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>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
53 #include <linux/swapops.h>
54 #include <linux/balloon_compaction.h>
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/vmscan.h>
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
65 /* This context's GFP mask */
68 /* Allocation order */
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup *target_mem_cgroup;
83 /* Scan (total_size >> priority) pages at once */
86 unsigned int may_writepage:1;
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
94 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash:1;
97 unsigned int hibernation_mode:1;
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned;
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed;
109 * Reclaim pages from a vma. If the page is shared by other tasks
110 * it is zapped from a vma without reclaim so it ends up remaining
111 * on memory until last task zap it.
113 struct vm_area_struct *target_vma;
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field) \
135 if ((_page)->lru.prev != _base) { \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 * From 0 .. 100. Higher means more swappy.
149 int vm_swappiness = 60;
151 * The total number of pages which are beyond the high watermark within all
154 unsigned long vm_total_pages;
156 #ifdef CONFIG_KSWAPD_CPU_AFFINITY_MASK
157 char *kswapd_cpu_mask = CONFIG_KSWAPD_CPU_AFFINITY_MASK;
159 char *kswapd_cpu_mask = NULL;
162 static LIST_HEAD(shrinker_list);
163 static DECLARE_RWSEM(shrinker_rwsem);
166 static bool global_reclaim(struct scan_control *sc)
168 return !sc->target_mem_cgroup;
172 * sane_reclaim - is the usual dirty throttling mechanism operational?
173 * @sc: scan_control in question
175 * The normal page dirty throttling mechanism in balance_dirty_pages() is
176 * completely broken with the legacy memcg and direct stalling in
177 * shrink_page_list() is used for throttling instead, which lacks all the
178 * niceties such as fairness, adaptive pausing, bandwidth proportional
179 * allocation and configurability.
181 * This function tests whether the vmscan currently in progress can assume
182 * that the normal dirty throttling mechanism is operational.
184 static bool sane_reclaim(struct scan_control *sc)
186 struct mem_cgroup *memcg = sc->target_mem_cgroup;
190 #ifdef CONFIG_CGROUP_WRITEBACK
191 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
197 static bool global_reclaim(struct scan_control *sc)
202 static bool sane_reclaim(struct scan_control *sc)
208 static unsigned long zone_reclaimable_pages(struct zone *zone)
212 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
213 zone_page_state(zone, NR_INACTIVE_FILE);
215 if (get_nr_swap_pages() > 0)
216 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
217 zone_page_state(zone, NR_INACTIVE_ANON);
222 bool zone_reclaimable(struct zone *zone)
224 return zone_page_state(zone, NR_PAGES_SCANNED) <
225 zone_reclaimable_pages(zone) * 6;
228 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
230 if (!mem_cgroup_disabled())
231 return mem_cgroup_get_lru_size(lruvec, lru);
233 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
237 * Add a shrinker callback to be called from the vm.
239 int register_shrinker(struct shrinker *shrinker)
241 size_t size = sizeof(*shrinker->nr_deferred);
244 * If we only have one possible node in the system anyway, save
245 * ourselves the trouble and disable NUMA aware behavior. This way we
246 * will save memory and some small loop time later.
248 if (nr_node_ids == 1)
249 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
251 if (shrinker->flags & SHRINKER_NUMA_AWARE)
254 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
255 if (!shrinker->nr_deferred)
258 down_write(&shrinker_rwsem);
259 list_add_tail(&shrinker->list, &shrinker_list);
260 up_write(&shrinker_rwsem);
263 EXPORT_SYMBOL(register_shrinker);
268 void unregister_shrinker(struct shrinker *shrinker)
270 down_write(&shrinker_rwsem);
271 list_del(&shrinker->list);
272 up_write(&shrinker_rwsem);
273 kfree(shrinker->nr_deferred);
275 EXPORT_SYMBOL(unregister_shrinker);
277 #define SHRINK_BATCH 128
279 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
280 struct shrinker *shrinker,
281 unsigned long nr_scanned,
282 unsigned long nr_eligible)
284 unsigned long freed = 0;
285 unsigned long long delta;
290 int nid = shrinkctl->nid;
291 long batch_size = shrinker->batch ? shrinker->batch
293 long scanned = 0, next_deferred;
294 long min_cache_size = batch_size;
296 if (current_is_kswapd())
299 freeable = shrinker->count_objects(shrinker, shrinkctl);
304 * copy the current shrinker scan count into a local variable
305 * and zero it so that other concurrent shrinker invocations
306 * don't also do this scanning work.
308 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
311 delta = (4 * nr_scanned) / shrinker->seeks;
313 do_div(delta, nr_eligible + 1);
315 if (total_scan < 0) {
316 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
317 shrinker->scan_objects, total_scan);
318 total_scan = freeable;
321 next_deferred = total_scan;
324 * We need to avoid excessive windup on filesystem shrinkers
325 * due to large numbers of GFP_NOFS allocations causing the
326 * shrinkers to return -1 all the time. This results in a large
327 * nr being built up so when a shrink that can do some work
328 * comes along it empties the entire cache due to nr >>>
329 * freeable. This is bad for sustaining a working set in
332 * Hence only allow the shrinker to scan the entire cache when
333 * a large delta change is calculated directly.
335 if (delta < freeable / 4)
336 total_scan = min(total_scan, freeable / 2);
339 * Avoid risking looping forever due to too large nr value:
340 * never try to free more than twice the estimate number of
343 if (total_scan > freeable * 2)
344 total_scan = freeable * 2;
346 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
347 nr_scanned, nr_eligible,
348 freeable, delta, total_scan);
351 * Normally, we should not scan less than batch_size objects in one
352 * pass to avoid too frequent shrinker calls, but if the slab has less
353 * than batch_size objects in total and we are really tight on memory,
354 * we will try to reclaim all available objects, otherwise we can end
355 * up failing allocations although there are plenty of reclaimable
356 * objects spread over several slabs with usage less than the
359 * We detect the "tight on memory" situations by looking at the total
360 * number of objects we want to scan (total_scan). If it is greater
361 * than the total number of objects on slab (freeable), we must be
362 * scanning at high prio and therefore should try to reclaim as much as
365 while (total_scan > min_cache_size ||
366 total_scan >= freeable) {
368 unsigned long nr_to_scan = min(batch_size, total_scan);
370 shrinkctl->nr_to_scan = nr_to_scan;
371 ret = shrinker->scan_objects(shrinker, shrinkctl);
372 if (ret == SHRINK_STOP)
376 count_vm_events(SLABS_SCANNED, nr_to_scan);
377 total_scan -= nr_to_scan;
378 scanned += nr_to_scan;
383 if (next_deferred >= scanned)
384 next_deferred -= scanned;
388 * move the unused scan count back into the shrinker in a
389 * manner that handles concurrent updates. If we exhausted the
390 * scan, there is no need to do an update.
392 if (next_deferred > 0)
393 new_nr = atomic_long_add_return(next_deferred,
394 &shrinker->nr_deferred[nid]);
396 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
398 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
402 static void shrink_slab_lmk(gfp_t gfp_mask, int nid,
403 struct mem_cgroup *memcg,
404 unsigned long nr_scanned,
405 unsigned long nr_eligible)
407 struct shrinker *shrinker;
410 nr_scanned = SWAP_CLUSTER_MAX;
412 if (!down_read_trylock(&shrinker_rwsem))
415 list_for_each_entry(shrinker, &shrinker_list, list) {
416 struct shrink_control sc = {
417 .gfp_mask = gfp_mask,
420 if (!(shrinker->flags & SHRINKER_LMK))
423 do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
426 up_read(&shrinker_rwsem);
432 * shrink_slab - shrink slab caches
433 * @gfp_mask: allocation context
434 * @nid: node whose slab caches to target
435 * @memcg: memory cgroup whose slab caches to target
436 * @nr_scanned: pressure numerator
437 * @nr_eligible: pressure denominator
439 * Call the shrink functions to age shrinkable caches.
441 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
442 * unaware shrinkers will receive a node id of 0 instead.
444 * @memcg specifies the memory cgroup to target. If it is not NULL,
445 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
446 * objects from the memory cgroup specified. Otherwise all shrinkers
447 * are called, and memcg aware shrinkers are supposed to scan the
450 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
451 * the available objects should be scanned. Page reclaim for example
452 * passes the number of pages scanned and the number of pages on the
453 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
454 * when it encountered mapped pages. The ratio is further biased by
455 * the ->seeks setting of the shrink function, which indicates the
456 * cost to recreate an object relative to that of an LRU page.
458 * Returns the number of reclaimed slab objects.
460 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
461 struct mem_cgroup *memcg,
462 unsigned long nr_scanned,
463 unsigned long nr_eligible)
465 struct shrinker *shrinker;
466 unsigned long freed = 0;
468 if (memcg && !memcg_kmem_is_active(memcg))
472 nr_scanned = SWAP_CLUSTER_MAX;
474 if (!down_read_trylock(&shrinker_rwsem)) {
476 * If we would return 0, our callers would understand that we
477 * have nothing else to shrink and give up trying. By returning
478 * 1 we keep it going and assume we'll be able to shrink next
485 list_for_each_entry(shrinker, &shrinker_list, list) {
486 struct shrink_control sc = {
487 .gfp_mask = gfp_mask,
492 if (shrinker->flags & SHRINKER_LMK)
495 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
498 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
501 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
504 up_read(&shrinker_rwsem);
510 void drop_slab_node(int nid)
515 struct mem_cgroup *memcg = NULL;
519 freed += shrink_slab(GFP_KERNEL, nid, memcg,
521 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
522 } while (freed > 10);
529 for_each_online_node(nid)
533 static inline int is_page_cache_freeable(struct page *page)
536 * A freeable page cache page is referenced only by the caller
537 * that isolated the page, the page cache radix tree and
538 * optional buffer heads at page->private.
540 return page_count(page) - page_has_private(page) == 2;
543 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
545 if (current->flags & PF_SWAPWRITE)
547 if (!inode_write_congested(inode))
549 if (inode_to_bdi(inode) == current->backing_dev_info)
555 * We detected a synchronous write error writing a page out. Probably
556 * -ENOSPC. We need to propagate that into the address_space for a subsequent
557 * fsync(), msync() or close().
559 * The tricky part is that after writepage we cannot touch the mapping: nothing
560 * prevents it from being freed up. But we have a ref on the page and once
561 * that page is locked, the mapping is pinned.
563 * We're allowed to run sleeping lock_page() here because we know the caller has
566 static void handle_write_error(struct address_space *mapping,
567 struct page *page, int error)
570 if (page_mapping(page) == mapping)
571 mapping_set_error(mapping, error);
575 /* possible outcome of pageout() */
577 /* failed to write page out, page is locked */
579 /* move page to the active list, page is locked */
581 /* page has been sent to the disk successfully, page is unlocked */
583 /* page is clean and locked */
588 * pageout is called by shrink_page_list() for each dirty page.
589 * Calls ->writepage().
591 static pageout_t pageout(struct page *page, struct address_space *mapping,
592 struct scan_control *sc)
595 * If the page is dirty, only perform writeback if that write
596 * will be non-blocking. To prevent this allocation from being
597 * stalled by pagecache activity. But note that there may be
598 * stalls if we need to run get_block(). We could test
599 * PagePrivate for that.
601 * If this process is currently in __generic_file_write_iter() against
602 * this page's queue, we can perform writeback even if that
605 * If the page is swapcache, write it back even if that would
606 * block, for some throttling. This happens by accident, because
607 * swap_backing_dev_info is bust: it doesn't reflect the
608 * congestion state of the swapdevs. Easy to fix, if needed.
610 if (!is_page_cache_freeable(page))
614 * Some data journaling orphaned pages can have
615 * page->mapping == NULL while being dirty with clean buffers.
617 if (page_has_private(page)) {
618 if (try_to_free_buffers(page)) {
619 ClearPageDirty(page);
620 pr_info("%s: orphaned page\n", __func__);
626 if (mapping->a_ops->writepage == NULL)
627 return PAGE_ACTIVATE;
628 if (!may_write_to_inode(mapping->host, sc))
631 if (clear_page_dirty_for_io(page)) {
633 struct writeback_control wbc = {
634 .sync_mode = WB_SYNC_NONE,
635 .nr_to_write = SWAP_CLUSTER_MAX,
637 .range_end = LLONG_MAX,
641 SetPageReclaim(page);
642 res = mapping->a_ops->writepage(page, &wbc);
644 handle_write_error(mapping, page, res);
645 if (res == AOP_WRITEPAGE_ACTIVATE) {
646 ClearPageReclaim(page);
647 return PAGE_ACTIVATE;
650 if (!PageWriteback(page)) {
651 /* synchronous write or broken a_ops? */
652 ClearPageReclaim(page);
654 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
655 inc_zone_page_state(page, NR_VMSCAN_WRITE);
663 * Same as remove_mapping, but if the page is removed from the mapping, it
664 * gets returned with a refcount of 0.
666 static int __remove_mapping(struct address_space *mapping, struct page *page,
670 struct mem_cgroup *memcg;
672 BUG_ON(!PageLocked(page));
673 BUG_ON(mapping != page_mapping(page));
675 memcg = mem_cgroup_begin_page_stat(page);
676 spin_lock_irqsave(&mapping->tree_lock, flags);
678 * The non racy check for a busy page.
680 * Must be careful with the order of the tests. When someone has
681 * a ref to the page, it may be possible that they dirty it then
682 * drop the reference. So if PageDirty is tested before page_count
683 * here, then the following race may occur:
685 * get_user_pages(&page);
686 * [user mapping goes away]
688 * !PageDirty(page) [good]
689 * SetPageDirty(page);
691 * !page_count(page) [good, discard it]
693 * [oops, our write_to data is lost]
695 * Reversing the order of the tests ensures such a situation cannot
696 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
697 * load is not satisfied before that of page->_count.
699 * Note that if SetPageDirty is always performed via set_page_dirty,
700 * and thus under tree_lock, then this ordering is not required.
702 if (!page_freeze_refs(page, 2))
704 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
705 if (unlikely(PageDirty(page))) {
706 page_unfreeze_refs(page, 2);
710 if (PageSwapCache(page)) {
711 swp_entry_t swap = { .val = page_private(page) };
712 mem_cgroup_swapout(page, swap);
713 __delete_from_swap_cache(page);
714 spin_unlock_irqrestore(&mapping->tree_lock, flags);
715 mem_cgroup_end_page_stat(memcg);
716 swapcache_free(swap);
718 void (*freepage)(struct page *);
721 freepage = mapping->a_ops->freepage;
723 * Remember a shadow entry for reclaimed file cache in
724 * order to detect refaults, thus thrashing, later on.
726 * But don't store shadows in an address space that is
727 * already exiting. This is not just an optizimation,
728 * inode reclaim needs to empty out the radix tree or
729 * the nodes are lost. Don't plant shadows behind its
732 if (reclaimed && page_is_file_cache(page) &&
733 !mapping_exiting(mapping))
734 shadow = workingset_eviction(mapping, page);
735 __delete_from_page_cache(page, shadow, memcg);
736 spin_unlock_irqrestore(&mapping->tree_lock, flags);
737 mem_cgroup_end_page_stat(memcg);
739 if (freepage != NULL)
746 spin_unlock_irqrestore(&mapping->tree_lock, flags);
747 mem_cgroup_end_page_stat(memcg);
752 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
753 * someone else has a ref on the page, abort and return 0. If it was
754 * successfully detached, return 1. Assumes the caller has a single ref on
757 int remove_mapping(struct address_space *mapping, struct page *page)
759 if (__remove_mapping(mapping, page, false)) {
761 * Unfreezing the refcount with 1 rather than 2 effectively
762 * drops the pagecache ref for us without requiring another
765 page_unfreeze_refs(page, 1);
772 * putback_lru_page - put previously isolated page onto appropriate LRU list
773 * @page: page to be put back to appropriate lru list
775 * Add previously isolated @page to appropriate LRU list.
776 * Page may still be unevictable for other reasons.
778 * lru_lock must not be held, interrupts must be enabled.
780 void putback_lru_page(struct page *page)
783 int was_unevictable = PageUnevictable(page);
785 VM_BUG_ON_PAGE(PageLRU(page), page);
788 ClearPageUnevictable(page);
790 if (page_evictable(page)) {
792 * For evictable pages, we can use the cache.
793 * In event of a race, worst case is we end up with an
794 * unevictable page on [in]active list.
795 * We know how to handle that.
797 is_unevictable = false;
801 * Put unevictable pages directly on zone's unevictable
804 is_unevictable = true;
805 add_page_to_unevictable_list(page);
807 * When racing with an mlock or AS_UNEVICTABLE clearing
808 * (page is unlocked) make sure that if the other thread
809 * does not observe our setting of PG_lru and fails
810 * isolation/check_move_unevictable_pages,
811 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
812 * the page back to the evictable list.
814 * The other side is TestClearPageMlocked() or shmem_lock().
820 * page's status can change while we move it among lru. If an evictable
821 * page is on unevictable list, it never be freed. To avoid that,
822 * check after we added it to the list, again.
824 if (is_unevictable && page_evictable(page)) {
825 if (!isolate_lru_page(page)) {
829 /* This means someone else dropped this page from LRU
830 * So, it will be freed or putback to LRU again. There is
831 * nothing to do here.
835 if (was_unevictable && !is_unevictable)
836 count_vm_event(UNEVICTABLE_PGRESCUED);
837 else if (!was_unevictable && is_unevictable)
838 count_vm_event(UNEVICTABLE_PGCULLED);
840 put_page(page); /* drop ref from isolate */
843 enum page_references {
845 PAGEREF_RECLAIM_CLEAN,
850 static enum page_references page_check_references(struct page *page,
851 struct scan_control *sc)
853 int referenced_ptes, referenced_page;
854 unsigned long vm_flags;
856 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
858 referenced_page = TestClearPageReferenced(page);
861 * Mlock lost the isolation race with us. Let try_to_unmap()
862 * move the page to the unevictable list.
864 if (vm_flags & VM_LOCKED)
865 return PAGEREF_RECLAIM;
867 if (referenced_ptes) {
868 if (PageSwapBacked(page))
869 return PAGEREF_ACTIVATE;
871 * All mapped pages start out with page table
872 * references from the instantiating fault, so we need
873 * to look twice if a mapped file page is used more
876 * Mark it and spare it for another trip around the
877 * inactive list. Another page table reference will
878 * lead to its activation.
880 * Note: the mark is set for activated pages as well
881 * so that recently deactivated but used pages are
884 SetPageReferenced(page);
886 if (referenced_page || referenced_ptes > 1)
887 return PAGEREF_ACTIVATE;
890 * Activate file-backed executable pages after first usage.
892 if (vm_flags & VM_EXEC)
893 return PAGEREF_ACTIVATE;
898 /* Reclaim if clean, defer dirty pages to writeback */
899 if (referenced_page && !PageSwapBacked(page))
900 return PAGEREF_RECLAIM_CLEAN;
902 return PAGEREF_RECLAIM;
905 /* Check if a page is dirty or under writeback */
906 static void page_check_dirty_writeback(struct page *page,
907 bool *dirty, bool *writeback)
909 struct address_space *mapping;
912 * Anonymous pages are not handled by flushers and must be written
913 * from reclaim context. Do not stall reclaim based on them
915 if (!page_is_file_cache(page)) {
921 /* By default assume that the page flags are accurate */
922 *dirty = PageDirty(page);
923 *writeback = PageWriteback(page);
925 /* Verify dirty/writeback state if the filesystem supports it */
926 if (!page_has_private(page))
929 mapping = page_mapping(page);
930 if (mapping && mapping->a_ops->is_dirty_writeback)
931 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
935 * shrink_page_list() returns the number of reclaimed pages
937 static unsigned long shrink_page_list(struct list_head *page_list,
939 struct scan_control *sc,
940 enum ttu_flags ttu_flags,
941 unsigned long *ret_nr_dirty,
942 unsigned long *ret_nr_unqueued_dirty,
943 unsigned long *ret_nr_congested,
944 unsigned long *ret_nr_writeback,
945 unsigned long *ret_nr_immediate,
948 LIST_HEAD(ret_pages);
949 LIST_HEAD(free_pages);
951 unsigned long nr_unqueued_dirty = 0;
952 unsigned long nr_dirty = 0;
953 unsigned long nr_congested = 0;
954 unsigned long nr_reclaimed = 0;
955 unsigned long nr_writeback = 0;
956 unsigned long nr_immediate = 0;
960 while (!list_empty(page_list)) {
961 struct address_space *mapping;
964 enum page_references references = PAGEREF_RECLAIM;
965 bool dirty, writeback;
969 page = lru_to_page(page_list);
970 list_del(&page->lru);
972 if (!trylock_page(page))
975 VM_BUG_ON_PAGE(PageActive(page), page);
977 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
981 if (unlikely(!page_evictable(page)))
984 if (!sc->may_unmap && page_mapped(page))
987 /* Double the slab pressure for mapped and swapcache pages */
988 if (page_mapped(page) || PageSwapCache(page))
991 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
992 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
995 * The number of dirty pages determines if a zone is marked
996 * reclaim_congested which affects wait_iff_congested. kswapd
997 * will stall and start writing pages if the tail of the LRU
998 * is all dirty unqueued pages.
1000 page_check_dirty_writeback(page, &dirty, &writeback);
1001 if (dirty || writeback)
1004 if (dirty && !writeback)
1005 nr_unqueued_dirty++;
1008 * Treat this page as congested if the underlying BDI is or if
1009 * pages are cycling through the LRU so quickly that the
1010 * pages marked for immediate reclaim are making it to the
1011 * end of the LRU a second time.
1013 mapping = page_mapping(page);
1014 if (((dirty || writeback) && mapping &&
1015 inode_write_congested(mapping->host)) ||
1016 (writeback && PageReclaim(page)))
1020 * If a page at the tail of the LRU is under writeback, there
1021 * are three cases to consider.
1023 * 1) If reclaim is encountering an excessive number of pages
1024 * under writeback and this page is both under writeback and
1025 * PageReclaim then it indicates that pages are being queued
1026 * for IO but are being recycled through the LRU before the
1027 * IO can complete. Waiting on the page itself risks an
1028 * indefinite stall if it is impossible to writeback the
1029 * page due to IO error or disconnected storage so instead
1030 * note that the LRU is being scanned too quickly and the
1031 * caller can stall after page list has been processed.
1033 * 2) Global or new memcg reclaim encounters a page that is
1034 * not marked for immediate reclaim, or the caller does not
1035 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1036 * not to fs). In this case mark the page for immediate
1037 * reclaim and continue scanning.
1039 * Require may_enter_fs because we would wait on fs, which
1040 * may not have submitted IO yet. And the loop driver might
1041 * enter reclaim, and deadlock if it waits on a page for
1042 * which it is needed to do the write (loop masks off
1043 * __GFP_IO|__GFP_FS for this reason); but more thought
1044 * would probably show more reasons.
1046 * 3) Legacy memcg encounters a page that is already marked
1047 * PageReclaim. memcg does not have any dirty pages
1048 * throttling so we could easily OOM just because too many
1049 * pages are in writeback and there is nothing else to
1050 * reclaim. Wait for the writeback to complete.
1052 if (PageWriteback(page)) {
1054 if (current_is_kswapd() &&
1055 PageReclaim(page) &&
1056 (zone && test_bit(ZONE_WRITEBACK, &zone->flags))) {
1061 } else if (sane_reclaim(sc) ||
1062 !PageReclaim(page) || !may_enter_fs) {
1064 * This is slightly racy - end_page_writeback()
1065 * might have just cleared PageReclaim, then
1066 * setting PageReclaim here end up interpreted
1067 * as PageReadahead - but that does not matter
1068 * enough to care. What we do want is for this
1069 * page to have PageReclaim set next time memcg
1070 * reclaim reaches the tests above, so it will
1071 * then wait_on_page_writeback() to avoid OOM;
1072 * and it's also appropriate in global reclaim.
1074 SetPageReclaim(page);
1081 wait_on_page_writeback(page);
1082 /* then go back and try same page again */
1083 list_add_tail(&page->lru, page_list);
1089 references = page_check_references(page, sc);
1091 switch (references) {
1092 case PAGEREF_ACTIVATE:
1093 goto activate_locked;
1096 case PAGEREF_RECLAIM:
1097 case PAGEREF_RECLAIM_CLEAN:
1098 ; /* try to reclaim the page below */
1102 * Anonymous process memory has backing store?
1103 * Try to allocate it some swap space here.
1105 if (PageAnon(page) && !PageSwapCache(page)) {
1106 if (!(sc->gfp_mask & __GFP_IO))
1108 if (!add_to_swap(page, page_list))
1109 goto activate_locked;
1112 /* Adding to swap updated mapping */
1113 mapping = page_mapping(page);
1117 * The page is mapped into the page tables of one or more
1118 * processes. Try to unmap it here.
1120 if (page_mapped(page) && mapping) {
1121 switch (try_to_unmap(page,
1122 ttu_flags|TTU_BATCH_FLUSH,
1125 goto activate_locked;
1131 ; /* try to free the page below */
1135 if (PageDirty(page)) {
1137 * Only kswapd can writeback filesystem pages to
1138 * avoid risk of stack overflow but only writeback
1139 * if many dirty pages have been encountered.
1141 if (page_is_file_cache(page) &&
1142 (!current_is_kswapd() ||
1144 !test_bit(ZONE_DIRTY, &zone->flags)))) {
1146 * Immediately reclaim when written back.
1147 * Similar in principal to deactivate_page()
1148 * except we already have the page isolated
1149 * and know it's dirty
1151 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1152 SetPageReclaim(page);
1157 if (references == PAGEREF_RECLAIM_CLEAN)
1161 if (!sc->may_writepage)
1165 * Page is dirty. Flush the TLB if a writable entry
1166 * potentially exists to avoid CPU writes after IO
1167 * starts and then write it out here.
1169 try_to_unmap_flush_dirty();
1170 switch (pageout(page, mapping, sc)) {
1174 goto activate_locked;
1176 if (PageWriteback(page))
1178 if (PageDirty(page))
1182 * A synchronous write - probably a ramdisk. Go
1183 * ahead and try to reclaim the page.
1185 if (!trylock_page(page))
1187 if (PageDirty(page) || PageWriteback(page))
1189 mapping = page_mapping(page);
1191 ; /* try to free the page below */
1196 * If the page has buffers, try to free the buffer mappings
1197 * associated with this page. If we succeed we try to free
1200 * We do this even if the page is PageDirty().
1201 * try_to_release_page() does not perform I/O, but it is
1202 * possible for a page to have PageDirty set, but it is actually
1203 * clean (all its buffers are clean). This happens if the
1204 * buffers were written out directly, with submit_bh(). ext3
1205 * will do this, as well as the blockdev mapping.
1206 * try_to_release_page() will discover that cleanness and will
1207 * drop the buffers and mark the page clean - it can be freed.
1209 * Rarely, pages can have buffers and no ->mapping. These are
1210 * the pages which were not successfully invalidated in
1211 * truncate_complete_page(). We try to drop those buffers here
1212 * and if that worked, and the page is no longer mapped into
1213 * process address space (page_count == 1) it can be freed.
1214 * Otherwise, leave the page on the LRU so it is swappable.
1216 if (page_has_private(page)) {
1217 if (!try_to_release_page(page, sc->gfp_mask))
1218 goto activate_locked;
1219 if (!mapping && page_count(page) == 1) {
1221 if (put_page_testzero(page))
1225 * rare race with speculative reference.
1226 * the speculative reference will free
1227 * this page shortly, so we may
1228 * increment nr_reclaimed here (and
1229 * leave it off the LRU).
1237 if (!mapping || !__remove_mapping(mapping, page, true))
1241 * At this point, we have no other references and there is
1242 * no way to pick any more up (removed from LRU, removed
1243 * from pagecache). Can use non-atomic bitops now (and
1244 * we obviously don't have to worry about waking up a process
1245 * waiting on the page lock, because there are no references.
1247 __clear_page_locked(page);
1252 * Is there need to periodically free_page_list? It would
1253 * appear not as the counts should be low
1255 list_add(&page->lru, &free_pages);
1257 * If pagelist are from multiple zones, we should decrease
1258 * NR_ISOLATED_ANON + x on freed pages in here.
1261 dec_zone_page_state(page, NR_ISOLATED_ANON +
1262 page_is_file_cache(page));
1266 if (PageSwapCache(page))
1267 try_to_free_swap(page);
1269 list_add(&page->lru, &ret_pages);
1273 /* Not a candidate for swapping, so reclaim swap space. */
1274 if (PageSwapCache(page) && vm_swap_full(page_swap_info(page)))
1275 try_to_free_swap(page);
1276 VM_BUG_ON_PAGE(PageActive(page), page);
1277 SetPageActive(page);
1282 list_add(&page->lru, &ret_pages);
1283 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1286 mem_cgroup_uncharge_list(&free_pages);
1287 try_to_unmap_flush();
1288 free_hot_cold_page_list(&free_pages, true);
1290 list_splice(&ret_pages, page_list);
1291 count_vm_events(PGACTIVATE, pgactivate);
1293 *ret_nr_dirty += nr_dirty;
1294 *ret_nr_congested += nr_congested;
1295 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1296 *ret_nr_writeback += nr_writeback;
1297 *ret_nr_immediate += nr_immediate;
1298 return nr_reclaimed;
1301 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1302 struct list_head *page_list)
1304 struct scan_control sc = {
1305 .gfp_mask = GFP_KERNEL,
1306 .priority = DEF_PRIORITY,
1308 /* Doesn't allow to write out dirty page */
1311 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1312 struct page *page, *next;
1313 LIST_HEAD(clean_pages);
1315 list_for_each_entry_safe(page, next, page_list, lru) {
1316 if (page_is_file_cache(page) && !PageDirty(page) &&
1317 !__PageMovable(page)) {
1318 ClearPageActive(page);
1319 list_move(&page->lru, &clean_pages);
1323 ret = shrink_page_list(&clean_pages, zone, &sc,
1324 TTU_UNMAP|TTU_IGNORE_ACCESS,
1325 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1326 list_splice(&clean_pages, page_list);
1327 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1331 #ifdef CONFIG_PROCESS_RECLAIM
1332 unsigned long reclaim_pages_from_list(struct list_head *page_list,
1333 struct vm_area_struct *vma)
1335 struct scan_control sc = {
1336 .gfp_mask = GFP_KERNEL,
1337 .priority = DEF_PRIORITY,
1344 unsigned long nr_reclaimed;
1346 unsigned long dummy1, dummy2, dummy3, dummy4, dummy5;
1348 list_for_each_entry(page, page_list, lru)
1349 ClearPageActive(page);
1351 nr_reclaimed = shrink_page_list(page_list, NULL, &sc,
1352 TTU_UNMAP|TTU_IGNORE_ACCESS,
1353 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1355 while (!list_empty(page_list)) {
1356 page = lru_to_page(page_list);
1357 list_del(&page->lru);
1358 dec_zone_page_state(page, NR_ISOLATED_ANON +
1359 page_is_file_cache(page));
1360 putback_lru_page(page);
1363 return nr_reclaimed;
1368 * Attempt to remove the specified page from its LRU. Only take this page
1369 * if it is of the appropriate PageActive status. Pages which are being
1370 * freed elsewhere are also ignored.
1372 * page: page to consider
1373 * mode: one of the LRU isolation modes defined above
1375 * returns 0 on success, -ve errno on failure.
1377 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1381 /* Only take pages on the LRU. */
1385 /* Compaction should not handle unevictable pages but CMA can do so */
1386 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1392 * To minimise LRU disruption, the caller can indicate that it only
1393 * wants to isolate pages it will be able to operate on without
1394 * blocking - clean pages for the most part.
1396 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1397 * is used by reclaim when it is cannot write to backing storage
1399 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1400 * that it is possible to migrate without blocking
1402 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1403 /* All the caller can do on PageWriteback is block */
1404 if (PageWriteback(page))
1407 if (PageDirty(page)) {
1408 struct address_space *mapping;
1410 /* ISOLATE_CLEAN means only clean pages */
1411 if (mode & ISOLATE_CLEAN)
1415 * Only pages without mappings or that have a
1416 * ->migratepage callback are possible to migrate
1419 mapping = page_mapping(page);
1420 if (mapping && !mapping->a_ops->migratepage)
1425 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1428 if (likely(get_page_unless_zero(page))) {
1430 * Be careful not to clear PageLRU until after we're
1431 * sure the page is not being freed elsewhere -- the
1432 * page release code relies on it.
1442 * zone->lru_lock is heavily contended. Some of the functions that
1443 * shrink the lists perform better by taking out a batch of pages
1444 * and working on them outside the LRU lock.
1446 * For pagecache intensive workloads, this function is the hottest
1447 * spot in the kernel (apart from copy_*_user functions).
1449 * Appropriate locks must be held before calling this function.
1451 * @nr_to_scan: The number of pages to look through on the list.
1452 * @lruvec: The LRU vector to pull pages from.
1453 * @dst: The temp list to put pages on to.
1454 * @nr_scanned: The number of pages that were scanned.
1455 * @sc: The scan_control struct for this reclaim session
1456 * @mode: One of the LRU isolation modes
1457 * @lru: LRU list id for isolating
1459 * returns how many pages were moved onto *@dst.
1461 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1462 struct lruvec *lruvec, struct list_head *dst,
1463 unsigned long *nr_scanned, struct scan_control *sc,
1464 isolate_mode_t mode, enum lru_list lru)
1466 struct list_head *src = &lruvec->lists[lru];
1467 unsigned long nr_taken = 0;
1470 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1471 !list_empty(src); scan++) {
1475 page = lru_to_page(src);
1476 prefetchw_prev_lru_page(page, src, flags);
1478 VM_BUG_ON_PAGE(!PageLRU(page), page);
1480 switch (__isolate_lru_page(page, mode)) {
1482 nr_pages = hpage_nr_pages(page);
1483 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1484 list_move(&page->lru, dst);
1485 nr_taken += nr_pages;
1489 /* else it is being freed elsewhere */
1490 list_move(&page->lru, src);
1499 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1500 nr_taken, mode, is_file_lru(lru));
1505 * isolate_lru_page - tries to isolate a page from its LRU list
1506 * @page: page to isolate from its LRU list
1508 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1509 * vmstat statistic corresponding to whatever LRU list the page was on.
1511 * Returns 0 if the page was removed from an LRU list.
1512 * Returns -EBUSY if the page was not on an LRU list.
1514 * The returned page will have PageLRU() cleared. If it was found on
1515 * the active list, it will have PageActive set. If it was found on
1516 * the unevictable list, it will have the PageUnevictable bit set. That flag
1517 * may need to be cleared by the caller before letting the page go.
1519 * The vmstat statistic corresponding to the list on which the page was
1520 * found will be decremented.
1523 * (1) Must be called with an elevated refcount on the page. This is a
1524 * fundamentnal difference from isolate_lru_pages (which is called
1525 * without a stable reference).
1526 * (2) the lru_lock must not be held.
1527 * (3) interrupts must be enabled.
1529 int isolate_lru_page(struct page *page)
1533 VM_BUG_ON_PAGE(!page_count(page), page);
1535 if (PageLRU(page)) {
1536 struct zone *zone = page_zone(page);
1537 struct lruvec *lruvec;
1539 spin_lock_irq(&zone->lru_lock);
1540 lruvec = mem_cgroup_page_lruvec(page, zone);
1541 if (PageLRU(page)) {
1542 int lru = page_lru(page);
1545 del_page_from_lru_list(page, lruvec, lru);
1548 spin_unlock_irq(&zone->lru_lock);
1553 static int __too_many_isolated(struct zone *zone, int file,
1554 struct scan_control *sc, int safe)
1556 unsigned long inactive, isolated;
1560 inactive = zone_page_state_snapshot(zone,
1562 isolated = zone_page_state_snapshot(zone,
1565 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1566 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1570 inactive = zone_page_state_snapshot(zone,
1572 isolated = zone_page_state_snapshot(zone,
1575 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1576 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1581 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1582 * won't get blocked by normal direct-reclaimers, forming a circular
1585 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1588 return isolated > inactive;
1592 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1593 * then get resheduled. When there are massive number of tasks doing page
1594 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1595 * the LRU list will go small and be scanned faster than necessary, leading to
1596 * unnecessary swapping, thrashing and OOM.
1598 static int too_many_isolated(struct zone *zone, int file,
1599 struct scan_control *sc, int safe)
1601 if (current_is_kswapd())
1604 if (!sane_reclaim(sc))
1607 if (unlikely(__too_many_isolated(zone, file, sc, 0))) {
1609 return __too_many_isolated(zone, file, sc, safe);
1617 static noinline_for_stack void
1618 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1620 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1621 struct zone *zone = lruvec_zone(lruvec);
1622 LIST_HEAD(pages_to_free);
1625 * Put back any unfreeable pages.
1627 while (!list_empty(page_list)) {
1628 struct page *page = lru_to_page(page_list);
1632 VM_BUG_ON_PAGE(PageLRU(page), page);
1633 list_del(&page->lru);
1634 if (unlikely(!page_evictable(page))) {
1635 spin_unlock_irq(&zone->lru_lock);
1636 putback_lru_page(page);
1637 spin_lock_irq(&zone->lru_lock);
1641 lruvec = mem_cgroup_page_lruvec(page, zone);
1644 lru = page_lru(page);
1645 add_page_to_lru_list(page, lruvec, lru);
1647 file = is_file_lru(lru);
1648 if (IS_ENABLED(CONFIG_ZCACHE))
1650 SetPageWasActive(page);
1651 if (is_active_lru(lru)) {
1652 int numpages = hpage_nr_pages(page);
1653 reclaim_stat->recent_rotated[file] += numpages;
1655 if (put_page_testzero(page)) {
1656 __ClearPageLRU(page);
1657 __ClearPageActive(page);
1658 del_page_from_lru_list(page, lruvec, lru);
1660 if (unlikely(PageCompound(page))) {
1661 spin_unlock_irq(&zone->lru_lock);
1662 mem_cgroup_uncharge(page);
1663 (*get_compound_page_dtor(page))(page);
1664 spin_lock_irq(&zone->lru_lock);
1666 list_add(&page->lru, &pages_to_free);
1671 * To save our caller's stack, now use input list for pages to free.
1673 list_splice(&pages_to_free, page_list);
1677 * If a kernel thread (such as nfsd for loop-back mounts) services
1678 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1679 * In that case we should only throttle if the backing device it is
1680 * writing to is congested. In other cases it is safe to throttle.
1682 static int current_may_throttle(void)
1684 return !(current->flags & PF_LESS_THROTTLE) ||
1685 current->backing_dev_info == NULL ||
1686 bdi_write_congested(current->backing_dev_info);
1690 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1691 * of reclaimed pages
1693 static noinline_for_stack unsigned long
1694 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1695 struct scan_control *sc, enum lru_list lru)
1697 LIST_HEAD(page_list);
1698 unsigned long nr_scanned;
1699 unsigned long nr_reclaimed = 0;
1700 unsigned long nr_taken;
1701 unsigned long nr_dirty = 0;
1702 unsigned long nr_congested = 0;
1703 unsigned long nr_unqueued_dirty = 0;
1704 unsigned long nr_writeback = 0;
1705 unsigned long nr_immediate = 0;
1706 isolate_mode_t isolate_mode = 0;
1707 int file = is_file_lru(lru);
1709 struct zone *zone = lruvec_zone(lruvec);
1710 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1712 while (unlikely(too_many_isolated(zone, file, sc, safe))) {
1713 congestion_wait(BLK_RW_ASYNC, HZ/10);
1715 /* We are about to die and free our memory. Return now. */
1716 if (fatal_signal_pending(current))
1717 return SWAP_CLUSTER_MAX;
1725 isolate_mode |= ISOLATE_UNMAPPED;
1726 if (!sc->may_writepage)
1727 isolate_mode |= ISOLATE_CLEAN;
1729 spin_lock_irq(&zone->lru_lock);
1731 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1732 &nr_scanned, sc, isolate_mode, lru);
1734 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1735 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1737 if (global_reclaim(sc)) {
1738 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1739 if (current_is_kswapd())
1740 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1742 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1744 spin_unlock_irq(&zone->lru_lock);
1749 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1750 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1751 &nr_writeback, &nr_immediate,
1754 spin_lock_irq(&zone->lru_lock);
1756 reclaim_stat->recent_scanned[file] += nr_taken;
1758 if (global_reclaim(sc)) {
1759 if (current_is_kswapd())
1760 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1763 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1767 putback_inactive_pages(lruvec, &page_list);
1769 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1771 spin_unlock_irq(&zone->lru_lock);
1773 mem_cgroup_uncharge_list(&page_list);
1774 free_hot_cold_page_list(&page_list, true);
1777 * If reclaim is isolating dirty pages under writeback, it implies
1778 * that the long-lived page allocation rate is exceeding the page
1779 * laundering rate. Either the global limits are not being effective
1780 * at throttling processes due to the page distribution throughout
1781 * zones or there is heavy usage of a slow backing device. The
1782 * only option is to throttle from reclaim context which is not ideal
1783 * as there is no guarantee the dirtying process is throttled in the
1784 * same way balance_dirty_pages() manages.
1786 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1787 * of pages under pages flagged for immediate reclaim and stall if any
1788 * are encountered in the nr_immediate check below.
1790 if (nr_writeback && nr_writeback == nr_taken)
1791 set_bit(ZONE_WRITEBACK, &zone->flags);
1794 * Legacy memcg will stall in page writeback so avoid forcibly
1797 if (sane_reclaim(sc)) {
1799 * Tag a zone as congested if all the dirty pages scanned were
1800 * backed by a congested BDI and wait_iff_congested will stall.
1802 if (nr_dirty && nr_dirty == nr_congested)
1803 set_bit(ZONE_CONGESTED, &zone->flags);
1806 * If dirty pages are scanned that are not queued for IO, it
1807 * implies that flushers are not keeping up. In this case, flag
1808 * the zone ZONE_DIRTY and kswapd will start writing pages from
1811 if (nr_unqueued_dirty == nr_taken)
1812 set_bit(ZONE_DIRTY, &zone->flags);
1815 * If kswapd scans pages marked marked for immediate
1816 * reclaim and under writeback (nr_immediate), it implies
1817 * that pages are cycling through the LRU faster than
1818 * they are written so also forcibly stall.
1820 if (nr_immediate && current_may_throttle())
1821 congestion_wait(BLK_RW_ASYNC, HZ/10);
1825 * Stall direct reclaim for IO completions if underlying BDIs or zone
1826 * is congested. Allow kswapd to continue until it starts encountering
1827 * unqueued dirty pages or cycling through the LRU too quickly.
1829 if (!sc->hibernation_mode && !current_is_kswapd() &&
1830 current_may_throttle())
1831 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1833 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1835 nr_scanned, nr_reclaimed,
1837 trace_shrink_flags(file));
1838 return nr_reclaimed;
1842 * This moves pages from the active list to the inactive list.
1844 * We move them the other way if the page is referenced by one or more
1845 * processes, from rmap.
1847 * If the pages are mostly unmapped, the processing is fast and it is
1848 * appropriate to hold zone->lru_lock across the whole operation. But if
1849 * the pages are mapped, the processing is slow (page_referenced()) so we
1850 * should drop zone->lru_lock around each page. It's impossible to balance
1851 * this, so instead we remove the pages from the LRU while processing them.
1852 * It is safe to rely on PG_active against the non-LRU pages in here because
1853 * nobody will play with that bit on a non-LRU page.
1855 * The downside is that we have to touch page->_count against each page.
1856 * But we had to alter page->flags anyway.
1859 static void move_active_pages_to_lru(struct lruvec *lruvec,
1860 struct list_head *list,
1861 struct list_head *pages_to_free,
1864 struct zone *zone = lruvec_zone(lruvec);
1865 unsigned long pgmoved = 0;
1869 while (!list_empty(list)) {
1870 page = lru_to_page(list);
1871 lruvec = mem_cgroup_page_lruvec(page, zone);
1873 VM_BUG_ON_PAGE(PageLRU(page), page);
1876 nr_pages = hpage_nr_pages(page);
1877 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1878 list_move(&page->lru, &lruvec->lists[lru]);
1879 pgmoved += nr_pages;
1881 if (put_page_testzero(page)) {
1882 __ClearPageLRU(page);
1883 __ClearPageActive(page);
1884 del_page_from_lru_list(page, lruvec, lru);
1886 if (unlikely(PageCompound(page))) {
1887 spin_unlock_irq(&zone->lru_lock);
1888 mem_cgroup_uncharge(page);
1889 (*get_compound_page_dtor(page))(page);
1890 spin_lock_irq(&zone->lru_lock);
1892 list_add(&page->lru, pages_to_free);
1895 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1896 if (!is_active_lru(lru))
1897 __count_vm_events(PGDEACTIVATE, pgmoved);
1900 static void shrink_active_list(unsigned long nr_to_scan,
1901 struct lruvec *lruvec,
1902 struct scan_control *sc,
1905 unsigned long nr_taken;
1906 unsigned long nr_scanned;
1907 unsigned long vm_flags;
1908 LIST_HEAD(l_hold); /* The pages which were snipped off */
1909 LIST_HEAD(l_active);
1910 LIST_HEAD(l_inactive);
1912 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1913 unsigned long nr_rotated = 0;
1914 isolate_mode_t isolate_mode = 0;
1915 int file = is_file_lru(lru);
1916 struct zone *zone = lruvec_zone(lruvec);
1921 isolate_mode |= ISOLATE_UNMAPPED;
1922 if (!sc->may_writepage)
1923 isolate_mode |= ISOLATE_CLEAN;
1925 spin_lock_irq(&zone->lru_lock);
1927 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1928 &nr_scanned, sc, isolate_mode, lru);
1929 if (global_reclaim(sc))
1930 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1932 reclaim_stat->recent_scanned[file] += nr_taken;
1934 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1935 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1936 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1937 spin_unlock_irq(&zone->lru_lock);
1939 while (!list_empty(&l_hold)) {
1941 page = lru_to_page(&l_hold);
1942 list_del(&page->lru);
1944 if (unlikely(!page_evictable(page))) {
1945 putback_lru_page(page);
1949 if (unlikely(buffer_heads_over_limit)) {
1950 if (page_has_private(page) && trylock_page(page)) {
1951 if (page_has_private(page))
1952 try_to_release_page(page, 0);
1957 if (page_referenced(page, 0, sc->target_mem_cgroup,
1959 nr_rotated += hpage_nr_pages(page);
1961 * Identify referenced, file-backed active pages and
1962 * give them one more trip around the active list. So
1963 * that executable code get better chances to stay in
1964 * memory under moderate memory pressure. Anon pages
1965 * are not likely to be evicted by use-once streaming
1966 * IO, plus JVM can create lots of anon VM_EXEC pages,
1967 * so we ignore them here.
1969 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1970 list_add(&page->lru, &l_active);
1975 ClearPageActive(page); /* we are de-activating */
1976 if (IS_ENABLED(CONFIG_ZCACHE))
1978 * For zcache to know whether the page is from active
1981 SetPageWasActive(page);
1982 list_add(&page->lru, &l_inactive);
1986 * Move pages back to the lru list.
1988 spin_lock_irq(&zone->lru_lock);
1990 * Count referenced pages from currently used mappings as rotated,
1991 * even though only some of them are actually re-activated. This
1992 * helps balance scan pressure between file and anonymous pages in
1995 reclaim_stat->recent_rotated[file] += nr_rotated;
1997 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1998 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1999 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
2000 spin_unlock_irq(&zone->lru_lock);
2002 mem_cgroup_uncharge_list(&l_hold);
2003 free_hot_cold_page_list(&l_hold, true);
2007 static bool inactive_anon_is_low_global(struct zone *zone)
2009 unsigned long active, inactive;
2011 active = zone_page_state(zone, NR_ACTIVE_ANON);
2012 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
2014 return inactive * zone->inactive_ratio < active;
2018 * inactive_anon_is_low - check if anonymous pages need to be deactivated
2019 * @lruvec: LRU vector to check
2021 * Returns true if the zone does not have enough inactive anon pages,
2022 * meaning some active anon pages need to be deactivated.
2024 static bool inactive_anon_is_low(struct lruvec *lruvec)
2027 * If we don't have swap space, anonymous page deactivation
2030 if (!total_swap_pages)
2033 if (!mem_cgroup_disabled())
2034 return mem_cgroup_inactive_anon_is_low(lruvec);
2036 return inactive_anon_is_low_global(lruvec_zone(lruvec));
2039 static inline bool inactive_anon_is_low(struct lruvec *lruvec)
2046 * inactive_file_is_low - check if file pages need to be deactivated
2047 * @lruvec: LRU vector to check
2049 * When the system is doing streaming IO, memory pressure here
2050 * ensures that active file pages get deactivated, until more
2051 * than half of the file pages are on the inactive list.
2053 * Once we get to that situation, protect the system's working
2054 * set from being evicted by disabling active file page aging.
2056 * This uses a different ratio than the anonymous pages, because
2057 * the page cache uses a use-once replacement algorithm.
2059 static bool inactive_file_is_low(struct lruvec *lruvec)
2061 unsigned long inactive;
2062 unsigned long active;
2064 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
2065 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
2067 return active > inactive;
2070 static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
2072 if (is_file_lru(lru))
2073 return inactive_file_is_low(lruvec);
2075 return inactive_anon_is_low(lruvec);
2078 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2079 struct lruvec *lruvec, struct scan_control *sc)
2081 if (is_active_lru(lru)) {
2082 if (inactive_list_is_low(lruvec, lru))
2083 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2087 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2098 * Determine how aggressively the anon and file LRU lists should be
2099 * scanned. The relative value of each set of LRU lists is determined
2100 * by looking at the fraction of the pages scanned we did rotate back
2101 * onto the active list instead of evict.
2103 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2104 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2106 static void get_scan_count(struct lruvec *lruvec, int swappiness,
2107 struct scan_control *sc, unsigned long *nr,
2108 unsigned long *lru_pages)
2110 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2112 u64 denominator = 0; /* gcc */
2113 struct zone *zone = lruvec_zone(lruvec);
2114 unsigned long anon_prio, file_prio;
2115 enum scan_balance scan_balance;
2116 unsigned long anon, file;
2117 bool force_scan = false;
2118 unsigned long ap, fp;
2124 * If the zone or memcg is small, nr[l] can be 0. This
2125 * results in no scanning on this priority and a potential
2126 * priority drop. Global direct reclaim can go to the next
2127 * zone and tends to have no problems. Global kswapd is for
2128 * zone balancing and it needs to scan a minimum amount. When
2129 * reclaiming for a memcg, a priority drop can cause high
2130 * latencies, so it's better to scan a minimum amount there as
2133 if (current_is_kswapd()) {
2134 if (!zone_reclaimable(zone))
2136 if (!mem_cgroup_lruvec_online(lruvec))
2139 if (!global_reclaim(sc))
2142 /* If we have no swap space, do not bother scanning anon pages. */
2143 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
2144 scan_balance = SCAN_FILE;
2149 * Global reclaim will swap to prevent OOM even with no
2150 * swappiness, but memcg users want to use this knob to
2151 * disable swapping for individual groups completely when
2152 * using the memory controller's swap limit feature would be
2155 if (!global_reclaim(sc) && !swappiness) {
2156 scan_balance = SCAN_FILE;
2161 * Do not apply any pressure balancing cleverness when the
2162 * system is close to OOM, scan both anon and file equally
2163 * (unless the swappiness setting disagrees with swapping).
2165 if (!sc->priority && swappiness) {
2166 scan_balance = SCAN_EQUAL;
2171 * Prevent the reclaimer from falling into the cache trap: as
2172 * cache pages start out inactive, every cache fault will tip
2173 * the scan balance towards the file LRU. And as the file LRU
2174 * shrinks, so does the window for rotation from references.
2175 * This means we have a runaway feedback loop where a tiny
2176 * thrashing file LRU becomes infinitely more attractive than
2177 * anon pages. Try to detect this based on file LRU size.
2179 if (global_reclaim(sc)) {
2180 unsigned long zonefile;
2181 unsigned long zonefree;
2183 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2184 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2185 zone_page_state(zone, NR_INACTIVE_FILE);
2187 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2188 scan_balance = SCAN_ANON;
2194 * There is enough inactive page cache, do not reclaim
2195 * anything from the anonymous working set right now.
2197 if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM) &&
2198 !inactive_file_is_low(lruvec)) {
2199 scan_balance = SCAN_FILE;
2203 scan_balance = SCAN_FRACT;
2206 * With swappiness at 100, anonymous and file have the same priority.
2207 * This scanning priority is essentially the inverse of IO cost.
2209 anon_prio = swappiness;
2210 file_prio = 200 - anon_prio;
2213 * OK, so we have swap space and a fair amount of page cache
2214 * pages. We use the recently rotated / recently scanned
2215 * ratios to determine how valuable each cache is.
2217 * Because workloads change over time (and to avoid overflow)
2218 * we keep these statistics as a floating average, which ends
2219 * up weighing recent references more than old ones.
2221 * anon in [0], file in [1]
2224 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2225 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2226 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2227 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2229 spin_lock_irq(&zone->lru_lock);
2230 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2231 reclaim_stat->recent_scanned[0] /= 2;
2232 reclaim_stat->recent_rotated[0] /= 2;
2235 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2236 reclaim_stat->recent_scanned[1] /= 2;
2237 reclaim_stat->recent_rotated[1] /= 2;
2241 * The amount of pressure on anon vs file pages is inversely
2242 * proportional to the fraction of recently scanned pages on
2243 * each list that were recently referenced and in active use.
2245 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2246 ap /= reclaim_stat->recent_rotated[0] + 1;
2248 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2249 fp /= reclaim_stat->recent_rotated[1] + 1;
2250 spin_unlock_irq(&zone->lru_lock);
2254 denominator = ap + fp + 1;
2256 some_scanned = false;
2257 /* Only use force_scan on second pass. */
2258 for (pass = 0; !some_scanned && pass < 2; pass++) {
2260 for_each_evictable_lru(lru) {
2261 int file = is_file_lru(lru);
2265 size = get_lru_size(lruvec, lru);
2266 scan = size >> sc->priority;
2268 if (!scan && pass && force_scan)
2269 scan = min(size, SWAP_CLUSTER_MAX);
2271 switch (scan_balance) {
2273 /* Scan lists relative to size */
2277 * Scan types proportional to swappiness and
2278 * their relative recent reclaim efficiency.
2280 scan = div64_u64(scan * fraction[file],
2285 /* Scan one type exclusively */
2286 if ((scan_balance == SCAN_FILE) != file) {
2292 /* Look ma, no brain */
2300 * Skip the second pass and don't force_scan,
2301 * if we found something to scan.
2303 some_scanned |= !!scan;
2309 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2311 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2312 struct scan_control *sc, unsigned long *lru_pages)
2314 unsigned long nr[NR_LRU_LISTS];
2315 unsigned long targets[NR_LRU_LISTS];
2316 unsigned long nr_to_scan;
2318 unsigned long nr_reclaimed = 0;
2319 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2320 struct blk_plug plug;
2323 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2325 /* Record the original scan target for proportional adjustments later */
2326 memcpy(targets, nr, sizeof(nr));
2329 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2330 * event that can occur when there is little memory pressure e.g.
2331 * multiple streaming readers/writers. Hence, we do not abort scanning
2332 * when the requested number of pages are reclaimed when scanning at
2333 * DEF_PRIORITY on the assumption that the fact we are direct
2334 * reclaiming implies that kswapd is not keeping up and it is best to
2335 * do a batch of work at once. For memcg reclaim one check is made to
2336 * abort proportional reclaim if either the file or anon lru has already
2337 * dropped to zero at the first pass.
2339 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2340 sc->priority == DEF_PRIORITY);
2342 blk_start_plug(&plug);
2343 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2344 nr[LRU_INACTIVE_FILE]) {
2345 unsigned long nr_anon, nr_file, percentage;
2346 unsigned long nr_scanned;
2348 for_each_evictable_lru(lru) {
2350 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2351 nr[lru] -= nr_to_scan;
2353 nr_reclaimed += shrink_list(lru, nr_to_scan,
2358 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2362 * For kswapd and memcg, reclaim at least the number of pages
2363 * requested. Ensure that the anon and file LRUs are scanned
2364 * proportionally what was requested by get_scan_count(). We
2365 * stop reclaiming one LRU and reduce the amount scanning
2366 * proportional to the original scan target.
2368 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2369 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2372 * It's just vindictive to attack the larger once the smaller
2373 * has gone to zero. And given the way we stop scanning the
2374 * smaller below, this makes sure that we only make one nudge
2375 * towards proportionality once we've got nr_to_reclaim.
2377 if (!nr_file || !nr_anon)
2380 if (nr_file > nr_anon) {
2381 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2382 targets[LRU_ACTIVE_ANON] + 1;
2384 percentage = nr_anon * 100 / scan_target;
2386 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2387 targets[LRU_ACTIVE_FILE] + 1;
2389 percentage = nr_file * 100 / scan_target;
2392 /* Stop scanning the smaller of the LRU */
2394 nr[lru + LRU_ACTIVE] = 0;
2397 * Recalculate the other LRU scan count based on its original
2398 * scan target and the percentage scanning already complete
2400 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2401 nr_scanned = targets[lru] - nr[lru];
2402 nr[lru] = targets[lru] * (100 - percentage) / 100;
2403 nr[lru] -= min(nr[lru], nr_scanned);
2406 nr_scanned = targets[lru] - nr[lru];
2407 nr[lru] = targets[lru] * (100 - percentage) / 100;
2408 nr[lru] -= min(nr[lru], nr_scanned);
2410 scan_adjusted = true;
2412 blk_finish_plug(&plug);
2413 sc->nr_reclaimed += nr_reclaimed;
2416 * Even if we did not try to evict anon pages at all, we want to
2417 * rebalance the anon lru active/inactive ratio.
2419 if (inactive_anon_is_low(lruvec))
2420 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2421 sc, LRU_ACTIVE_ANON);
2423 throttle_vm_writeout(sc->gfp_mask);
2426 /* Use reclaim/compaction for costly allocs or under memory pressure */
2427 static bool in_reclaim_compaction(struct scan_control *sc)
2429 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2430 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2431 sc->priority < DEF_PRIORITY - 2))
2438 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2439 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2440 * true if more pages should be reclaimed such that when the page allocator
2441 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2442 * It will give up earlier than that if there is difficulty reclaiming pages.
2444 static inline bool should_continue_reclaim(struct zone *zone,
2445 unsigned long nr_reclaimed,
2446 unsigned long nr_scanned,
2447 struct scan_control *sc)
2449 unsigned long pages_for_compaction;
2450 unsigned long inactive_lru_pages;
2452 /* If not in reclaim/compaction mode, stop */
2453 if (!in_reclaim_compaction(sc))
2456 /* Consider stopping depending on scan and reclaim activity */
2457 if (sc->gfp_mask & __GFP_REPEAT) {
2459 * For __GFP_REPEAT allocations, stop reclaiming if the
2460 * full LRU list has been scanned and we are still failing
2461 * to reclaim pages. This full LRU scan is potentially
2462 * expensive but a __GFP_REPEAT caller really wants to succeed
2464 if (!nr_reclaimed && !nr_scanned)
2468 * For non-__GFP_REPEAT allocations which can presumably
2469 * fail without consequence, stop if we failed to reclaim
2470 * any pages from the last SWAP_CLUSTER_MAX number of
2471 * pages that were scanned. This will return to the
2472 * caller faster at the risk reclaim/compaction and
2473 * the resulting allocation attempt fails
2480 * If we have not reclaimed enough pages for compaction and the
2481 * inactive lists are large enough, continue reclaiming
2483 pages_for_compaction = (2UL << sc->order);
2484 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2485 if (get_nr_swap_pages() > 0)
2486 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2487 if (sc->nr_reclaimed < pages_for_compaction &&
2488 inactive_lru_pages > pages_for_compaction)
2491 /* If compaction would go ahead or the allocation would succeed, stop */
2492 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2493 case COMPACT_PARTIAL:
2494 case COMPACT_CONTINUE:
2501 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2504 struct reclaim_state *reclaim_state = current->reclaim_state;
2505 unsigned long nr_reclaimed, nr_scanned;
2506 bool reclaimable = false;
2509 struct mem_cgroup *root = sc->target_mem_cgroup;
2510 struct mem_cgroup_reclaim_cookie reclaim = {
2512 .priority = sc->priority,
2514 unsigned long zone_lru_pages = 0;
2515 struct mem_cgroup *memcg;
2517 nr_reclaimed = sc->nr_reclaimed;
2518 nr_scanned = sc->nr_scanned;
2520 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2522 unsigned long lru_pages;
2523 unsigned long scanned;
2524 struct lruvec *lruvec;
2527 if (mem_cgroup_low(root, memcg)) {
2528 if (!sc->may_thrash)
2530 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2533 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2534 swappiness = mem_cgroup_swappiness(memcg);
2535 scanned = sc->nr_scanned;
2537 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2538 zone_lru_pages += lru_pages;
2540 if (memcg && is_classzone)
2541 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2542 memcg, sc->nr_scanned - scanned,
2546 * Direct reclaim and kswapd have to scan all memory
2547 * cgroups to fulfill the overall scan target for the
2550 * Limit reclaim, on the other hand, only cares about
2551 * nr_to_reclaim pages to be reclaimed and it will
2552 * retry with decreasing priority if one round over the
2553 * whole hierarchy is not sufficient.
2555 if (!global_reclaim(sc) &&
2556 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2557 mem_cgroup_iter_break(root, memcg);
2560 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2563 * Shrink the slab caches in the same proportion that
2564 * the eligible LRU pages were scanned.
2566 if (global_reclaim(sc) && is_classzone)
2567 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2568 sc->nr_scanned - nr_scanned,
2572 * Record the subtree's reclaim efficiency. The reclaimed
2573 * pages from slab is excluded here because the corresponding
2574 * scanned pages is not accounted. Moreover, freeing a page
2575 * by slab shrinking depends on each slab's object population,
2576 * making the cost model (i.e. scan:free) different from that
2579 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2580 sc->nr_scanned - nr_scanned,
2581 sc->nr_reclaimed - nr_reclaimed);
2583 if (reclaim_state) {
2584 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2585 reclaim_state->reclaimed_slab = 0;
2588 if (sc->nr_reclaimed - nr_reclaimed)
2591 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2592 sc->nr_scanned - nr_scanned, sc));
2598 * Returns true if compaction should go ahead for a high-order request, or
2599 * the high-order allocation would succeed without compaction.
2601 static inline bool compaction_ready(struct zone *zone, int order)
2603 unsigned long balance_gap, watermark;
2607 * Compaction takes time to run and there are potentially other
2608 * callers using the pages just freed. Continue reclaiming until
2609 * there is a buffer of free pages available to give compaction
2610 * a reasonable chance of completing and allocating the page
2612 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2613 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2614 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2615 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2618 * If compaction is deferred, reclaim up to a point where
2619 * compaction will have a chance of success when re-enabled
2621 if (compaction_deferred(zone, order))
2622 return watermark_ok;
2625 * If compaction is not ready to start and allocation is not likely
2626 * to succeed without it, then keep reclaiming.
2628 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2631 return watermark_ok;
2635 * This is the direct reclaim path, for page-allocating processes. We only
2636 * try to reclaim pages from zones which will satisfy the caller's allocation
2639 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2641 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2643 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2644 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2645 * zone defense algorithm.
2647 * If a zone is deemed to be full of pinned pages then just give it a light
2648 * scan then give up on it.
2650 * Returns true if a zone was reclaimable.
2652 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2656 unsigned long nr_soft_reclaimed;
2657 unsigned long nr_soft_scanned;
2659 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2660 bool reclaimable = false;
2661 unsigned long lru_pages = 0;
2664 * If the number of buffer_heads in the machine exceeds the maximum
2665 * allowed level, force direct reclaim to scan the highmem zone as
2666 * highmem pages could be pinning lowmem pages storing buffer_heads
2668 orig_mask = sc->gfp_mask;
2669 if (buffer_heads_over_limit)
2670 sc->gfp_mask |= __GFP_HIGHMEM;
2672 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2673 gfp_zone(sc->gfp_mask), sc->nodemask) {
2674 enum zone_type classzone_idx;
2676 if (!populated_zone(zone))
2679 classzone_idx = gfp_zone(sc->gfp_mask);
2680 while (!populated_zone(zone->zone_pgdat->node_zones +
2685 * Take care memory controller reclaiming has small influence
2688 if (global_reclaim(sc)) {
2689 lru_pages += zone_reclaimable_pages(zone);
2690 if (!cpuset_zone_allowed(zone,
2691 GFP_KERNEL | __GFP_HARDWALL))
2694 if (sc->priority != DEF_PRIORITY &&
2695 !zone_reclaimable(zone))
2696 continue; /* Let kswapd poll it */
2699 * If we already have plenty of memory free for
2700 * compaction in this zone, don't free any more.
2701 * Even though compaction is invoked for any
2702 * non-zero order, only frequent costly order
2703 * reclamation is disruptive enough to become a
2704 * noticeable problem, like transparent huge
2707 if (IS_ENABLED(CONFIG_COMPACTION) &&
2708 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2709 zonelist_zone_idx(z) <= requested_highidx &&
2710 compaction_ready(zone, sc->order)) {
2711 sc->compaction_ready = true;
2716 * This steals pages from memory cgroups over softlimit
2717 * and returns the number of reclaimed pages and
2718 * scanned pages. This works for global memory pressure
2719 * and balancing, not for a memcg's limit.
2721 nr_soft_scanned = 0;
2722 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2723 sc->order, sc->gfp_mask,
2725 sc->nr_reclaimed += nr_soft_reclaimed;
2726 sc->nr_scanned += nr_soft_scanned;
2727 if (nr_soft_reclaimed)
2729 /* need some check for avoid more shrink_zone() */
2732 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2735 if (global_reclaim(sc) &&
2736 !reclaimable && zone_reclaimable(zone))
2740 if (global_reclaim(sc))
2741 shrink_slab_lmk(sc->gfp_mask, 0, NULL,
2742 sc->nr_scanned, lru_pages);
2744 * Restore to original mask to avoid the impact on the caller if we
2745 * promoted it to __GFP_HIGHMEM.
2747 sc->gfp_mask = orig_mask;
2753 * This is the main entry point to direct page reclaim.
2755 * If a full scan of the inactive list fails to free enough memory then we
2756 * are "out of memory" and something needs to be killed.
2758 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2759 * high - the zone may be full of dirty or under-writeback pages, which this
2760 * caller can't do much about. We kick the writeback threads and take explicit
2761 * naps in the hope that some of these pages can be written. But if the
2762 * allocating task holds filesystem locks which prevent writeout this might not
2763 * work, and the allocation attempt will fail.
2765 * returns: 0, if no pages reclaimed
2766 * else, the number of pages reclaimed
2768 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2769 struct scan_control *sc)
2771 int initial_priority = sc->priority;
2772 unsigned long total_scanned = 0;
2773 unsigned long writeback_threshold;
2774 bool zones_reclaimable;
2776 delayacct_freepages_start();
2778 if (global_reclaim(sc))
2779 count_vm_event(ALLOCSTALL);
2782 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2785 zones_reclaimable = shrink_zones(zonelist, sc);
2787 total_scanned += sc->nr_scanned;
2788 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2791 if (sc->compaction_ready)
2795 * If we're getting trouble reclaiming, start doing
2796 * writepage even in laptop mode.
2798 if (sc->priority < DEF_PRIORITY - 2)
2799 sc->may_writepage = 1;
2802 * Try to write back as many pages as we just scanned. This
2803 * tends to cause slow streaming writers to write data to the
2804 * disk smoothly, at the dirtying rate, which is nice. But
2805 * that's undesirable in laptop mode, where we *want* lumpy
2806 * writeout. So in laptop mode, write out the whole world.
2808 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2809 if (total_scanned > writeback_threshold) {
2810 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2811 WB_REASON_TRY_TO_FREE_PAGES);
2812 sc->may_writepage = 1;
2814 } while (--sc->priority >= 0);
2816 delayacct_freepages_end();
2818 if (sc->nr_reclaimed)
2819 return sc->nr_reclaimed;
2821 /* Aborted reclaim to try compaction? don't OOM, then */
2822 if (sc->compaction_ready)
2825 /* Untapped cgroup reserves? Don't OOM, retry. */
2826 if (!sc->may_thrash) {
2827 sc->priority = initial_priority;
2832 /* Any of the zones still reclaimable? Don't OOM. */
2833 if (zones_reclaimable)
2839 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2842 unsigned long pfmemalloc_reserve = 0;
2843 unsigned long free_pages = 0;
2847 for (i = 0; i <= ZONE_NORMAL; i++) {
2848 zone = &pgdat->node_zones[i];
2849 if (!populated_zone(zone) ||
2850 zone_reclaimable_pages(zone) == 0)
2853 pfmemalloc_reserve += min_wmark_pages(zone);
2854 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2857 /* If there are no reserves (unexpected config) then do not throttle */
2858 if (!pfmemalloc_reserve)
2861 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2863 /* kswapd must be awake if processes are being throttled */
2864 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2865 pgdat->classzone_idx = min(pgdat->classzone_idx,
2866 (enum zone_type)ZONE_NORMAL);
2867 wake_up_interruptible(&pgdat->kswapd_wait);
2874 * Throttle direct reclaimers if backing storage is backed by the network
2875 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2876 * depleted. kswapd will continue to make progress and wake the processes
2877 * when the low watermark is reached.
2879 * Returns true if a fatal signal was delivered during throttling. If this
2880 * happens, the page allocator should not consider triggering the OOM killer.
2882 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2883 nodemask_t *nodemask)
2887 pg_data_t *pgdat = NULL;
2890 * Kernel threads should not be throttled as they may be indirectly
2891 * responsible for cleaning pages necessary for reclaim to make forward
2892 * progress. kjournald for example may enter direct reclaim while
2893 * committing a transaction where throttling it could forcing other
2894 * processes to block on log_wait_commit().
2896 if (current->flags & PF_KTHREAD)
2900 * If a fatal signal is pending, this process should not throttle.
2901 * It should return quickly so it can exit and free its memory
2903 if (fatal_signal_pending(current))
2907 * Check if the pfmemalloc reserves are ok by finding the first node
2908 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2909 * GFP_KERNEL will be required for allocating network buffers when
2910 * swapping over the network so ZONE_HIGHMEM is unusable.
2912 * Throttling is based on the first usable node and throttled processes
2913 * wait on a queue until kswapd makes progress and wakes them. There
2914 * is an affinity then between processes waking up and where reclaim
2915 * progress has been made assuming the process wakes on the same node.
2916 * More importantly, processes running on remote nodes will not compete
2917 * for remote pfmemalloc reserves and processes on different nodes
2918 * should make reasonable progress.
2920 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2921 gfp_zone(gfp_mask), nodemask) {
2922 if (zone_idx(zone) > ZONE_NORMAL)
2925 /* Throttle based on the first usable node */
2926 pgdat = zone->zone_pgdat;
2927 if (pfmemalloc_watermark_ok(pgdat))
2932 /* If no zone was usable by the allocation flags then do not throttle */
2936 /* Account for the throttling */
2937 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2940 * If the caller cannot enter the filesystem, it's possible that it
2941 * is due to the caller holding an FS lock or performing a journal
2942 * transaction in the case of a filesystem like ext[3|4]. In this case,
2943 * it is not safe to block on pfmemalloc_wait as kswapd could be
2944 * blocked waiting on the same lock. Instead, throttle for up to a
2945 * second before continuing.
2947 if (!(gfp_mask & __GFP_FS)) {
2948 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2949 pfmemalloc_watermark_ok(pgdat), HZ);
2954 /* Throttle until kswapd wakes the process */
2955 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2956 pfmemalloc_watermark_ok(pgdat));
2959 if (fatal_signal_pending(current))
2966 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2967 gfp_t gfp_mask, nodemask_t *nodemask)
2969 unsigned long nr_reclaimed;
2970 struct scan_control sc = {
2971 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2972 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2974 .nodemask = nodemask,
2975 .priority = DEF_PRIORITY,
2976 .may_writepage = !laptop_mode,
2982 * Do not enter reclaim if fatal signal was delivered while throttled.
2983 * 1 is returned so that the page allocator does not OOM kill at this
2986 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2989 trace_mm_vmscan_direct_reclaim_begin(order,
2993 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2995 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2997 return nr_reclaimed;
3002 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
3003 gfp_t gfp_mask, bool noswap,
3005 unsigned long *nr_scanned)
3007 struct scan_control sc = {
3008 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3009 .target_mem_cgroup = memcg,
3010 .may_writepage = !laptop_mode,
3012 .may_swap = !noswap,
3014 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3015 int swappiness = mem_cgroup_swappiness(memcg);
3016 unsigned long lru_pages;
3018 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3019 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3021 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3026 * NOTE: Although we can get the priority field, using it
3027 * here is not a good idea, since it limits the pages we can scan.
3028 * if we don't reclaim here, the shrink_zone from balance_pgdat
3029 * will pick up pages from other mem cgroup's as well. We hack
3030 * the priority and make it zero.
3032 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
3034 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3036 *nr_scanned = sc.nr_scanned;
3037 return sc.nr_reclaimed;
3040 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3041 unsigned long nr_pages,
3045 struct zonelist *zonelist;
3046 unsigned long nr_reclaimed;
3048 struct scan_control sc = {
3049 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3050 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3051 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3052 .target_mem_cgroup = memcg,
3053 .priority = DEF_PRIORITY,
3054 .may_writepage = !laptop_mode,
3056 .may_swap = may_swap,
3060 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3061 * take care of from where we get pages. So the node where we start the
3062 * scan does not need to be the current node.
3064 nid = mem_cgroup_select_victim_node(memcg);
3066 zonelist = NODE_DATA(nid)->node_zonelists;
3068 trace_mm_vmscan_memcg_reclaim_begin(0,
3072 current->flags |= PF_MEMALLOC;
3073 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3074 current->flags &= ~PF_MEMALLOC;
3076 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3078 return nr_reclaimed;
3082 static void age_active_anon(struct zone *zone, struct scan_control *sc)
3084 struct mem_cgroup *memcg;
3086 if (!total_swap_pages)
3089 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3091 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3093 if (inactive_anon_is_low(lruvec))
3094 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3095 sc, LRU_ACTIVE_ANON);
3097 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3101 static bool zone_balanced(struct zone *zone, int order, bool highorder,
3102 unsigned long balance_gap, int classzone_idx)
3104 unsigned long mark = high_wmark_pages(zone) + balance_gap;
3107 * When checking from pgdat_balanced(), kswapd should stop and sleep
3108 * when it reaches the high order-0 watermark and let kcompactd take
3109 * over. Other callers such as wakeup_kswapd() want to determine the
3110 * true high-order watermark.
3112 if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) {
3113 mark += (1UL << order);
3117 return zone_watermark_ok_safe(zone, order, mark, classzone_idx);
3121 * pgdat_balanced() is used when checking if a node is balanced.
3123 * For order-0, all zones must be balanced!
3125 * For high-order allocations only zones that meet watermarks and are in a
3126 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3127 * total of balanced pages must be at least 25% of the zones allowed by
3128 * classzone_idx for the node to be considered balanced. Forcing all zones to
3129 * be balanced for high orders can cause excessive reclaim when there are
3131 * The choice of 25% is due to
3132 * o a 16M DMA zone that is balanced will not balance a zone on any
3133 * reasonable sized machine
3134 * o On all other machines, the top zone must be at least a reasonable
3135 * percentage of the middle zones. For example, on 32-bit x86, highmem
3136 * would need to be at least 256M for it to be balance a whole node.
3137 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3138 * to balance a node on its own. These seemed like reasonable ratios.
3140 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3142 unsigned long managed_pages = 0;
3143 unsigned long balanced_pages = 0;
3146 /* Check the watermark levels */
3147 for (i = 0; i <= classzone_idx; i++) {
3148 struct zone *zone = pgdat->node_zones + i;
3150 if (!populated_zone(zone))
3153 managed_pages += zone->managed_pages;
3156 * A special case here:
3158 * balance_pgdat() skips over all_unreclaimable after
3159 * DEF_PRIORITY. Effectively, it considers them balanced so
3160 * they must be considered balanced here as well!
3162 if (!zone_reclaimable(zone)) {
3163 balanced_pages += zone->managed_pages;
3167 if (zone_balanced(zone, order, false, 0, i))
3168 balanced_pages += zone->managed_pages;
3174 return balanced_pages >= (managed_pages >> 2);
3180 * Prepare kswapd for sleeping. This verifies that there are no processes
3181 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3183 * Returns true if kswapd is ready to sleep
3185 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3188 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3193 * The throttled processes are normally woken up in balance_pgdat() as
3194 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3195 * race between when kswapd checks the watermarks and a process gets
3196 * throttled. There is also a potential race if processes get
3197 * throttled, kswapd wakes, a large process exits thereby balancing the
3198 * zones, which causes kswapd to exit balance_pgdat() before reaching
3199 * the wake up checks. If kswapd is going to sleep, no process should
3200 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3201 * the wake up is premature, processes will wake kswapd and get
3202 * throttled again. The difference from wake ups in balance_pgdat() is
3203 * that here we are under prepare_to_wait().
3205 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3206 wake_up_all(&pgdat->pfmemalloc_wait);
3208 return pgdat_balanced(pgdat, order, classzone_idx);
3212 * kswapd shrinks the zone by the number of pages required to reach
3213 * the high watermark.
3215 * Returns true if kswapd scanned at least the requested number of pages to
3216 * reclaim or if the lack of progress was due to pages under writeback.
3217 * This is used to determine if the scanning priority needs to be raised.
3219 static bool kswapd_shrink_zone(struct zone *zone,
3221 struct scan_control *sc,
3222 unsigned long lru_pages)
3224 unsigned long balance_gap;
3225 bool lowmem_pressure;
3227 /* Reclaim above the high watermark. */
3228 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3231 * We put equal pressure on every zone, unless one zone has way too
3232 * many pages free already. The "too many pages" is defined as the
3233 * high wmark plus a "gap" where the gap is either the low
3234 * watermark or 1% of the zone, whichever is smaller.
3236 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3237 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3240 * If there is no low memory pressure or the zone is balanced then no
3241 * reclaim is necessary
3243 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3244 if (!lowmem_pressure && zone_balanced(zone, sc->order, false,
3245 balance_gap, classzone_idx))
3248 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3249 shrink_slab_lmk(sc->gfp_mask, zone_to_nid(zone), NULL,
3250 sc->nr_scanned, lru_pages);
3252 clear_bit(ZONE_WRITEBACK, &zone->flags);
3255 * If a zone reaches its high watermark, consider it to be no longer
3256 * congested. It's possible there are dirty pages backed by congested
3257 * BDIs but as pressure is relieved, speculatively avoid congestion
3260 if (zone_reclaimable(zone) &&
3261 zone_balanced(zone, sc->order, false, 0, classzone_idx)) {
3262 clear_bit(ZONE_CONGESTED, &zone->flags);
3263 clear_bit(ZONE_DIRTY, &zone->flags);
3266 return sc->nr_scanned >= sc->nr_to_reclaim;
3270 * For kswapd, balance_pgdat() will work across all this node's zones until
3271 * they are all at high_wmark_pages(zone).
3273 * Returns the highest zone idx kswapd was reclaiming at
3275 * There is special handling here for zones which are full of pinned pages.
3276 * This can happen if the pages are all mlocked, or if they are all used by
3277 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3278 * What we do is to detect the case where all pages in the zone have been
3279 * scanned twice and there has been zero successful reclaim. Mark the zone as
3280 * dead and from now on, only perform a short scan. Basically we're polling
3281 * the zone for when the problem goes away.
3283 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3284 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3285 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3286 * lower zones regardless of the number of free pages in the lower zones. This
3287 * interoperates with the page allocator fallback scheme to ensure that aging
3288 * of pages is balanced across the zones.
3290 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3293 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3294 unsigned long nr_soft_reclaimed;
3295 unsigned long nr_soft_scanned;
3296 struct scan_control sc = {
3297 .gfp_mask = GFP_KERNEL,
3299 .priority = DEF_PRIORITY,
3300 .may_writepage = !laptop_mode,
3304 count_vm_event(PAGEOUTRUN);
3307 bool raise_priority = true;
3308 unsigned long lru_pages = 0;
3310 sc.nr_reclaimed = 0;
3313 * Scan in the highmem->dma direction for the highest
3314 * zone which needs scanning
3316 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3317 struct zone *zone = pgdat->node_zones + i;
3319 if (!populated_zone(zone))
3322 if (sc.priority != DEF_PRIORITY &&
3323 !zone_reclaimable(zone))
3327 * Do some background aging of the anon list, to give
3328 * pages a chance to be referenced before reclaiming.
3330 age_active_anon(zone, &sc);
3333 * If the number of buffer_heads in the machine
3334 * exceeds the maximum allowed level and this node
3335 * has a highmem zone, force kswapd to reclaim from
3336 * it to relieve lowmem pressure.
3338 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3343 if (!zone_balanced(zone, order, false, 0, 0)) {
3348 * If balanced, clear the dirty and congested
3351 clear_bit(ZONE_CONGESTED, &zone->flags);
3352 clear_bit(ZONE_DIRTY, &zone->flags);
3360 * If we're getting trouble reclaiming, start doing writepage
3361 * even in laptop mode.
3363 if (sc.priority < DEF_PRIORITY - 2)
3364 sc.may_writepage = 1;
3366 for (i = 0; i <= end_zone; i++) {
3367 struct zone *zone = pgdat->node_zones + i;
3369 if (!populated_zone(zone))
3372 lru_pages += zone_reclaimable_pages(zone);
3376 * Now scan the zone in the dma->highmem direction, stopping
3377 * at the last zone which needs scanning.
3379 * We do this because the page allocator works in the opposite
3380 * direction. This prevents the page allocator from allocating
3381 * pages behind kswapd's direction of progress, which would
3382 * cause too much scanning of the lower zones.
3384 for (i = 0; i <= end_zone; i++) {
3385 struct zone *zone = pgdat->node_zones + i;
3387 if (!populated_zone(zone))
3390 if (sc.priority != DEF_PRIORITY &&
3391 !zone_reclaimable(zone))
3396 nr_soft_scanned = 0;
3398 * Call soft limit reclaim before calling shrink_zone.
3400 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3403 sc.nr_reclaimed += nr_soft_reclaimed;
3406 * There should be no need to raise the scanning
3407 * priority if enough pages are already being scanned
3408 * that that high watermark would be met at 100%
3411 if (kswapd_shrink_zone(zone, end_zone, &sc, lru_pages))
3412 raise_priority = false;
3416 * If the low watermark is met there is no need for processes
3417 * to be throttled on pfmemalloc_wait as they should not be
3418 * able to safely make forward progress. Wake them
3420 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3421 pfmemalloc_watermark_ok(pgdat))
3422 wake_up_all(&pgdat->pfmemalloc_wait);
3424 /* Check if kswapd should be suspending */
3425 if (try_to_freeze() || kthread_should_stop())
3429 * Raise priority if scanning rate is too low or there was no
3430 * progress in reclaiming pages
3432 if (raise_priority || !sc.nr_reclaimed)
3434 } while (sc.priority >= 1 &&
3435 !pgdat_balanced(pgdat, order, classzone_idx));
3439 * Return the highest zone idx we were reclaiming at so
3440 * prepare_kswapd_sleep() makes the same decisions as here.
3445 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order,
3446 int classzone_idx, int balanced_classzone_idx)
3451 if (freezing(current) || kthread_should_stop())
3454 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3456 /* Try to sleep for a short interval */
3457 if (prepare_kswapd_sleep(pgdat, order, remaining,
3458 balanced_classzone_idx)) {
3460 * Compaction records what page blocks it recently failed to
3461 * isolate pages from and skips them in the future scanning.
3462 * When kswapd is going to sleep, it is reasonable to assume
3463 * that pages and compaction may succeed so reset the cache.
3465 reset_isolation_suitable(pgdat);
3468 * We have freed the memory, now we should compact it to make
3469 * allocation of the requested order possible.
3471 wakeup_kcompactd(pgdat, order, classzone_idx);
3473 remaining = schedule_timeout(HZ/10);
3474 finish_wait(&pgdat->kswapd_wait, &wait);
3475 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3479 * After a short sleep, check if it was a premature sleep. If not, then
3480 * go fully to sleep until explicitly woken up.
3482 if (prepare_kswapd_sleep(pgdat, order, remaining,
3483 balanced_classzone_idx)) {
3484 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3487 * vmstat counters are not perfectly accurate and the estimated
3488 * value for counters such as NR_FREE_PAGES can deviate from the
3489 * true value by nr_online_cpus * threshold. To avoid the zone
3490 * watermarks being breached while under pressure, we reduce the
3491 * per-cpu vmstat threshold while kswapd is awake and restore
3492 * them before going back to sleep.
3494 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3496 if (!kthread_should_stop())
3499 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3502 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3504 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3506 finish_wait(&pgdat->kswapd_wait, &wait);
3510 * The background pageout daemon, started as a kernel thread
3511 * from the init process.
3513 * This basically trickles out pages so that we have _some_
3514 * free memory available even if there is no other activity
3515 * that frees anything up. This is needed for things like routing
3516 * etc, where we otherwise might have all activity going on in
3517 * asynchronous contexts that cannot page things out.
3519 * If there are applications that are active memory-allocators
3520 * (most normal use), this basically shouldn't matter.
3522 static int kswapd(void *p)
3524 unsigned long order, new_order;
3525 int classzone_idx, new_classzone_idx;
3526 int balanced_classzone_idx;
3527 pg_data_t *pgdat = (pg_data_t*)p;
3528 struct task_struct *tsk = current;
3530 struct reclaim_state reclaim_state = {
3531 .reclaimed_slab = 0,
3533 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3535 lockdep_set_current_reclaim_state(GFP_KERNEL);
3537 if (kswapd_cpu_mask == NULL && !cpumask_empty(cpumask))
3538 set_cpus_allowed_ptr(tsk, cpumask);
3539 current->reclaim_state = &reclaim_state;
3542 * Tell the memory management that we're a "memory allocator",
3543 * and that if we need more memory we should get access to it
3544 * regardless (see "__alloc_pages()"). "kswapd" should
3545 * never get caught in the normal page freeing logic.
3547 * (Kswapd normally doesn't need memory anyway, but sometimes
3548 * you need a small amount of memory in order to be able to
3549 * page out something else, and this flag essentially protects
3550 * us from recursively trying to free more memory as we're
3551 * trying to free the first piece of memory in the first place).
3553 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3556 order = new_order = 0;
3557 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3558 balanced_classzone_idx = classzone_idx;
3563 * While we were reclaiming, there might have been another
3564 * wakeup, so check the values.
3566 new_order = pgdat->kswapd_max_order;
3567 new_classzone_idx = pgdat->classzone_idx;
3568 pgdat->kswapd_max_order = 0;
3569 pgdat->classzone_idx = pgdat->nr_zones - 1;
3571 if (order < new_order || classzone_idx > new_classzone_idx) {
3573 * Don't sleep if someone wants a larger 'order'
3574 * allocation or has tigher zone constraints
3577 classzone_idx = new_classzone_idx;
3579 kswapd_try_to_sleep(pgdat, order, classzone_idx,
3580 balanced_classzone_idx);
3581 order = pgdat->kswapd_max_order;
3582 classzone_idx = pgdat->classzone_idx;
3584 new_classzone_idx = classzone_idx;
3585 pgdat->kswapd_max_order = 0;
3586 pgdat->classzone_idx = pgdat->nr_zones - 1;
3589 ret = try_to_freeze();
3590 if (kthread_should_stop())
3594 * We can speed up thawing tasks if we don't call balance_pgdat
3595 * after returning from the refrigerator
3598 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3599 balanced_classzone_idx = balance_pgdat(pgdat, order,
3604 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3605 current->reclaim_state = NULL;
3606 lockdep_clear_current_reclaim_state();
3612 * A zone is low on free memory, so wake its kswapd task to service it.
3614 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3618 if (!populated_zone(zone))
3621 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3623 pgdat = zone->zone_pgdat;
3624 if (pgdat->kswapd_max_order < order) {
3625 pgdat->kswapd_max_order = order;
3626 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3628 if (!waitqueue_active(&pgdat->kswapd_wait))
3630 if (zone_balanced(zone, order, true, 0, 0))
3633 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3634 wake_up_interruptible(&pgdat->kswapd_wait);
3637 #ifdef CONFIG_HIBERNATION
3639 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3642 * Rather than trying to age LRUs the aim is to preserve the overall
3643 * LRU order by reclaiming preferentially
3644 * inactive > active > active referenced > active mapped
3646 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3648 struct reclaim_state reclaim_state;
3649 struct scan_control sc = {
3650 .nr_to_reclaim = nr_to_reclaim,
3651 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3652 .priority = DEF_PRIORITY,
3656 .hibernation_mode = 1,
3658 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3659 struct task_struct *p = current;
3660 unsigned long nr_reclaimed;
3662 p->flags |= PF_MEMALLOC;
3663 lockdep_set_current_reclaim_state(sc.gfp_mask);
3664 reclaim_state.reclaimed_slab = 0;
3665 p->reclaim_state = &reclaim_state;
3667 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3669 p->reclaim_state = NULL;
3670 lockdep_clear_current_reclaim_state();
3671 p->flags &= ~PF_MEMALLOC;
3673 return nr_reclaimed;
3675 #endif /* CONFIG_HIBERNATION */
3677 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3678 not required for correctness. So if the last cpu in a node goes
3679 away, we get changed to run anywhere: as the first one comes back,
3680 restore their cpu bindings. */
3681 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3686 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3687 for_each_node_state(nid, N_MEMORY) {
3688 pg_data_t *pgdat = NODE_DATA(nid);
3689 const struct cpumask *mask;
3691 mask = cpumask_of_node(pgdat->node_id);
3693 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3694 /* One of our CPUs online: restore mask */
3695 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3701 static int set_kswapd_cpu_mask(pg_data_t *pgdat)
3706 if (!kswapd_cpu_mask)
3709 cpumask_clear(&tmask);
3710 ret = cpumask_parse(kswapd_cpu_mask, &tmask);
3714 return set_cpus_allowed_ptr(pgdat->kswapd, &tmask);
3718 * This kswapd start function will be called by init and node-hot-add.
3719 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3721 int kswapd_run(int nid)
3723 pg_data_t *pgdat = NODE_DATA(nid);
3729 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3730 if (IS_ERR(pgdat->kswapd)) {
3731 /* failure at boot is fatal */
3732 BUG_ON(system_state == SYSTEM_BOOTING);
3733 pr_err("Failed to start kswapd on node %d\n", nid);
3734 ret = PTR_ERR(pgdat->kswapd);
3735 pgdat->kswapd = NULL;
3736 } else if (kswapd_cpu_mask) {
3737 if (set_kswapd_cpu_mask(pgdat))
3738 pr_warn("error setting kswapd cpu affinity mask\n");
3744 * Called by memory hotplug when all memory in a node is offlined. Caller must
3745 * hold mem_hotplug_begin/end().
3747 void kswapd_stop(int nid)
3749 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3752 kthread_stop(kswapd);
3753 NODE_DATA(nid)->kswapd = NULL;
3757 static int __init kswapd_init(void)
3762 for_each_node_state(nid, N_MEMORY)
3764 if (kswapd_cpu_mask == NULL)
3765 hotcpu_notifier(cpu_callback, 0);
3769 module_init(kswapd_init)
3775 * If non-zero call zone_reclaim when the number of free pages falls below
3778 int zone_reclaim_mode __read_mostly;
3780 #define RECLAIM_OFF 0
3781 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3782 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3783 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3786 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3787 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3790 #define ZONE_RECLAIM_PRIORITY 4
3793 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3796 int sysctl_min_unmapped_ratio = 1;
3799 * If the number of slab pages in a zone grows beyond this percentage then
3800 * slab reclaim needs to occur.
3802 int sysctl_min_slab_ratio = 5;
3804 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3806 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3807 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3808 zone_page_state(zone, NR_ACTIVE_FILE);
3811 * It's possible for there to be more file mapped pages than
3812 * accounted for by the pages on the file LRU lists because
3813 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3815 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3818 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3819 static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3821 unsigned long nr_pagecache_reclaimable;
3822 unsigned long delta = 0;
3825 * If RECLAIM_UNMAP is set, then all file pages are considered
3826 * potentially reclaimable. Otherwise, we have to worry about
3827 * pages like swapcache and zone_unmapped_file_pages() provides
3830 if (zone_reclaim_mode & RECLAIM_UNMAP)
3831 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3833 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3835 /* If we can't clean pages, remove dirty pages from consideration */
3836 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3837 delta += zone_page_state(zone, NR_FILE_DIRTY);
3839 /* Watch for any possible underflows due to delta */
3840 if (unlikely(delta > nr_pagecache_reclaimable))
3841 delta = nr_pagecache_reclaimable;
3843 return nr_pagecache_reclaimable - delta;
3847 * Try to free up some pages from this zone through reclaim.
3849 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3851 /* Minimum pages needed in order to stay on node */
3852 const unsigned long nr_pages = 1 << order;
3853 struct task_struct *p = current;
3854 struct reclaim_state reclaim_state;
3855 struct scan_control sc = {
3856 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3857 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3859 .priority = ZONE_RECLAIM_PRIORITY,
3860 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3861 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3867 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3868 * and we also need to be able to write out pages for RECLAIM_WRITE
3869 * and RECLAIM_UNMAP.
3871 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3872 lockdep_set_current_reclaim_state(gfp_mask);
3873 reclaim_state.reclaimed_slab = 0;
3874 p->reclaim_state = &reclaim_state;
3876 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3878 * Free memory by calling shrink zone with increasing
3879 * priorities until we have enough memory freed.
3882 shrink_zone(zone, &sc, true);
3883 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3886 p->reclaim_state = NULL;
3887 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3888 lockdep_clear_current_reclaim_state();
3889 return sc.nr_reclaimed >= nr_pages;
3892 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3898 * Zone reclaim reclaims unmapped file backed pages and
3899 * slab pages if we are over the defined limits.
3901 * A small portion of unmapped file backed pages is needed for
3902 * file I/O otherwise pages read by file I/O will be immediately
3903 * thrown out if the zone is overallocated. So we do not reclaim
3904 * if less than a specified percentage of the zone is used by
3905 * unmapped file backed pages.
3907 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3908 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3909 return ZONE_RECLAIM_FULL;
3911 if (!zone_reclaimable(zone))
3912 return ZONE_RECLAIM_FULL;
3915 * Do not scan if the allocation should not be delayed.
3917 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3918 return ZONE_RECLAIM_NOSCAN;
3921 * Only run zone reclaim on the local zone or on zones that do not
3922 * have associated processors. This will favor the local processor
3923 * over remote processors and spread off node memory allocations
3924 * as wide as possible.
3926 node_id = zone_to_nid(zone);
3927 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3928 return ZONE_RECLAIM_NOSCAN;
3930 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3931 return ZONE_RECLAIM_NOSCAN;
3933 ret = __zone_reclaim(zone, gfp_mask, order);
3934 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3937 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3944 * page_evictable - test whether a page is evictable
3945 * @page: the page to test
3947 * Test whether page is evictable--i.e., should be placed on active/inactive
3948 * lists vs unevictable list.
3950 * Reasons page might not be evictable:
3951 * (1) page's mapping marked unevictable
3952 * (2) page is part of an mlocked VMA
3955 int page_evictable(struct page *page)
3957 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3962 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3963 * @pages: array of pages to check
3964 * @nr_pages: number of pages to check
3966 * Checks pages for evictability and moves them to the appropriate lru list.
3968 * This function is only used for SysV IPC SHM_UNLOCK.
3970 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3972 struct lruvec *lruvec;
3973 struct zone *zone = NULL;
3978 for (i = 0; i < nr_pages; i++) {
3979 struct page *page = pages[i];
3980 struct zone *pagezone;
3983 pagezone = page_zone(page);
3984 if (pagezone != zone) {
3986 spin_unlock_irq(&zone->lru_lock);
3988 spin_lock_irq(&zone->lru_lock);
3990 lruvec = mem_cgroup_page_lruvec(page, zone);
3992 if (!PageLRU(page) || !PageUnevictable(page))
3995 if (page_evictable(page)) {
3996 enum lru_list lru = page_lru_base_type(page);
3998 VM_BUG_ON_PAGE(PageActive(page), page);
3999 ClearPageUnevictable(page);
4000 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4001 add_page_to_lru_list(page, lruvec, lru);
4007 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4008 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4009 spin_unlock_irq(&zone->lru_lock);
4012 #endif /* CONFIG_SHMEM */