
867 * so that recently deactivated but used pages are
870 SetPageReferenced(page);
872 if (referenced_page || referenced_ptes > 1)
873 return PAGEREF_ACTIVATE;
876 * Activate file-backed executable pages after first usage.
878 if (vm_flags & VM_EXEC)
879 return PAGEREF_ACTIVATE;
884 /* Reclaim if clean, defer dirty pages to writeback */
885 if (referenced_page && !PageSwapBacked(page))
886 return PAGEREF_RECLAIM_CLEAN;
888 return PAGEREF_RECLAIM;
891 /* Check if a page is dirty or under writeback */
892 static void page_check_dirty_writeback(struct page *page,
893 bool *dirty, bool *writeback)
895 struct address_space *mapping;
898 * Anonymous pages are not handled by flushers and must be written
899 * from reclaim context. Do not stall reclaim based on them
901 if (!page_is_file_cache(page)) {
907 /* By default assume that the page flags are accurate */
908 *dirty = PageDirty(page);
909 *writeback = PageWriteback(page);
911 /* Verify dirty/writeback state if the filesystem supports it */
912 if (!page_has_private(page))
915 mapping = page_mapping(page);
916 if (mapping && mapping->a_ops->is_dirty_writeback)
917 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
921 * shrink_page_list() returns the number of reclaimed pages
923 static unsigned long shrink_page_list(struct list_head *page_list,
925 struct scan_control *sc,
926 enum ttu_flags ttu_flags,
927 unsigned long *ret_nr_dirty,
928 unsigned long *ret_nr_unqueued_dirty,
929 unsigned long *ret_nr_congested,
930 unsigned long *ret_nr_writeback,
931 unsigned long *ret_nr_immediate,
934 LIST_HEAD(ret_pages);
935 LIST_HEAD(free_pages);
937 unsigned long nr_unqueued_dirty = 0;
938 unsigned long nr_dirty = 0;
939 unsigned long nr_congested = 0;
940 unsigned long nr_reclaimed = 0;
941 unsigned long nr_writeback = 0;
942 unsigned long nr_immediate = 0;
946 while (!list_empty(page_list)) {
947 struct address_space *mapping;
950 enum page_references references = PAGEREF_RECLAIM_CLEAN;
951 bool dirty, writeback;
955 page = lru_to_page(page_list);
956 list_del(&page->lru);
958 if (!trylock_page(page))
961 VM_BUG_ON_PAGE(PageActive(page), page);
962 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
966 if (unlikely(!page_evictable(page)))
969 if (!sc->may_unmap && page_mapped(page))
972 /* Double the slab pressure for mapped and swapcache pages */
973 if (page_mapped(page) || PageSwapCache(page))
976 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
977 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
980 * The number of dirty pages determines if a zone is marked
981 * reclaim_congested which affects wait_iff_congested. kswapd
982 * will stall and start writing pages if the tail of the LRU
983 * is all dirty unqueued pages.
985 page_check_dirty_writeback(page, &dirty, &writeback);
986 if (dirty || writeback)
989 if (dirty && !writeback)
993 * Treat this page as congested if the underlying BDI is or if
994 * pages are cycling through the LRU so quickly that the
995 * pages marked for immediate reclaim are making it to the
996 * end of the LRU a second time.
998 mapping = page_mapping(page);
999 if (((dirty || writeback) && mapping &&
1000 inode_write_congested(mapping->host)) ||
1001 (writeback && PageReclaim(page)))
1005 * If a page at the tail of the LRU is under writeback, there
1006 * are three cases to consider.
1008 * 1) If reclaim is encountering an excessive number of pages
1009 * under writeback and this page is both under writeback and
1010 * PageReclaim then it indicates that pages are being queued
1011 * for IO but are being recycled through the LRU before the
1012 * IO can complete. Waiting on the page itself risks an
1013 * indefinite stall if it is impossible to writeback the
1014 * page due to IO error or disconnected storage so instead
1015 * note that the LRU is being scanned too quickly and the
1016 * caller can stall after page list has been processed.
1018 * 2) Global or new memcg reclaim encounters a page that is
1019 * not marked for immediate reclaim, or the caller does not
1020 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1021 * not to fs). In this case mark the page for immediate
1022 * reclaim and continue scanning.
1024 * Require may_enter_fs because we would wait on fs, which
1025 * may not have submitted IO yet. And the loop driver might
1026 * enter reclaim, and deadlock if it waits on a page for
1027 * which it is needed to do the write (loop masks off
1028 * __GFP_IO|__GFP_FS for this reason); but more thought
1029 * would probably show more reasons.
1031 * 3) Legacy memcg encounters a page that is already marked
1032 * PageReclaim. memcg does not have any dirty pages
1033 * throttling so we could easily OOM just because too many
1034 * pages are in writeback and there is nothing else to
1035 * reclaim. Wait for the writeback to complete.
1037 if (PageWriteback(page)) {
1039 if (current_is_kswapd() &&
1040 PageReclaim(page) &&
1041 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1046 } else if (sane_reclaim(sc) ||
1047 !PageReclaim(page) || !may_enter_fs) {
1049 * This is slightly racy - end_page_writeback()
1050 * might have just cleared PageReclaim, then
1051 * setting PageReclaim here end up interpreted
1052 * as PageReadahead - but that does not matter
1053 * enough to care. What we do want is for this
1054 * page to have PageReclaim set next time memcg
1055 * reclaim reaches the tests above, so it will
1056 * then wait_on_page_writeback() to avoid OOM;
1057 * and it's also appropriate in global reclaim.
1059 SetPageReclaim(page);
1066 wait_on_page_writeback(page);
1067 /* then go back and try same page again */
1068 list_add_tail(&page->lru, page_list);
1074 references = page_check_references(page, sc);
1076 switch (references) {
1077 case PAGEREF_ACTIVATE:
1078 goto activate_locked;
1081 case PAGEREF_RECLAIM:
1082 case PAGEREF_RECLAIM_CLEAN:
1083 ; /* try to reclaim the page below */
1087 * Anonymous process memory has backing store?
1088 * Try to allocate it some swap space here.
1090 if (PageAnon(page) && !PageSwapCache(page)) {
1091 if (!(sc->gfp_mask & __GFP_IO))
1093 if (!add_to_swap(page, page_list))
1094 goto activate_locked;
1097 /* Adding to swap updated mapping */
1098 mapping = page_mapping(page);
1102 * The page is mapped into the page tables of one or more
1103 * processes. Try to unmap it here.
1105 if (page_mapped(page) && mapping) {
1106 switch (try_to_unmap(page,
1107 ttu_flags|TTU_BATCH_FLUSH)) {
1109 goto activate_locked;
1115 ; /* try to free the page below */
1119 if (PageDirty(page)) {
1121 * Only kswapd can writeback filesystem pages to
1122 * avoid risk of stack overflow but only writeback
1123 * if many dirty pages have been encountered.
1125 if (page_is_file_cache(page) &&
1126 (!current_is_kswapd() ||
1127 !test_bit(ZONE_DIRTY, &zone->flags))) {
1129 * Immediately reclaim when written back.
1130 * Similar in principal to deactivate_page()
1131 * except we already have the page isolated
1132 * and know it's dirty
1134 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1135 SetPageReclaim(page);
1140 if (references == PAGEREF_RECLAIM_CLEAN)
1144 if (!sc->may_writepage)
1148 * Page is dirty. Flush the TLB if a writable entry
1149 * potentially exists to avoid CPU writes after IO
1150 * starts and then write it out here.
1152 try_to_unmap_flush_dirty();
1153 switch (pageout(page, mapping, sc)) {
1157 goto activate_locked;
1159 if (PageWriteback(page))
1161 if (PageDirty(page))
1165 * A synchronous write - probably a ramdisk. Go
1166 * ahead and try to reclaim the page.
1168 if (!trylock_page(page))
1170 if (PageDirty(page) || PageWriteback(page))
1172 mapping = page_mapping(page);
1174 ; /* try to free the page below */
1179 * If the page has buffers, try to free the buffer mappings
1180 * associated with this page. If we succeed we try to free
1183 * We do this even if the page is PageDirty().
1184 * try_to_release_page() does not perform I/O, but it is
1185 * possible for a page to have PageDirty set, but it is actually
1186 * clean (all its buffers are clean). This happens if the
1187 * buffers were written out directly, with submit_bh(). ext3
1188 * will do this, as well as the blockdev mapping.
1189 * try_to_release_page() will discover that cleanness and will
1190 * drop the buffers and mark the page clean - it can be freed.
1192 * Rarely, pages can have buffers and no ->mapping. These are
1193 * the pages which were not successfully invalidated in
1194 * truncate_complete_page(). We try to drop those buffers here
1195 * and if that worked, and the page is no longer mapped into
1196 * process address space (page_count == 1) it can be freed.
1197 * Otherwise, leave the page on the LRU so it is swappable.
1199 if (page_has_private(page)) {
1200 if (!try_to_release_page(page, sc->gfp_mask))
1201 goto activate_locked;
1202 if (!mapping && page_count(page) == 1) {
1204 if (put_page_testzero(page))
1208 * rare race with speculative reference.
1209 * the speculative reference will free
1210 * this page shortly, so we may
1211 * increment nr_reclaimed here (and
1212 * leave it off the LRU).
1220 if (!mapping || !__remove_mapping(mapping, page, true))
1224 * At this point, we have no other references and there is
1225 * no way to pick any more up (removed from LRU, removed
1226 * from pagecache). Can use non-atomic bitops now (and
1227 * we obviously don't have to worry about waking up a process
1228 * waiting on the page lock, because there are no references.
1230 __clear_page_locked(page);
1235 * Is there need to periodically free_page_list? It would
1236 * appear not as the counts should be low
1238 list_add(&page->lru, &free_pages);
1242 if (PageSwapCache(page))
1243 try_to_free_swap(page);
1245 list_add(&page->lru, &ret_pages);
1249 /* Not a candidate for swapping, so reclaim swap space. */
1250 if (PageSwapCache(page) && vm_swap_full())
1251 try_to_free_swap(page);
1252 VM_BUG_ON_PAGE(PageActive(page), page);
1253 SetPageActive(page);
1258 list_add(&page->lru, &ret_pages);
1259 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1262 mem_cgroup_uncharge_list(&free_pages);
1263 try_to_unmap_flush();
1264 free_hot_cold_page_list(&free_pages, true);
1266 list_splice(&ret_pages, page_list);
1267 count_vm_events(PGACTIVATE, pgactivate);
1269 *ret_nr_dirty += nr_dirty;
1270 *ret_nr_congested += nr_congested;
1271 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1272 *ret_nr_writeback += nr_writeback;
1273 *ret_nr_immediate += nr_immediate;
1274 return nr_reclaimed;
1277 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1278 struct list_head *page_list)
1280 struct scan_control sc = {
1281 .gfp_mask = GFP_KERNEL,
1282 .priority = DEF_PRIORITY,
1285 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1286 struct page *page, *next;
1287 LIST_HEAD(clean_pages);
1289 list_for_each_entry_safe(page, next, page_list, lru) {
1290 if (page_is_file_cache(page) && !PageDirty(page) &&
1291 !isolated_balloon_page(page)) {
1292 ClearPageActive(page);
1293 list_move(&page->lru, &clean_pages);
1297 ret = shrink_page_list(&clean_pages, zone, &sc,
1298 TTU_UNMAP|TTU_IGNORE_ACCESS,
1299 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1300 list_splice(&clean_pages, page_list);
1301 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1306 * Attempt to remove the specified page from its LRU. Only take this page
1307 * if it is of the appropriate PageActive status. Pages which are being
1308 * freed elsewhere are also ignored.
1310 * page: page to consider
1311 * mode: one of the LRU isolation modes defined above
1313 * returns 0 on success, -ve errno on failure.
1315 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1319 /* Only take pages on the LRU. */
1323 /* Compaction should not handle unevictable pages but CMA can do so */
1324 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1330 * To minimise LRU disruption, the caller can indicate that it only
1331 * wants to isolate pages it will be able to operate on without
1332 * blocking - clean pages for the most part.
1334 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1335 * is used by reclaim when it is cannot write to backing storage
1337 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1338 * that it is possible to migrate without blocking
1340 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1341 /* All the caller can do on PageWriteback is block */
1342 if (PageWriteback(page))
1345 if (PageDirty(page)) {
1346 struct address_space *mapping;
1348 /* ISOLATE_CLEAN means only clean pages */
1349 if (mode & ISOLATE_CLEAN)
1353 * Only pages without mappings or that have a
1354 * ->migratepage callback are possible to migrate
1357 mapping = page_mapping(page);
1358 if (mapping && !mapping->a_ops->migratepage)
1363 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1366 if (likely(get_page_unless_zero(page))) {
1368 * Be careful not to clear PageLRU until after we're
1369 * sure the page is not being freed elsewhere -- the
1370 * page release code relies on it.
1380 * zone->lru_lock is heavily contended. Some of the functions that
1381 * shrink the lists perform better by taking out a batch of pages
1382 * and working on them outside the LRU lock.
1384 * For pagecache intensive workloads, this function is the hottest
1385 * spot in the kernel (apart from copy_*_user functions).
1387 * Appropriate locks must be held before calling this function.
1389 * @nr_to_scan: The number of pages to look through on the list.
1390 * @lruvec: The LRU vector to pull pages from.
1391 * @dst: The temp list to put pages on to.
1392 * @nr_scanned: The number of pages that were scanned.
1393 * @sc: The scan_control struct for this reclaim session
1394 * @mode: One of the LRU isolation modes
1395 * @lru: LRU list id for isolating
1397 * returns how many pages were moved onto *@dst.
1399 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1400 struct lruvec *lruvec, struct list_head *dst,
1401 unsigned long *nr_scanned, struct scan_control *sc,
1402 isolate_mode_t mode, enum lru_list lru)
1404 struct list_head *src = &lruvec->lists[lru];
1405 unsigned long nr_taken = 0;
1408 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1409 !list_empty(src); scan++) {
1413 page = lru_to_page(src);
1414 prefetchw_prev_lru_page(page, src, flags);
1416 VM_BUG_ON_PAGE(!PageLRU(page), page);
1418 switch (__isolate_lru_page(page, mode)) {
1420 nr_pages = hpage_nr_pages(page);
1421 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1422 list_move(&page->lru, dst);
1423 nr_taken += nr_pages;
1427 /* else it is being freed elsewhere */
1428 list_move(&page->lru, src);
1437 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1438 nr_taken, mode, is_file_lru(lru));
1443 * isolate_lru_page - tries to isolate a page from its LRU list
1444 * @page: page to isolate from its LRU list
1446 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1447 * vmstat statistic corresponding to whatever LRU list the page was on.
1449 * Returns 0 if the page was removed from an LRU list.
1450 * Returns -EBUSY if the page was not on an LRU list.
1452 * The returned page will have PageLRU() cleared. If it was found on
1453 * the active list, it will have PageActive set. If it was found on
1454 * the unevictable list, it will have the PageUnevictable bit set. That flag
1455 * may need to be cleared by the caller before letting the page go.
1457 * The vmstat statistic corresponding to the list on which the page was
1458 * found will be decremented.
1461 * (1) Must be called with an elevated refcount on the page. This is a
1462 * fundamentnal difference from isolate_lru_pages (which is called
1463 * without a stable reference).
1464 * (2) the lru_lock must not be held.
1465 * (3) interrupts must be enabled.
1467 int isolate_lru_page(struct page *page)
1471 VM_BUG_ON_PAGE(!page_count(page), page);
1473 if (PageLRU(page)) {
1474 struct zone *zone = page_zone(page);
1475 struct lruvec *lruvec;
1477 spin_lock_irq(&zone->lru_lock);
1478 lruvec = mem_cgroup_page_lruvec(page, zone);
1479 if (PageLRU(page)) {
1480 int lru = page_lru(page);
1483 del_page_from_lru_list(page, lruvec, lru);
1486 spin_unlock_irq(&zone->lru_lock);
1492 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1493 * then get resheduled. When there are massive number of tasks doing page
1494 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1495 * the LRU list will go small and be scanned faster than necessary, leading to
1496 * unnecessary swapping, thrashing and OOM.
1498 static int too_many_isolated(struct zone *zone, int file,
1499 struct scan_control *sc)
1501 unsigned long inactive, isolated;
1503 if (current_is_kswapd())
1506 if (!sane_reclaim(sc))
1510 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1511 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1513 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1514 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1518 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1519 * won't get blocked by normal direct-reclaimers, forming a circular
1522 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1525 return isolated > inactive;
1528 static noinline_for_stack void
1529 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1531 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1532 struct zone *zone = lruvec_zone(lruvec);
1533 LIST_HEAD(pages_to_free);
1536 * Put back any unfreeable pages.
1538 while (!list_empty(page_list)) {
1539 struct page *page = lru_to_page(page_list);
1542 VM_BUG_ON_PAGE(PageLRU(page), page);
1543 list_del(&page->lru);
1544 if (unlikely(!page_evictable(page))) {
1545 spin_unlock_irq(&zone->lru_lock);
1546 putback_lru_page(page);
1547 spin_lock_irq(&zone->lru_lock);
1551 lruvec = mem_cgroup_page_lruvec(page, zone);
1554 lru = page_lru(page);
1555 add_page_to_lru_list(page, lruvec, lru);
1557 if (is_active_lru(lru)) {
1558 int file = is_file_lru(lru);
1559 int numpages = hpage_nr_pages(page);
1560 reclaim_stat->recent_rotated[file] += numpages;
1562 if (put_page_testzero(page)) {
1563 __ClearPageLRU(page);
1564 __ClearPageActive(page);
1565 del_page_from_lru_list(page, lruvec, lru);
1567 if (unlikely(PageCompound(page))) {
1568 spin_unlock_irq(&zone->lru_lock);
1569 mem_cgroup_uncharge(page);
1570 (*get_compound_page_dtor(page))(page);
1571 spin_lock_irq(&zone->lru_lock);
1573 list_add(&page->lru, &pages_to_free);
1578 * To save our caller's stack, now use input list for pages to free.
1580 list_splice(&pages_to_free, page_list);
1584 * If a kernel thread (such as nfsd for loop-back mounts) services
1585 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1586 * In that case we should only throttle if the backing device it is
1587 * writing to is congested. In other cases it is safe to throttle.
1589 static int current_may_throttle(void)
1591 return !(current->flags & PF_LESS_THROTTLE) ||
1592 current->backing_dev_info == NULL ||
1593 bdi_write_congested(current->backing_dev_info);
1597 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1598 * of reclaimed pages
1600 static noinline_for_stack unsigned long
1601 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1602 struct scan_control *sc, enum lru_list lru)
1604 LIST_HEAD(page_list);
1605 unsigned long nr_scanned;
1606 unsigned long nr_reclaimed = 0;
1607 unsigned long nr_taken;
1608 unsigned long nr_dirty = 0;
1609 unsigned long nr_congested = 0;
1610 unsigned long nr_unqueued_dirty = 0;
1611 unsigned long nr_writeback = 0;
1612 unsigned long nr_immediate = 0;
1613 isolate_mode_t isolate_mode = 0;
1614 int file = is_file_lru(lru);
1615 struct zone *zone = lruvec_zone(lruvec);
1616 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1618 while (unlikely(too_many_isolated(zone, file, sc))) {
1619 congestion_wait(BLK_RW_ASYNC, HZ/10);
1621 /* We are about to die and free our memory. Return now. */
1622 if (fatal_signal_pending(current))
1623 return SWAP_CLUSTER_MAX;
1629 isolate_mode |= ISOLATE_UNMAPPED;
1630 if (!sc->may_writepage)
1631 isolate_mode |= ISOLATE_CLEAN;
1633 spin_lock_irq(&zone->lru_lock);
1635 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1636 &nr_scanned, sc, isolate_mode, lru);
1638 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1639 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1641 if (global_reclaim(sc)) {
1642 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1643 if (current_is_kswapd())
1644 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1646 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1648 spin_unlock_irq(&zone->lru_lock);
1653 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1654 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1655 &nr_writeback, &nr_immediate,
1658 spin_lock_irq(&zone->lru_lock);
1660 reclaim_stat->recent_scanned[file] += nr_taken;
1662 if (global_reclaim(sc)) {
1663 if (current_is_kswapd())
1664 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1667 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1671 putback_inactive_pages(lruvec, &page_list);
1673 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1675 spin_unlock_irq(&zone->lru_lock);
1677 mem_cgroup_uncharge_list(&page_list);
1678 free_hot_cold_page_list(&page_list, true);
1681 * If reclaim is isolating dirty pages under writeback, it implies
1682 * that the long-lived page allocation rate is exceeding the page
1683 * laundering rate. Either the global limits are not being effective
1684 * at throttling processes due to the page distribution throughout
1685 * zones or there is heavy usage of a slow backing device. The
1686 * only option is to throttle from reclaim context which is not ideal
1687 * as there is no guarantee the dirtying process is throttled in the
1688 * same way balance_dirty_pages() manages.
1690 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1691 * of pages under pages flagged for immediate reclaim and stall if any
1692 * are encountered in the nr_immediate check below.
1694 if (nr_writeback && nr_writeback == nr_taken)
1695 set_bit(ZONE_WRITEBACK, &zone->flags);
1698 * Legacy memcg will stall in page writeback so avoid forcibly
1701 if (sane_reclaim(sc)) {
1703 * Tag a zone as congested if all the dirty pages scanned were
1704 * backed by a congested BDI and wait_iff_congested will stall.
1706 if (nr_dirty && nr_dirty == nr_congested)
1707 set_bit(ZONE_CONGESTED, &zone->flags);
1710 * If dirty pages are scanned that are not queued for IO, it
1711 * implies that flushers are not keeping up. In this case, flag
1712 * the zone ZONE_DIRTY and kswapd will start writing pages from
1715 if (nr_unqueued_dirty == nr_taken)
1716 set_bit(ZONE_DIRTY, &zone->flags);
1719 * If kswapd scans pages marked marked for immediate
1720 * reclaim and under writeback (nr_immediate), it implies
1721 * that pages are cycling through the LRU faster than
1722 * they are written so also forcibly stall.
1724 if (nr_immediate && current_may_throttle())
1725 congestion_wait(BLK_RW_ASYNC, HZ/10);
1729 * Stall direct reclaim for IO completions if underlying BDIs or zone
1730 * is congested. Allow kswapd to continue until it starts encountering
1731 * unqueued dirty pages or cycling through the LRU too quickly.
1733 if (!sc->hibernation_mode && !current_is_kswapd() &&
1734 current_may_throttle())
1735 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1737 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1739 nr_scanned, nr_reclaimed,
1741 trace_shrink_flags(file));
1742 return nr_reclaimed;
1746 * This moves pages from the active list to the inactive list.
1748 * We move them the other way if the page is referenced by one or more
1749 * processes, from rmap.
1751 * If the pages are mostly unmapped, the processing is fast and it is
1752 * appropriate to hold zone->lru_lock across the whole operation. But if
1753 * the pages are mapped, the processing is slow (page_referenced()) so we
1754 * should drop zone->lru_lock around each page. It's impossible to balance
1755 * this, so instead we remove the pages from the LRU while processing them.
1756 * It is safe to rely on PG_active against the non-LRU pages in here because
1757 * nobody will play with that bit on a non-LRU page.
1759 * The downside is that we have to touch page->_count against each page.
1760 * But we had to alter page->flags anyway.
1763 static void move_active_pages_to_lru(struct lruvec *lruvec,
1764 struct list_head *list,
1765 struct list_head *pages_to_free,
1768 struct zone *zone = lruvec_zone(lruvec);
1769 unsigned long pgmoved = 0;
1773 while (!list_empty(list)) {
1774 page = lru_to_page(list);
1775 lruvec = mem_cgroup_page_lruvec(page, zone);
1777 VM_BUG_ON_PAGE(PageLRU(page), page);
1780 nr_pages = hpage_nr_pages(page);
1781 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1782 list_move(&page->lru, &lruvec->lists[lru]);
1783 pgmoved += nr_pages;
1785 if (put_page_testzero(page)) {
1786 __ClearPageLRU(page);
1787 __ClearPageActive(page);
1788 del_page_from_lru_list(page, lruvec, lru);
1790 if (unlikely(PageCompound(page))) {
1791 spin_unlock_irq(&zone->lru_lock);
1792 mem_cgroup_uncharge(page);
1793 (*get_compound_page_dtor(page))(page);
1794 spin_lock_irq(&zone->lru_lock);
1796 list_add(&page->lru, pages_to_free);
1799 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1800 if (!is_active_lru(lru))
1801 __count_vm_events(PGDEACTIVATE, pgmoved);
1804 static void shrink_active_list(unsigned long nr_to_scan,
1805 struct lruvec *lruvec,
1806 struct scan_control *sc,
1809 unsigned long nr_taken;
1810 unsigned long nr_scanned;
1811 unsigned long vm_flags;
1812 LIST_HEAD(l_hold); /* The pages which were snipped off */
1813 LIST_HEAD(l_active);
1814 LIST_HEAD(l_inactive);
1816 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1817 unsigned long nr_rotated = 0;
1818 isolate_mode_t isolate_mode = 0;
1819 int file = is_file_lru(lru);
1820 struct zone *zone = lruvec_zone(lruvec);
1825 isolate_mode |= ISOLATE_UNMAPPED;
1826 if (!sc->may_writepage)
1827 isolate_mode |= ISOLATE_CLEAN;
1829 spin_lock_irq(&zone->lru_lock);
1831 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1832 &nr_scanned, sc, isolate_mode, lru);
1833 if (global_reclaim(sc))
1834 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1836 reclaim_stat->recent_scanned[file] += nr_taken;
1838 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1839 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1840 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1841 spin_unlock_irq(&zone->lru_lock);
1843 while (!list_empty(&l_hold)) {
1845 page = lru_to_page(&l_hold);
1846 list_del(&page->lru);
1848 if (unlikely(!page_evictable(page))) {
1849 putback_lru_page(page);
1853 if (unlikely(buffer_heads_over_limit)) {
1854 if (page_has_private(page) && trylock_page(page)) {
1855 if (page_has_private(page))
1856 try_to_release_page(page, 0);
1861 if (page_referenced(page, 0, sc->target_mem_cgroup,
1863 nr_rotated += hpage_nr_pages(page);
1865 * Identify referenced, file-backed active pages and
1866 * give them one more trip around the active list. So
1867 * that executable code get better chances to stay in
1868 * memory under moderate memory pressure. Anon pages
1869 * are not likely to be evicted by use-once streaming
1870 * IO, plus JVM can create lots of anon VM_EXEC pages,
1871 * so we ignore them here.
1873 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1874 list_add(&page->lru, &l_active);
1879 ClearPageActive(page); /* we are de-activating */
1880 list_add(&page->lru, &l_inactive);
1884 * Move pages back to the lru list.
1886 spin_lock_irq(&zone->lru_lock);
1888 * Count referenced pages from currently used mappings as rotated,
1889 * even though only some of them are actually re-activated. This
1890 * helps balance scan pressure between file and anonymous pages in
1893 reclaim_stat->recent_rotated[file] += nr_rotated;
1895 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1896 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1897 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1898 spin_unlock_irq(&zone->lru_lock);
1900 mem_cgroup_uncharge_list(&l_hold);
1901 free_hot_cold_page_list(&l_hold, true);
1905 static bool inactive_anon_is_low_global(struct zone *zone)
1907 unsigned long active, inactive;
1909 active = zone_page_state(zone, NR_ACTIVE_ANON);
1910 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1912 return inactive * zone->inactive_ratio < active;
1916 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1917 * @lruvec: LRU vector to check
1919 * Returns true if the zone does not have enough inactive anon pages,
1920 * meaning some active anon pages need to be deactivated.
1922 static bool inactive_anon_is_low(struct lruvec *lruvec)
1925 * If we don't have swap space, anonymous page deactivation
1928 if (!total_swap_pages)
1931 if (!mem_cgroup_disabled())
1932 return mem_cgroup_inactive_anon_is_low(lruvec);
1934 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1937 static inline bool inactive_anon_is_low(struct lruvec *lruvec)
1944 * inactive_file_is_low - check if file pages need to be deactivated
1945 * @lruvec: LRU vector to check
1947 * When the system is doing streaming IO, memory pressure here
1948 * ensures that active file pages get deactivated, until more
1949 * than half of the file pages are on the inactive list.
1951 * Once we get to that situation, protect the system's working
1952 * set from being evicted by disabling active file page aging.
1954 * This uses a different ratio than the anonymous pages, because
1955 * the page cache uses a use-once replacement algorithm.
1957 static bool inactive_file_is_low(struct lruvec *lruvec)
1959 unsigned long inactive;
1960 unsigned long active;
1962 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1963 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1965 return active > inactive;
1968 static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1970 if (is_file_lru(lru))
1971 return inactive_file_is_low(lruvec);
1973 return inactive_anon_is_low(lruvec);
1976 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1977 struct lruvec *lruvec, struct scan_control *sc)
1979 if (is_active_lru(lru)) {
1980 if (inactive_list_is_low(lruvec, lru))
1981 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1985 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1996 * Determine how aggressively the anon and file LRU lists should be
1997 * scanned. The relative value of each set of LRU lists is determined
1998 * by looking at the fraction of the pages scanned we did rotate back
1999 * onto the active list instead of evict.
2001 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2002 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2004 static void get_scan_count(struct lruvec *lruvec, int swappiness,
2005 struct scan_control *sc, unsigned long *nr,
2006 unsigned long *lru_pages)
2008 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2010 u64 denominator = 0; /* gcc */
2011 struct zone *zone = lruvec_zone(lruvec);
2012 unsigned long anon_prio, file_prio;
2013 enum scan_balance scan_balance;
2014 unsigned long anon, file;
2015 bool force_scan = false;
2016 unsigned long ap, fp;
2022 * If the zone or memcg is small, nr[l] can be 0. This
2023 * results in no scanning on this priority and a potential
2024 * priority drop. Global direct reclaim can go to the next
2025 * zone and tends to have no problems. Global kswapd is for
2026 * zone balancing and it needs to scan a minimum amount. When
2027 * reclaiming for a memcg, a priority drop can cause high
2028 * latencies, so it's better to scan a minimum amount there as
2031 if (current_is_kswapd()) {
2032 if (!zone_reclaimable(zone))
2034 if (!mem_cgroup_lruvec_online(lruvec))
2037 if (!global_reclaim(sc))
2040 /* If we have no swap space, do not bother scanning anon pages. */
2041 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
2042 scan_balance = SCAN_FILE;
2047 * Global reclaim will swap to prevent OOM even with no
2048 * swappiness, but memcg users want to use this knob to
2049 * disable swapping for individual groups completely when
2050 * using the memory controller's swap limit feature would be
2053 if (!global_reclaim(sc) && !swappiness) {
2054 scan_balance = SCAN_FILE;
2059 * Do not apply any pressure balancing cleverness when the
2060 * system is close to OOM, scan both anon and file equally
2061 * (unless the swappiness setting disagrees with swapping).
2063 if (!sc->priority && swappiness) {
2064 scan_balance = SCAN_EQUAL;
2069 * Prevent the reclaimer from falling into the cache trap: as
2070 * cache pages start out inactive, every cache fault will tip
2071 * the scan balance towards the file LRU. And as the file LRU
2072 * shrinks, so does the window for rotation from references.
2073 * This means we have a runaway feedback loop where a tiny
2074 * thrashing file LRU becomes infinitely more attractive than
2075 * anon pages. Try to detect this based on file LRU size.
2077 if (global_reclaim(sc)) {
2078 unsigned long zonefile;
2079 unsigned long zonefree;
2081 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2082 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2083 zone_page_state(zone, NR_INACTIVE_FILE);
2085 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2086 scan_balance = SCAN_ANON;
2092 * There is enough inactive page cache, do not reclaim
2093 * anything from the anonymous working set right now.
2095 if (!inactive_file_is_low(lruvec)) {
2096 scan_balance = SCAN_FILE;
2100 scan_balance = SCAN_FRACT;
2103 * With swappiness at 100, anonymous and file have the same priority.
2104 * This scanning priority is essentially the inverse of IO cost.
2106 anon_prio = swappiness;
2107 file_prio = 200 - anon_prio;
2110 * OK, so we have swap space and a fair amount of page cache
2111 * pages. We use the recently rotated / recently scanned
2112 * ratios to determine how valuable each cache is.
2114 * Because workloads change over time (and to avoid overflow)
2115 * we keep these statistics as a floating average, which ends
2116 * up weighing recent references more than old ones.
2118 * anon in [0], file in [1]
2121 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2122 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2123 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2124 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2126 spin_lock_irq(&zone->lru_lock);
2127 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2128 reclaim_stat->recent_scanned[0] /= 2;
2129 reclaim_stat->recent_rotated[0] /= 2;
2132 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2133 reclaim_stat->recent_scanned[1] /= 2;
2134 reclaim_stat->recent_rotated[1] /= 2;
2138 * The amount of pressure on anon vs file pages is inversely
2139 * proportional to the fraction of recently scanned pages on
2140 * each list that were recently referenced and in active use.
2142 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2143 ap /= reclaim_stat->recent_rotated[0] + 1;
2145 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2146 fp /= reclaim_stat->recent_rotated[1] + 1;
2147 spin_unlock_irq(&zone->lru_lock);
2151 denominator = ap + fp + 1;
2153 some_scanned = false;
2154 /* Only use force_scan on second pass. */
2155 for (pass = 0; !some_scanned && pass < 2; pass++) {
2157 for_each_evictable_lru(lru) {
2158 int file = is_file_lru(lru);
2162 size = get_lru_size(lruvec, lru);
2163 scan = size >> sc->priority;
2165 if (!scan && pass && force_scan)
2166 scan = min(size, SWAP_CLUSTER_MAX);
2168 switch (scan_balance) {
2170 /* Scan lists relative to size */
2174 * Scan types proportional to swappiness and
2175 * their relative recent reclaim efficiency.
2177 scan = div64_u64(scan * fraction[file],
2182 /* Scan one type exclusively */
2183 if ((scan_balance == SCAN_FILE) != file) {
2189 /* Look ma, no brain */
2197 * Skip the second pass and don't force_scan,
2198 * if we found something to scan.
2200 some_scanned |= !!scan;
2205 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2206 static void init_tlb_ubc(void)
2209 * This deliberately does not clear the cpumask as it's expensive
2210 * and unnecessary. If there happens to be data in there then the
2211 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2212 * then will be cleared.
2214 current->tlb_ubc.flush_required = false;
2217 static inline void init_tlb_ubc(void)
2220 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2223 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2225 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2226 struct scan_control *sc, unsigned long *lru_pages)
2228 unsigned long nr[NR_LRU_LISTS];
2229 unsigned long targets[NR_LRU_LISTS];
2230 unsigned long nr_to_scan;
2232 unsigned long nr_reclaimed = 0;
2233 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2234 struct blk_plug plug;
2237 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2239 /* Record the original scan target for proportional adjustments later */
2240 memcpy(targets, nr, sizeof(nr));
2243 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2244 * event that can occur when there is little memory pressure e.g.
2245 * multiple streaming readers/writers. Hence, we do not abort scanning
2246 * when the requested number of pages are reclaimed when scanning at
2247 * DEF_PRIORITY on the assumption that the fact we are direct
2248 * reclaiming implies that kswapd is not keeping up and it is best to
2249 * do a batch of work at once. For memcg reclaim one check is made to
2250 * abort proportional reclaim if either the file or anon lru has already
2251 * dropped to zero at the first pass.
2253 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2254 sc->priority == DEF_PRIORITY);
2258 blk_start_plug(&plug);
2259 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2260 nr[LRU_INACTIVE_FILE]) {
2261 unsigned long nr_anon, nr_file, percentage;
2262 unsigned long nr_scanned;
2264 for_each_evictable_lru(lru) {
2266 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2267 nr[lru] -= nr_to_scan;
2269 nr_reclaimed += shrink_list(lru, nr_to_scan,
2274 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2278 * For kswapd and memcg, reclaim at least the number of pages
2279 * requested. Ensure that the anon and file LRUs are scanned
2280 * proportionally what was requested by get_scan_count(). We
2281 * stop reclaiming one LRU and reduce the amount scanning
2282 * proportional to the original scan target.
2284 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2285 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2288 * It's just vindictive to attack the larger once the smaller
2289 * has gone to zero. And given the way we stop scanning the
2290 * smaller below, this makes sure that we only make one nudge
2291 * towards proportionality once we've got nr_to_reclaim.
2293 if (!nr_file || !nr_anon)
2296 if (nr_file > nr_anon) {
2297 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2298 targets[LRU_ACTIVE_ANON] + 1;
2300 percentage = nr_anon * 100 / scan_target;
2302 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2303 targets[LRU_ACTIVE_FILE] + 1;
2305 percentage = nr_file * 100 / scan_target;
2308 /* Stop scanning the smaller of the LRU */
2310 nr[lru + LRU_ACTIVE] = 0;
2313 * Recalculate the other LRU scan count based on its original
2314 * scan target and the percentage scanning already complete
2316 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2317 nr_scanned = targets[lru] - nr[lru];
2318 nr[lru] = targets[lru] * (100 - percentage) / 100;
2319 nr[lru] -= min(nr[lru], nr_scanned);
2322 nr_scanned = targets[lru] - nr[lru];
2323 nr[lru] = targets[lru] * (100 - percentage) / 100;
2324 nr[lru] -= min(nr[lru], nr_scanned);
2326 scan_adjusted = true;
2328 blk_finish_plug(&plug);
2329 sc->nr_reclaimed += nr_reclaimed;
2332 * Even if we did not try to evict anon pages at all, we want to
2333 * rebalance the anon lru active/inactive ratio.
2335 if (inactive_anon_is_low(lruvec))
2336 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2337 sc, LRU_ACTIVE_ANON);
2339 throttle_vm_writeout(sc->gfp_mask);
2342 /* Use reclaim/compaction for costly allocs or under memory pressure */
2343 static bool in_reclaim_compaction(struct scan_control *sc)
2345 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2346 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2347 sc->priority < DEF_PRIORITY - 2))
2354 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2355 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2356 * true if more pages should be reclaimed such that when the page allocator
2357 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2358 * It will give up earlier than that if there is difficulty reclaiming pages.
2360 static inline bool should_continue_reclaim(struct zone *zone,
2361 unsigned long nr_reclaimed,
2362 unsigned long nr_scanned,
2363 struct scan_control *sc)
2365 unsigned long pages_for_compaction;
2366 unsigned long inactive_lru_pages;
2368 /* If not in reclaim/compaction mode, stop */
2369 if (!in_reclaim_compaction(sc))
2372 /* Consider stopping depending on scan and reclaim activity */
2373 if (sc->gfp_mask & __GFP_REPEAT) {
2375 * For __GFP_REPEAT allocations, stop reclaiming if the
2376 * full LRU list has been scanned and we are still failing
2377 * to reclaim pages. This full LRU scan is potentially
2378 * expensive but a __GFP_REPEAT caller really wants to succeed
2380 if (!nr_reclaimed && !nr_scanned)
2384 * For non-__GFP_REPEAT allocations which can presumably
2385 * fail without consequence, stop if we failed to reclaim
2386 * any pages from the last SWAP_CLUSTER_MAX number of
2387 * pages that were scanned. This will return to the
2388 * caller faster at the risk reclaim/compaction and
2389 * the resulting allocation attempt fails
2396 * If we have not reclaimed enough pages for compaction and the
2397 * inactive lists are large enough, continue reclaiming
2399 pages_for_compaction = (2UL << sc->order);
2400 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2401 if (get_nr_swap_pages() > 0)
2402 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2403 if (sc->nr_reclaimed < pages_for_compaction &&
2404 inactive_lru_pages > pages_for_compaction)
2407 /* If compaction would go ahead or the allocation would succeed, stop */
2408 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2409 case COMPACT_PARTIAL:
2410 case COMPACT_CONTINUE:
2417 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2420 struct reclaim_state *reclaim_state = current->reclaim_state;
2421 unsigned long nr_reclaimed, nr_scanned;
2422 bool reclaimable = false;
2425 struct mem_cgroup *root = sc->target_mem_cgroup;
2426 struct mem_cgroup_reclaim_cookie reclaim = {
2428 .priority = sc->priority,
2430 unsigned long zone_lru_pages = 0;
2431 struct mem_cgroup *memcg;
2433 nr_reclaimed = sc->nr_reclaimed;
2434 nr_scanned = sc->nr_scanned;
2436 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2438 unsigned long lru_pages;
2439 unsigned long scanned;
2440 struct lruvec *lruvec;
2443 if (mem_cgroup_low(root, memcg)) {
2444 if (!sc->may_thrash)
2446 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2449 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2450 swappiness = mem_cgroup_swappiness(memcg);
2451 scanned = sc->nr_scanned;
2453 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2454 zone_lru_pages += lru_pages;
2456 if (memcg && is_classzone)
2457 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2458 memcg, sc->nr_scanned - scanned,
2462 * Direct reclaim and kswapd have to scan all memory
2463 * cgroups to fulfill the overall scan target for the
2466 * Limit reclaim, on the other hand, only cares about
2467 * nr_to_reclaim pages to be reclaimed and it will
2468 * retry with decreasing priority if one round over the
2469 * whole hierarchy is not sufficient.
2471 if (!global_reclaim(sc) &&
2472 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2473 mem_cgroup_iter_break(root, memcg);
2476 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2479 * Shrink the slab caches in the same proportion that
2480 * the eligible LRU pages were scanned.
2482 if (global_reclaim(sc) && is_classzone)
2483 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2484 sc->nr_scanned - nr_scanned,
2487 if (reclaim_state) {
2488 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2489 reclaim_state->reclaimed_slab = 0;
2492 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2493 sc->nr_scanned - nr_scanned,
2494 sc->nr_reclaimed - nr_reclaimed);
2496 if (sc->nr_reclaimed - nr_reclaimed)
2499 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2500 sc->nr_scanned - nr_scanned, sc));
2506 * Returns true if compaction should go ahead for a high-order request, or
2507 * the high-order allocation would succeed without compaction.
2509 static inline bool compaction_ready(struct zone *zone, int order)
2511 unsigned long balance_gap, watermark;
2515 * Compaction takes time to run and there are potentially other
2516 * callers using the pages just freed. Continue reclaiming until
2517 * there is a buffer of free pages available to give compaction
2518 * a reasonable chance of completing and allocating the page
2520 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2521 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2522 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2523 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2526 * If compaction is deferred, reclaim up to a point where
2527 * compaction will have a chance of success when re-enabled
2529 if (compaction_deferred(zone, order))
2530 return watermark_ok;
2533 * If compaction is not ready to start and allocation is not likely
2534 * to succeed without it, then keep reclaiming.
2536 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2539 return watermark_ok;
2543 * This is the direct reclaim path, for page-allocating processes. We only
2544 * try to reclaim pages from zones which will satisfy the caller's allocation
2547 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2549 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2551 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2552 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2553 * zone defense algorithm.
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 * Returns true if a zone was reclaimable.
2560 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2564 unsigned long nr_soft_reclaimed;
2565 unsigned long nr_soft_scanned;
2567 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2568 bool reclaimable = false;
2571 * If the number of buffer_heads in the machine exceeds the maximum
2572 * allowed level, force direct reclaim to scan the highmem zone as
2573 * highmem pages could be pinning lowmem pages storing buffer_heads
2575 orig_mask = sc->gfp_mask;
2576 if (buffer_heads_over_limit)
2577 sc->gfp_mask |= __GFP_HIGHMEM;
2579 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2580 requested_highidx, sc->nodemask) {
2581 enum zone_type classzone_idx;
2583 if (!populated_zone(zone))
2586 classzone_idx = requested_highidx;
2587 while (!populated_zone(zone->zone_pgdat->node_zones +
2592 * Take care memory controller reclaiming has small influence
2595 if (global_reclaim(sc)) {
2596 if (!cpuset_zone_allowed(zone,
2597 GFP_KERNEL | __GFP_HARDWALL))
2600 if (sc->priority != DEF_PRIORITY &&
2601 !zone_reclaimable(zone))
2602 continue; /* Let kswapd poll it */
2605 * If we already have plenty of memory free for
2606 * compaction in this zone, don't free any more.
2607 * Even though compaction is invoked for any
2608 * non-zero order, only frequent costly order
2609 * reclamation is disruptive enough to become a
2610 * noticeable problem, like transparent huge
2613 if (IS_ENABLED(CONFIG_COMPACTION) &&
2614 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2615 zonelist_zone_idx(z) <= requested_highidx &&
2616 compaction_ready(zone, sc->order)) {
2617 sc->compaction_ready = true;
2622 * This steals pages from memory cgroups over softlimit
2623 * and returns the number of reclaimed pages and
2624 * scanned pages. This works for global memory pressure
2625 * and balancing, not for a memcg's limit.
2627 nr_soft_scanned = 0;
2628 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2629 sc->order, sc->gfp_mask,
2631 sc->nr_reclaimed += nr_soft_reclaimed;
2632 sc->nr_scanned += nr_soft_scanned;
2633 if (nr_soft_reclaimed)
2635 /* need some check for avoid more shrink_zone() */
2638 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2641 if (global_reclaim(sc) &&
2642 !reclaimable && zone_reclaimable(zone))
2647 * Restore to original mask to avoid the impact on the caller if we
2648 * promoted it to __GFP_HIGHMEM.
2650 sc->gfp_mask = orig_mask;
2656 * This is the main entry point to direct page reclaim.
2658 * If a full scan of the inactive list fails to free enough memory then we
2659 * are "out of memory" and something needs to be killed.
2661 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2662 * high - the zone may be full of dirty or under-writeback pages, which this
2663 * caller can't do much about. We kick the writeback threads and take explicit
2664 * naps in the hope that some of these pages can be written. But if the
2665 * allocating task holds filesystem locks which prevent writeout this might not
2666 * work, and the allocation attempt will fail.
2668 * returns: 0, if no pages reclaimed
2669 * else, the number of pages reclaimed
2671 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2672 struct scan_control *sc)
2674 int initial_priority = sc->priority;
2675 unsigned long total_scanned = 0;
2676 unsigned long writeback_threshold;
2677 bool zones_reclaimable;
2679 delayacct_freepages_start();
2681 if (global_reclaim(sc))
2682 count_vm_event(ALLOCSTALL);
2685 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2688 zones_reclaimable = shrink_zones(zonelist, sc);
2690 total_scanned += sc->nr_scanned;
2691 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2694 if (sc->compaction_ready)
2698 * If we're getting trouble reclaiming, start doing
2699 * writepage even in laptop mode.
2701 if (sc->priority < DEF_PRIORITY - 2)
2702 sc->may_writepage = 1;
2705 * Try to write back as many pages as we just scanned. This
2706 * tends to cause slow streaming writers to write data to the
2707 * disk smoothly, at the dirtying rate, which is nice. But
2708 * that's undesirable in laptop mode, where we *want* lumpy
2709 * writeout. So in laptop mode, write out the whole world.
2711 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2712 if (total_scanned > writeback_threshold) {
2713 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2714 WB_REASON_TRY_TO_FREE_PAGES);
2715 sc->may_writepage = 1;
2717 } while (--sc->priority >= 0);
2719 delayacct_freepages_end();
2721 if (sc->nr_reclaimed)
2722 return sc->nr_reclaimed;
2724 /* Aborted reclaim to try compaction? don't OOM, then */
2725 if (sc->compaction_ready)
2728 /* Untapped cgroup reserves? Don't OOM, retry. */
2729 if (!sc->may_thrash) {
2730 sc->priority = initial_priority;
2735 /* Any of the zones still reclaimable? Don't OOM. */
2736 if (zones_reclaimable)
2742 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2745 unsigned long pfmemalloc_reserve = 0;
2746 unsigned long free_pages = 0;
2750 for (i = 0; i <= ZONE_NORMAL; i++) {
2751 zone = &pgdat->node_zones[i];
2752 if (!populated_zone(zone) ||
2753 zone_reclaimable_pages(zone) == 0)
2756 pfmemalloc_reserve += min_wmark_pages(zone);
2757 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2760 /* If there are no reserves (unexpected config) then do not throttle */
2761 if (!pfmemalloc_reserve)
2764 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2766 /* kswapd must be awake if processes are being throttled */
2767 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2768 pgdat->classzone_idx = min(pgdat->classzone_idx,
2769 (enum zone_type)ZONE_NORMAL);
2770 wake_up_interruptible(&pgdat->kswapd_wait);
2777 * Throttle direct reclaimers if backing storage is backed by the network
2778 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2779 * depleted. kswapd will continue to make progress and wake the processes
2780 * when the low watermark is reached.
2782 * Returns true if a fatal signal was delivered during throttling. If this
2783 * happens, the page allocator should not consider triggering the OOM killer.
2785 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2786 nodemask_t *nodemask)
2790 pg_data_t *pgdat = NULL;
2793 * Kernel threads should not be throttled as they may be indirectly
2794 * responsible for cleaning pages necessary for reclaim to make forward
2795 * progress. kjournald for example may enter direct reclaim while
2796 * committing a transaction where throttling it could forcing other
2797 * processes to block on log_wait_commit().
2799 if (current->flags & PF_KTHREAD)
2803 * If a fatal signal is pending, this process should not throttle.
2804 * It should return quickly so it can exit and free its memory
2806 if (fatal_signal_pending(current))
2810 * Check if the pfmemalloc reserves are ok by finding the first node
2811 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2812 * GFP_KERNEL will be required for allocating network buffers when
2813 * swapping over the network so ZONE_HIGHMEM is unusable.
2815 * Throttling is based on the first usable node and throttled processes
2816 * wait on a queue until kswapd makes progress and wakes them. There
2817 * is an affinity then between processes waking up and where reclaim
2818 * progress has been made assuming the process wakes on the same node.
2819 * More importantly, processes running on remote nodes will not compete
2820 * for remote pfmemalloc reserves and processes on different nodes
2821 * should make reasonable progress.
2823 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2824 gfp_zone(gfp_mask), nodemask) {
2825 if (zone_idx(zone) > ZONE_NORMAL)
2828 /* Throttle based on the first usable node */
2829 pgdat = zone->zone_pgdat;
2830 if (pfmemalloc_watermark_ok(pgdat))
2835 /* If no zone was usable by the allocation flags then do not throttle */
2839 /* Account for the throttling */
2840 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2843 * If the caller cannot enter the filesystem, it's possible that it
2844 * is due to the caller holding an FS lock or performing a journal
2845 * transaction in the case of a filesystem like ext[3|4]. In this case,
2846 * it is not safe to block on pfmemalloc_wait as kswapd could be
2847 * blocked waiting on the same lock. Instead, throttle for up to a
2848 * second before continuing.
2850 if (!(gfp_mask & __GFP_FS)) {
2851 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2852 pfmemalloc_watermark_ok(pgdat), HZ);
2857 /* Throttle until kswapd wakes the process */
2858 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2859 pfmemalloc_watermark_ok(pgdat));
2862 if (fatal_signal_pending(current))
2869 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2870 gfp_t gfp_mask, nodemask_t *nodemask)
2872 unsigned long nr_reclaimed;
2873 struct scan_control sc = {
2874 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2875 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2877 .nodemask = nodemask,
2878 .priority = DEF_PRIORITY,
2879 .may_writepage = !laptop_mode,
2885 * Do not enter reclaim if fatal signal was delivered while throttled.
2886 * 1 is returned so that the page allocator does not OOM kill at this
2889 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2892 trace_mm_vmscan_direct_reclaim_begin(order,
2896 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2898 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2900 return nr_reclaimed;
2905 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2906 gfp_t gfp_mask, bool noswap,
2908 unsigned long *nr_scanned)
2910 struct scan_control sc = {
2911 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2912 .target_mem_cgroup = memcg,
2913 .may_writepage = !laptop_mode,
2915 .may_swap = !noswap,
2917 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2918 int swappiness = mem_cgroup_swappiness(memcg);
2919 unsigned long lru_pages;
2921 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2922 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2924 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2929 * NOTE: Although we can get the priority field, using it
2930 * here is not a good idea, since it limits the pages we can scan.
2931 * if we don't reclaim here, the shrink_zone from balance_pgdat
2932 * will pick up pages from other mem cgroup's as well. We hack
2933 * the priority and make it zero.
2935 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
2937 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2939 *nr_scanned = sc.nr_scanned;
2940 return sc.nr_reclaimed;
2943 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2944 unsigned long nr_pages,
2948 struct zonelist *zonelist;
2949 unsigned long nr_reclaimed;
2951 struct scan_control sc = {
2952 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2953 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2954 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2955 .target_mem_cgroup = memcg,
2956 .priority = DEF_PRIORITY,
2957 .may_writepage = !laptop_mode,
2959 .may_swap = may_swap,
2963 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2964 * take care of from where we get pages. So the node where we start the
2965 * scan does not need to be the current node.
2967 nid = mem_cgroup_select_victim_node(memcg);
2969 zonelist = NODE_DATA(nid)->node_zonelists;
2971 trace_mm_vmscan_memcg_reclaim_begin(0,
2975 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2977 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2979 return nr_reclaimed;
2983 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2985 struct mem_cgroup *memcg;
2987 if (!total_swap_pages)
2990 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2992 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2994 if (inactive_anon_is_low(lruvec))
2995 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2996 sc, LRU_ACTIVE_ANON);
2998 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3002 static bool zone_balanced(struct zone *zone, int order,
3003 unsigned long balance_gap, int classzone_idx)
3005 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
3006 balance_gap, classzone_idx))
3009 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
3010 order, 0, classzone_idx) == COMPACT_SKIPPED)
3017 * pgdat_balanced() is used when checking if a node is balanced.
3019 * For order-0, all zones must be balanced!
3021 * For high-order allocations only zones that meet watermarks and are in a
3022 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3023 * total of balanced pages must be at least 25% of the zones allowed by
3024 * classzone_idx for the node to be considered balanced. Forcing all zones to
3025 * be balanced for high orders can cause excessive reclaim when there are
3027 * The choice of 25% is due to
3028 * o a 16M DMA zone that is balanced will not balance a zone on any
3029 * reasonable sized machine
3030 * o On all other machines, the top zone must be at least a reasonable
3031 * percentage of the middle zones. For example, on 32-bit x86, highmem
3032 * would need to be at least 256M for it to be balance a whole node.
3033 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3034 * to balance a node on its own. These seemed like reasonable ratios.
3036 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3038 unsigned long managed_pages = 0;
3039 unsigned long balanced_pages = 0;
3042 /* Check the watermark levels */
3043 for (i = 0; i <= classzone_idx; i++) {
3044 struct zone *zone = pgdat->node_zones + i;
3046 if (!populated_zone(zone))
3049 managed_pages += zone->managed_pages;
3052 * A special case here:
3054 * balance_pgdat() skips over all_unreclaimable after
3055 * DEF_PRIORITY. Effectively, it considers them balanced so
3056 * they must be considered balanced here as well!
3058 if (!zone_reclaimable(zone)) {
3059 balanced_pages += zone->managed_pages;
3063 if (zone_balanced(zone, order, 0, i))
3064 balanced_pages += zone->managed_pages;
3070 return balanced_pages >= (managed_pages >> 2);
3076 * Prepare kswapd for sleeping. This verifies that there are no processes
3077 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3079 * Returns true if kswapd is ready to sleep
3081 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3084 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3089 * The throttled processes are normally woken up in balance_pgdat() as
3090 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3091 * race between when kswapd checks the watermarks and a process gets
3092 * throttled. There is also a potential race if processes get
3093 * throttled, kswapd wakes, a large process exits thereby balancing the
3094 * zones, which causes kswapd to exit balance_pgdat() before reaching
3095 * the wake up checks. If kswapd is going to sleep, no process should
3096 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3097 * the wake up is premature, processes will wake kswapd and get
3098 * throttled again. The difference from wake ups in balance_pgdat() is
3099 * that here we are under prepare_to_wait().
3101 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3102 wake_up_all(&pgdat->pfmemalloc_wait);
3104 return pgdat_balanced(pgdat, order, classzone_idx);
3108 * kswapd shrinks the zone by the number of pages required to reach
3109 * the high watermark.
3111 * Returns true if kswapd scanned at least the requested number of pages to
3112 * reclaim or if the lack of progress was due to pages under writeback.
3113 * This is used to determine if the scanning priority needs to be raised.
3115 static bool kswapd_shrink_zone(struct zone *zone,
3117 struct scan_control *sc,
3118 unsigned long *nr_attempted)
3120 int testorder = sc->order;
3121 unsigned long balance_gap;
3122 bool lowmem_pressure;
3124 /* Reclaim above the high watermark. */
3125 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3128 * Kswapd reclaims only single pages with compaction enabled. Trying
3129 * too hard to reclaim until contiguous free pages have become
3130 * available can hurt performance by evicting too much useful data
3131 * from memory. Do not reclaim more than needed for compaction.
3133 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3134 compaction_suitable(zone, sc->order, 0, classzone_idx)
3139 * We put equal pressure on every zone, unless one zone has way too
3140 * many pages free already. The "too many pages" is defined as the
3141 * high wmark plus a "gap" where the gap is either the low
3142 * watermark or 1% of the zone, whichever is smaller.
3144 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3145 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3148 * If there is no low memory pressure or the zone is balanced then no
3149 * reclaim is necessary
3151 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3152 if (!lowmem_pressure && zone_balanced(zone, testorder,
3153 balance_gap, classzone_idx))
3156 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3158 /* Account for the number of pages attempted to reclaim */
3159 *nr_attempted += sc->nr_to_reclaim;
3161 clear_bit(ZONE_WRITEBACK, &zone->flags);
3164 * If a zone reaches its high watermark, consider it to be no longer
3165 * congested. It's possible there are dirty pages backed by congested
3166 * BDIs but as pressure is relieved, speculatively avoid congestion
3169 if (zone_reclaimable(zone) &&
3170 zone_balanced(zone, testorder, 0, classzone_idx)) {
3171 clear_bit(ZONE_CONGESTED, &zone->flags);
3172 clear_bit(ZONE_DIRTY, &zone->flags);
3175 return sc->nr_scanned >= sc->nr_to_reclaim;
3179 * For kswapd, balance_pgdat() will work across all this node's zones until
3180 * they are all at high_wmark_pages(zone).
3182 * Returns the final order kswapd was reclaiming at
3184 * There is special handling here for zones which are full of pinned pages.
3185 * This can happen if the pages are all mlocked, or if they are all used by
3186 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3187 * What we do is to detect the case where all pages in the zone have been
3188 * scanned twice and there has been zero successful reclaim. Mark the zone as
3189 * dead and from now on, only perform a short scan. Basically we're polling
3190 * the zone for when the problem goes away.
3192 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3193 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3194 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3195 * lower zones regardless of the number of free pages in the lower zones. This
3196 * interoperates with the page allocator fallback scheme to ensure that aging
3197 * of pages is balanced across the zones.
3199 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3203 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3204 unsigned long nr_soft_reclaimed;
3205 unsigned long nr_soft_scanned;
3206 struct scan_control sc = {
3207 .gfp_mask = GFP_KERNEL,
3209 .priority = DEF_PRIORITY,
3210 .may_writepage = !laptop_mode,
3214 count_vm_event(PAGEOUTRUN);
3217 unsigned long nr_attempted = 0;
3218 bool raise_priority = true;
3219 bool pgdat_needs_compaction = (order > 0);
3221 sc.nr_reclaimed = 0;
3224 * Scan in the highmem->dma direction for the highest
3225 * zone which needs scanning
3227 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3228 struct zone *zone = pgdat->node_zones + i;
3230 if (!populated_zone(zone))
3233 if (sc.priority != DEF_PRIORITY &&
3234 !zone_reclaimable(zone))
3238 * Do some background aging of the anon list, to give
3239 * pages a chance to be referenced before reclaiming.
3241 age_active_anon(zone, &sc);
3244 * If the number of buffer_heads in the machine
3245 * exceeds the maximum allowed level and this node
3246 * has a highmem zone, force kswapd to reclaim from
3247 * it to relieve lowmem pressure.
3249 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3254 if (!zone_balanced(zone, order, 0, 0)) {
3259 * If balanced, clear the dirty and congested
3262 clear_bit(ZONE_CONGESTED, &zone->flags);
3263 clear_bit(ZONE_DIRTY, &zone->flags);
3270 for (i = 0; i <= end_zone; i++) {
3271 struct zone *zone = pgdat->node_zones + i;
3273 if (!populated_zone(zone))
3277 * If any zone is currently balanced then kswapd will
3278 * not call compaction as it is expected that the
3279 * necessary pages are already available.
3281 if (pgdat_needs_compaction &&
3282 zone_watermark_ok(zone, order,
3283 low_wmark_pages(zone),
3285 pgdat_needs_compaction = false;
3289 * If we're getting trouble reclaiming, start doing writepage
3290 * even in laptop mode.
3292 if (sc.priority < DEF_PRIORITY - 2)
3293 sc.may_writepage = 1;
3296 * Now scan the zone in the dma->highmem direction, stopping
3297 * at the last zone which needs scanning.
3299 * We do this because the page allocator works in the opposite
3300 * direction. This prevents the page allocator from allocating
3301 * pages behind kswapd's direction of progress, which would
3302 * cause too much scanning of the lower zones.
3304 for (i = 0; i <= end_zone; i++) {
3305 struct zone *zone = pgdat->node_zones + i;
3307 if (!populated_zone(zone))
3310 if (sc.priority != DEF_PRIORITY &&
3311 !zone_reclaimable(zone))
3316 nr_soft_scanned = 0;
3318 * Call soft limit reclaim before calling shrink_zone.
3320 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3323 sc.nr_reclaimed += nr_soft_reclaimed;
3326 * There should be no need to raise the scanning
3327 * priority if enough pages are already being scanned
3328 * that that high watermark would be met at 100%
3331 if (kswapd_shrink_zone(zone, end_zone,
3332 &sc, &nr_attempted))
3333 raise_priority = false;
3337 * If the low watermark is met there is no need for processes
3338 * to be throttled on pfmemalloc_wait as they should not be
3339 * able to safely make forward progress. Wake them
3341 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3342 pfmemalloc_watermark_ok(pgdat))
3343 wake_up_all(&pgdat->pfmemalloc_wait);
3346 * Fragmentation may mean that the system cannot be rebalanced
3347 * for high-order allocations in all zones. If twice the
3348 * allocation size has been reclaimed and the zones are still
3349 * not balanced then recheck the watermarks at order-0 to
3350 * prevent kswapd reclaiming excessively. Assume that a
3351 * process requested a high-order can direct reclaim/compact.
3353 if (order && sc.nr_reclaimed >= 2UL << order)
3354 order = sc.order = 0;
3356 /* Check if kswapd should be suspending */
3357 if (try_to_freeze() || kthread_should_stop())
3361 * Compact if necessary and kswapd is reclaiming at least the
3362 * high watermark number of pages as requsted
3364 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3365 compact_pgdat(pgdat, order);
3368 * Raise priority if scanning rate is too low or there was no
3369 * progress in reclaiming pages
3371 if (raise_priority || !sc.nr_reclaimed)
3373 } while (sc.priority >= 1 &&
3374 !pgdat_balanced(pgdat, order, *classzone_idx));
3378 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3379 * makes a decision on the order we were last reclaiming at. However,
3380 * if another caller entered the allocator slow path while kswapd
3381 * was awake, order will remain at the higher level
3383 *classzone_idx = end_zone;
3387 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3392 if (freezing(current) || kthread_should_stop())
3395 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3397 /* Try to sleep for a short interval */
3398 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3399 remaining = schedule_timeout(HZ/10);
3400 finish_wait(&pgdat->kswapd_wait, &wait);
3401 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3405 * After a short sleep, check if it was a premature sleep. If not, then
3406 * go fully to sleep until explicitly woken up.
3408 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3409 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3412 * vmstat counters are not perfectly accurate and the estimated
3413 * value for counters such as NR_FREE_PAGES can deviate from the
3414 * true value by nr_online_cpus * threshold. To avoid the zone
3415 * watermarks being breached while under pressure, we reduce the
3416 * per-cpu vmstat threshold while kswapd is awake and restore
3417 * them before going back to sleep.
3419 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3422 * Compaction records what page blocks it recently failed to
3423 * isolate pages from and skips them in the future scanning.
3424 * When kswapd is going to sleep, it is reasonable to assume
3425 * that pages and compaction may succeed so reset the cache.
3427 reset_isolation_suitable(pgdat);
3429 if (!kthread_should_stop())
3432 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3435 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3437 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3439 finish_wait(&pgdat->kswapd_wait, &wait);
3443 * The background pageout daemon, started as a kernel thread
3444 * from the init process.
3446 * This basically trickles out pages so that we have _some_
3447 * free memory available even if there is no other activity
3448 * that frees anything up. This is needed for things like routing
3449 * etc, where we otherwise might have all activity going on in
3450 * asynchronous contexts that cannot page things out.
3452 * If there are applications that are active memory-allocators
3453 * (most normal use), this basically shouldn't matter.
3455 static int kswapd(void *p)
3457 unsigned long order, new_order;
3458 unsigned balanced_order;
3459 int classzone_idx, new_classzone_idx;
3460 int balanced_classzone_idx;
3461 pg_data_t *pgdat = (pg_data_t*)p;
3462 struct task_struct *tsk = current;
3464 struct reclaim_state reclaim_state = {
3465 .reclaimed_slab = 0,
3467 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3469 lockdep_set_current_reclaim_state(GFP_KERNEL);
3471 if (!cpumask_empty(cpumask))
3472 set_cpus_allowed_ptr(tsk, cpumask);
3473 current->reclaim_state = &reclaim_state;
3476 * Tell the memory management that we're a "memory allocator",
3477 * and that if we need more memory we should get access to it
3478 * regardless (see "__alloc_pages()"). "kswapd" should
3479 * never get caught in the normal page freeing logic.
3481 * (Kswapd normally doesn't need memory anyway, but sometimes
3482 * you need a small amount of memory in order to be able to
3483 * page out something else, and this flag essentially protects
3484 * us from recursively trying to free more memory as we're
3485 * trying to free the first piece of memory in the first place).
3487 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3490 order = new_order = 0;
3492 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3493 balanced_classzone_idx = classzone_idx;
3498 * If the last balance_pgdat was unsuccessful it's unlikely a
3499 * new request of a similar or harder type will succeed soon
3500 * so consider going to sleep on the basis we reclaimed at
3502 if (balanced_classzone_idx >= new_classzone_idx &&
3503 balanced_order == new_order) {
3504 new_order = pgdat->kswapd_max_order;
3505 new_classzone_idx = pgdat->classzone_idx;
3506 pgdat->kswapd_max_order = 0;
3507 pgdat->classzone_idx = pgdat->nr_zones - 1;
3510 if (order < new_order || classzone_idx > new_classzone_idx) {
3512 * Don't sleep if someone wants a larger 'order'
3513 * allocation or has tigher zone constraints
3516 classzone_idx = new_classzone_idx;
3518 kswapd_try_to_sleep(pgdat, balanced_order,
3519 balanced_classzone_idx);
3520 order = pgdat->kswapd_max_order;
3521 classzone_idx = pgdat->classzone_idx;
3523 new_classzone_idx = classzone_idx;
3524 pgdat->kswapd_max_order = 0;
3525 pgdat->classzone_idx = pgdat->nr_zones - 1;
3528 ret = try_to_freeze();
3529 if (kthread_should_stop())
3533 * We can speed up thawing tasks if we don't call balance_pgdat
3534 * after returning from the refrigerator
3537 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3538 balanced_classzone_idx = classzone_idx;
3539 balanced_order = balance_pgdat(pgdat, order,
3540 &balanced_classzone_idx);
3544 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3545 current->reclaim_state = NULL;
3546 lockdep_clear_current_reclaim_state();
3552 * A zone is low on free memory, so wake its kswapd task to service it.
3554 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3558 if (!populated_zone(zone))
3561 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3563 pgdat = zone->zone_pgdat;
3564 if (pgdat->kswapd_max_order < order) {
3565 pgdat->kswapd_max_order = order;
3566 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3568 if (!waitqueue_active(&pgdat->kswapd_wait))
3570 if (zone_balanced(zone, order, 0, 0))
3573 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3574 wake_up_interruptible(&pgdat->kswapd_wait);
3577 #ifdef CONFIG_HIBERNATION
3579 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3582 * Rather than trying to age LRUs the aim is to preserve the overall
3583 * LRU order by reclaiming preferentially
3584 * inactive > active > active referenced > active mapped
3586 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3588 struct reclaim_state reclaim_state;
3589 struct scan_control sc = {
3590 .nr_to_reclaim = nr_to_reclaim,
3591 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3592 .priority = DEF_PRIORITY,
3596 .hibernation_mode = 1,
3598 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3599 struct task_struct *p = current;
3600 unsigned long nr_reclaimed;
3602 p->flags |= PF_MEMALLOC;
3603 lockdep_set_current_reclaim_state(sc.gfp_mask);
3604 reclaim_state.reclaimed_slab = 0;
3605 p->reclaim_state = &reclaim_state;
3607 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3609 p->reclaim_state = NULL;
3610 lockdep_clear_current_reclaim_state();
3611 p->flags &= ~PF_MEMALLOC;
3613 return nr_reclaimed;
3615 #endif /* CONFIG_HIBERNATION */
3617 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3618 not required for correctness. So if the last cpu in a node goes
3619 away, we get changed to run anywhere: as the first one comes back,
3620 restore their cpu bindings. */
3621 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3626 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3627 for_each_node_state(nid, N_MEMORY) {
3628 pg_data_t *pgdat = NODE_DATA(nid);
3629 const struct cpumask *mask;
3631 mask = cpumask_of_node(pgdat->node_id);
3633 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3634 /* One of our CPUs online: restore mask */
3635 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3642 * This kswapd start function will be called by init and node-hot-add.
3643 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3645 int kswapd_run(int nid)
3647 pg_data_t *pgdat = NODE_DATA(nid);
3653 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3654 if (IS_ERR(pgdat->kswapd)) {
3655 /* failure at boot is fatal */
3656 BUG_ON(system_state == SYSTEM_BOOTING);
3657 pr_err("Failed to start kswapd on node %d\n", nid);
3658 ret = PTR_ERR(pgdat->kswapd);
3659 pgdat->kswapd = NULL;
3665 * Called by memory hotplug when all memory in a node is offlined. Caller must
3666 * hold mem_hotplug_begin/end().
3668 void kswapd_stop(int nid)
3670 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3673 kthread_stop(kswapd);
3674 NODE_DATA(nid)->kswapd = NULL;
3678 static int __init kswapd_init(void)
3683 for_each_node_state(nid, N_MEMORY)
3685 hotcpu_notifier(cpu_callback, 0);
3689 module_init(kswapd_init)
3695 * If non-zero call zone_reclaim when the number of free pages falls below
3698 int zone_reclaim_mode __read_mostly;
3700 #define RECLAIM_OFF 0
3701 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3702 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3703 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3706 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3707 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3710 #define ZONE_RECLAIM_PRIORITY 4
3713 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3716 int sysctl_min_unmapped_ratio = 1;
3719 * If the number of slab pages in a zone grows beyond this percentage then
3720 * slab reclaim needs to occur.
3722 int sysctl_min_slab_ratio = 5;
3724 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3726 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3727 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3728 zone_page_state(zone, NR_ACTIVE_FILE);
3731 * It's possible for there to be more file mapped pages than
3732 * accounted for by the pages on the file LRU lists because
3733 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3735 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3738 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3739 static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3741 unsigned long nr_pagecache_reclaimable;
3742 unsigned long delta = 0;
3745 * If RECLAIM_UNMAP is set, then all file pages are considered
3746 * potentially reclaimable. Otherwise, we have to worry about
3747 * pages like swapcache and zone_unmapped_file_pages() provides
3750 if (zone_reclaim_mode & RECLAIM_UNMAP)
3751 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3753 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3755 /* If we can't clean pages, remove dirty pages from consideration */
3756 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3757 delta += zone_page_state(zone, NR_FILE_DIRTY);
3759 /* Watch for any possible underflows due to delta */
3760 if (unlikely(delta > nr_pagecache_reclaimable))
3761 delta = nr_pagecache_reclaimable;
3763 return nr_pagecache_reclaimable - delta;
3767 * Try to free up some pages from this zone through reclaim.
3769 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3771 /* Minimum pages needed in order to stay on node */
3772 const unsigned long nr_pages = 1 << order;
3773 struct task_struct *p = current;
3774 struct reclaim_state reclaim_state;
3775 struct scan_control sc = {
3776 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3777 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3779 .priority = ZONE_RECLAIM_PRIORITY,
3780 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3781 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3787 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3788 * and we also need to be able to write out pages for RECLAIM_WRITE
3789 * and RECLAIM_UNMAP.
3791 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3792 lockdep_set_current_reclaim_state(gfp_mask);
3793 reclaim_state.reclaimed_slab = 0;
3794 p->reclaim_state = &reclaim_state;
3796 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3798 * Free memory by calling shrink zone with increasing
3799 * priorities until we have enough memory freed.
3802 shrink_zone(zone, &sc, true);
3803 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3806 p->reclaim_state = NULL;
3807 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3808 lockdep_clear_current_reclaim_state();
3809 return sc.nr_reclaimed >= nr_pages;
3812 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3818 * Zone reclaim reclaims unmapped file backed pages and
3819 * slab pages if we are over the defined limits.
3821 * A small portion of unmapped file backed pages is needed for
3822 * file I/O otherwise pages read by file I/O will be immediately
3823 * thrown out if the zone is overallocated. So we do not reclaim
3824 * if less than a specified percentage of the zone is used by
3825 * unmapped file backed pages.
3827 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3828 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3829 return ZONE_RECLAIM_FULL;
3831 if (!zone_reclaimable(zone))
3832 return ZONE_RECLAIM_FULL;
3835 * Do not scan if the allocation should not be delayed.
3837 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3838 return ZONE_RECLAIM_NOSCAN;
3841 * Only run zone reclaim on the local zone or on zones that do not
3842 * have associated processors. This will favor the local processor
3843 * over remote processors and spread off node memory allocations
3844 * as wide as possible.
3846 node_id = zone_to_nid(zone);
3847 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3848 return ZONE_RECLAIM_NOSCAN;
3850 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3851 return ZONE_RECLAIM_NOSCAN;
3853 ret = __zone_reclaim(zone, gfp_mask, order);
3854 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3857 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3864 * page_evictable - test whether a page is evictable
3865 * @page: the page to test
3867 * Test whether page is evictable--i.e., should be placed on active/inactive
3868 * lists vs unevictable list.
3870 * Reasons page might not be evictable:
3871 * (1) page's mapping marked unevictable
3872 * (2) page is part of an mlocked VMA
3875 int page_evictable(struct page *page)
3877 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3882 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3883 * @pages: array of pages to check
3884 * @nr_pages: number of pages to check
3886 * Checks pages for evictability and moves them to the appropriate lru list.
3888 * This function is only used for SysV IPC SHM_UNLOCK.
3890 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3892 struct lruvec *lruvec;
3893 struct zone *zone = NULL;
3898 for (i = 0; i < nr_pages; i++) {
3899 struct page *page = pages[i];
3900 struct zone *pagezone;
3903 pagezone = page_zone(page);
3904 if (pagezone != zone) {
3906 spin_unlock_irq(&zone->lru_lock);
3908 spin_lock_irq(&zone->lru_lock);
3910 lruvec = mem_cgroup_page_lruvec(page, zone);
3912 if (!PageLRU(page) || !PageUnevictable(page))
3915 if (page_evictable(page)) {
3916 enum lru_list lru = page_lru_base_type(page);
3918 VM_BUG_ON_PAGE(PageActive(page), page);
3919 ClearPageUnevictable(page);
3920 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3921 add_page_to_lru_list(page, lruvec, lru);
3927 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3928 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3929 spin_unlock_irq(&zone->lru_lock);
3932 #endif /* CONFIG_SHMEM */