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.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29 #include <linux/mm_inline.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
51 #include <linux/balloon_compaction.h>
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/vmscan.h>
59 /* Incremented by the number of inactive pages that were scanned */
60 unsigned long nr_scanned;
62 /* Number of pages freed so far during a call to shrink_zones() */
63 unsigned long nr_reclaimed;
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
68 unsigned long hibernation_mode;
70 /* This context's GFP mask */
75 /* Can mapped pages be reclaimed? */
78 /* Can pages be swapped as part of reclaim? */
83 /* Scan (total_size >> priority) pages at once */
86 /* anon vs. file LRUs scanning "ratio" */
90 * The memory cgroup that hit its limit and as a result is the
91 * primary target of this reclaim invocation.
93 struct mem_cgroup *target_mem_cgroup;
96 * Nodemask of nodes allowed by the caller. If NULL, all nodes
102 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104 #ifdef ARCH_HAS_PREFETCH
105 #define prefetch_prev_lru_page(_page, _base, _field) \
107 if ((_page)->lru.prev != _base) { \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetch(&prev->_field); \
115 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
118 #ifdef ARCH_HAS_PREFETCHW
119 #define prefetchw_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetchw(&prev->_field); \
129 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
133 * From 0 .. 100. Higher means more swappy.
135 int vm_swappiness = 60;
136 unsigned long vm_total_pages; /* The total number of pages which the VM controls */
138 static LIST_HEAD(shrinker_list);
139 static DECLARE_RWSEM(shrinker_rwsem);
142 static bool global_reclaim(struct scan_control *sc)
144 return !sc->target_mem_cgroup;
147 static bool global_reclaim(struct scan_control *sc)
153 static unsigned long zone_reclaimable_pages(struct zone *zone)
157 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
158 zone_page_state(zone, NR_INACTIVE_FILE);
160 if (get_nr_swap_pages() > 0)
161 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
162 zone_page_state(zone, NR_INACTIVE_ANON);
167 bool zone_reclaimable(struct zone *zone)
169 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
172 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
174 if (!mem_cgroup_disabled())
175 return mem_cgroup_get_lru_size(lruvec, lru);
177 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
181 * Add a shrinker callback to be called from the vm.
183 int register_shrinker(struct shrinker *shrinker)
185 size_t size = sizeof(*shrinker->nr_deferred);
188 * If we only have one possible node in the system anyway, save
189 * ourselves the trouble and disable NUMA aware behavior. This way we
190 * will save memory and some small loop time later.
192 if (nr_node_ids == 1)
193 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
195 if (shrinker->flags & SHRINKER_NUMA_AWARE)
198 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
199 if (!shrinker->nr_deferred)
202 down_write(&shrinker_rwsem);
203 list_add_tail(&shrinker->list, &shrinker_list);
204 up_write(&shrinker_rwsem);
207 EXPORT_SYMBOL(register_shrinker);
212 void unregister_shrinker(struct shrinker *shrinker)
214 down_write(&shrinker_rwsem);
215 list_del(&shrinker->list);
216 up_write(&shrinker_rwsem);
217 kfree(shrinker->nr_deferred);
219 EXPORT_SYMBOL(unregister_shrinker);
221 #define SHRINK_BATCH 128
224 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
225 unsigned long nr_pages_scanned, unsigned long lru_pages)
227 unsigned long freed = 0;
228 unsigned long long delta;
233 int nid = shrinkctl->nid;
234 long batch_size = shrinker->batch ? shrinker->batch
237 freeable = shrinker->count_objects(shrinker, shrinkctl);
242 * copy the current shrinker scan count into a local variable
243 * and zero it so that other concurrent shrinker invocations
244 * don't also do this scanning work.
246 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
249 delta = (4 * nr_pages_scanned) / shrinker->seeks;
251 do_div(delta, lru_pages + 1);
253 if (total_scan < 0) {
255 "shrink_slab: %pF negative objects to delete nr=%ld\n",
256 shrinker->scan_objects, total_scan);
257 total_scan = freeable;
261 * We need to avoid excessive windup on filesystem shrinkers
262 * due to large numbers of GFP_NOFS allocations causing the
263 * shrinkers to return -1 all the time. This results in a large
264 * nr being built up so when a shrink that can do some work
265 * comes along it empties the entire cache due to nr >>>
266 * freeable. This is bad for sustaining a working set in
269 * Hence only allow the shrinker to scan the entire cache when
270 * a large delta change is calculated directly.
272 if (delta < freeable / 4)
273 total_scan = min(total_scan, freeable / 2);
276 * Avoid risking looping forever due to too large nr value:
277 * never try to free more than twice the estimate number of
280 if (total_scan > freeable * 2)
281 total_scan = freeable * 2;
283 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
284 nr_pages_scanned, lru_pages,
285 freeable, delta, total_scan);
288 * Normally, we should not scan less than batch_size objects in one
289 * pass to avoid too frequent shrinker calls, but if the slab has less
290 * than batch_size objects in total and we are really tight on memory,
291 * we will try to reclaim all available objects, otherwise we can end
292 * up failing allocations although there are plenty of reclaimable
293 * objects spread over several slabs with usage less than the
296 * We detect the "tight on memory" situations by looking at the total
297 * number of objects we want to scan (total_scan). If it is greater
298 * than the total number of objects on slab (freeable), we must be
299 * scanning at high prio and therefore should try to reclaim as much as
302 while (total_scan >= batch_size ||
303 total_scan >= freeable) {
305 unsigned long nr_to_scan = min(batch_size, total_scan);
307 shrinkctl->nr_to_scan = nr_to_scan;
308 ret = shrinker->scan_objects(shrinker, shrinkctl);
309 if (ret == SHRINK_STOP)
313 count_vm_events(SLABS_SCANNED, nr_to_scan);
314 total_scan -= nr_to_scan;
320 * move the unused scan count back into the shrinker in a
321 * manner that handles concurrent updates. If we exhausted the
322 * scan, there is no need to do an update.
325 new_nr = atomic_long_add_return(total_scan,
326 &shrinker->nr_deferred[nid]);
328 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
330 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
335 * Call the shrink functions to age shrinkable caches
337 * Here we assume it costs one seek to replace a lru page and that it also
338 * takes a seek to recreate a cache object. With this in mind we age equal
339 * percentages of the lru and ageable caches. This should balance the seeks
340 * generated by these structures.
342 * If the vm encountered mapped pages on the LRU it increase the pressure on
343 * slab to avoid swapping.
345 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
347 * `lru_pages' represents the number of on-LRU pages in all the zones which
348 * are eligible for the caller's allocation attempt. It is used for balancing
349 * slab reclaim versus page reclaim.
351 * Returns the number of slab objects which we shrunk.
353 unsigned long shrink_slab(struct shrink_control *shrinkctl,
354 unsigned long nr_pages_scanned,
355 unsigned long lru_pages)
357 struct shrinker *shrinker;
358 unsigned long freed = 0;
360 if (nr_pages_scanned == 0)
361 nr_pages_scanned = SWAP_CLUSTER_MAX;
363 if (!down_read_trylock(&shrinker_rwsem)) {
365 * If we would return 0, our callers would understand that we
366 * have nothing else to shrink and give up trying. By returning
367 * 1 we keep it going and assume we'll be able to shrink next
374 list_for_each_entry(shrinker, &shrinker_list, list) {
375 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
377 freed += shrink_slab_node(shrinkctl, shrinker,
378 nr_pages_scanned, lru_pages);
382 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
383 if (node_online(shrinkctl->nid))
384 freed += shrink_slab_node(shrinkctl, shrinker,
385 nr_pages_scanned, lru_pages);
389 up_read(&shrinker_rwsem);
395 static inline int is_page_cache_freeable(struct page *page)
398 * A freeable page cache page is referenced only by the caller
399 * that isolated the page, the page cache radix tree and
400 * optional buffer heads at page->private.
402 return page_count(page) - page_has_private(page) == 2;
405 static int may_write_to_queue(struct backing_dev_info *bdi,
406 struct scan_control *sc)
408 if (current->flags & PF_SWAPWRITE)
410 if (!bdi_write_congested(bdi))
412 if (bdi == current->backing_dev_info)
418 * We detected a synchronous write error writing a page out. Probably
419 * -ENOSPC. We need to propagate that into the address_space for a subsequent
420 * fsync(), msync() or close().
422 * The tricky part is that after writepage we cannot touch the mapping: nothing
423 * prevents it from being freed up. But we have a ref on the page and once
424 * that page is locked, the mapping is pinned.
426 * We're allowed to run sleeping lock_page() here because we know the caller has
429 static void handle_write_error(struct address_space *mapping,
430 struct page *page, int error)
433 if (page_mapping(page) == mapping)
434 mapping_set_error(mapping, error);
438 /* possible outcome of pageout() */
440 /* failed to write page out, page is locked */
442 /* move page to the active list, page is locked */
444 /* page has been sent to the disk successfully, page is unlocked */
446 /* page is clean and locked */
451 * pageout is called by shrink_page_list() for each dirty page.
452 * Calls ->writepage().
454 static pageout_t pageout(struct page *page, struct address_space *mapping,
455 struct scan_control *sc)
458 * If the page is dirty, only perform writeback if that write
459 * will be non-blocking. To prevent this allocation from being
460 * stalled by pagecache activity. But note that there may be
461 * stalls if we need to run get_block(). We could test
462 * PagePrivate for that.
464 * If this process is currently in __generic_file_aio_write() against
465 * this page's queue, we can perform writeback even if that
468 * If the page is swapcache, write it back even if that would
469 * block, for some throttling. This happens by accident, because
470 * swap_backing_dev_info is bust: it doesn't reflect the
471 * congestion state of the swapdevs. Easy to fix, if needed.
473 if (!is_page_cache_freeable(page))
477 * Some data journaling orphaned pages can have
478 * page->mapping == NULL while being dirty with clean buffers.
480 if (page_has_private(page)) {
481 if (try_to_free_buffers(page)) {
482 ClearPageDirty(page);
483 printk("%s: orphaned page\n", __func__);
489 if (mapping->a_ops->writepage == NULL)
490 return PAGE_ACTIVATE;
491 if (!may_write_to_queue(mapping->backing_dev_info, sc))
494 if (clear_page_dirty_for_io(page)) {
496 struct writeback_control wbc = {
497 .sync_mode = WB_SYNC_NONE,
498 .nr_to_write = SWAP_CLUSTER_MAX,
500 .range_end = LLONG_MAX,
504 SetPageReclaim(page);
505 res = mapping->a_ops->writepage(page, &wbc);
507 handle_write_error(mapping, page, res);
508 if (res == AOP_WRITEPAGE_ACTIVATE) {
509 ClearPageReclaim(page);
510 return PAGE_ACTIVATE;
513 if (!PageWriteback(page)) {
514 /* synchronous write or broken a_ops? */
515 ClearPageReclaim(page);
517 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
518 inc_zone_page_state(page, NR_VMSCAN_WRITE);
526 * Same as remove_mapping, but if the page is removed from the mapping, it
527 * gets returned with a refcount of 0.
529 static int __remove_mapping(struct address_space *mapping, struct page *page,
532 BUG_ON(!PageLocked(page));
533 BUG_ON(mapping != page_mapping(page));
535 spin_lock_irq(&mapping->tree_lock);
537 * The non racy check for a busy page.
539 * Must be careful with the order of the tests. When someone has
540 * a ref to the page, it may be possible that they dirty it then
541 * drop the reference. So if PageDirty is tested before page_count
542 * here, then the following race may occur:
544 * get_user_pages(&page);
545 * [user mapping goes away]
547 * !PageDirty(page) [good]
548 * SetPageDirty(page);
550 * !page_count(page) [good, discard it]
552 * [oops, our write_to data is lost]
554 * Reversing the order of the tests ensures such a situation cannot
555 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
556 * load is not satisfied before that of page->_count.
558 * Note that if SetPageDirty is always performed via set_page_dirty,
559 * and thus under tree_lock, then this ordering is not required.
561 if (!page_freeze_refs(page, 2))
563 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
564 if (unlikely(PageDirty(page))) {
565 page_unfreeze_refs(page, 2);
569 if (PageSwapCache(page)) {
570 swp_entry_t swap = { .val = page_private(page) };
571 __delete_from_swap_cache(page);
572 spin_unlock_irq(&mapping->tree_lock);
573 swapcache_free(swap, page);
575 void (*freepage)(struct page *);
578 freepage = mapping->a_ops->freepage;
580 * Remember a shadow entry for reclaimed file cache in
581 * order to detect refaults, thus thrashing, later on.
583 * But don't store shadows in an address space that is
584 * already exiting. This is not just an optizimation,
585 * inode reclaim needs to empty out the radix tree or
586 * the nodes are lost. Don't plant shadows behind its
589 if (reclaimed && page_is_file_cache(page) &&
590 !mapping_exiting(mapping))
591 shadow = workingset_eviction(mapping, page);
592 __delete_from_page_cache(page, shadow);
593 spin_unlock_irq(&mapping->tree_lock);
594 mem_cgroup_uncharge_cache_page(page);
596 if (freepage != NULL)
603 spin_unlock_irq(&mapping->tree_lock);
608 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
609 * someone else has a ref on the page, abort and return 0. If it was
610 * successfully detached, return 1. Assumes the caller has a single ref on
613 int remove_mapping(struct address_space *mapping, struct page *page)
615 if (__remove_mapping(mapping, page, false)) {
617 * Unfreezing the refcount with 1 rather than 2 effectively
618 * drops the pagecache ref for us without requiring another
621 page_unfreeze_refs(page, 1);
628 * putback_lru_page - put previously isolated page onto appropriate LRU list
629 * @page: page to be put back to appropriate lru list
631 * Add previously isolated @page to appropriate LRU list.
632 * Page may still be unevictable for other reasons.
634 * lru_lock must not be held, interrupts must be enabled.
636 void putback_lru_page(struct page *page)
639 int was_unevictable = PageUnevictable(page);
641 VM_BUG_ON_PAGE(PageLRU(page), page);
644 ClearPageUnevictable(page);
646 if (page_evictable(page)) {
648 * For evictable pages, we can use the cache.
649 * In event of a race, worst case is we end up with an
650 * unevictable page on [in]active list.
651 * We know how to handle that.
653 is_unevictable = false;
657 * Put unevictable pages directly on zone's unevictable
660 is_unevictable = true;
661 add_page_to_unevictable_list(page);
663 * When racing with an mlock or AS_UNEVICTABLE clearing
664 * (page is unlocked) make sure that if the other thread
665 * does not observe our setting of PG_lru and fails
666 * isolation/check_move_unevictable_pages,
667 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
668 * the page back to the evictable list.
670 * The other side is TestClearPageMlocked() or shmem_lock().
676 * page's status can change while we move it among lru. If an evictable
677 * page is on unevictable list, it never be freed. To avoid that,
678 * check after we added it to the list, again.
680 if (is_unevictable && page_evictable(page)) {
681 if (!isolate_lru_page(page)) {
685 /* This means someone else dropped this page from LRU
686 * So, it will be freed or putback to LRU again. There is
687 * nothing to do here.
691 if (was_unevictable && !is_unevictable)
692 count_vm_event(UNEVICTABLE_PGRESCUED);
693 else if (!was_unevictable && is_unevictable)
694 count_vm_event(UNEVICTABLE_PGCULLED);
696 put_page(page); /* drop ref from isolate */
699 enum page_references {
701 PAGEREF_RECLAIM_CLEAN,
706 static enum page_references page_check_references(struct page *page,
707 struct scan_control *sc)
709 int referenced_ptes, referenced_page;
710 unsigned long vm_flags;
712 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
714 referenced_page = TestClearPageReferenced(page);
717 * Mlock lost the isolation race with us. Let try_to_unmap()
718 * move the page to the unevictable list.
720 if (vm_flags & VM_LOCKED)
721 return PAGEREF_RECLAIM;
723 if (referenced_ptes) {
724 if (PageSwapBacked(page))
725 return PAGEREF_ACTIVATE;
727 * All mapped pages start out with page table
728 * references from the instantiating fault, so we need
729 * to look twice if a mapped file page is used more
732 * Mark it and spare it for another trip around the
733 * inactive list. Another page table reference will
734 * lead to its activation.
736 * Note: the mark is set for activated pages as well
737 * so that recently deactivated but used pages are
740 SetPageReferenced(page);
742 if (referenced_page || referenced_ptes > 1)
743 return PAGEREF_ACTIVATE;
746 * Activate file-backed executable pages after first usage.
748 if (vm_flags & VM_EXEC)
749 return PAGEREF_ACTIVATE;
754 /* Reclaim if clean, defer dirty pages to writeback */
755 if (referenced_page && !PageSwapBacked(page))
756 return PAGEREF_RECLAIM_CLEAN;
758 return PAGEREF_RECLAIM;
761 /* Check if a page is dirty or under writeback */
762 static void page_check_dirty_writeback(struct page *page,
763 bool *dirty, bool *writeback)
765 struct address_space *mapping;
768 * Anonymous pages are not handled by flushers and must be written
769 * from reclaim context. Do not stall reclaim based on them
771 if (!page_is_file_cache(page)) {
777 /* By default assume that the page flags are accurate */
778 *dirty = PageDirty(page);
779 *writeback = PageWriteback(page);
781 /* Verify dirty/writeback state if the filesystem supports it */
782 if (!page_has_private(page))
785 mapping = page_mapping(page);
786 if (mapping && mapping->a_ops->is_dirty_writeback)
787 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
791 * shrink_page_list() returns the number of reclaimed pages
793 static unsigned long shrink_page_list(struct list_head *page_list,
795 struct scan_control *sc,
796 enum ttu_flags ttu_flags,
797 unsigned long *ret_nr_dirty,
798 unsigned long *ret_nr_unqueued_dirty,
799 unsigned long *ret_nr_congested,
800 unsigned long *ret_nr_writeback,
801 unsigned long *ret_nr_immediate,
804 LIST_HEAD(ret_pages);
805 LIST_HEAD(free_pages);
807 unsigned long nr_unqueued_dirty = 0;
808 unsigned long nr_dirty = 0;
809 unsigned long nr_congested = 0;
810 unsigned long nr_reclaimed = 0;
811 unsigned long nr_writeback = 0;
812 unsigned long nr_immediate = 0;
816 mem_cgroup_uncharge_start();
817 while (!list_empty(page_list)) {
818 struct address_space *mapping;
821 enum page_references references = PAGEREF_RECLAIM_CLEAN;
822 bool dirty, writeback;
826 page = lru_to_page(page_list);
827 list_del(&page->lru);
829 if (!trylock_page(page))
832 VM_BUG_ON_PAGE(PageActive(page), page);
833 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
837 if (unlikely(!page_evictable(page)))
840 if (!sc->may_unmap && page_mapped(page))
843 /* Double the slab pressure for mapped and swapcache pages */
844 if (page_mapped(page) || PageSwapCache(page))
847 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
848 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
851 * The number of dirty pages determines if a zone is marked
852 * reclaim_congested which affects wait_iff_congested. kswapd
853 * will stall and start writing pages if the tail of the LRU
854 * is all dirty unqueued pages.
856 page_check_dirty_writeback(page, &dirty, &writeback);
857 if (dirty || writeback)
860 if (dirty && !writeback)
864 * Treat this page as congested if the underlying BDI is or if
865 * pages are cycling through the LRU so quickly that the
866 * pages marked for immediate reclaim are making it to the
867 * end of the LRU a second time.
869 mapping = page_mapping(page);
870 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
871 (writeback && PageReclaim(page)))
875 * If a page at the tail of the LRU is under writeback, there
876 * are three cases to consider.
878 * 1) If reclaim is encountering an excessive number of pages
879 * under writeback and this page is both under writeback and
880 * PageReclaim then it indicates that pages are being queued
881 * for IO but are being recycled through the LRU before the
882 * IO can complete. Waiting on the page itself risks an
883 * indefinite stall if it is impossible to writeback the
884 * page due to IO error or disconnected storage so instead
885 * note that the LRU is being scanned too quickly and the
886 * caller can stall after page list has been processed.
888 * 2) Global reclaim encounters a page, memcg encounters a
889 * page that is not marked for immediate reclaim or
890 * the caller does not have __GFP_IO. In this case mark
891 * the page for immediate reclaim and continue scanning.
893 * __GFP_IO is checked because a loop driver thread might
894 * enter reclaim, and deadlock if it waits on a page for
895 * which it is needed to do the write (loop masks off
896 * __GFP_IO|__GFP_FS for this reason); but more thought
897 * would probably show more reasons.
899 * Don't require __GFP_FS, since we're not going into the
900 * FS, just waiting on its writeback completion. Worryingly,
901 * ext4 gfs2 and xfs allocate pages with
902 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
903 * may_enter_fs here is liable to OOM on them.
905 * 3) memcg encounters a page that is not already marked
906 * PageReclaim. memcg does not have any dirty pages
907 * throttling so we could easily OOM just because too many
908 * pages are in writeback and there is nothing else to
909 * reclaim. Wait for the writeback to complete.
911 if (PageWriteback(page)) {
913 if (current_is_kswapd() &&
915 zone_is_reclaim_writeback(zone)) {
920 } else if (global_reclaim(sc) ||
921 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
923 * This is slightly racy - end_page_writeback()
924 * might have just cleared PageReclaim, then
925 * setting PageReclaim here end up interpreted
926 * as PageReadahead - but that does not matter
927 * enough to care. What we do want is for this
928 * page to have PageReclaim set next time memcg
929 * reclaim reaches the tests above, so it will
930 * then wait_on_page_writeback() to avoid OOM;
931 * and it's also appropriate in global reclaim.
933 SetPageReclaim(page);
940 wait_on_page_writeback(page);
945 references = page_check_references(page, sc);
947 switch (references) {
948 case PAGEREF_ACTIVATE:
949 goto activate_locked;
952 case PAGEREF_RECLAIM:
953 case PAGEREF_RECLAIM_CLEAN:
954 ; /* try to reclaim the page below */
958 * Anonymous process memory has backing store?
959 * Try to allocate it some swap space here.
961 if (PageAnon(page) && !PageSwapCache(page)) {
962 if (!(sc->gfp_mask & __GFP_IO))
964 if (!add_to_swap(page, page_list))
965 goto activate_locked;
968 /* Adding to swap updated mapping */
969 mapping = page_mapping(page);
973 * The page is mapped into the page tables of one or more
974 * processes. Try to unmap it here.
976 if (page_mapped(page) && mapping) {
977 switch (try_to_unmap(page, ttu_flags)) {
979 goto activate_locked;
985 ; /* try to free the page below */
989 if (PageDirty(page)) {
991 * Only kswapd can writeback filesystem pages to
992 * avoid risk of stack overflow but only writeback
993 * if many dirty pages have been encountered.
995 if (page_is_file_cache(page) &&
996 (!current_is_kswapd() ||
997 !zone_is_reclaim_dirty(zone))) {
999 * Immediately reclaim when written back.
1000 * Similar in principal to deactivate_page()
1001 * except we already have the page isolated
1002 * and know it's dirty
1004 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1005 SetPageReclaim(page);
1010 if (references == PAGEREF_RECLAIM_CLEAN)
1014 if (!sc->may_writepage)
1017 /* Page is dirty, try to write it out here */
1018 switch (pageout(page, mapping, sc)) {
1022 goto activate_locked;
1024 if (PageWriteback(page))
1026 if (PageDirty(page))
1030 * A synchronous write - probably a ramdisk. Go
1031 * ahead and try to reclaim the page.
1033 if (!trylock_page(page))
1035 if (PageDirty(page) || PageWriteback(page))
1037 mapping = page_mapping(page);
1039 ; /* try to free the page below */
1044 * If the page has buffers, try to free the buffer mappings
1045 * associated with this page. If we succeed we try to free
1048 * We do this even if the page is PageDirty().
1049 * try_to_release_page() does not perform I/O, but it is
1050 * possible for a page to have PageDirty set, but it is actually
1051 * clean (all its buffers are clean). This happens if the
1052 * buffers were written out directly, with submit_bh(). ext3
1053 * will do this, as well as the blockdev mapping.
1054 * try_to_release_page() will discover that cleanness and will
1055 * drop the buffers and mark the page clean - it can be freed.
1057 * Rarely, pages can have buffers and no ->mapping. These are
1058 * the pages which were not successfully invalidated in
1059 * truncate_complete_page(). We try to drop those buffers here
1060 * and if that worked, and the page is no longer mapped into
1061 * process address space (page_count == 1) it can be freed.
1062 * Otherwise, leave the page on the LRU so it is swappable.
1064 if (page_has_private(page)) {
1065 if (!try_to_release_page(page, sc->gfp_mask))
1066 goto activate_locked;
1067 if (!mapping && page_count(page) == 1) {
1069 if (put_page_testzero(page))
1073 * rare race with speculative reference.
1074 * the speculative reference will free
1075 * this page shortly, so we may
1076 * increment nr_reclaimed here (and
1077 * leave it off the LRU).
1085 if (!mapping || !__remove_mapping(mapping, page, true))
1089 * At this point, we have no other references and there is
1090 * no way to pick any more up (removed from LRU, removed
1091 * from pagecache). Can use non-atomic bitops now (and
1092 * we obviously don't have to worry about waking up a process
1093 * waiting on the page lock, because there are no references.
1095 __clear_page_locked(page);
1100 * Is there need to periodically free_page_list? It would
1101 * appear not as the counts should be low
1103 list_add(&page->lru, &free_pages);
1107 if (PageSwapCache(page))
1108 try_to_free_swap(page);
1110 putback_lru_page(page);
1114 /* Not a candidate for swapping, so reclaim swap space. */
1115 if (PageSwapCache(page) && vm_swap_full())
1116 try_to_free_swap(page);
1117 VM_BUG_ON_PAGE(PageActive(page), page);
1118 SetPageActive(page);
1123 list_add(&page->lru, &ret_pages);
1124 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1127 free_hot_cold_page_list(&free_pages, true);
1129 list_splice(&ret_pages, page_list);
1130 count_vm_events(PGACTIVATE, pgactivate);
1131 mem_cgroup_uncharge_end();
1132 *ret_nr_dirty += nr_dirty;
1133 *ret_nr_congested += nr_congested;
1134 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1135 *ret_nr_writeback += nr_writeback;
1136 *ret_nr_immediate += nr_immediate;
1137 return nr_reclaimed;
1140 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1141 struct list_head *page_list)
1143 struct scan_control sc = {
1144 .gfp_mask = GFP_KERNEL,
1145 .priority = DEF_PRIORITY,
1148 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1149 struct page *page, *next;
1150 LIST_HEAD(clean_pages);
1152 list_for_each_entry_safe(page, next, page_list, lru) {
1153 if (page_is_file_cache(page) && !PageDirty(page) &&
1154 !isolated_balloon_page(page)) {
1155 ClearPageActive(page);
1156 list_move(&page->lru, &clean_pages);
1160 ret = shrink_page_list(&clean_pages, zone, &sc,
1161 TTU_UNMAP|TTU_IGNORE_ACCESS,
1162 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1163 list_splice(&clean_pages, page_list);
1164 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1169 * Attempt to remove the specified page from its LRU. Only take this page
1170 * if it is of the appropriate PageActive status. Pages which are being
1171 * freed elsewhere are also ignored.
1173 * page: page to consider
1174 * mode: one of the LRU isolation modes defined above
1176 * returns 0 on success, -ve errno on failure.
1178 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1182 /* Only take pages on the LRU. */
1186 /* Compaction should not handle unevictable pages but CMA can do so */
1187 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1193 * To minimise LRU disruption, the caller can indicate that it only
1194 * wants to isolate pages it will be able to operate on without
1195 * blocking - clean pages for the most part.
1197 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1198 * is used by reclaim when it is cannot write to backing storage
1200 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1201 * that it is possible to migrate without blocking
1203 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1204 /* All the caller can do on PageWriteback is block */
1205 if (PageWriteback(page))
1208 if (PageDirty(page)) {
1209 struct address_space *mapping;
1211 /* ISOLATE_CLEAN means only clean pages */
1212 if (mode & ISOLATE_CLEAN)
1216 * Only pages without mappings or that have a
1217 * ->migratepage callback are possible to migrate
1220 mapping = page_mapping(page);
1221 if (mapping && !mapping->a_ops->migratepage)
1226 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1229 if (likely(get_page_unless_zero(page))) {
1231 * Be careful not to clear PageLRU until after we're
1232 * sure the page is not being freed elsewhere -- the
1233 * page release code relies on it.
1243 * zone->lru_lock is heavily contended. Some of the functions that
1244 * shrink the lists perform better by taking out a batch of pages
1245 * and working on them outside the LRU lock.
1247 * For pagecache intensive workloads, this function is the hottest
1248 * spot in the kernel (apart from copy_*_user functions).
1250 * Appropriate locks must be held before calling this function.
1252 * @nr_to_scan: The number of pages to look through on the list.
1253 * @lruvec: The LRU vector to pull pages from.
1254 * @dst: The temp list to put pages on to.
1255 * @nr_scanned: The number of pages that were scanned.
1256 * @sc: The scan_control struct for this reclaim session
1257 * @mode: One of the LRU isolation modes
1258 * @lru: LRU list id for isolating
1260 * returns how many pages were moved onto *@dst.
1262 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1263 struct lruvec *lruvec, struct list_head *dst,
1264 unsigned long *nr_scanned, struct scan_control *sc,
1265 isolate_mode_t mode, enum lru_list lru)
1267 struct list_head *src = &lruvec->lists[lru];
1268 unsigned long nr_taken = 0;
1271 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1275 page = lru_to_page(src);
1276 prefetchw_prev_lru_page(page, src, flags);
1278 VM_BUG_ON_PAGE(!PageLRU(page), page);
1280 switch (__isolate_lru_page(page, mode)) {
1282 nr_pages = hpage_nr_pages(page);
1283 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1284 list_move(&page->lru, dst);
1285 nr_taken += nr_pages;
1289 /* else it is being freed elsewhere */
1290 list_move(&page->lru, src);
1299 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1300 nr_taken, mode, is_file_lru(lru));
1305 * isolate_lru_page - tries to isolate a page from its LRU list
1306 * @page: page to isolate from its LRU list
1308 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1309 * vmstat statistic corresponding to whatever LRU list the page was on.
1311 * Returns 0 if the page was removed from an LRU list.
1312 * Returns -EBUSY if the page was not on an LRU list.
1314 * The returned page will have PageLRU() cleared. If it was found on
1315 * the active list, it will have PageActive set. If it was found on
1316 * the unevictable list, it will have the PageUnevictable bit set. That flag
1317 * may need to be cleared by the caller before letting the page go.
1319 * The vmstat statistic corresponding to the list on which the page was
1320 * found will be decremented.
1323 * (1) Must be called with an elevated refcount on the page. This is a
1324 * fundamentnal difference from isolate_lru_pages (which is called
1325 * without a stable reference).
1326 * (2) the lru_lock must not be held.
1327 * (3) interrupts must be enabled.
1329 int isolate_lru_page(struct page *page)
1333 VM_BUG_ON_PAGE(!page_count(page), page);
1335 if (PageLRU(page)) {
1336 struct zone *zone = page_zone(page);
1337 struct lruvec *lruvec;
1339 spin_lock_irq(&zone->lru_lock);
1340 lruvec = mem_cgroup_page_lruvec(page, zone);
1341 if (PageLRU(page)) {
1342 int lru = page_lru(page);
1345 del_page_from_lru_list(page, lruvec, lru);
1348 spin_unlock_irq(&zone->lru_lock);
1354 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1355 * then get resheduled. When there are massive number of tasks doing page
1356 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1357 * the LRU list will go small and be scanned faster than necessary, leading to
1358 * unnecessary swapping, thrashing and OOM.
1360 static int too_many_isolated(struct zone *zone, int file,
1361 struct scan_control *sc)
1363 unsigned long inactive, isolated;
1365 if (current_is_kswapd())
1368 if (!global_reclaim(sc))
1372 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1373 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1375 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1376 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1380 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1381 * won't get blocked by normal direct-reclaimers, forming a circular
1384 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1387 return isolated > inactive;
1390 static noinline_for_stack void
1391 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1393 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1394 struct zone *zone = lruvec_zone(lruvec);
1395 LIST_HEAD(pages_to_free);
1398 * Put back any unfreeable pages.
1400 while (!list_empty(page_list)) {
1401 struct page *page = lru_to_page(page_list);
1404 VM_BUG_ON_PAGE(PageLRU(page), page);
1405 list_del(&page->lru);
1406 if (unlikely(!page_evictable(page))) {
1407 spin_unlock_irq(&zone->lru_lock);
1408 putback_lru_page(page);
1409 spin_lock_irq(&zone->lru_lock);
1413 lruvec = mem_cgroup_page_lruvec(page, zone);
1416 lru = page_lru(page);
1417 add_page_to_lru_list(page, lruvec, lru);
1419 if (is_active_lru(lru)) {
1420 int file = is_file_lru(lru);
1421 int numpages = hpage_nr_pages(page);
1422 reclaim_stat->recent_rotated[file] += numpages;
1424 if (put_page_testzero(page)) {
1425 __ClearPageLRU(page);
1426 __ClearPageActive(page);
1427 del_page_from_lru_list(page, lruvec, lru);
1429 if (unlikely(PageCompound(page))) {
1430 spin_unlock_irq(&zone->lru_lock);
1431 (*get_compound_page_dtor(page))(page);
1432 spin_lock_irq(&zone->lru_lock);
1434 list_add(&page->lru, &pages_to_free);
1439 * To save our caller's stack, now use input list for pages to free.
1441 list_splice(&pages_to_free, page_list);
1445 * If a kernel thread (such as nfsd for loop-back mounts) services
1446 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1447 * In that case we should only throttle if the backing device it is
1448 * writing to is congested. In other cases it is safe to throttle.
1450 static int current_may_throttle(void)
1452 return !(current->flags & PF_LESS_THROTTLE) ||
1453 current->backing_dev_info == NULL ||
1454 bdi_write_congested(current->backing_dev_info);
1458 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1459 * of reclaimed pages
1461 static noinline_for_stack unsigned long
1462 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1463 struct scan_control *sc, enum lru_list lru)
1465 LIST_HEAD(page_list);
1466 unsigned long nr_scanned;
1467 unsigned long nr_reclaimed = 0;
1468 unsigned long nr_taken;
1469 unsigned long nr_dirty = 0;
1470 unsigned long nr_congested = 0;
1471 unsigned long nr_unqueued_dirty = 0;
1472 unsigned long nr_writeback = 0;
1473 unsigned long nr_immediate = 0;
1474 isolate_mode_t isolate_mode = 0;
1475 int file = is_file_lru(lru);
1476 struct zone *zone = lruvec_zone(lruvec);
1477 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1479 while (unlikely(too_many_isolated(zone, file, sc))) {
1480 congestion_wait(BLK_RW_ASYNC, HZ/10);
1482 /* We are about to die and free our memory. Return now. */
1483 if (fatal_signal_pending(current))
1484 return SWAP_CLUSTER_MAX;
1490 isolate_mode |= ISOLATE_UNMAPPED;
1491 if (!sc->may_writepage)
1492 isolate_mode |= ISOLATE_CLEAN;
1494 spin_lock_irq(&zone->lru_lock);
1496 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1497 &nr_scanned, sc, isolate_mode, lru);
1499 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1500 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1502 if (global_reclaim(sc)) {
1503 zone->pages_scanned += nr_scanned;
1504 if (current_is_kswapd())
1505 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1507 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1509 spin_unlock_irq(&zone->lru_lock);
1514 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1515 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1516 &nr_writeback, &nr_immediate,
1519 spin_lock_irq(&zone->lru_lock);
1521 reclaim_stat->recent_scanned[file] += nr_taken;
1523 if (global_reclaim(sc)) {
1524 if (current_is_kswapd())
1525 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1528 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1532 putback_inactive_pages(lruvec, &page_list);
1534 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1536 spin_unlock_irq(&zone->lru_lock);
1538 free_hot_cold_page_list(&page_list, true);
1541 * If reclaim is isolating dirty pages under writeback, it implies
1542 * that the long-lived page allocation rate is exceeding the page
1543 * laundering rate. Either the global limits are not being effective
1544 * at throttling processes due to the page distribution throughout
1545 * zones or there is heavy usage of a slow backing device. The
1546 * only option is to throttle from reclaim context which is not ideal
1547 * as there is no guarantee the dirtying process is throttled in the
1548 * same way balance_dirty_pages() manages.
1550 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1551 * of pages under pages flagged for immediate reclaim and stall if any
1552 * are encountered in the nr_immediate check below.
1554 if (nr_writeback && nr_writeback == nr_taken)
1555 zone_set_flag(zone, ZONE_WRITEBACK);
1558 * memcg will stall in page writeback so only consider forcibly
1559 * stalling for global reclaim
1561 if (global_reclaim(sc)) {
1563 * Tag a zone as congested if all the dirty pages scanned were
1564 * backed by a congested BDI and wait_iff_congested will stall.
1566 if (nr_dirty && nr_dirty == nr_congested)
1567 zone_set_flag(zone, ZONE_CONGESTED);
1570 * If dirty pages are scanned that are not queued for IO, it
1571 * implies that flushers are not keeping up. In this case, flag
1572 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1573 * pages from reclaim context. It will forcibly stall in the
1576 if (nr_unqueued_dirty == nr_taken)
1577 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1580 * In addition, if kswapd scans pages marked marked for
1581 * immediate reclaim and under writeback (nr_immediate), it
1582 * implies that pages are cycling through the LRU faster than
1583 * they are written so also forcibly stall.
1585 if ((nr_unqueued_dirty == nr_taken || nr_immediate) &&
1586 current_may_throttle())
1587 congestion_wait(BLK_RW_ASYNC, HZ/10);
1591 * Stall direct reclaim for IO completions if underlying BDIs or zone
1592 * is congested. Allow kswapd to continue until it starts encountering
1593 * unqueued dirty pages or cycling through the LRU too quickly.
1595 if (!sc->hibernation_mode && !current_is_kswapd() &&
1596 current_may_throttle())
1597 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1599 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1601 nr_scanned, nr_reclaimed,
1603 trace_shrink_flags(file));
1604 return nr_reclaimed;
1608 * This moves pages from the active list to the inactive list.
1610 * We move them the other way if the page is referenced by one or more
1611 * processes, from rmap.
1613 * If the pages are mostly unmapped, the processing is fast and it is
1614 * appropriate to hold zone->lru_lock across the whole operation. But if
1615 * the pages are mapped, the processing is slow (page_referenced()) so we
1616 * should drop zone->lru_lock around each page. It's impossible to balance
1617 * this, so instead we remove the pages from the LRU while processing them.
1618 * It is safe to rely on PG_active against the non-LRU pages in here because
1619 * nobody will play with that bit on a non-LRU page.
1621 * The downside is that we have to touch page->_count against each page.
1622 * But we had to alter page->flags anyway.
1625 static void move_active_pages_to_lru(struct lruvec *lruvec,
1626 struct list_head *list,
1627 struct list_head *pages_to_free,
1630 struct zone *zone = lruvec_zone(lruvec);
1631 unsigned long pgmoved = 0;
1635 while (!list_empty(list)) {
1636 page = lru_to_page(list);
1637 lruvec = mem_cgroup_page_lruvec(page, zone);
1639 VM_BUG_ON_PAGE(PageLRU(page), page);
1642 nr_pages = hpage_nr_pages(page);
1643 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1644 list_move(&page->lru, &lruvec->lists[lru]);
1645 pgmoved += nr_pages;
1647 if (put_page_testzero(page)) {
1648 __ClearPageLRU(page);
1649 __ClearPageActive(page);
1650 del_page_from_lru_list(page, lruvec, lru);
1652 if (unlikely(PageCompound(page))) {
1653 spin_unlock_irq(&zone->lru_lock);
1654 (*get_compound_page_dtor(page))(page);
1655 spin_lock_irq(&zone->lru_lock);
1657 list_add(&page->lru, pages_to_free);
1660 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1661 if (!is_active_lru(lru))
1662 __count_vm_events(PGDEACTIVATE, pgmoved);
1665 static void shrink_active_list(unsigned long nr_to_scan,
1666 struct lruvec *lruvec,
1667 struct scan_control *sc,
1670 unsigned long nr_taken;
1671 unsigned long nr_scanned;
1672 unsigned long vm_flags;
1673 LIST_HEAD(l_hold); /* The pages which were snipped off */
1674 LIST_HEAD(l_active);
1675 LIST_HEAD(l_inactive);
1677 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1678 unsigned long nr_rotated = 0;
1679 isolate_mode_t isolate_mode = 0;
1680 int file = is_file_lru(lru);
1681 struct zone *zone = lruvec_zone(lruvec);
1686 isolate_mode |= ISOLATE_UNMAPPED;
1687 if (!sc->may_writepage)
1688 isolate_mode |= ISOLATE_CLEAN;
1690 spin_lock_irq(&zone->lru_lock);
1692 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1693 &nr_scanned, sc, isolate_mode, lru);
1694 if (global_reclaim(sc))
1695 zone->pages_scanned += nr_scanned;
1697 reclaim_stat->recent_scanned[file] += nr_taken;
1699 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1700 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1701 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1702 spin_unlock_irq(&zone->lru_lock);
1704 while (!list_empty(&l_hold)) {
1706 page = lru_to_page(&l_hold);
1707 list_del(&page->lru);
1709 if (unlikely(!page_evictable(page))) {
1710 putback_lru_page(page);
1714 if (unlikely(buffer_heads_over_limit)) {
1715 if (page_has_private(page) && trylock_page(page)) {
1716 if (page_has_private(page))
1717 try_to_release_page(page, 0);
1722 if (page_referenced(page, 0, sc->target_mem_cgroup,
1724 nr_rotated += hpage_nr_pages(page);
1726 * Identify referenced, file-backed active pages and
1727 * give them one more trip around the active list. So
1728 * that executable code get better chances to stay in
1729 * memory under moderate memory pressure. Anon pages
1730 * are not likely to be evicted by use-once streaming
1731 * IO, plus JVM can create lots of anon VM_EXEC pages,
1732 * so we ignore them here.
1734 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1735 list_add(&page->lru, &l_active);
1740 ClearPageActive(page); /* we are de-activating */
1741 list_add(&page->lru, &l_inactive);
1745 * Move pages back to the lru list.
1747 spin_lock_irq(&zone->lru_lock);
1749 * Count referenced pages from currently used mappings as rotated,
1750 * even though only some of them are actually re-activated. This
1751 * helps balance scan pressure between file and anonymous pages in
1754 reclaim_stat->recent_rotated[file] += nr_rotated;
1756 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1757 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1758 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1759 spin_unlock_irq(&zone->lru_lock);
1761 free_hot_cold_page_list(&l_hold, true);
1765 static int inactive_anon_is_low_global(struct zone *zone)
1767 unsigned long active, inactive;
1769 active = zone_page_state(zone, NR_ACTIVE_ANON);
1770 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1772 if (inactive * zone->inactive_ratio < active)
1779 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1780 * @lruvec: LRU vector to check
1782 * Returns true if the zone does not have enough inactive anon pages,
1783 * meaning some active anon pages need to be deactivated.
1785 static int inactive_anon_is_low(struct lruvec *lruvec)
1788 * If we don't have swap space, anonymous page deactivation
1791 if (!total_swap_pages)
1794 if (!mem_cgroup_disabled())
1795 return mem_cgroup_inactive_anon_is_low(lruvec);
1797 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1800 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1807 * inactive_file_is_low - check if file pages need to be deactivated
1808 * @lruvec: LRU vector to check
1810 * When the system is doing streaming IO, memory pressure here
1811 * ensures that active file pages get deactivated, until more
1812 * than half of the file pages are on the inactive list.
1814 * Once we get to that situation, protect the system's working
1815 * set from being evicted by disabling active file page aging.
1817 * This uses a different ratio than the anonymous pages, because
1818 * the page cache uses a use-once replacement algorithm.
1820 static int inactive_file_is_low(struct lruvec *lruvec)
1822 unsigned long inactive;
1823 unsigned long active;
1825 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1826 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1828 return active > inactive;
1831 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1833 if (is_file_lru(lru))
1834 return inactive_file_is_low(lruvec);
1836 return inactive_anon_is_low(lruvec);
1839 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1840 struct lruvec *lruvec, struct scan_control *sc)
1842 if (is_active_lru(lru)) {
1843 if (inactive_list_is_low(lruvec, lru))
1844 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1848 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1859 * Determine how aggressively the anon and file LRU lists should be
1860 * scanned. The relative value of each set of LRU lists is determined
1861 * by looking at the fraction of the pages scanned we did rotate back
1862 * onto the active list instead of evict.
1864 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1865 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1867 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1870 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1872 u64 denominator = 0; /* gcc */
1873 struct zone *zone = lruvec_zone(lruvec);
1874 unsigned long anon_prio, file_prio;
1875 enum scan_balance scan_balance;
1876 unsigned long anon, file;
1877 bool force_scan = false;
1878 unsigned long ap, fp;
1884 * If the zone or memcg is small, nr[l] can be 0. This
1885 * results in no scanning on this priority and a potential
1886 * priority drop. Global direct reclaim can go to the next
1887 * zone and tends to have no problems. Global kswapd is for
1888 * zone balancing and it needs to scan a minimum amount. When
1889 * reclaiming for a memcg, a priority drop can cause high
1890 * latencies, so it's better to scan a minimum amount there as
1893 if (current_is_kswapd() && !zone_reclaimable(zone))
1895 if (!global_reclaim(sc))
1898 /* If we have no swap space, do not bother scanning anon pages. */
1899 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1900 scan_balance = SCAN_FILE;
1905 * Global reclaim will swap to prevent OOM even with no
1906 * swappiness, but memcg users want to use this knob to
1907 * disable swapping for individual groups completely when
1908 * using the memory controller's swap limit feature would be
1911 if (!global_reclaim(sc) && !sc->swappiness) {
1912 scan_balance = SCAN_FILE;
1917 * Do not apply any pressure balancing cleverness when the
1918 * system is close to OOM, scan both anon and file equally
1919 * (unless the swappiness setting disagrees with swapping).
1921 if (!sc->priority && sc->swappiness) {
1922 scan_balance = SCAN_EQUAL;
1926 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1927 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1928 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1929 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1932 * Prevent the reclaimer from falling into the cache trap: as
1933 * cache pages start out inactive, every cache fault will tip
1934 * the scan balance towards the file LRU. And as the file LRU
1935 * shrinks, so does the window for rotation from references.
1936 * This means we have a runaway feedback loop where a tiny
1937 * thrashing file LRU becomes infinitely more attractive than
1938 * anon pages. Try to detect this based on file LRU size.
1940 if (global_reclaim(sc)) {
1941 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1943 if (unlikely(file + free <= high_wmark_pages(zone))) {
1944 scan_balance = SCAN_ANON;
1950 * There is enough inactive page cache, do not reclaim
1951 * anything from the anonymous working set right now.
1953 if (!inactive_file_is_low(lruvec)) {
1954 scan_balance = SCAN_FILE;
1958 scan_balance = SCAN_FRACT;
1961 * With swappiness at 100, anonymous and file have the same priority.
1962 * This scanning priority is essentially the inverse of IO cost.
1964 anon_prio = sc->swappiness;
1965 file_prio = 200 - anon_prio;
1968 * OK, so we have swap space and a fair amount of page cache
1969 * pages. We use the recently rotated / recently scanned
1970 * ratios to determine how valuable each cache is.
1972 * Because workloads change over time (and to avoid overflow)
1973 * we keep these statistics as a floating average, which ends
1974 * up weighing recent references more than old ones.
1976 * anon in [0], file in [1]
1978 spin_lock_irq(&zone->lru_lock);
1979 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1980 reclaim_stat->recent_scanned[0] /= 2;
1981 reclaim_stat->recent_rotated[0] /= 2;
1984 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1985 reclaim_stat->recent_scanned[1] /= 2;
1986 reclaim_stat->recent_rotated[1] /= 2;
1990 * The amount of pressure on anon vs file pages is inversely
1991 * proportional to the fraction of recently scanned pages on
1992 * each list that were recently referenced and in active use.
1994 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1995 ap /= reclaim_stat->recent_rotated[0] + 1;
1997 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1998 fp /= reclaim_stat->recent_rotated[1] + 1;
1999 spin_unlock_irq(&zone->lru_lock);
2003 denominator = ap + fp + 1;
2005 some_scanned = false;
2006 /* Only use force_scan on second pass. */
2007 for (pass = 0; !some_scanned && pass < 2; pass++) {
2008 for_each_evictable_lru(lru) {
2009 int file = is_file_lru(lru);
2013 size = get_lru_size(lruvec, lru);
2014 scan = size >> sc->priority;
2016 if (!scan && pass && force_scan)
2017 scan = min(size, SWAP_CLUSTER_MAX);
2019 switch (scan_balance) {
2021 /* Scan lists relative to size */
2025 * Scan types proportional to swappiness and
2026 * their relative recent reclaim efficiency.
2028 scan = div64_u64(scan * fraction[file],
2033 /* Scan one type exclusively */
2034 if ((scan_balance == SCAN_FILE) != file)
2038 /* Look ma, no brain */
2043 * Skip the second pass and don't force_scan,
2044 * if we found something to scan.
2046 some_scanned |= !!scan;
2052 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2054 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2056 unsigned long nr[NR_LRU_LISTS];
2057 unsigned long targets[NR_LRU_LISTS];
2058 unsigned long nr_to_scan;
2060 unsigned long nr_reclaimed = 0;
2061 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2062 struct blk_plug plug;
2065 get_scan_count(lruvec, sc, nr);
2067 /* Record the original scan target for proportional adjustments later */
2068 memcpy(targets, nr, sizeof(nr));
2071 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2072 * event that can occur when there is little memory pressure e.g.
2073 * multiple streaming readers/writers. Hence, we do not abort scanning
2074 * when the requested number of pages are reclaimed when scanning at
2075 * DEF_PRIORITY on the assumption that the fact we are direct
2076 * reclaiming implies that kswapd is not keeping up and it is best to
2077 * do a batch of work at once. For memcg reclaim one check is made to
2078 * abort proportional reclaim if either the file or anon lru has already
2079 * dropped to zero at the first pass.
2081 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2082 sc->priority == DEF_PRIORITY);
2084 blk_start_plug(&plug);
2085 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2086 nr[LRU_INACTIVE_FILE]) {
2087 unsigned long nr_anon, nr_file, percentage;
2088 unsigned long nr_scanned;
2090 for_each_evictable_lru(lru) {
2092 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2093 nr[lru] -= nr_to_scan;
2095 nr_reclaimed += shrink_list(lru, nr_to_scan,
2100 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2104 * For kswapd and memcg, reclaim at least the number of pages
2105 * requested. Ensure that the anon and file LRUs are scanned
2106 * proportionally what was requested by get_scan_count(). We
2107 * stop reclaiming one LRU and reduce the amount scanning
2108 * proportional to the original scan target.
2110 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2111 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2114 * It's just vindictive to attack the larger once the smaller
2115 * has gone to zero. And given the way we stop scanning the
2116 * smaller below, this makes sure that we only make one nudge
2117 * towards proportionality once we've got nr_to_reclaim.
2119 if (!nr_file || !nr_anon)
2122 if (nr_file > nr_anon) {
2123 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2124 targets[LRU_ACTIVE_ANON] + 1;
2126 percentage = nr_anon * 100 / scan_target;
2128 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2129 targets[LRU_ACTIVE_FILE] + 1;
2131 percentage = nr_file * 100 / scan_target;
2134 /* Stop scanning the smaller of the LRU */
2136 nr[lru + LRU_ACTIVE] = 0;
2139 * Recalculate the other LRU scan count based on its original
2140 * scan target and the percentage scanning already complete
2142 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2143 nr_scanned = targets[lru] - nr[lru];
2144 nr[lru] = targets[lru] * (100 - percentage) / 100;
2145 nr[lru] -= min(nr[lru], nr_scanned);
2148 nr_scanned = targets[lru] - nr[lru];
2149 nr[lru] = targets[lru] * (100 - percentage) / 100;
2150 nr[lru] -= min(nr[lru], nr_scanned);
2152 scan_adjusted = true;
2154 blk_finish_plug(&plug);
2155 sc->nr_reclaimed += nr_reclaimed;
2158 * Even if we did not try to evict anon pages at all, we want to
2159 * rebalance the anon lru active/inactive ratio.
2161 if (inactive_anon_is_low(lruvec))
2162 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2163 sc, LRU_ACTIVE_ANON);
2165 throttle_vm_writeout(sc->gfp_mask);
2168 /* Use reclaim/compaction for costly allocs or under memory pressure */
2169 static bool in_reclaim_compaction(struct scan_control *sc)
2171 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2172 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2173 sc->priority < DEF_PRIORITY - 2))
2180 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2181 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2182 * true if more pages should be reclaimed such that when the page allocator
2183 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2184 * It will give up earlier than that if there is difficulty reclaiming pages.
2186 static inline bool should_continue_reclaim(struct zone *zone,
2187 unsigned long nr_reclaimed,
2188 unsigned long nr_scanned,
2189 struct scan_control *sc)
2191 unsigned long pages_for_compaction;
2192 unsigned long inactive_lru_pages;
2194 /* If not in reclaim/compaction mode, stop */
2195 if (!in_reclaim_compaction(sc))
2198 /* Consider stopping depending on scan and reclaim activity */
2199 if (sc->gfp_mask & __GFP_REPEAT) {
2201 * For __GFP_REPEAT allocations, stop reclaiming if the
2202 * full LRU list has been scanned and we are still failing
2203 * to reclaim pages. This full LRU scan is potentially
2204 * expensive but a __GFP_REPEAT caller really wants to succeed
2206 if (!nr_reclaimed && !nr_scanned)
2210 * For non-__GFP_REPEAT allocations which can presumably
2211 * fail without consequence, stop if we failed to reclaim
2212 * any pages from the last SWAP_CLUSTER_MAX number of
2213 * pages that were scanned. This will return to the
2214 * caller faster at the risk reclaim/compaction and
2215 * the resulting allocation attempt fails
2222 * If we have not reclaimed enough pages for compaction and the
2223 * inactive lists are large enough, continue reclaiming
2225 pages_for_compaction = (2UL << sc->order);
2226 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2227 if (get_nr_swap_pages() > 0)
2228 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2229 if (sc->nr_reclaimed < pages_for_compaction &&
2230 inactive_lru_pages > pages_for_compaction)
2233 /* If compaction would go ahead or the allocation would succeed, stop */
2234 switch (compaction_suitable(zone, sc->order)) {
2235 case COMPACT_PARTIAL:
2236 case COMPACT_CONTINUE:
2243 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2245 unsigned long nr_reclaimed, nr_scanned;
2248 struct mem_cgroup *root = sc->target_mem_cgroup;
2249 struct mem_cgroup_reclaim_cookie reclaim = {
2251 .priority = sc->priority,
2253 struct mem_cgroup *memcg;
2255 nr_reclaimed = sc->nr_reclaimed;
2256 nr_scanned = sc->nr_scanned;
2258 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2260 struct lruvec *lruvec;
2262 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2264 sc->swappiness = mem_cgroup_swappiness(memcg);
2265 shrink_lruvec(lruvec, sc);
2268 * Direct reclaim and kswapd have to scan all memory
2269 * cgroups to fulfill the overall scan target for the
2272 * Limit reclaim, on the other hand, only cares about
2273 * nr_to_reclaim pages to be reclaimed and it will
2274 * retry with decreasing priority if one round over the
2275 * whole hierarchy is not sufficient.
2277 if (!global_reclaim(sc) &&
2278 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2279 mem_cgroup_iter_break(root, memcg);
2282 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2285 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2286 sc->nr_scanned - nr_scanned,
2287 sc->nr_reclaimed - nr_reclaimed);
2289 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2290 sc->nr_scanned - nr_scanned, sc));
2293 /* Returns true if compaction should go ahead for a high-order request */
2294 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2296 unsigned long balance_gap, watermark;
2299 /* Do not consider compaction for orders reclaim is meant to satisfy */
2300 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2304 * Compaction takes time to run and there are potentially other
2305 * callers using the pages just freed. Continue reclaiming until
2306 * there is a buffer of free pages available to give compaction
2307 * a reasonable chance of completing and allocating the page
2309 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2310 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2311 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2312 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2315 * If compaction is deferred, reclaim up to a point where
2316 * compaction will have a chance of success when re-enabled
2318 if (compaction_deferred(zone, sc->order))
2319 return watermark_ok;
2321 /* If compaction is not ready to start, keep reclaiming */
2322 if (!compaction_suitable(zone, sc->order))
2325 return watermark_ok;
2329 * This is the direct reclaim path, for page-allocating processes. We only
2330 * try to reclaim pages from zones which will satisfy the caller's allocation
2333 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2335 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2337 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2338 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2339 * zone defense algorithm.
2341 * If a zone is deemed to be full of pinned pages then just give it a light
2342 * scan then give up on it.
2344 * This function returns true if a zone is being reclaimed for a costly
2345 * high-order allocation and compaction is ready to begin. This indicates to
2346 * the caller that it should consider retrying the allocation instead of
2349 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2353 unsigned long nr_soft_reclaimed;
2354 unsigned long nr_soft_scanned;
2355 unsigned long lru_pages = 0;
2356 bool aborted_reclaim = false;
2357 struct reclaim_state *reclaim_state = current->reclaim_state;
2359 struct shrink_control shrink = {
2360 .gfp_mask = sc->gfp_mask,
2362 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2365 * If the number of buffer_heads in the machine exceeds the maximum
2366 * allowed level, force direct reclaim to scan the highmem zone as
2367 * highmem pages could be pinning lowmem pages storing buffer_heads
2369 orig_mask = sc->gfp_mask;
2370 if (buffer_heads_over_limit)
2371 sc->gfp_mask |= __GFP_HIGHMEM;
2373 nodes_clear(shrink.nodes_to_scan);
2375 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2376 gfp_zone(sc->gfp_mask), sc->nodemask) {
2377 if (!populated_zone(zone))
2380 * Take care memory controller reclaiming has small influence
2383 if (global_reclaim(sc)) {
2384 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2387 lru_pages += zone_reclaimable_pages(zone);
2388 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2390 if (sc->priority != DEF_PRIORITY &&
2391 !zone_reclaimable(zone))
2392 continue; /* Let kswapd poll it */
2393 if (IS_ENABLED(CONFIG_COMPACTION)) {
2395 * If we already have plenty of memory free for
2396 * compaction in this zone, don't free any more.
2397 * Even though compaction is invoked for any
2398 * non-zero order, only frequent costly order
2399 * reclamation is disruptive enough to become a
2400 * noticeable problem, like transparent huge
2403 if ((zonelist_zone_idx(z) <= requested_highidx)
2404 && compaction_ready(zone, sc)) {
2405 aborted_reclaim = true;
2410 * This steals pages from memory cgroups over softlimit
2411 * and returns the number of reclaimed pages and
2412 * scanned pages. This works for global memory pressure
2413 * and balancing, not for a memcg's limit.
2415 nr_soft_scanned = 0;
2416 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2417 sc->order, sc->gfp_mask,
2419 sc->nr_reclaimed += nr_soft_reclaimed;
2420 sc->nr_scanned += nr_soft_scanned;
2421 /* need some check for avoid more shrink_zone() */
2424 shrink_zone(zone, sc);
2428 * Don't shrink slabs when reclaiming memory from over limit cgroups
2429 * but do shrink slab at least once when aborting reclaim for
2430 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2433 if (global_reclaim(sc)) {
2434 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2435 if (reclaim_state) {
2436 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2437 reclaim_state->reclaimed_slab = 0;
2442 * Restore to original mask to avoid the impact on the caller if we
2443 * promoted it to __GFP_HIGHMEM.
2445 sc->gfp_mask = orig_mask;
2447 return aborted_reclaim;
2450 /* All zones in zonelist are unreclaimable? */
2451 static bool all_unreclaimable(struct zonelist *zonelist,
2452 struct scan_control *sc)
2457 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2458 gfp_zone(sc->gfp_mask), sc->nodemask) {
2459 if (!populated_zone(zone))
2461 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2463 if (zone_reclaimable(zone))
2471 * This is the main entry point to direct page reclaim.
2473 * If a full scan of the inactive list fails to free enough memory then we
2474 * are "out of memory" and something needs to be killed.
2476 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2477 * high - the zone may be full of dirty or under-writeback pages, which this
2478 * caller can't do much about. We kick the writeback threads and take explicit
2479 * naps in the hope that some of these pages can be written. But if the
2480 * allocating task holds filesystem locks which prevent writeout this might not
2481 * work, and the allocation attempt will fail.
2483 * returns: 0, if no pages reclaimed
2484 * else, the number of pages reclaimed
2486 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2487 struct scan_control *sc)
2489 unsigned long total_scanned = 0;
2490 unsigned long writeback_threshold;
2491 bool aborted_reclaim;
2493 delayacct_freepages_start();
2495 if (global_reclaim(sc))
2496 count_vm_event(ALLOCSTALL);
2499 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2502 aborted_reclaim = shrink_zones(zonelist, sc);
2504 total_scanned += sc->nr_scanned;
2505 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2509 * If we're getting trouble reclaiming, start doing
2510 * writepage even in laptop mode.
2512 if (sc->priority < DEF_PRIORITY - 2)
2513 sc->may_writepage = 1;
2516 * Try to write back as many pages as we just scanned. This
2517 * tends to cause slow streaming writers to write data to the
2518 * disk smoothly, at the dirtying rate, which is nice. But
2519 * that's undesirable in laptop mode, where we *want* lumpy
2520 * writeout. So in laptop mode, write out the whole world.
2522 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2523 if (total_scanned > writeback_threshold) {
2524 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2525 WB_REASON_TRY_TO_FREE_PAGES);
2526 sc->may_writepage = 1;
2528 } while (--sc->priority >= 0 && !aborted_reclaim);
2531 delayacct_freepages_end();
2533 if (sc->nr_reclaimed)
2534 return sc->nr_reclaimed;
2537 * As hibernation is going on, kswapd is freezed so that it can't mark
2538 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2541 if (oom_killer_disabled)
2544 /* Aborted reclaim to try compaction? don't OOM, then */
2545 if (aborted_reclaim)
2548 /* top priority shrink_zones still had more to do? don't OOM, then */
2549 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2555 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2558 unsigned long pfmemalloc_reserve = 0;
2559 unsigned long free_pages = 0;
2563 for (i = 0; i <= ZONE_NORMAL; i++) {
2564 zone = &pgdat->node_zones[i];
2565 if (!populated_zone(zone))
2568 pfmemalloc_reserve += min_wmark_pages(zone);
2569 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2572 /* If there are no reserves (unexpected config) then do not throttle */
2573 if (!pfmemalloc_reserve)
2576 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2578 /* kswapd must be awake if processes are being throttled */
2579 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2580 pgdat->classzone_idx = min(pgdat->classzone_idx,
2581 (enum zone_type)ZONE_NORMAL);
2582 wake_up_interruptible(&pgdat->kswapd_wait);
2589 * Throttle direct reclaimers if backing storage is backed by the network
2590 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2591 * depleted. kswapd will continue to make progress and wake the processes
2592 * when the low watermark is reached.
2594 * Returns true if a fatal signal was delivered during throttling. If this
2595 * happens, the page allocator should not consider triggering the OOM killer.
2597 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2598 nodemask_t *nodemask)
2602 pg_data_t *pgdat = NULL;
2605 * Kernel threads should not be throttled as they may be indirectly
2606 * responsible for cleaning pages necessary for reclaim to make forward
2607 * progress. kjournald for example may enter direct reclaim while
2608 * committing a transaction where throttling it could forcing other
2609 * processes to block on log_wait_commit().
2611 if (current->flags & PF_KTHREAD)
2615 * If a fatal signal is pending, this process should not throttle.
2616 * It should return quickly so it can exit and free its memory
2618 if (fatal_signal_pending(current))
2622 * Check if the pfmemalloc reserves are ok by finding the first node
2623 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2624 * GFP_KERNEL will be required for allocating network buffers when
2625 * swapping over the network so ZONE_HIGHMEM is unusable.
2627 * Throttling is based on the first usable node and throttled processes
2628 * wait on a queue until kswapd makes progress and wakes them. There
2629 * is an affinity then between processes waking up and where reclaim
2630 * progress has been made assuming the process wakes on the same node.
2631 * More importantly, processes running on remote nodes will not compete
2632 * for remote pfmemalloc reserves and processes on different nodes
2633 * should make reasonable progress.
2635 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2636 gfp_mask, nodemask) {
2637 if (zone_idx(zone) > ZONE_NORMAL)
2640 /* Throttle based on the first usable node */
2641 pgdat = zone->zone_pgdat;
2642 if (pfmemalloc_watermark_ok(pgdat))
2647 /* If no zone was usable by the allocation flags then do not throttle */
2651 /* Account for the throttling */
2652 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2655 * If the caller cannot enter the filesystem, it's possible that it
2656 * is due to the caller holding an FS lock or performing a journal
2657 * transaction in the case of a filesystem like ext[3|4]. In this case,
2658 * it is not safe to block on pfmemalloc_wait as kswapd could be
2659 * blocked waiting on the same lock. Instead, throttle for up to a
2660 * second before continuing.
2662 if (!(gfp_mask & __GFP_FS)) {
2663 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2664 pfmemalloc_watermark_ok(pgdat), HZ);
2669 /* Throttle until kswapd wakes the process */
2670 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2671 pfmemalloc_watermark_ok(pgdat));
2674 if (fatal_signal_pending(current))
2681 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2682 gfp_t gfp_mask, nodemask_t *nodemask)
2684 unsigned long nr_reclaimed;
2685 struct scan_control sc = {
2686 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2687 .may_writepage = !laptop_mode,
2688 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2692 .priority = DEF_PRIORITY,
2693 .target_mem_cgroup = NULL,
2694 .nodemask = nodemask,
2698 * Do not enter reclaim if fatal signal was delivered while throttled.
2699 * 1 is returned so that the page allocator does not OOM kill at this
2702 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2705 trace_mm_vmscan_direct_reclaim_begin(order,
2709 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2711 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2713 return nr_reclaimed;
2718 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2719 gfp_t gfp_mask, bool noswap,
2721 unsigned long *nr_scanned)
2723 struct scan_control sc = {
2725 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2726 .may_writepage = !laptop_mode,
2728 .may_swap = !noswap,
2731 .swappiness = mem_cgroup_swappiness(memcg),
2732 .target_mem_cgroup = memcg,
2734 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2736 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2737 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2739 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2744 * NOTE: Although we can get the priority field, using it
2745 * here is not a good idea, since it limits the pages we can scan.
2746 * if we don't reclaim here, the shrink_zone from balance_pgdat
2747 * will pick up pages from other mem cgroup's as well. We hack
2748 * the priority and make it zero.
2750 shrink_lruvec(lruvec, &sc);
2752 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2754 *nr_scanned = sc.nr_scanned;
2755 return sc.nr_reclaimed;
2758 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2762 struct zonelist *zonelist;
2763 unsigned long nr_reclaimed;
2765 struct scan_control sc = {
2766 .may_writepage = !laptop_mode,
2768 .may_swap = !noswap,
2769 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2771 .priority = DEF_PRIORITY,
2772 .target_mem_cgroup = memcg,
2773 .nodemask = NULL, /* we don't care the placement */
2774 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2775 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2779 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2780 * take care of from where we get pages. So the node where we start the
2781 * scan does not need to be the current node.
2783 nid = mem_cgroup_select_victim_node(memcg);
2785 zonelist = NODE_DATA(nid)->node_zonelists;
2787 trace_mm_vmscan_memcg_reclaim_begin(0,
2791 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2793 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2795 return nr_reclaimed;
2799 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2801 struct mem_cgroup *memcg;
2803 if (!total_swap_pages)
2806 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2808 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2810 if (inactive_anon_is_low(lruvec))
2811 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2812 sc, LRU_ACTIVE_ANON);
2814 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2818 static bool zone_balanced(struct zone *zone, int order,
2819 unsigned long balance_gap, int classzone_idx)
2821 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2822 balance_gap, classzone_idx, 0))
2825 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2826 !compaction_suitable(zone, order))
2833 * pgdat_balanced() is used when checking if a node is balanced.
2835 * For order-0, all zones must be balanced!
2837 * For high-order allocations only zones that meet watermarks and are in a
2838 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2839 * total of balanced pages must be at least 25% of the zones allowed by
2840 * classzone_idx for the node to be considered balanced. Forcing all zones to
2841 * be balanced for high orders can cause excessive reclaim when there are
2843 * The choice of 25% is due to
2844 * o a 16M DMA zone that is balanced will not balance a zone on any
2845 * reasonable sized machine
2846 * o On all other machines, the top zone must be at least a reasonable
2847 * percentage of the middle zones. For example, on 32-bit x86, highmem
2848 * would need to be at least 256M for it to be balance a whole node.
2849 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2850 * to balance a node on its own. These seemed like reasonable ratios.
2852 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2854 unsigned long managed_pages = 0;
2855 unsigned long balanced_pages = 0;
2858 /* Check the watermark levels */
2859 for (i = 0; i <= classzone_idx; i++) {
2860 struct zone *zone = pgdat->node_zones + i;
2862 if (!populated_zone(zone))
2865 managed_pages += zone->managed_pages;
2868 * A special case here:
2870 * balance_pgdat() skips over all_unreclaimable after
2871 * DEF_PRIORITY. Effectively, it considers them balanced so
2872 * they must be considered balanced here as well!
2874 if (!zone_reclaimable(zone)) {
2875 balanced_pages += zone->managed_pages;
2879 if (zone_balanced(zone, order, 0, i))
2880 balanced_pages += zone->managed_pages;
2886 return balanced_pages >= (managed_pages >> 2);
2892 * Prepare kswapd for sleeping. This verifies that there are no processes
2893 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2895 * Returns true if kswapd is ready to sleep
2897 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2900 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2905 * There is a potential race between when kswapd checks its watermarks
2906 * and a process gets throttled. There is also a potential race if
2907 * processes get throttled, kswapd wakes, a large process exits therby
2908 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2909 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2910 * so wake them now if necessary. If necessary, processes will wake
2911 * kswapd and get throttled again
2913 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2914 wake_up(&pgdat->pfmemalloc_wait);
2918 return pgdat_balanced(pgdat, order, classzone_idx);
2922 * kswapd shrinks the zone by the number of pages required to reach
2923 * the high watermark.
2925 * Returns true if kswapd scanned at least the requested number of pages to
2926 * reclaim or if the lack of progress was due to pages under writeback.
2927 * This is used to determine if the scanning priority needs to be raised.
2929 static bool kswapd_shrink_zone(struct zone *zone,
2931 struct scan_control *sc,
2932 unsigned long lru_pages,
2933 unsigned long *nr_attempted)
2935 int testorder = sc->order;
2936 unsigned long balance_gap;
2937 struct reclaim_state *reclaim_state = current->reclaim_state;
2938 struct shrink_control shrink = {
2939 .gfp_mask = sc->gfp_mask,
2941 bool lowmem_pressure;
2943 /* Reclaim above the high watermark. */
2944 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2947 * Kswapd reclaims only single pages with compaction enabled. Trying
2948 * too hard to reclaim until contiguous free pages have become
2949 * available can hurt performance by evicting too much useful data
2950 * from memory. Do not reclaim more than needed for compaction.
2952 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2953 compaction_suitable(zone, sc->order) !=
2958 * We put equal pressure on every zone, unless one zone has way too
2959 * many pages free already. The "too many pages" is defined as the
2960 * high wmark plus a "gap" where the gap is either the low
2961 * watermark or 1% of the zone, whichever is smaller.
2963 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2964 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2967 * If there is no low memory pressure or the zone is balanced then no
2968 * reclaim is necessary
2970 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2971 if (!lowmem_pressure && zone_balanced(zone, testorder,
2972 balance_gap, classzone_idx))
2975 shrink_zone(zone, sc);
2976 nodes_clear(shrink.nodes_to_scan);
2977 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2979 reclaim_state->reclaimed_slab = 0;
2980 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2981 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2983 /* Account for the number of pages attempted to reclaim */
2984 *nr_attempted += sc->nr_to_reclaim;
2986 zone_clear_flag(zone, ZONE_WRITEBACK);
2989 * If a zone reaches its high watermark, consider it to be no longer
2990 * congested. It's possible there are dirty pages backed by congested
2991 * BDIs but as pressure is relieved, speculatively avoid congestion
2994 if (zone_reclaimable(zone) &&
2995 zone_balanced(zone, testorder, 0, classzone_idx)) {
2996 zone_clear_flag(zone, ZONE_CONGESTED);
2997 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3000 return sc->nr_scanned >= sc->nr_to_reclaim;
3004 * For kswapd, balance_pgdat() will work across all this node's zones until
3005 * they are all at high_wmark_pages(zone).
3007 * Returns the final order kswapd was reclaiming at
3009 * There is special handling here for zones which are full of pinned pages.
3010 * This can happen if the pages are all mlocked, or if they are all used by
3011 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3012 * What we do is to detect the case where all pages in the zone have been
3013 * scanned twice and there has been zero successful reclaim. Mark the zone as
3014 * dead and from now on, only perform a short scan. Basically we're polling
3015 * the zone for when the problem goes away.
3017 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3018 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3019 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3020 * lower zones regardless of the number of free pages in the lower zones. This
3021 * interoperates with the page allocator fallback scheme to ensure that aging
3022 * of pages is balanced across the zones.
3024 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3028 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3029 unsigned long nr_soft_reclaimed;
3030 unsigned long nr_soft_scanned;
3031 struct scan_control sc = {
3032 .gfp_mask = GFP_KERNEL,
3033 .priority = DEF_PRIORITY,
3036 .may_writepage = !laptop_mode,
3038 .target_mem_cgroup = NULL,
3040 count_vm_event(PAGEOUTRUN);
3043 unsigned long lru_pages = 0;
3044 unsigned long nr_attempted = 0;
3045 bool raise_priority = true;
3046 bool pgdat_needs_compaction = (order > 0);
3048 sc.nr_reclaimed = 0;
3051 * Scan in the highmem->dma direction for the highest
3052 * zone which needs scanning
3054 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3055 struct zone *zone = pgdat->node_zones + i;
3057 if (!populated_zone(zone))
3060 if (sc.priority != DEF_PRIORITY &&
3061 !zone_reclaimable(zone))
3065 * Do some background aging of the anon list, to give
3066 * pages a chance to be referenced before reclaiming.
3068 age_active_anon(zone, &sc);
3071 * If the number of buffer_heads in the machine
3072 * exceeds the maximum allowed level and this node
3073 * has a highmem zone, force kswapd to reclaim from
3074 * it to relieve lowmem pressure.
3076 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3081 if (!zone_balanced(zone, order, 0, 0)) {
3086 * If balanced, clear the dirty and congested
3089 zone_clear_flag(zone, ZONE_CONGESTED);
3090 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3097 for (i = 0; i <= end_zone; i++) {
3098 struct zone *zone = pgdat->node_zones + i;
3100 if (!populated_zone(zone))
3103 lru_pages += zone_reclaimable_pages(zone);
3106 * If any zone is currently balanced then kswapd will
3107 * not call compaction as it is expected that the
3108 * necessary pages are already available.
3110 if (pgdat_needs_compaction &&
3111 zone_watermark_ok(zone, order,
3112 low_wmark_pages(zone),
3114 pgdat_needs_compaction = false;
3118 * If we're getting trouble reclaiming, start doing writepage
3119 * even in laptop mode.
3121 if (sc.priority < DEF_PRIORITY - 2)
3122 sc.may_writepage = 1;
3125 * Now scan the zone in the dma->highmem direction, stopping
3126 * at the last zone which needs scanning.
3128 * We do this because the page allocator works in the opposite
3129 * direction. This prevents the page allocator from allocating
3130 * pages behind kswapd's direction of progress, which would
3131 * cause too much scanning of the lower zones.
3133 for (i = 0; i <= end_zone; i++) {
3134 struct zone *zone = pgdat->node_zones + i;
3136 if (!populated_zone(zone))
3139 if (sc.priority != DEF_PRIORITY &&
3140 !zone_reclaimable(zone))
3145 nr_soft_scanned = 0;
3147 * Call soft limit reclaim before calling shrink_zone.
3149 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3152 sc.nr_reclaimed += nr_soft_reclaimed;
3155 * There should be no need to raise the scanning
3156 * priority if enough pages are already being scanned
3157 * that that high watermark would be met at 100%
3160 if (kswapd_shrink_zone(zone, end_zone, &sc,
3161 lru_pages, &nr_attempted))
3162 raise_priority = false;
3166 * If the low watermark is met there is no need for processes
3167 * to be throttled on pfmemalloc_wait as they should not be
3168 * able to safely make forward progress. Wake them
3170 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3171 pfmemalloc_watermark_ok(pgdat))
3172 wake_up(&pgdat->pfmemalloc_wait);
3175 * Fragmentation may mean that the system cannot be rebalanced
3176 * for high-order allocations in all zones. If twice the
3177 * allocation size has been reclaimed and the zones are still
3178 * not balanced then recheck the watermarks at order-0 to
3179 * prevent kswapd reclaiming excessively. Assume that a
3180 * process requested a high-order can direct reclaim/compact.
3182 if (order && sc.nr_reclaimed >= 2UL << order)
3183 order = sc.order = 0;
3185 /* Check if kswapd should be suspending */
3186 if (try_to_freeze() || kthread_should_stop())
3190 * Compact if necessary and kswapd is reclaiming at least the
3191 * high watermark number of pages as requsted
3193 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3194 compact_pgdat(pgdat, order);
3197 * Raise priority if scanning rate is too low or there was no
3198 * progress in reclaiming pages
3200 if (raise_priority || !sc.nr_reclaimed)
3202 } while (sc.priority >= 1 &&
3203 !pgdat_balanced(pgdat, order, *classzone_idx));
3207 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3208 * makes a decision on the order we were last reclaiming at. However,
3209 * if another caller entered the allocator slow path while kswapd
3210 * was awake, order will remain at the higher level
3212 *classzone_idx = end_zone;
3216 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3221 if (freezing(current) || kthread_should_stop())
3224 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3226 /* Try to sleep for a short interval */
3227 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3228 remaining = schedule_timeout(HZ/10);
3229 finish_wait(&pgdat->kswapd_wait, &wait);
3230 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3234 * After a short sleep, check if it was a premature sleep. If not, then
3235 * go fully to sleep until explicitly woken up.
3237 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3238 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3241 * vmstat counters are not perfectly accurate and the estimated
3242 * value for counters such as NR_FREE_PAGES can deviate from the
3243 * true value by nr_online_cpus * threshold. To avoid the zone
3244 * watermarks being breached while under pressure, we reduce the
3245 * per-cpu vmstat threshold while kswapd is awake and restore
3246 * them before going back to sleep.
3248 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3251 * Compaction records what page blocks it recently failed to
3252 * isolate pages from and skips them in the future scanning.
3253 * When kswapd is going to sleep, it is reasonable to assume
3254 * that pages and compaction may succeed so reset the cache.
3256 reset_isolation_suitable(pgdat);
3258 if (!kthread_should_stop())
3261 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3264 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3266 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3268 finish_wait(&pgdat->kswapd_wait, &wait);
3272 * The background pageout daemon, started as a kernel thread
3273 * from the init process.
3275 * This basically trickles out pages so that we have _some_
3276 * free memory available even if there is no other activity
3277 * that frees anything up. This is needed for things like routing
3278 * etc, where we otherwise might have all activity going on in
3279 * asynchronous contexts that cannot page things out.
3281 * If there are applications that are active memory-allocators
3282 * (most normal use), this basically shouldn't matter.
3284 static int kswapd(void *p)
3286 unsigned long order, new_order;
3287 unsigned balanced_order;
3288 int classzone_idx, new_classzone_idx;
3289 int balanced_classzone_idx;
3290 pg_data_t *pgdat = (pg_data_t*)p;
3291 struct task_struct *tsk = current;
3293 struct reclaim_state reclaim_state = {
3294 .reclaimed_slab = 0,
3296 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3298 lockdep_set_current_reclaim_state(GFP_KERNEL);
3300 if (!cpumask_empty(cpumask))
3301 set_cpus_allowed_ptr(tsk, cpumask);
3302 current->reclaim_state = &reclaim_state;
3305 * Tell the memory management that we're a "memory allocator",
3306 * and that if we need more memory we should get access to it
3307 * regardless (see "__alloc_pages()"). "kswapd" should
3308 * never get caught in the normal page freeing logic.
3310 * (Kswapd normally doesn't need memory anyway, but sometimes
3311 * you need a small amount of memory in order to be able to
3312 * page out something else, and this flag essentially protects
3313 * us from recursively trying to free more memory as we're
3314 * trying to free the first piece of memory in the first place).
3316 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3319 order = new_order = 0;
3321 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3322 balanced_classzone_idx = classzone_idx;
3327 * If the last balance_pgdat was unsuccessful it's unlikely a
3328 * new request of a similar or harder type will succeed soon
3329 * so consider going to sleep on the basis we reclaimed at
3331 if (balanced_classzone_idx >= new_classzone_idx &&
3332 balanced_order == new_order) {
3333 new_order = pgdat->kswapd_max_order;
3334 new_classzone_idx = pgdat->classzone_idx;
3335 pgdat->kswapd_max_order = 0;
3336 pgdat->classzone_idx = pgdat->nr_zones - 1;
3339 if (order < new_order || classzone_idx > new_classzone_idx) {
3341 * Don't sleep if someone wants a larger 'order'
3342 * allocation or has tigher zone constraints
3345 classzone_idx = new_classzone_idx;
3347 kswapd_try_to_sleep(pgdat, balanced_order,
3348 balanced_classzone_idx);
3349 order = pgdat->kswapd_max_order;
3350 classzone_idx = pgdat->classzone_idx;
3352 new_classzone_idx = classzone_idx;
3353 pgdat->kswapd_max_order = 0;
3354 pgdat->classzone_idx = pgdat->nr_zones - 1;
3357 ret = try_to_freeze();
3358 if (kthread_should_stop())
3362 * We can speed up thawing tasks if we don't call balance_pgdat
3363 * after returning from the refrigerator
3366 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3367 balanced_classzone_idx = classzone_idx;
3368 balanced_order = balance_pgdat(pgdat, order,
3369 &balanced_classzone_idx);
3373 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3374 current->reclaim_state = NULL;
3375 lockdep_clear_current_reclaim_state();
3381 * A zone is low on free memory, so wake its kswapd task to service it.
3383 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3387 if (!populated_zone(zone))
3390 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3392 pgdat = zone->zone_pgdat;
3393 if (pgdat->kswapd_max_order < order) {
3394 pgdat->kswapd_max_order = order;
3395 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3397 if (!waitqueue_active(&pgdat->kswapd_wait))
3399 if (zone_balanced(zone, order, 0, 0))
3402 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3403 wake_up_interruptible(&pgdat->kswapd_wait);
3406 #ifdef CONFIG_HIBERNATION
3408 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3411 * Rather than trying to age LRUs the aim is to preserve the overall
3412 * LRU order by reclaiming preferentially
3413 * inactive > active > active referenced > active mapped
3415 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3417 struct reclaim_state reclaim_state;
3418 struct scan_control sc = {
3419 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3423 .nr_to_reclaim = nr_to_reclaim,
3424 .hibernation_mode = 1,
3426 .priority = DEF_PRIORITY,
3428 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3429 struct task_struct *p = current;
3430 unsigned long nr_reclaimed;
3432 p->flags |= PF_MEMALLOC;
3433 lockdep_set_current_reclaim_state(sc.gfp_mask);
3434 reclaim_state.reclaimed_slab = 0;
3435 p->reclaim_state = &reclaim_state;
3437 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3439 p->reclaim_state = NULL;
3440 lockdep_clear_current_reclaim_state();
3441 p->flags &= ~PF_MEMALLOC;
3443 return nr_reclaimed;
3445 #endif /* CONFIG_HIBERNATION */
3447 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3448 not required for correctness. So if the last cpu in a node goes
3449 away, we get changed to run anywhere: as the first one comes back,
3450 restore their cpu bindings. */
3451 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3456 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3457 for_each_node_state(nid, N_MEMORY) {
3458 pg_data_t *pgdat = NODE_DATA(nid);
3459 const struct cpumask *mask;
3461 mask = cpumask_of_node(pgdat->node_id);
3463 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3464 /* One of our CPUs online: restore mask */
3465 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3472 * This kswapd start function will be called by init and node-hot-add.
3473 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3475 int kswapd_run(int nid)
3477 pg_data_t *pgdat = NODE_DATA(nid);
3483 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3484 if (IS_ERR(pgdat->kswapd)) {
3485 /* failure at boot is fatal */
3486 BUG_ON(system_state == SYSTEM_BOOTING);
3487 pr_err("Failed to start kswapd on node %d\n", nid);
3488 ret = PTR_ERR(pgdat->kswapd);
3489 pgdat->kswapd = NULL;
3495 * Called by memory hotplug when all memory in a node is offlined. Caller must
3496 * hold mem_hotplug_begin/end().
3498 void kswapd_stop(int nid)
3500 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3503 kthread_stop(kswapd);
3504 NODE_DATA(nid)->kswapd = NULL;
3508 static int __init kswapd_init(void)
3513 for_each_node_state(nid, N_MEMORY)
3515 hotcpu_notifier(cpu_callback, 0);
3519 module_init(kswapd_init)
3525 * If non-zero call zone_reclaim when the number of free pages falls below
3528 int zone_reclaim_mode __read_mostly;
3530 #define RECLAIM_OFF 0
3531 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3532 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3533 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3536 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3537 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3540 #define ZONE_RECLAIM_PRIORITY 4
3543 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3546 int sysctl_min_unmapped_ratio = 1;
3549 * If the number of slab pages in a zone grows beyond this percentage then
3550 * slab reclaim needs to occur.
3552 int sysctl_min_slab_ratio = 5;
3554 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3556 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3557 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3558 zone_page_state(zone, NR_ACTIVE_FILE);
3561 * It's possible for there to be more file mapped pages than
3562 * accounted for by the pages on the file LRU lists because
3563 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3565 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3568 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3569 static long zone_pagecache_reclaimable(struct zone *zone)
3571 long nr_pagecache_reclaimable;
3575 * If RECLAIM_SWAP is set, then all file pages are considered
3576 * potentially reclaimable. Otherwise, we have to worry about
3577 * pages like swapcache and zone_unmapped_file_pages() provides
3580 if (zone_reclaim_mode & RECLAIM_SWAP)
3581 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3583 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3585 /* If we can't clean pages, remove dirty pages from consideration */
3586 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3587 delta += zone_page_state(zone, NR_FILE_DIRTY);
3589 /* Watch for any possible underflows due to delta */
3590 if (unlikely(delta > nr_pagecache_reclaimable))
3591 delta = nr_pagecache_reclaimable;
3593 return nr_pagecache_reclaimable - delta;
3597 * Try to free up some pages from this zone through reclaim.
3599 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3601 /* Minimum pages needed in order to stay on node */
3602 const unsigned long nr_pages = 1 << order;
3603 struct task_struct *p = current;
3604 struct reclaim_state reclaim_state;
3605 struct scan_control sc = {
3606 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3607 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3609 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3610 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3612 .priority = ZONE_RECLAIM_PRIORITY,
3614 struct shrink_control shrink = {
3615 .gfp_mask = sc.gfp_mask,
3617 unsigned long nr_slab_pages0, nr_slab_pages1;
3621 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3622 * and we also need to be able to write out pages for RECLAIM_WRITE
3625 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3626 lockdep_set_current_reclaim_state(gfp_mask);
3627 reclaim_state.reclaimed_slab = 0;
3628 p->reclaim_state = &reclaim_state;
3630 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3632 * Free memory by calling shrink zone with increasing
3633 * priorities until we have enough memory freed.
3636 shrink_zone(zone, &sc);
3637 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3640 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3641 if (nr_slab_pages0 > zone->min_slab_pages) {
3643 * shrink_slab() does not currently allow us to determine how
3644 * many pages were freed in this zone. So we take the current
3645 * number of slab pages and shake the slab until it is reduced
3646 * by the same nr_pages that we used for reclaiming unmapped
3649 nodes_clear(shrink.nodes_to_scan);
3650 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3652 unsigned long lru_pages = zone_reclaimable_pages(zone);
3654 /* No reclaimable slab or very low memory pressure */
3655 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3658 /* Freed enough memory */
3659 nr_slab_pages1 = zone_page_state(zone,
3660 NR_SLAB_RECLAIMABLE);
3661 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3666 * Update nr_reclaimed by the number of slab pages we
3667 * reclaimed from this zone.
3669 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3670 if (nr_slab_pages1 < nr_slab_pages0)
3671 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3674 p->reclaim_state = NULL;
3675 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3676 lockdep_clear_current_reclaim_state();
3677 return sc.nr_reclaimed >= nr_pages;
3680 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3686 * Zone reclaim reclaims unmapped file backed pages and
3687 * slab pages if we are over the defined limits.
3689 * A small portion of unmapped file backed pages is needed for
3690 * file I/O otherwise pages read by file I/O will be immediately
3691 * thrown out if the zone is overallocated. So we do not reclaim
3692 * if less than a specified percentage of the zone is used by
3693 * unmapped file backed pages.
3695 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3696 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3697 return ZONE_RECLAIM_FULL;
3699 if (!zone_reclaimable(zone))
3700 return ZONE_RECLAIM_FULL;
3703 * Do not scan if the allocation should not be delayed.
3705 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3706 return ZONE_RECLAIM_NOSCAN;
3709 * Only run zone reclaim on the local zone or on zones that do not
3710 * have associated processors. This will favor the local processor
3711 * over remote processors and spread off node memory allocations
3712 * as wide as possible.
3714 node_id = zone_to_nid(zone);
3715 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3716 return ZONE_RECLAIM_NOSCAN;
3718 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3719 return ZONE_RECLAIM_NOSCAN;
3721 ret = __zone_reclaim(zone, gfp_mask, order);
3722 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3725 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3732 * page_evictable - test whether a page is evictable
3733 * @page: the page to test
3735 * Test whether page is evictable--i.e., should be placed on active/inactive
3736 * lists vs unevictable list.
3738 * Reasons page might not be evictable:
3739 * (1) page's mapping marked unevictable
3740 * (2) page is part of an mlocked VMA
3743 int page_evictable(struct page *page)
3745 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3750 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3751 * @pages: array of pages to check
3752 * @nr_pages: number of pages to check
3754 * Checks pages for evictability and moves them to the appropriate lru list.
3756 * This function is only used for SysV IPC SHM_UNLOCK.
3758 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3760 struct lruvec *lruvec;
3761 struct zone *zone = NULL;
3766 for (i = 0; i < nr_pages; i++) {
3767 struct page *page = pages[i];
3768 struct zone *pagezone;
3771 pagezone = page_zone(page);
3772 if (pagezone != zone) {
3774 spin_unlock_irq(&zone->lru_lock);
3776 spin_lock_irq(&zone->lru_lock);
3778 lruvec = mem_cgroup_page_lruvec(page, zone);
3780 if (!PageLRU(page) || !PageUnevictable(page))
3783 if (page_evictable(page)) {
3784 enum lru_list lru = page_lru_base_type(page);
3786 VM_BUG_ON_PAGE(PageActive(page), page);
3787 ClearPageUnevictable(page);
3788 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3789 add_page_to_lru_list(page, lruvec, lru);
3795 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3796 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3797 spin_unlock_irq(&zone->lru_lock);
3800 #endif /* CONFIG_SHMEM */
3802 static void warn_scan_unevictable_pages(void)
3804 printk_once(KERN_WARNING
3805 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3806 "disabled for lack of a legitimate use case. If you have "
3807 "one, please send an email to linux-mm@kvack.org.\n",
3812 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3813 * all nodes' unevictable lists for evictable pages
3815 unsigned long scan_unevictable_pages;
3817 int scan_unevictable_handler(struct ctl_table *table, int write,
3818 void __user *buffer,
3819 size_t *length, loff_t *ppos)
3821 warn_scan_unevictable_pages();
3822 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3823 scan_unevictable_pages = 0;
3829 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3830 * a specified node's per zone unevictable lists for evictable pages.
3833 static ssize_t read_scan_unevictable_node(struct device *dev,
3834 struct device_attribute *attr,
3837 warn_scan_unevictable_pages();
3838 return sprintf(buf, "0\n"); /* always zero; should fit... */
3841 static ssize_t write_scan_unevictable_node(struct device *dev,
3842 struct device_attribute *attr,
3843 const char *buf, size_t count)
3845 warn_scan_unevictable_pages();
3850 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3851 read_scan_unevictable_node,
3852 write_scan_unevictable_node);
3854 int scan_unevictable_register_node(struct node *node)
3856 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3859 void scan_unevictable_unregister_node(struct node *node)
3861 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);