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>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned;
61 /* Number of pages freed so far during a call to shrink_zones() */
62 unsigned long nr_reclaimed;
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim;
67 unsigned long hibernation_mode;
69 /* This context's GFP mask */
74 /* Can mapped pages be reclaimed? */
77 /* Can pages be swapped as part of reclaim? */
82 /* Scan (total_size >> priority) pages at once */
86 * The memory cgroup that hit its limit and as a result is the
87 * primary target of this reclaim invocation.
89 struct mem_cgroup *target_mem_cgroup;
92 * Nodemask of nodes allowed by the caller. If NULL, all nodes
98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100 #ifdef ARCH_HAS_PREFETCH
101 #define prefetch_prev_lru_page(_page, _base, _field) \
103 if ((_page)->lru.prev != _base) { \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
114 #ifdef ARCH_HAS_PREFETCHW
115 #define prefetchw_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
129 * From 0 .. 100. Higher means more swappy.
131 int vm_swappiness = 60;
132 unsigned long vm_total_pages; /* The total number of pages which the VM controls */
134 static LIST_HEAD(shrinker_list);
135 static DECLARE_RWSEM(shrinker_rwsem);
138 static bool global_reclaim(struct scan_control *sc)
140 return !sc->target_mem_cgroup;
143 static bool mem_cgroup_should_soft_reclaim(struct scan_control *sc)
145 return !mem_cgroup_disabled() && global_reclaim(sc);
148 static bool global_reclaim(struct scan_control *sc)
153 static bool mem_cgroup_should_soft_reclaim(struct scan_control *sc)
159 unsigned long zone_reclaimable_pages(struct zone *zone)
163 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
164 zone_page_state(zone, NR_INACTIVE_FILE);
166 if (get_nr_swap_pages() > 0)
167 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
168 zone_page_state(zone, NR_INACTIVE_ANON);
173 bool zone_reclaimable(struct zone *zone)
175 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
178 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
180 if (!mem_cgroup_disabled())
181 return mem_cgroup_get_lru_size(lruvec, lru);
183 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
187 * Add a shrinker callback to be called from the vm.
189 int register_shrinker(struct shrinker *shrinker)
191 size_t size = sizeof(*shrinker->nr_deferred);
194 * If we only have one possible node in the system anyway, save
195 * ourselves the trouble and disable NUMA aware behavior. This way we
196 * will save memory and some small loop time later.
198 if (nr_node_ids == 1)
199 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
201 if (shrinker->flags & SHRINKER_NUMA_AWARE)
204 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
205 if (!shrinker->nr_deferred)
208 down_write(&shrinker_rwsem);
209 list_add_tail(&shrinker->list, &shrinker_list);
210 up_write(&shrinker_rwsem);
213 EXPORT_SYMBOL(register_shrinker);
218 void unregister_shrinker(struct shrinker *shrinker)
220 down_write(&shrinker_rwsem);
221 list_del(&shrinker->list);
222 up_write(&shrinker_rwsem);
224 EXPORT_SYMBOL(unregister_shrinker);
226 #define SHRINK_BATCH 128
229 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
230 unsigned long nr_pages_scanned, unsigned long lru_pages)
232 unsigned long freed = 0;
233 unsigned long long delta;
238 int nid = shrinkctl->nid;
239 long batch_size = shrinker->batch ? shrinker->batch
242 max_pass = shrinker->count_objects(shrinker, shrinkctl);
247 * copy the current shrinker scan count into a local variable
248 * and zero it so that other concurrent shrinker invocations
249 * don't also do this scanning work.
251 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
254 delta = (4 * nr_pages_scanned) / shrinker->seeks;
256 do_div(delta, lru_pages + 1);
258 if (total_scan < 0) {
260 "shrink_slab: %pF negative objects to delete nr=%ld\n",
261 shrinker->scan_objects, total_scan);
262 total_scan = max_pass;
266 * We need to avoid excessive windup on filesystem shrinkers
267 * due to large numbers of GFP_NOFS allocations causing the
268 * shrinkers to return -1 all the time. This results in a large
269 * nr being built up so when a shrink that can do some work
270 * comes along it empties the entire cache due to nr >>>
271 * max_pass. This is bad for sustaining a working set in
274 * Hence only allow the shrinker to scan the entire cache when
275 * a large delta change is calculated directly.
277 if (delta < max_pass / 4)
278 total_scan = min(total_scan, max_pass / 2);
281 * Avoid risking looping forever due to too large nr value:
282 * never try to free more than twice the estimate number of
285 if (total_scan > max_pass * 2)
286 total_scan = max_pass * 2;
288 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
289 nr_pages_scanned, lru_pages,
290 max_pass, delta, total_scan);
292 while (total_scan >= batch_size) {
295 shrinkctl->nr_to_scan = batch_size;
296 ret = shrinker->scan_objects(shrinker, shrinkctl);
297 if (ret == SHRINK_STOP)
301 count_vm_events(SLABS_SCANNED, batch_size);
302 total_scan -= batch_size;
308 * move the unused scan count back into the shrinker in a
309 * manner that handles concurrent updates. If we exhausted the
310 * scan, there is no need to do an update.
313 new_nr = atomic_long_add_return(total_scan,
314 &shrinker->nr_deferred[nid]);
316 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
318 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
323 * Call the shrink functions to age shrinkable caches
325 * Here we assume it costs one seek to replace a lru page and that it also
326 * takes a seek to recreate a cache object. With this in mind we age equal
327 * percentages of the lru and ageable caches. This should balance the seeks
328 * generated by these structures.
330 * If the vm encountered mapped pages on the LRU it increase the pressure on
331 * slab to avoid swapping.
333 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
335 * `lru_pages' represents the number of on-LRU pages in all the zones which
336 * are eligible for the caller's allocation attempt. It is used for balancing
337 * slab reclaim versus page reclaim.
339 * Returns the number of slab objects which we shrunk.
341 unsigned long shrink_slab(struct shrink_control *shrinkctl,
342 unsigned long nr_pages_scanned,
343 unsigned long lru_pages)
345 struct shrinker *shrinker;
346 unsigned long freed = 0;
348 if (nr_pages_scanned == 0)
349 nr_pages_scanned = SWAP_CLUSTER_MAX;
351 if (!down_read_trylock(&shrinker_rwsem)) {
353 * If we would return 0, our callers would understand that we
354 * have nothing else to shrink and give up trying. By returning
355 * 1 we keep it going and assume we'll be able to shrink next
362 list_for_each_entry(shrinker, &shrinker_list, list) {
363 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
364 if (!node_online(shrinkctl->nid))
367 if (!(shrinker->flags & SHRINKER_NUMA_AWARE) &&
368 (shrinkctl->nid != 0))
371 freed += shrink_slab_node(shrinkctl, shrinker,
372 nr_pages_scanned, lru_pages);
376 up_read(&shrinker_rwsem);
382 static inline int is_page_cache_freeable(struct page *page)
385 * A freeable page cache page is referenced only by the caller
386 * that isolated the page, the page cache radix tree and
387 * optional buffer heads at page->private.
389 return page_count(page) - page_has_private(page) == 2;
392 static int may_write_to_queue(struct backing_dev_info *bdi,
393 struct scan_control *sc)
395 if (current->flags & PF_SWAPWRITE)
397 if (!bdi_write_congested(bdi))
399 if (bdi == current->backing_dev_info)
405 * We detected a synchronous write error writing a page out. Probably
406 * -ENOSPC. We need to propagate that into the address_space for a subsequent
407 * fsync(), msync() or close().
409 * The tricky part is that after writepage we cannot touch the mapping: nothing
410 * prevents it from being freed up. But we have a ref on the page and once
411 * that page is locked, the mapping is pinned.
413 * We're allowed to run sleeping lock_page() here because we know the caller has
416 static void handle_write_error(struct address_space *mapping,
417 struct page *page, int error)
420 if (page_mapping(page) == mapping)
421 mapping_set_error(mapping, error);
425 /* possible outcome of pageout() */
427 /* failed to write page out, page is locked */
429 /* move page to the active list, page is locked */
431 /* page has been sent to the disk successfully, page is unlocked */
433 /* page is clean and locked */
438 * pageout is called by shrink_page_list() for each dirty page.
439 * Calls ->writepage().
441 static pageout_t pageout(struct page *page, struct address_space *mapping,
442 struct scan_control *sc)
445 * If the page is dirty, only perform writeback if that write
446 * will be non-blocking. To prevent this allocation from being
447 * stalled by pagecache activity. But note that there may be
448 * stalls if we need to run get_block(). We could test
449 * PagePrivate for that.
451 * If this process is currently in __generic_file_aio_write() against
452 * this page's queue, we can perform writeback even if that
455 * If the page is swapcache, write it back even if that would
456 * block, for some throttling. This happens by accident, because
457 * swap_backing_dev_info is bust: it doesn't reflect the
458 * congestion state of the swapdevs. Easy to fix, if needed.
460 if (!is_page_cache_freeable(page))
464 * Some data journaling orphaned pages can have
465 * page->mapping == NULL while being dirty with clean buffers.
467 if (page_has_private(page)) {
468 if (try_to_free_buffers(page)) {
469 ClearPageDirty(page);
470 printk("%s: orphaned page\n", __func__);
476 if (mapping->a_ops->writepage == NULL)
477 return PAGE_ACTIVATE;
478 if (!may_write_to_queue(mapping->backing_dev_info, sc))
481 if (clear_page_dirty_for_io(page)) {
483 struct writeback_control wbc = {
484 .sync_mode = WB_SYNC_NONE,
485 .nr_to_write = SWAP_CLUSTER_MAX,
487 .range_end = LLONG_MAX,
491 SetPageReclaim(page);
492 res = mapping->a_ops->writepage(page, &wbc);
494 handle_write_error(mapping, page, res);
495 if (res == AOP_WRITEPAGE_ACTIVATE) {
496 ClearPageReclaim(page);
497 return PAGE_ACTIVATE;
500 if (!PageWriteback(page)) {
501 /* synchronous write or broken a_ops? */
502 ClearPageReclaim(page);
504 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
505 inc_zone_page_state(page, NR_VMSCAN_WRITE);
513 * Same as remove_mapping, but if the page is removed from the mapping, it
514 * gets returned with a refcount of 0.
516 static int __remove_mapping(struct address_space *mapping, struct page *page)
518 BUG_ON(!PageLocked(page));
519 BUG_ON(mapping != page_mapping(page));
521 spin_lock_irq(&mapping->tree_lock);
523 * The non racy check for a busy page.
525 * Must be careful with the order of the tests. When someone has
526 * a ref to the page, it may be possible that they dirty it then
527 * drop the reference. So if PageDirty is tested before page_count
528 * here, then the following race may occur:
530 * get_user_pages(&page);
531 * [user mapping goes away]
533 * !PageDirty(page) [good]
534 * SetPageDirty(page);
536 * !page_count(page) [good, discard it]
538 * [oops, our write_to data is lost]
540 * Reversing the order of the tests ensures such a situation cannot
541 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
542 * load is not satisfied before that of page->_count.
544 * Note that if SetPageDirty is always performed via set_page_dirty,
545 * and thus under tree_lock, then this ordering is not required.
547 if (!page_freeze_refs(page, 2))
549 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
550 if (unlikely(PageDirty(page))) {
551 page_unfreeze_refs(page, 2);
555 if (PageSwapCache(page)) {
556 swp_entry_t swap = { .val = page_private(page) };
557 __delete_from_swap_cache(page);
558 spin_unlock_irq(&mapping->tree_lock);
559 swapcache_free(swap, page);
561 void (*freepage)(struct page *);
563 freepage = mapping->a_ops->freepage;
565 __delete_from_page_cache(page);
566 spin_unlock_irq(&mapping->tree_lock);
567 mem_cgroup_uncharge_cache_page(page);
569 if (freepage != NULL)
576 spin_unlock_irq(&mapping->tree_lock);
581 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
582 * someone else has a ref on the page, abort and return 0. If it was
583 * successfully detached, return 1. Assumes the caller has a single ref on
586 int remove_mapping(struct address_space *mapping, struct page *page)
588 if (__remove_mapping(mapping, page)) {
590 * Unfreezing the refcount with 1 rather than 2 effectively
591 * drops the pagecache ref for us without requiring another
594 page_unfreeze_refs(page, 1);
601 * putback_lru_page - put previously isolated page onto appropriate LRU list
602 * @page: page to be put back to appropriate lru list
604 * Add previously isolated @page to appropriate LRU list.
605 * Page may still be unevictable for other reasons.
607 * lru_lock must not be held, interrupts must be enabled.
609 void putback_lru_page(struct page *page)
612 int was_unevictable = PageUnevictable(page);
614 VM_BUG_ON(PageLRU(page));
617 ClearPageUnevictable(page);
619 if (page_evictable(page)) {
621 * For evictable pages, we can use the cache.
622 * In event of a race, worst case is we end up with an
623 * unevictable page on [in]active list.
624 * We know how to handle that.
626 is_unevictable = false;
630 * Put unevictable pages directly on zone's unevictable
633 is_unevictable = true;
634 add_page_to_unevictable_list(page);
636 * When racing with an mlock or AS_UNEVICTABLE clearing
637 * (page is unlocked) make sure that if the other thread
638 * does not observe our setting of PG_lru and fails
639 * isolation/check_move_unevictable_pages,
640 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641 * the page back to the evictable list.
643 * The other side is TestClearPageMlocked() or shmem_lock().
649 * page's status can change while we move it among lru. If an evictable
650 * page is on unevictable list, it never be freed. To avoid that,
651 * check after we added it to the list, again.
653 if (is_unevictable && page_evictable(page)) {
654 if (!isolate_lru_page(page)) {
658 /* This means someone else dropped this page from LRU
659 * So, it will be freed or putback to LRU again. There is
660 * nothing to do here.
664 if (was_unevictable && !is_unevictable)
665 count_vm_event(UNEVICTABLE_PGRESCUED);
666 else if (!was_unevictable && is_unevictable)
667 count_vm_event(UNEVICTABLE_PGCULLED);
669 put_page(page); /* drop ref from isolate */
672 enum page_references {
674 PAGEREF_RECLAIM_CLEAN,
679 static enum page_references page_check_references(struct page *page,
680 struct scan_control *sc)
682 int referenced_ptes, referenced_page;
683 unsigned long vm_flags;
685 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
687 referenced_page = TestClearPageReferenced(page);
690 * Mlock lost the isolation race with us. Let try_to_unmap()
691 * move the page to the unevictable list.
693 if (vm_flags & VM_LOCKED)
694 return PAGEREF_RECLAIM;
696 if (referenced_ptes) {
697 if (PageSwapBacked(page))
698 return PAGEREF_ACTIVATE;
700 * All mapped pages start out with page table
701 * references from the instantiating fault, so we need
702 * to look twice if a mapped file page is used more
705 * Mark it and spare it for another trip around the
706 * inactive list. Another page table reference will
707 * lead to its activation.
709 * Note: the mark is set for activated pages as well
710 * so that recently deactivated but used pages are
713 SetPageReferenced(page);
715 if (referenced_page || referenced_ptes > 1)
716 return PAGEREF_ACTIVATE;
719 * Activate file-backed executable pages after first usage.
721 if (vm_flags & VM_EXEC)
722 return PAGEREF_ACTIVATE;
727 /* Reclaim if clean, defer dirty pages to writeback */
728 if (referenced_page && !PageSwapBacked(page))
729 return PAGEREF_RECLAIM_CLEAN;
731 return PAGEREF_RECLAIM;
734 /* Check if a page is dirty or under writeback */
735 static void page_check_dirty_writeback(struct page *page,
736 bool *dirty, bool *writeback)
738 struct address_space *mapping;
741 * Anonymous pages are not handled by flushers and must be written
742 * from reclaim context. Do not stall reclaim based on them
744 if (!page_is_file_cache(page)) {
750 /* By default assume that the page flags are accurate */
751 *dirty = PageDirty(page);
752 *writeback = PageWriteback(page);
754 /* Verify dirty/writeback state if the filesystem supports it */
755 if (!page_has_private(page))
758 mapping = page_mapping(page);
759 if (mapping && mapping->a_ops->is_dirty_writeback)
760 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
764 * shrink_page_list() returns the number of reclaimed pages
766 static unsigned long shrink_page_list(struct list_head *page_list,
768 struct scan_control *sc,
769 enum ttu_flags ttu_flags,
770 unsigned long *ret_nr_dirty,
771 unsigned long *ret_nr_unqueued_dirty,
772 unsigned long *ret_nr_congested,
773 unsigned long *ret_nr_writeback,
774 unsigned long *ret_nr_immediate,
777 LIST_HEAD(ret_pages);
778 LIST_HEAD(free_pages);
780 unsigned long nr_unqueued_dirty = 0;
781 unsigned long nr_dirty = 0;
782 unsigned long nr_congested = 0;
783 unsigned long nr_reclaimed = 0;
784 unsigned long nr_writeback = 0;
785 unsigned long nr_immediate = 0;
789 mem_cgroup_uncharge_start();
790 while (!list_empty(page_list)) {
791 struct address_space *mapping;
794 enum page_references references = PAGEREF_RECLAIM_CLEAN;
795 bool dirty, writeback;
799 page = lru_to_page(page_list);
800 list_del(&page->lru);
802 if (!trylock_page(page))
805 VM_BUG_ON(PageActive(page));
806 VM_BUG_ON(page_zone(page) != zone);
810 if (unlikely(!page_evictable(page)))
813 if (!sc->may_unmap && page_mapped(page))
816 /* Double the slab pressure for mapped and swapcache pages */
817 if (page_mapped(page) || PageSwapCache(page))
820 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
821 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
824 * The number of dirty pages determines if a zone is marked
825 * reclaim_congested which affects wait_iff_congested. kswapd
826 * will stall and start writing pages if the tail of the LRU
827 * is all dirty unqueued pages.
829 page_check_dirty_writeback(page, &dirty, &writeback);
830 if (dirty || writeback)
833 if (dirty && !writeback)
837 * Treat this page as congested if the underlying BDI is or if
838 * pages are cycling through the LRU so quickly that the
839 * pages marked for immediate reclaim are making it to the
840 * end of the LRU a second time.
842 mapping = page_mapping(page);
843 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
844 (writeback && PageReclaim(page)))
848 * If a page at the tail of the LRU is under writeback, there
849 * are three cases to consider.
851 * 1) If reclaim is encountering an excessive number of pages
852 * under writeback and this page is both under writeback and
853 * PageReclaim then it indicates that pages are being queued
854 * for IO but are being recycled through the LRU before the
855 * IO can complete. Waiting on the page itself risks an
856 * indefinite stall if it is impossible to writeback the
857 * page due to IO error or disconnected storage so instead
858 * note that the LRU is being scanned too quickly and the
859 * caller can stall after page list has been processed.
861 * 2) Global reclaim encounters a page, memcg encounters a
862 * page that is not marked for immediate reclaim or
863 * the caller does not have __GFP_IO. In this case mark
864 * the page for immediate reclaim and continue scanning.
866 * __GFP_IO is checked because a loop driver thread might
867 * enter reclaim, and deadlock if it waits on a page for
868 * which it is needed to do the write (loop masks off
869 * __GFP_IO|__GFP_FS for this reason); but more thought
870 * would probably show more reasons.
872 * Don't require __GFP_FS, since we're not going into the
873 * FS, just waiting on its writeback completion. Worryingly,
874 * ext4 gfs2 and xfs allocate pages with
875 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
876 * may_enter_fs here is liable to OOM on them.
878 * 3) memcg encounters a page that is not already marked
879 * PageReclaim. memcg does not have any dirty pages
880 * throttling so we could easily OOM just because too many
881 * pages are in writeback and there is nothing else to
882 * reclaim. Wait for the writeback to complete.
884 if (PageWriteback(page)) {
886 if (current_is_kswapd() &&
888 zone_is_reclaim_writeback(zone)) {
893 } else if (global_reclaim(sc) ||
894 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
896 * This is slightly racy - end_page_writeback()
897 * might have just cleared PageReclaim, then
898 * setting PageReclaim here end up interpreted
899 * as PageReadahead - but that does not matter
900 * enough to care. What we do want is for this
901 * page to have PageReclaim set next time memcg
902 * reclaim reaches the tests above, so it will
903 * then wait_on_page_writeback() to avoid OOM;
904 * and it's also appropriate in global reclaim.
906 SetPageReclaim(page);
913 wait_on_page_writeback(page);
918 references = page_check_references(page, sc);
920 switch (references) {
921 case PAGEREF_ACTIVATE:
922 goto activate_locked;
925 case PAGEREF_RECLAIM:
926 case PAGEREF_RECLAIM_CLEAN:
927 ; /* try to reclaim the page below */
931 * Anonymous process memory has backing store?
932 * Try to allocate it some swap space here.
934 if (PageAnon(page) && !PageSwapCache(page)) {
935 if (!(sc->gfp_mask & __GFP_IO))
937 if (!add_to_swap(page, page_list))
938 goto activate_locked;
941 /* Adding to swap updated mapping */
942 mapping = page_mapping(page);
946 * The page is mapped into the page tables of one or more
947 * processes. Try to unmap it here.
949 if (page_mapped(page) && mapping) {
950 switch (try_to_unmap(page, ttu_flags)) {
952 goto activate_locked;
958 ; /* try to free the page below */
962 if (PageDirty(page)) {
964 * Only kswapd can writeback filesystem pages to
965 * avoid risk of stack overflow but only writeback
966 * if many dirty pages have been encountered.
968 if (page_is_file_cache(page) &&
969 (!current_is_kswapd() ||
970 !zone_is_reclaim_dirty(zone))) {
972 * Immediately reclaim when written back.
973 * Similar in principal to deactivate_page()
974 * except we already have the page isolated
975 * and know it's dirty
977 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
978 SetPageReclaim(page);
983 if (references == PAGEREF_RECLAIM_CLEAN)
987 if (!sc->may_writepage)
990 /* Page is dirty, try to write it out here */
991 switch (pageout(page, mapping, sc)) {
995 goto activate_locked;
997 if (PageWriteback(page))
1003 * A synchronous write - probably a ramdisk. Go
1004 * ahead and try to reclaim the page.
1006 if (!trylock_page(page))
1008 if (PageDirty(page) || PageWriteback(page))
1010 mapping = page_mapping(page);
1012 ; /* try to free the page below */
1017 * If the page has buffers, try to free the buffer mappings
1018 * associated with this page. If we succeed we try to free
1021 * We do this even if the page is PageDirty().
1022 * try_to_release_page() does not perform I/O, but it is
1023 * possible for a page to have PageDirty set, but it is actually
1024 * clean (all its buffers are clean). This happens if the
1025 * buffers were written out directly, with submit_bh(). ext3
1026 * will do this, as well as the blockdev mapping.
1027 * try_to_release_page() will discover that cleanness and will
1028 * drop the buffers and mark the page clean - it can be freed.
1030 * Rarely, pages can have buffers and no ->mapping. These are
1031 * the pages which were not successfully invalidated in
1032 * truncate_complete_page(). We try to drop those buffers here
1033 * and if that worked, and the page is no longer mapped into
1034 * process address space (page_count == 1) it can be freed.
1035 * Otherwise, leave the page on the LRU so it is swappable.
1037 if (page_has_private(page)) {
1038 if (!try_to_release_page(page, sc->gfp_mask))
1039 goto activate_locked;
1040 if (!mapping && page_count(page) == 1) {
1042 if (put_page_testzero(page))
1046 * rare race with speculative reference.
1047 * the speculative reference will free
1048 * this page shortly, so we may
1049 * increment nr_reclaimed here (and
1050 * leave it off the LRU).
1058 if (!mapping || !__remove_mapping(mapping, page))
1062 * At this point, we have no other references and there is
1063 * no way to pick any more up (removed from LRU, removed
1064 * from pagecache). Can use non-atomic bitops now (and
1065 * we obviously don't have to worry about waking up a process
1066 * waiting on the page lock, because there are no references.
1068 __clear_page_locked(page);
1073 * Is there need to periodically free_page_list? It would
1074 * appear not as the counts should be low
1076 list_add(&page->lru, &free_pages);
1080 if (PageSwapCache(page))
1081 try_to_free_swap(page);
1083 putback_lru_page(page);
1087 /* Not a candidate for swapping, so reclaim swap space. */
1088 if (PageSwapCache(page) && vm_swap_full())
1089 try_to_free_swap(page);
1090 VM_BUG_ON(PageActive(page));
1091 SetPageActive(page);
1096 list_add(&page->lru, &ret_pages);
1097 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1100 free_hot_cold_page_list(&free_pages, 1);
1102 list_splice(&ret_pages, page_list);
1103 count_vm_events(PGACTIVATE, pgactivate);
1104 mem_cgroup_uncharge_end();
1105 *ret_nr_dirty += nr_dirty;
1106 *ret_nr_congested += nr_congested;
1107 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1108 *ret_nr_writeback += nr_writeback;
1109 *ret_nr_immediate += nr_immediate;
1110 return nr_reclaimed;
1113 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1114 struct list_head *page_list)
1116 struct scan_control sc = {
1117 .gfp_mask = GFP_KERNEL,
1118 .priority = DEF_PRIORITY,
1121 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1122 struct page *page, *next;
1123 LIST_HEAD(clean_pages);
1125 list_for_each_entry_safe(page, next, page_list, lru) {
1126 if (page_is_file_cache(page) && !PageDirty(page)) {
1127 ClearPageActive(page);
1128 list_move(&page->lru, &clean_pages);
1132 ret = shrink_page_list(&clean_pages, zone, &sc,
1133 TTU_UNMAP|TTU_IGNORE_ACCESS,
1134 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1135 list_splice(&clean_pages, page_list);
1136 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1141 * Attempt to remove the specified page from its LRU. Only take this page
1142 * if it is of the appropriate PageActive status. Pages which are being
1143 * freed elsewhere are also ignored.
1145 * page: page to consider
1146 * mode: one of the LRU isolation modes defined above
1148 * returns 0 on success, -ve errno on failure.
1150 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1154 /* Only take pages on the LRU. */
1158 /* Compaction should not handle unevictable pages but CMA can do so */
1159 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1165 * To minimise LRU disruption, the caller can indicate that it only
1166 * wants to isolate pages it will be able to operate on without
1167 * blocking - clean pages for the most part.
1169 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1170 * is used by reclaim when it is cannot write to backing storage
1172 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1173 * that it is possible to migrate without blocking
1175 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1176 /* All the caller can do on PageWriteback is block */
1177 if (PageWriteback(page))
1180 if (PageDirty(page)) {
1181 struct address_space *mapping;
1183 /* ISOLATE_CLEAN means only clean pages */
1184 if (mode & ISOLATE_CLEAN)
1188 * Only pages without mappings or that have a
1189 * ->migratepage callback are possible to migrate
1192 mapping = page_mapping(page);
1193 if (mapping && !mapping->a_ops->migratepage)
1198 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1201 if (likely(get_page_unless_zero(page))) {
1203 * Be careful not to clear PageLRU until after we're
1204 * sure the page is not being freed elsewhere -- the
1205 * page release code relies on it.
1215 * zone->lru_lock is heavily contended. Some of the functions that
1216 * shrink the lists perform better by taking out a batch of pages
1217 * and working on them outside the LRU lock.
1219 * For pagecache intensive workloads, this function is the hottest
1220 * spot in the kernel (apart from copy_*_user functions).
1222 * Appropriate locks must be held before calling this function.
1224 * @nr_to_scan: The number of pages to look through on the list.
1225 * @lruvec: The LRU vector to pull pages from.
1226 * @dst: The temp list to put pages on to.
1227 * @nr_scanned: The number of pages that were scanned.
1228 * @sc: The scan_control struct for this reclaim session
1229 * @mode: One of the LRU isolation modes
1230 * @lru: LRU list id for isolating
1232 * returns how many pages were moved onto *@dst.
1234 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1235 struct lruvec *lruvec, struct list_head *dst,
1236 unsigned long *nr_scanned, struct scan_control *sc,
1237 isolate_mode_t mode, enum lru_list lru)
1239 struct list_head *src = &lruvec->lists[lru];
1240 unsigned long nr_taken = 0;
1243 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1247 page = lru_to_page(src);
1248 prefetchw_prev_lru_page(page, src, flags);
1250 VM_BUG_ON(!PageLRU(page));
1252 switch (__isolate_lru_page(page, mode)) {
1254 nr_pages = hpage_nr_pages(page);
1255 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1256 list_move(&page->lru, dst);
1257 nr_taken += nr_pages;
1261 /* else it is being freed elsewhere */
1262 list_move(&page->lru, src);
1271 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1272 nr_taken, mode, is_file_lru(lru));
1277 * isolate_lru_page - tries to isolate a page from its LRU list
1278 * @page: page to isolate from its LRU list
1280 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1281 * vmstat statistic corresponding to whatever LRU list the page was on.
1283 * Returns 0 if the page was removed from an LRU list.
1284 * Returns -EBUSY if the page was not on an LRU list.
1286 * The returned page will have PageLRU() cleared. If it was found on
1287 * the active list, it will have PageActive set. If it was found on
1288 * the unevictable list, it will have the PageUnevictable bit set. That flag
1289 * may need to be cleared by the caller before letting the page go.
1291 * The vmstat statistic corresponding to the list on which the page was
1292 * found will be decremented.
1295 * (1) Must be called with an elevated refcount on the page. This is a
1296 * fundamentnal difference from isolate_lru_pages (which is called
1297 * without a stable reference).
1298 * (2) the lru_lock must not be held.
1299 * (3) interrupts must be enabled.
1301 int isolate_lru_page(struct page *page)
1305 VM_BUG_ON(!page_count(page));
1307 if (PageLRU(page)) {
1308 struct zone *zone = page_zone(page);
1309 struct lruvec *lruvec;
1311 spin_lock_irq(&zone->lru_lock);
1312 lruvec = mem_cgroup_page_lruvec(page, zone);
1313 if (PageLRU(page)) {
1314 int lru = page_lru(page);
1317 del_page_from_lru_list(page, lruvec, lru);
1320 spin_unlock_irq(&zone->lru_lock);
1326 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1327 * then get resheduled. When there are massive number of tasks doing page
1328 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1329 * the LRU list will go small and be scanned faster than necessary, leading to
1330 * unnecessary swapping, thrashing and OOM.
1332 static int too_many_isolated(struct zone *zone, int file,
1333 struct scan_control *sc)
1335 unsigned long inactive, isolated;
1337 if (current_is_kswapd())
1340 if (!global_reclaim(sc))
1344 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1345 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1347 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1348 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1352 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1353 * won't get blocked by normal direct-reclaimers, forming a circular
1356 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1359 return isolated > inactive;
1362 static noinline_for_stack void
1363 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1365 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1366 struct zone *zone = lruvec_zone(lruvec);
1367 LIST_HEAD(pages_to_free);
1370 * Put back any unfreeable pages.
1372 while (!list_empty(page_list)) {
1373 struct page *page = lru_to_page(page_list);
1376 VM_BUG_ON(PageLRU(page));
1377 list_del(&page->lru);
1378 if (unlikely(!page_evictable(page))) {
1379 spin_unlock_irq(&zone->lru_lock);
1380 putback_lru_page(page);
1381 spin_lock_irq(&zone->lru_lock);
1385 lruvec = mem_cgroup_page_lruvec(page, zone);
1388 lru = page_lru(page);
1389 add_page_to_lru_list(page, lruvec, lru);
1391 if (is_active_lru(lru)) {
1392 int file = is_file_lru(lru);
1393 int numpages = hpage_nr_pages(page);
1394 reclaim_stat->recent_rotated[file] += numpages;
1396 if (put_page_testzero(page)) {
1397 __ClearPageLRU(page);
1398 __ClearPageActive(page);
1399 del_page_from_lru_list(page, lruvec, lru);
1401 if (unlikely(PageCompound(page))) {
1402 spin_unlock_irq(&zone->lru_lock);
1403 (*get_compound_page_dtor(page))(page);
1404 spin_lock_irq(&zone->lru_lock);
1406 list_add(&page->lru, &pages_to_free);
1411 * To save our caller's stack, now use input list for pages to free.
1413 list_splice(&pages_to_free, page_list);
1417 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1418 * of reclaimed pages
1420 static noinline_for_stack unsigned long
1421 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1422 struct scan_control *sc, enum lru_list lru)
1424 LIST_HEAD(page_list);
1425 unsigned long nr_scanned;
1426 unsigned long nr_reclaimed = 0;
1427 unsigned long nr_taken;
1428 unsigned long nr_dirty = 0;
1429 unsigned long nr_congested = 0;
1430 unsigned long nr_unqueued_dirty = 0;
1431 unsigned long nr_writeback = 0;
1432 unsigned long nr_immediate = 0;
1433 isolate_mode_t isolate_mode = 0;
1434 int file = is_file_lru(lru);
1435 struct zone *zone = lruvec_zone(lruvec);
1436 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1438 while (unlikely(too_many_isolated(zone, file, sc))) {
1439 congestion_wait(BLK_RW_ASYNC, HZ/10);
1441 /* We are about to die and free our memory. Return now. */
1442 if (fatal_signal_pending(current))
1443 return SWAP_CLUSTER_MAX;
1449 isolate_mode |= ISOLATE_UNMAPPED;
1450 if (!sc->may_writepage)
1451 isolate_mode |= ISOLATE_CLEAN;
1453 spin_lock_irq(&zone->lru_lock);
1455 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1456 &nr_scanned, sc, isolate_mode, lru);
1458 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1459 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1461 if (global_reclaim(sc)) {
1462 zone->pages_scanned += nr_scanned;
1463 if (current_is_kswapd())
1464 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1466 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1468 spin_unlock_irq(&zone->lru_lock);
1473 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1474 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1475 &nr_writeback, &nr_immediate,
1478 spin_lock_irq(&zone->lru_lock);
1480 reclaim_stat->recent_scanned[file] += nr_taken;
1482 if (global_reclaim(sc)) {
1483 if (current_is_kswapd())
1484 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1487 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1491 putback_inactive_pages(lruvec, &page_list);
1493 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1495 spin_unlock_irq(&zone->lru_lock);
1497 free_hot_cold_page_list(&page_list, 1);
1500 * If reclaim is isolating dirty pages under writeback, it implies
1501 * that the long-lived page allocation rate is exceeding the page
1502 * laundering rate. Either the global limits are not being effective
1503 * at throttling processes due to the page distribution throughout
1504 * zones or there is heavy usage of a slow backing device. The
1505 * only option is to throttle from reclaim context which is not ideal
1506 * as there is no guarantee the dirtying process is throttled in the
1507 * same way balance_dirty_pages() manages.
1509 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1510 * of pages under pages flagged for immediate reclaim and stall if any
1511 * are encountered in the nr_immediate check below.
1513 if (nr_writeback && nr_writeback == nr_taken)
1514 zone_set_flag(zone, ZONE_WRITEBACK);
1517 * memcg will stall in page writeback so only consider forcibly
1518 * stalling for global reclaim
1520 if (global_reclaim(sc)) {
1522 * Tag a zone as congested if all the dirty pages scanned were
1523 * backed by a congested BDI and wait_iff_congested will stall.
1525 if (nr_dirty && nr_dirty == nr_congested)
1526 zone_set_flag(zone, ZONE_CONGESTED);
1529 * If dirty pages are scanned that are not queued for IO, it
1530 * implies that flushers are not keeping up. In this case, flag
1531 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1532 * pages from reclaim context. It will forcibly stall in the
1535 if (nr_unqueued_dirty == nr_taken)
1536 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1539 * In addition, if kswapd scans pages marked marked for
1540 * immediate reclaim and under writeback (nr_immediate), it
1541 * implies that pages are cycling through the LRU faster than
1542 * they are written so also forcibly stall.
1544 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1545 congestion_wait(BLK_RW_ASYNC, HZ/10);
1549 * Stall direct reclaim for IO completions if underlying BDIs or zone
1550 * is congested. Allow kswapd to continue until it starts encountering
1551 * unqueued dirty pages or cycling through the LRU too quickly.
1553 if (!sc->hibernation_mode && !current_is_kswapd())
1554 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1556 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1558 nr_scanned, nr_reclaimed,
1560 trace_shrink_flags(file));
1561 return nr_reclaimed;
1565 * This moves pages from the active list to the inactive list.
1567 * We move them the other way if the page is referenced by one or more
1568 * processes, from rmap.
1570 * If the pages are mostly unmapped, the processing is fast and it is
1571 * appropriate to hold zone->lru_lock across the whole operation. But if
1572 * the pages are mapped, the processing is slow (page_referenced()) so we
1573 * should drop zone->lru_lock around each page. It's impossible to balance
1574 * this, so instead we remove the pages from the LRU while processing them.
1575 * It is safe to rely on PG_active against the non-LRU pages in here because
1576 * nobody will play with that bit on a non-LRU page.
1578 * The downside is that we have to touch page->_count against each page.
1579 * But we had to alter page->flags anyway.
1582 static void move_active_pages_to_lru(struct lruvec *lruvec,
1583 struct list_head *list,
1584 struct list_head *pages_to_free,
1587 struct zone *zone = lruvec_zone(lruvec);
1588 unsigned long pgmoved = 0;
1592 while (!list_empty(list)) {
1593 page = lru_to_page(list);
1594 lruvec = mem_cgroup_page_lruvec(page, zone);
1596 VM_BUG_ON(PageLRU(page));
1599 nr_pages = hpage_nr_pages(page);
1600 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1601 list_move(&page->lru, &lruvec->lists[lru]);
1602 pgmoved += nr_pages;
1604 if (put_page_testzero(page)) {
1605 __ClearPageLRU(page);
1606 __ClearPageActive(page);
1607 del_page_from_lru_list(page, lruvec, lru);
1609 if (unlikely(PageCompound(page))) {
1610 spin_unlock_irq(&zone->lru_lock);
1611 (*get_compound_page_dtor(page))(page);
1612 spin_lock_irq(&zone->lru_lock);
1614 list_add(&page->lru, pages_to_free);
1617 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1618 if (!is_active_lru(lru))
1619 __count_vm_events(PGDEACTIVATE, pgmoved);
1622 static void shrink_active_list(unsigned long nr_to_scan,
1623 struct lruvec *lruvec,
1624 struct scan_control *sc,
1627 unsigned long nr_taken;
1628 unsigned long nr_scanned;
1629 unsigned long vm_flags;
1630 LIST_HEAD(l_hold); /* The pages which were snipped off */
1631 LIST_HEAD(l_active);
1632 LIST_HEAD(l_inactive);
1634 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1635 unsigned long nr_rotated = 0;
1636 isolate_mode_t isolate_mode = 0;
1637 int file = is_file_lru(lru);
1638 struct zone *zone = lruvec_zone(lruvec);
1643 isolate_mode |= ISOLATE_UNMAPPED;
1644 if (!sc->may_writepage)
1645 isolate_mode |= ISOLATE_CLEAN;
1647 spin_lock_irq(&zone->lru_lock);
1649 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1650 &nr_scanned, sc, isolate_mode, lru);
1651 if (global_reclaim(sc))
1652 zone->pages_scanned += nr_scanned;
1654 reclaim_stat->recent_scanned[file] += nr_taken;
1656 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1657 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1658 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1659 spin_unlock_irq(&zone->lru_lock);
1661 while (!list_empty(&l_hold)) {
1663 page = lru_to_page(&l_hold);
1664 list_del(&page->lru);
1666 if (unlikely(!page_evictable(page))) {
1667 putback_lru_page(page);
1671 if (unlikely(buffer_heads_over_limit)) {
1672 if (page_has_private(page) && trylock_page(page)) {
1673 if (page_has_private(page))
1674 try_to_release_page(page, 0);
1679 if (page_referenced(page, 0, sc->target_mem_cgroup,
1681 nr_rotated += hpage_nr_pages(page);
1683 * Identify referenced, file-backed active pages and
1684 * give them one more trip around the active list. So
1685 * that executable code get better chances to stay in
1686 * memory under moderate memory pressure. Anon pages
1687 * are not likely to be evicted by use-once streaming
1688 * IO, plus JVM can create lots of anon VM_EXEC pages,
1689 * so we ignore them here.
1691 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1692 list_add(&page->lru, &l_active);
1697 ClearPageActive(page); /* we are de-activating */
1698 list_add(&page->lru, &l_inactive);
1702 * Move pages back to the lru list.
1704 spin_lock_irq(&zone->lru_lock);
1706 * Count referenced pages from currently used mappings as rotated,
1707 * even though only some of them are actually re-activated. This
1708 * helps balance scan pressure between file and anonymous pages in
1711 reclaim_stat->recent_rotated[file] += nr_rotated;
1713 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1714 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1715 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1716 spin_unlock_irq(&zone->lru_lock);
1718 free_hot_cold_page_list(&l_hold, 1);
1722 static int inactive_anon_is_low_global(struct zone *zone)
1724 unsigned long active, inactive;
1726 active = zone_page_state(zone, NR_ACTIVE_ANON);
1727 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1729 if (inactive * zone->inactive_ratio < active)
1736 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1737 * @lruvec: LRU vector to check
1739 * Returns true if the zone does not have enough inactive anon pages,
1740 * meaning some active anon pages need to be deactivated.
1742 static int inactive_anon_is_low(struct lruvec *lruvec)
1745 * If we don't have swap space, anonymous page deactivation
1748 if (!total_swap_pages)
1751 if (!mem_cgroup_disabled())
1752 return mem_cgroup_inactive_anon_is_low(lruvec);
1754 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1757 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1764 * inactive_file_is_low - check if file pages need to be deactivated
1765 * @lruvec: LRU vector to check
1767 * When the system is doing streaming IO, memory pressure here
1768 * ensures that active file pages get deactivated, until more
1769 * than half of the file pages are on the inactive list.
1771 * Once we get to that situation, protect the system's working
1772 * set from being evicted by disabling active file page aging.
1774 * This uses a different ratio than the anonymous pages, because
1775 * the page cache uses a use-once replacement algorithm.
1777 static int inactive_file_is_low(struct lruvec *lruvec)
1779 unsigned long inactive;
1780 unsigned long active;
1782 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1783 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1785 return active > inactive;
1788 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1790 if (is_file_lru(lru))
1791 return inactive_file_is_low(lruvec);
1793 return inactive_anon_is_low(lruvec);
1796 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1797 struct lruvec *lruvec, struct scan_control *sc)
1799 if (is_active_lru(lru)) {
1800 if (inactive_list_is_low(lruvec, lru))
1801 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1805 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1808 static int vmscan_swappiness(struct scan_control *sc)
1810 if (global_reclaim(sc))
1811 return vm_swappiness;
1812 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1823 * Determine how aggressively the anon and file LRU lists should be
1824 * scanned. The relative value of each set of LRU lists is determined
1825 * by looking at the fraction of the pages scanned we did rotate back
1826 * onto the active list instead of evict.
1828 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1829 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1831 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1834 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1836 u64 denominator = 0; /* gcc */
1837 struct zone *zone = lruvec_zone(lruvec);
1838 unsigned long anon_prio, file_prio;
1839 enum scan_balance scan_balance;
1840 unsigned long anon, file, free;
1841 bool force_scan = false;
1842 unsigned long ap, fp;
1846 * If the zone or memcg is small, nr[l] can be 0. This
1847 * results in no scanning on this priority and a potential
1848 * priority drop. Global direct reclaim can go to the next
1849 * zone and tends to have no problems. Global kswapd is for
1850 * zone balancing and it needs to scan a minimum amount. When
1851 * reclaiming for a memcg, a priority drop can cause high
1852 * latencies, so it's better to scan a minimum amount there as
1855 if (current_is_kswapd() && !zone_reclaimable(zone))
1857 if (!global_reclaim(sc))
1860 /* If we have no swap space, do not bother scanning anon pages. */
1861 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1862 scan_balance = SCAN_FILE;
1867 * Global reclaim will swap to prevent OOM even with no
1868 * swappiness, but memcg users want to use this knob to
1869 * disable swapping for individual groups completely when
1870 * using the memory controller's swap limit feature would be
1873 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1874 scan_balance = SCAN_FILE;
1879 * Do not apply any pressure balancing cleverness when the
1880 * system is close to OOM, scan both anon and file equally
1881 * (unless the swappiness setting disagrees with swapping).
1883 if (!sc->priority && vmscan_swappiness(sc)) {
1884 scan_balance = SCAN_EQUAL;
1888 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1889 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1890 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1891 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1894 * If it's foreseeable that reclaiming the file cache won't be
1895 * enough to get the zone back into a desirable shape, we have
1896 * to swap. Better start now and leave the - probably heavily
1897 * thrashing - remaining file pages alone.
1899 if (global_reclaim(sc)) {
1900 free = zone_page_state(zone, NR_FREE_PAGES);
1901 if (unlikely(file + free <= high_wmark_pages(zone))) {
1902 scan_balance = SCAN_ANON;
1908 * There is enough inactive page cache, do not reclaim
1909 * anything from the anonymous working set right now.
1911 if (!inactive_file_is_low(lruvec)) {
1912 scan_balance = SCAN_FILE;
1916 scan_balance = SCAN_FRACT;
1919 * With swappiness at 100, anonymous and file have the same priority.
1920 * This scanning priority is essentially the inverse of IO cost.
1922 anon_prio = vmscan_swappiness(sc);
1923 file_prio = 200 - anon_prio;
1926 * OK, so we have swap space and a fair amount of page cache
1927 * pages. We use the recently rotated / recently scanned
1928 * ratios to determine how valuable each cache is.
1930 * Because workloads change over time (and to avoid overflow)
1931 * we keep these statistics as a floating average, which ends
1932 * up weighing recent references more than old ones.
1934 * anon in [0], file in [1]
1936 spin_lock_irq(&zone->lru_lock);
1937 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1938 reclaim_stat->recent_scanned[0] /= 2;
1939 reclaim_stat->recent_rotated[0] /= 2;
1942 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1943 reclaim_stat->recent_scanned[1] /= 2;
1944 reclaim_stat->recent_rotated[1] /= 2;
1948 * The amount of pressure on anon vs file pages is inversely
1949 * proportional to the fraction of recently scanned pages on
1950 * each list that were recently referenced and in active use.
1952 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1953 ap /= reclaim_stat->recent_rotated[0] + 1;
1955 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1956 fp /= reclaim_stat->recent_rotated[1] + 1;
1957 spin_unlock_irq(&zone->lru_lock);
1961 denominator = ap + fp + 1;
1963 for_each_evictable_lru(lru) {
1964 int file = is_file_lru(lru);
1968 size = get_lru_size(lruvec, lru);
1969 scan = size >> sc->priority;
1971 if (!scan && force_scan)
1972 scan = min(size, SWAP_CLUSTER_MAX);
1974 switch (scan_balance) {
1976 /* Scan lists relative to size */
1980 * Scan types proportional to swappiness and
1981 * their relative recent reclaim efficiency.
1983 scan = div64_u64(scan * fraction[file], denominator);
1987 /* Scan one type exclusively */
1988 if ((scan_balance == SCAN_FILE) != file)
1992 /* Look ma, no brain */
2000 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2002 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2004 unsigned long nr[NR_LRU_LISTS];
2005 unsigned long targets[NR_LRU_LISTS];
2006 unsigned long nr_to_scan;
2008 unsigned long nr_reclaimed = 0;
2009 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2010 struct blk_plug plug;
2011 bool scan_adjusted = false;
2013 get_scan_count(lruvec, sc, nr);
2015 /* Record the original scan target for proportional adjustments later */
2016 memcpy(targets, nr, sizeof(nr));
2018 blk_start_plug(&plug);
2019 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2020 nr[LRU_INACTIVE_FILE]) {
2021 unsigned long nr_anon, nr_file, percentage;
2022 unsigned long nr_scanned;
2024 for_each_evictable_lru(lru) {
2026 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2027 nr[lru] -= nr_to_scan;
2029 nr_reclaimed += shrink_list(lru, nr_to_scan,
2034 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2038 * For global direct reclaim, reclaim only the number of pages
2039 * requested. Less care is taken to scan proportionally as it
2040 * is more important to minimise direct reclaim stall latency
2041 * than it is to properly age the LRU lists.
2043 if (global_reclaim(sc) && !current_is_kswapd())
2047 * For kswapd and memcg, reclaim at least the number of pages
2048 * requested. Ensure that the anon and file LRUs shrink
2049 * proportionally what was requested by get_scan_count(). We
2050 * stop reclaiming one LRU and reduce the amount scanning
2051 * proportional to the original scan target.
2053 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2054 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2056 if (nr_file > nr_anon) {
2057 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2058 targets[LRU_ACTIVE_ANON] + 1;
2060 percentage = nr_anon * 100 / scan_target;
2062 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2063 targets[LRU_ACTIVE_FILE] + 1;
2065 percentage = nr_file * 100 / scan_target;
2068 /* Stop scanning the smaller of the LRU */
2070 nr[lru + LRU_ACTIVE] = 0;
2073 * Recalculate the other LRU scan count based on its original
2074 * scan target and the percentage scanning already complete
2076 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2077 nr_scanned = targets[lru] - nr[lru];
2078 nr[lru] = targets[lru] * (100 - percentage) / 100;
2079 nr[lru] -= min(nr[lru], nr_scanned);
2082 nr_scanned = targets[lru] - nr[lru];
2083 nr[lru] = targets[lru] * (100 - percentage) / 100;
2084 nr[lru] -= min(nr[lru], nr_scanned);
2086 scan_adjusted = true;
2088 blk_finish_plug(&plug);
2089 sc->nr_reclaimed += nr_reclaimed;
2092 * Even if we did not try to evict anon pages at all, we want to
2093 * rebalance the anon lru active/inactive ratio.
2095 if (inactive_anon_is_low(lruvec))
2096 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2097 sc, LRU_ACTIVE_ANON);
2099 throttle_vm_writeout(sc->gfp_mask);
2102 /* Use reclaim/compaction for costly allocs or under memory pressure */
2103 static bool in_reclaim_compaction(struct scan_control *sc)
2105 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2106 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2107 sc->priority < DEF_PRIORITY - 2))
2114 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2115 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2116 * true if more pages should be reclaimed such that when the page allocator
2117 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2118 * It will give up earlier than that if there is difficulty reclaiming pages.
2120 static inline bool should_continue_reclaim(struct zone *zone,
2121 unsigned long nr_reclaimed,
2122 unsigned long nr_scanned,
2123 struct scan_control *sc)
2125 unsigned long pages_for_compaction;
2126 unsigned long inactive_lru_pages;
2128 /* If not in reclaim/compaction mode, stop */
2129 if (!in_reclaim_compaction(sc))
2132 /* Consider stopping depending on scan and reclaim activity */
2133 if (sc->gfp_mask & __GFP_REPEAT) {
2135 * For __GFP_REPEAT allocations, stop reclaiming if the
2136 * full LRU list has been scanned and we are still failing
2137 * to reclaim pages. This full LRU scan is potentially
2138 * expensive but a __GFP_REPEAT caller really wants to succeed
2140 if (!nr_reclaimed && !nr_scanned)
2144 * For non-__GFP_REPEAT allocations which can presumably
2145 * fail without consequence, stop if we failed to reclaim
2146 * any pages from the last SWAP_CLUSTER_MAX number of
2147 * pages that were scanned. This will return to the
2148 * caller faster at the risk reclaim/compaction and
2149 * the resulting allocation attempt fails
2156 * If we have not reclaimed enough pages for compaction and the
2157 * inactive lists are large enough, continue reclaiming
2159 pages_for_compaction = (2UL << sc->order);
2160 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2161 if (get_nr_swap_pages() > 0)
2162 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2163 if (sc->nr_reclaimed < pages_for_compaction &&
2164 inactive_lru_pages > pages_for_compaction)
2167 /* If compaction would go ahead or the allocation would succeed, stop */
2168 switch (compaction_suitable(zone, sc->order)) {
2169 case COMPACT_PARTIAL:
2170 case COMPACT_CONTINUE:
2178 __shrink_zone(struct zone *zone, struct scan_control *sc, bool soft_reclaim)
2180 unsigned long nr_reclaimed, nr_scanned;
2183 struct mem_cgroup *root = sc->target_mem_cgroup;
2184 struct mem_cgroup_reclaim_cookie reclaim = {
2186 .priority = sc->priority,
2188 struct mem_cgroup *memcg;
2190 nr_reclaimed = sc->nr_reclaimed;
2191 nr_scanned = sc->nr_scanned;
2193 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2195 struct lruvec *lruvec;
2198 !mem_cgroup_soft_reclaim_eligible(memcg)) {
2199 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2203 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2205 shrink_lruvec(lruvec, sc);
2208 * Direct reclaim and kswapd have to scan all memory
2209 * cgroups to fulfill the overall scan target for the
2212 * Limit reclaim, on the other hand, only cares about
2213 * nr_to_reclaim pages to be reclaimed and it will
2214 * retry with decreasing priority if one round over the
2215 * whole hierarchy is not sufficient.
2217 if (!global_reclaim(sc) &&
2218 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2219 mem_cgroup_iter_break(root, memcg);
2222 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2225 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2226 sc->nr_scanned - nr_scanned,
2227 sc->nr_reclaimed - nr_reclaimed);
2229 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2230 sc->nr_scanned - nr_scanned, sc));
2234 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2236 bool do_soft_reclaim = mem_cgroup_should_soft_reclaim(sc);
2237 unsigned long nr_scanned = sc->nr_scanned;
2239 __shrink_zone(zone, sc, do_soft_reclaim);
2242 * No group is over the soft limit or those that are do not have
2243 * pages in the zone we are reclaiming so we have to reclaim everybody
2245 if (do_soft_reclaim && (sc->nr_scanned == nr_scanned)) {
2246 __shrink_zone(zone, sc, false);
2251 /* Returns true if compaction should go ahead for a high-order request */
2252 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2254 unsigned long balance_gap, watermark;
2257 /* Do not consider compaction for orders reclaim is meant to satisfy */
2258 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2262 * Compaction takes time to run and there are potentially other
2263 * callers using the pages just freed. Continue reclaiming until
2264 * there is a buffer of free pages available to give compaction
2265 * a reasonable chance of completing and allocating the page
2267 balance_gap = min(low_wmark_pages(zone),
2268 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2269 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2270 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2271 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2274 * If compaction is deferred, reclaim up to a point where
2275 * compaction will have a chance of success when re-enabled
2277 if (compaction_deferred(zone, sc->order))
2278 return watermark_ok;
2280 /* If compaction is not ready to start, keep reclaiming */
2281 if (!compaction_suitable(zone, sc->order))
2284 return watermark_ok;
2288 * This is the direct reclaim path, for page-allocating processes. We only
2289 * try to reclaim pages from zones which will satisfy the caller's allocation
2292 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2294 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2296 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2297 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2298 * zone defense algorithm.
2300 * If a zone is deemed to be full of pinned pages then just give it a light
2301 * scan then give up on it.
2303 * This function returns true if a zone is being reclaimed for a costly
2304 * high-order allocation and compaction is ready to begin. This indicates to
2305 * the caller that it should consider retrying the allocation instead of
2308 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2312 bool aborted_reclaim = false;
2315 * If the number of buffer_heads in the machine exceeds the maximum
2316 * allowed level, force direct reclaim to scan the highmem zone as
2317 * highmem pages could be pinning lowmem pages storing buffer_heads
2319 if (buffer_heads_over_limit)
2320 sc->gfp_mask |= __GFP_HIGHMEM;
2322 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2323 gfp_zone(sc->gfp_mask), sc->nodemask) {
2324 if (!populated_zone(zone))
2327 * Take care memory controller reclaiming has small influence
2330 if (global_reclaim(sc)) {
2331 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2333 if (sc->priority != DEF_PRIORITY &&
2334 !zone_reclaimable(zone))
2335 continue; /* Let kswapd poll it */
2336 if (IS_ENABLED(CONFIG_COMPACTION)) {
2338 * If we already have plenty of memory free for
2339 * compaction in this zone, don't free any more.
2340 * Even though compaction is invoked for any
2341 * non-zero order, only frequent costly order
2342 * reclamation is disruptive enough to become a
2343 * noticeable problem, like transparent huge
2346 if (compaction_ready(zone, sc)) {
2347 aborted_reclaim = true;
2351 /* need some check for avoid more shrink_zone() */
2354 shrink_zone(zone, sc);
2357 return aborted_reclaim;
2360 /* All zones in zonelist are unreclaimable? */
2361 static bool all_unreclaimable(struct zonelist *zonelist,
2362 struct scan_control *sc)
2367 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2368 gfp_zone(sc->gfp_mask), sc->nodemask) {
2369 if (!populated_zone(zone))
2371 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2373 if (zone_reclaimable(zone))
2381 * This is the main entry point to direct page reclaim.
2383 * If a full scan of the inactive list fails to free enough memory then we
2384 * are "out of memory" and something needs to be killed.
2386 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2387 * high - the zone may be full of dirty or under-writeback pages, which this
2388 * caller can't do much about. We kick the writeback threads and take explicit
2389 * naps in the hope that some of these pages can be written. But if the
2390 * allocating task holds filesystem locks which prevent writeout this might not
2391 * work, and the allocation attempt will fail.
2393 * returns: 0, if no pages reclaimed
2394 * else, the number of pages reclaimed
2396 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2397 struct scan_control *sc,
2398 struct shrink_control *shrink)
2400 unsigned long total_scanned = 0;
2401 struct reclaim_state *reclaim_state = current->reclaim_state;
2404 unsigned long writeback_threshold;
2405 bool aborted_reclaim;
2407 delayacct_freepages_start();
2409 if (global_reclaim(sc))
2410 count_vm_event(ALLOCSTALL);
2413 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2416 aborted_reclaim = shrink_zones(zonelist, sc);
2419 * Don't shrink slabs when reclaiming memory from over limit
2420 * cgroups but do shrink slab at least once when aborting
2421 * reclaim for compaction to avoid unevenly scanning file/anon
2422 * LRU pages over slab pages.
2424 if (global_reclaim(sc)) {
2425 unsigned long lru_pages = 0;
2427 nodes_clear(shrink->nodes_to_scan);
2428 for_each_zone_zonelist(zone, z, zonelist,
2429 gfp_zone(sc->gfp_mask)) {
2430 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2433 lru_pages += zone_reclaimable_pages(zone);
2434 node_set(zone_to_nid(zone),
2435 shrink->nodes_to_scan);
2438 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2439 if (reclaim_state) {
2440 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2441 reclaim_state->reclaimed_slab = 0;
2444 total_scanned += sc->nr_scanned;
2445 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2449 * If we're getting trouble reclaiming, start doing
2450 * writepage even in laptop mode.
2452 if (sc->priority < DEF_PRIORITY - 2)
2453 sc->may_writepage = 1;
2456 * Try to write back as many pages as we just scanned. This
2457 * tends to cause slow streaming writers to write data to the
2458 * disk smoothly, at the dirtying rate, which is nice. But
2459 * that's undesirable in laptop mode, where we *want* lumpy
2460 * writeout. So in laptop mode, write out the whole world.
2462 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2463 if (total_scanned > writeback_threshold) {
2464 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2465 WB_REASON_TRY_TO_FREE_PAGES);
2466 sc->may_writepage = 1;
2468 } while (--sc->priority >= 0 && !aborted_reclaim);
2471 delayacct_freepages_end();
2473 if (sc->nr_reclaimed)
2474 return sc->nr_reclaimed;
2477 * As hibernation is going on, kswapd is freezed so that it can't mark
2478 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2481 if (oom_killer_disabled)
2484 /* Aborted reclaim to try compaction? don't OOM, then */
2485 if (aborted_reclaim)
2488 /* top priority shrink_zones still had more to do? don't OOM, then */
2489 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2495 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2498 unsigned long pfmemalloc_reserve = 0;
2499 unsigned long free_pages = 0;
2503 for (i = 0; i <= ZONE_NORMAL; i++) {
2504 zone = &pgdat->node_zones[i];
2505 pfmemalloc_reserve += min_wmark_pages(zone);
2506 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2509 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2511 /* kswapd must be awake if processes are being throttled */
2512 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2513 pgdat->classzone_idx = min(pgdat->classzone_idx,
2514 (enum zone_type)ZONE_NORMAL);
2515 wake_up_interruptible(&pgdat->kswapd_wait);
2522 * Throttle direct reclaimers if backing storage is backed by the network
2523 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2524 * depleted. kswapd will continue to make progress and wake the processes
2525 * when the low watermark is reached.
2527 * Returns true if a fatal signal was delivered during throttling. If this
2528 * happens, the page allocator should not consider triggering the OOM killer.
2530 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2531 nodemask_t *nodemask)
2534 int high_zoneidx = gfp_zone(gfp_mask);
2538 * Kernel threads should not be throttled as they may be indirectly
2539 * responsible for cleaning pages necessary for reclaim to make forward
2540 * progress. kjournald for example may enter direct reclaim while
2541 * committing a transaction where throttling it could forcing other
2542 * processes to block on log_wait_commit().
2544 if (current->flags & PF_KTHREAD)
2548 * If a fatal signal is pending, this process should not throttle.
2549 * It should return quickly so it can exit and free its memory
2551 if (fatal_signal_pending(current))
2554 /* Check if the pfmemalloc reserves are ok */
2555 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2556 pgdat = zone->zone_pgdat;
2557 if (pfmemalloc_watermark_ok(pgdat))
2560 /* Account for the throttling */
2561 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2564 * If the caller cannot enter the filesystem, it's possible that it
2565 * is due to the caller holding an FS lock or performing a journal
2566 * transaction in the case of a filesystem like ext[3|4]. In this case,
2567 * it is not safe to block on pfmemalloc_wait as kswapd could be
2568 * blocked waiting on the same lock. Instead, throttle for up to a
2569 * second before continuing.
2571 if (!(gfp_mask & __GFP_FS)) {
2572 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2573 pfmemalloc_watermark_ok(pgdat), HZ);
2578 /* Throttle until kswapd wakes the process */
2579 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2580 pfmemalloc_watermark_ok(pgdat));
2583 if (fatal_signal_pending(current))
2590 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2591 gfp_t gfp_mask, nodemask_t *nodemask)
2593 unsigned long nr_reclaimed;
2594 struct scan_control sc = {
2595 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2596 .may_writepage = !laptop_mode,
2597 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2601 .priority = DEF_PRIORITY,
2602 .target_mem_cgroup = NULL,
2603 .nodemask = nodemask,
2605 struct shrink_control shrink = {
2606 .gfp_mask = sc.gfp_mask,
2610 * Do not enter reclaim if fatal signal was delivered while throttled.
2611 * 1 is returned so that the page allocator does not OOM kill at this
2614 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2617 trace_mm_vmscan_direct_reclaim_begin(order,
2621 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2623 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2625 return nr_reclaimed;
2630 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2631 gfp_t gfp_mask, bool noswap,
2633 unsigned long *nr_scanned)
2635 struct scan_control sc = {
2637 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2638 .may_writepage = !laptop_mode,
2640 .may_swap = !noswap,
2643 .target_mem_cgroup = memcg,
2645 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2647 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2648 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2650 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2655 * NOTE: Although we can get the priority field, using it
2656 * here is not a good idea, since it limits the pages we can scan.
2657 * if we don't reclaim here, the shrink_zone from balance_pgdat
2658 * will pick up pages from other mem cgroup's as well. We hack
2659 * the priority and make it zero.
2661 shrink_lruvec(lruvec, &sc);
2663 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2665 *nr_scanned = sc.nr_scanned;
2666 return sc.nr_reclaimed;
2669 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2673 struct zonelist *zonelist;
2674 unsigned long nr_reclaimed;
2676 struct scan_control sc = {
2677 .may_writepage = !laptop_mode,
2679 .may_swap = !noswap,
2680 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2682 .priority = DEF_PRIORITY,
2683 .target_mem_cgroup = memcg,
2684 .nodemask = NULL, /* we don't care the placement */
2685 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2686 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2688 struct shrink_control shrink = {
2689 .gfp_mask = sc.gfp_mask,
2693 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2694 * take care of from where we get pages. So the node where we start the
2695 * scan does not need to be the current node.
2697 nid = mem_cgroup_select_victim_node(memcg);
2699 zonelist = NODE_DATA(nid)->node_zonelists;
2701 trace_mm_vmscan_memcg_reclaim_begin(0,
2705 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2707 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2709 return nr_reclaimed;
2713 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2715 struct mem_cgroup *memcg;
2717 if (!total_swap_pages)
2720 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2722 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2724 if (inactive_anon_is_low(lruvec))
2725 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2726 sc, LRU_ACTIVE_ANON);
2728 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2732 static bool zone_balanced(struct zone *zone, int order,
2733 unsigned long balance_gap, int classzone_idx)
2735 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2736 balance_gap, classzone_idx, 0))
2739 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2740 !compaction_suitable(zone, order))
2747 * pgdat_balanced() is used when checking if a node is balanced.
2749 * For order-0, all zones must be balanced!
2751 * For high-order allocations only zones that meet watermarks and are in a
2752 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2753 * total of balanced pages must be at least 25% of the zones allowed by
2754 * classzone_idx for the node to be considered balanced. Forcing all zones to
2755 * be balanced for high orders can cause excessive reclaim when there are
2757 * The choice of 25% is due to
2758 * o a 16M DMA zone that is balanced will not balance a zone on any
2759 * reasonable sized machine
2760 * o On all other machines, the top zone must be at least a reasonable
2761 * percentage of the middle zones. For example, on 32-bit x86, highmem
2762 * would need to be at least 256M for it to be balance a whole node.
2763 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2764 * to balance a node on its own. These seemed like reasonable ratios.
2766 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2768 unsigned long managed_pages = 0;
2769 unsigned long balanced_pages = 0;
2772 /* Check the watermark levels */
2773 for (i = 0; i <= classzone_idx; i++) {
2774 struct zone *zone = pgdat->node_zones + i;
2776 if (!populated_zone(zone))
2779 managed_pages += zone->managed_pages;
2782 * A special case here:
2784 * balance_pgdat() skips over all_unreclaimable after
2785 * DEF_PRIORITY. Effectively, it considers them balanced so
2786 * they must be considered balanced here as well!
2788 if (!zone_reclaimable(zone)) {
2789 balanced_pages += zone->managed_pages;
2793 if (zone_balanced(zone, order, 0, i))
2794 balanced_pages += zone->managed_pages;
2800 return balanced_pages >= (managed_pages >> 2);
2806 * Prepare kswapd for sleeping. This verifies that there are no processes
2807 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2809 * Returns true if kswapd is ready to sleep
2811 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2814 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2819 * There is a potential race between when kswapd checks its watermarks
2820 * and a process gets throttled. There is also a potential race if
2821 * processes get throttled, kswapd wakes, a large process exits therby
2822 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2823 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2824 * so wake them now if necessary. If necessary, processes will wake
2825 * kswapd and get throttled again
2827 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2828 wake_up(&pgdat->pfmemalloc_wait);
2832 return pgdat_balanced(pgdat, order, classzone_idx);
2836 * kswapd shrinks the zone by the number of pages required to reach
2837 * the high watermark.
2839 * Returns true if kswapd scanned at least the requested number of pages to
2840 * reclaim or if the lack of progress was due to pages under writeback.
2841 * This is used to determine if the scanning priority needs to be raised.
2843 static bool kswapd_shrink_zone(struct zone *zone,
2845 struct scan_control *sc,
2846 unsigned long lru_pages,
2847 unsigned long *nr_attempted)
2849 int testorder = sc->order;
2850 unsigned long balance_gap;
2851 struct reclaim_state *reclaim_state = current->reclaim_state;
2852 struct shrink_control shrink = {
2853 .gfp_mask = sc->gfp_mask,
2855 bool lowmem_pressure;
2857 /* Reclaim above the high watermark. */
2858 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2861 * Kswapd reclaims only single pages with compaction enabled. Trying
2862 * too hard to reclaim until contiguous free pages have become
2863 * available can hurt performance by evicting too much useful data
2864 * from memory. Do not reclaim more than needed for compaction.
2866 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2867 compaction_suitable(zone, sc->order) !=
2872 * We put equal pressure on every zone, unless one zone has way too
2873 * many pages free already. The "too many pages" is defined as the
2874 * high wmark plus a "gap" where the gap is either the low
2875 * watermark or 1% of the zone, whichever is smaller.
2877 balance_gap = min(low_wmark_pages(zone),
2878 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2879 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2882 * If there is no low memory pressure or the zone is balanced then no
2883 * reclaim is necessary
2885 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2886 if (!lowmem_pressure && zone_balanced(zone, testorder,
2887 balance_gap, classzone_idx))
2890 shrink_zone(zone, sc);
2891 nodes_clear(shrink.nodes_to_scan);
2892 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2894 reclaim_state->reclaimed_slab = 0;
2895 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2896 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2898 /* Account for the number of pages attempted to reclaim */
2899 *nr_attempted += sc->nr_to_reclaim;
2901 zone_clear_flag(zone, ZONE_WRITEBACK);
2904 * If a zone reaches its high watermark, consider it to be no longer
2905 * congested. It's possible there are dirty pages backed by congested
2906 * BDIs but as pressure is relieved, speculatively avoid congestion
2909 if (zone_reclaimable(zone) &&
2910 zone_balanced(zone, testorder, 0, classzone_idx)) {
2911 zone_clear_flag(zone, ZONE_CONGESTED);
2912 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2915 return sc->nr_scanned >= sc->nr_to_reclaim;
2919 * For kswapd, balance_pgdat() will work across all this node's zones until
2920 * they are all at high_wmark_pages(zone).
2922 * Returns the final order kswapd was reclaiming at
2924 * There is special handling here for zones which are full of pinned pages.
2925 * This can happen if the pages are all mlocked, or if they are all used by
2926 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2927 * What we do is to detect the case where all pages in the zone have been
2928 * scanned twice and there has been zero successful reclaim. Mark the zone as
2929 * dead and from now on, only perform a short scan. Basically we're polling
2930 * the zone for when the problem goes away.
2932 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2933 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2934 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2935 * lower zones regardless of the number of free pages in the lower zones. This
2936 * interoperates with the page allocator fallback scheme to ensure that aging
2937 * of pages is balanced across the zones.
2939 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2943 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2944 struct scan_control sc = {
2945 .gfp_mask = GFP_KERNEL,
2946 .priority = DEF_PRIORITY,
2949 .may_writepage = !laptop_mode,
2951 .target_mem_cgroup = NULL,
2953 count_vm_event(PAGEOUTRUN);
2956 unsigned long lru_pages = 0;
2957 unsigned long nr_attempted = 0;
2958 bool raise_priority = true;
2959 bool pgdat_needs_compaction = (order > 0);
2961 sc.nr_reclaimed = 0;
2964 * Scan in the highmem->dma direction for the highest
2965 * zone which needs scanning
2967 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2968 struct zone *zone = pgdat->node_zones + i;
2970 if (!populated_zone(zone))
2973 if (sc.priority != DEF_PRIORITY &&
2974 !zone_reclaimable(zone))
2978 * Do some background aging of the anon list, to give
2979 * pages a chance to be referenced before reclaiming.
2981 age_active_anon(zone, &sc);
2984 * If the number of buffer_heads in the machine
2985 * exceeds the maximum allowed level and this node
2986 * has a highmem zone, force kswapd to reclaim from
2987 * it to relieve lowmem pressure.
2989 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2994 if (!zone_balanced(zone, order, 0, 0)) {
2999 * If balanced, clear the dirty and congested
3002 zone_clear_flag(zone, ZONE_CONGESTED);
3003 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3010 for (i = 0; i <= end_zone; i++) {
3011 struct zone *zone = pgdat->node_zones + i;
3013 if (!populated_zone(zone))
3016 lru_pages += zone_reclaimable_pages(zone);
3019 * If any zone is currently balanced then kswapd will
3020 * not call compaction as it is expected that the
3021 * necessary pages are already available.
3023 if (pgdat_needs_compaction &&
3024 zone_watermark_ok(zone, order,
3025 low_wmark_pages(zone),
3027 pgdat_needs_compaction = false;
3031 * If we're getting trouble reclaiming, start doing writepage
3032 * even in laptop mode.
3034 if (sc.priority < DEF_PRIORITY - 2)
3035 sc.may_writepage = 1;
3038 * Now scan the zone in the dma->highmem direction, stopping
3039 * at the last zone which needs scanning.
3041 * We do this because the page allocator works in the opposite
3042 * direction. This prevents the page allocator from allocating
3043 * pages behind kswapd's direction of progress, which would
3044 * cause too much scanning of the lower zones.
3046 for (i = 0; i <= end_zone; i++) {
3047 struct zone *zone = pgdat->node_zones + i;
3049 if (!populated_zone(zone))
3052 if (sc.priority != DEF_PRIORITY &&
3053 !zone_reclaimable(zone))
3059 * There should be no need to raise the scanning
3060 * priority if enough pages are already being scanned
3061 * that that high watermark would be met at 100%
3064 if (kswapd_shrink_zone(zone, end_zone, &sc,
3065 lru_pages, &nr_attempted))
3066 raise_priority = false;
3070 * If the low watermark is met there is no need for processes
3071 * to be throttled on pfmemalloc_wait as they should not be
3072 * able to safely make forward progress. Wake them
3074 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3075 pfmemalloc_watermark_ok(pgdat))
3076 wake_up(&pgdat->pfmemalloc_wait);
3079 * Fragmentation may mean that the system cannot be rebalanced
3080 * for high-order allocations in all zones. If twice the
3081 * allocation size has been reclaimed and the zones are still
3082 * not balanced then recheck the watermarks at order-0 to
3083 * prevent kswapd reclaiming excessively. Assume that a
3084 * process requested a high-order can direct reclaim/compact.
3086 if (order && sc.nr_reclaimed >= 2UL << order)
3087 order = sc.order = 0;
3089 /* Check if kswapd should be suspending */
3090 if (try_to_freeze() || kthread_should_stop())
3094 * Compact if necessary and kswapd is reclaiming at least the
3095 * high watermark number of pages as requsted
3097 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3098 compact_pgdat(pgdat, order);
3101 * Raise priority if scanning rate is too low or there was no
3102 * progress in reclaiming pages
3104 if (raise_priority || !sc.nr_reclaimed)
3106 } while (sc.priority >= 1 &&
3107 !pgdat_balanced(pgdat, order, *classzone_idx));
3111 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3112 * makes a decision on the order we were last reclaiming at. However,
3113 * if another caller entered the allocator slow path while kswapd
3114 * was awake, order will remain at the higher level
3116 *classzone_idx = end_zone;
3120 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3125 if (freezing(current) || kthread_should_stop())
3128 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3130 /* Try to sleep for a short interval */
3131 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3132 remaining = schedule_timeout(HZ/10);
3133 finish_wait(&pgdat->kswapd_wait, &wait);
3134 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3138 * After a short sleep, check if it was a premature sleep. If not, then
3139 * go fully to sleep until explicitly woken up.
3141 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3142 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3145 * vmstat counters are not perfectly accurate and the estimated
3146 * value for counters such as NR_FREE_PAGES can deviate from the
3147 * true value by nr_online_cpus * threshold. To avoid the zone
3148 * watermarks being breached while under pressure, we reduce the
3149 * per-cpu vmstat threshold while kswapd is awake and restore
3150 * them before going back to sleep.
3152 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3155 * Compaction records what page blocks it recently failed to
3156 * isolate pages from and skips them in the future scanning.
3157 * When kswapd is going to sleep, it is reasonable to assume
3158 * that pages and compaction may succeed so reset the cache.
3160 reset_isolation_suitable(pgdat);
3162 if (!kthread_should_stop())
3165 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3168 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3170 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3172 finish_wait(&pgdat->kswapd_wait, &wait);
3176 * The background pageout daemon, started as a kernel thread
3177 * from the init process.
3179 * This basically trickles out pages so that we have _some_
3180 * free memory available even if there is no other activity
3181 * that frees anything up. This is needed for things like routing
3182 * etc, where we otherwise might have all activity going on in
3183 * asynchronous contexts that cannot page things out.
3185 * If there are applications that are active memory-allocators
3186 * (most normal use), this basically shouldn't matter.
3188 static int kswapd(void *p)
3190 unsigned long order, new_order;
3191 unsigned balanced_order;
3192 int classzone_idx, new_classzone_idx;
3193 int balanced_classzone_idx;
3194 pg_data_t *pgdat = (pg_data_t*)p;
3195 struct task_struct *tsk = current;
3197 struct reclaim_state reclaim_state = {
3198 .reclaimed_slab = 0,
3200 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3202 lockdep_set_current_reclaim_state(GFP_KERNEL);
3204 if (!cpumask_empty(cpumask))
3205 set_cpus_allowed_ptr(tsk, cpumask);
3206 current->reclaim_state = &reclaim_state;
3209 * Tell the memory management that we're a "memory allocator",
3210 * and that if we need more memory we should get access to it
3211 * regardless (see "__alloc_pages()"). "kswapd" should
3212 * never get caught in the normal page freeing logic.
3214 * (Kswapd normally doesn't need memory anyway, but sometimes
3215 * you need a small amount of memory in order to be able to
3216 * page out something else, and this flag essentially protects
3217 * us from recursively trying to free more memory as we're
3218 * trying to free the first piece of memory in the first place).
3220 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3223 order = new_order = 0;
3225 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3226 balanced_classzone_idx = classzone_idx;
3231 * If the last balance_pgdat was unsuccessful it's unlikely a
3232 * new request of a similar or harder type will succeed soon
3233 * so consider going to sleep on the basis we reclaimed at
3235 if (balanced_classzone_idx >= new_classzone_idx &&
3236 balanced_order == new_order) {
3237 new_order = pgdat->kswapd_max_order;
3238 new_classzone_idx = pgdat->classzone_idx;
3239 pgdat->kswapd_max_order = 0;
3240 pgdat->classzone_idx = pgdat->nr_zones - 1;
3243 if (order < new_order || classzone_idx > new_classzone_idx) {
3245 * Don't sleep if someone wants a larger 'order'
3246 * allocation or has tigher zone constraints
3249 classzone_idx = new_classzone_idx;
3251 kswapd_try_to_sleep(pgdat, balanced_order,
3252 balanced_classzone_idx);
3253 order = pgdat->kswapd_max_order;
3254 classzone_idx = pgdat->classzone_idx;
3256 new_classzone_idx = classzone_idx;
3257 pgdat->kswapd_max_order = 0;
3258 pgdat->classzone_idx = pgdat->nr_zones - 1;
3261 ret = try_to_freeze();
3262 if (kthread_should_stop())
3266 * We can speed up thawing tasks if we don't call balance_pgdat
3267 * after returning from the refrigerator
3270 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3271 balanced_classzone_idx = classzone_idx;
3272 balanced_order = balance_pgdat(pgdat, order,
3273 &balanced_classzone_idx);
3277 current->reclaim_state = NULL;
3282 * A zone is low on free memory, so wake its kswapd task to service it.
3284 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3288 if (!populated_zone(zone))
3291 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3293 pgdat = zone->zone_pgdat;
3294 if (pgdat->kswapd_max_order < order) {
3295 pgdat->kswapd_max_order = order;
3296 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3298 if (!waitqueue_active(&pgdat->kswapd_wait))
3300 if (zone_balanced(zone, order, 0, 0))
3303 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3304 wake_up_interruptible(&pgdat->kswapd_wait);
3308 * The reclaimable count would be mostly accurate.
3309 * The less reclaimable pages may be
3310 * - mlocked pages, which will be moved to unevictable list when encountered
3311 * - mapped pages, which may require several travels to be reclaimed
3312 * - dirty pages, which is not "instantly" reclaimable
3314 unsigned long global_reclaimable_pages(void)
3318 nr = global_page_state(NR_ACTIVE_FILE) +
3319 global_page_state(NR_INACTIVE_FILE);
3321 if (get_nr_swap_pages() > 0)
3322 nr += global_page_state(NR_ACTIVE_ANON) +
3323 global_page_state(NR_INACTIVE_ANON);
3328 #ifdef CONFIG_HIBERNATION
3330 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3333 * Rather than trying to age LRUs the aim is to preserve the overall
3334 * LRU order by reclaiming preferentially
3335 * inactive > active > active referenced > active mapped
3337 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3339 struct reclaim_state reclaim_state;
3340 struct scan_control sc = {
3341 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3345 .nr_to_reclaim = nr_to_reclaim,
3346 .hibernation_mode = 1,
3348 .priority = DEF_PRIORITY,
3350 struct shrink_control shrink = {
3351 .gfp_mask = sc.gfp_mask,
3353 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3354 struct task_struct *p = current;
3355 unsigned long nr_reclaimed;
3357 p->flags |= PF_MEMALLOC;
3358 lockdep_set_current_reclaim_state(sc.gfp_mask);
3359 reclaim_state.reclaimed_slab = 0;
3360 p->reclaim_state = &reclaim_state;
3362 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3364 p->reclaim_state = NULL;
3365 lockdep_clear_current_reclaim_state();
3366 p->flags &= ~PF_MEMALLOC;
3368 return nr_reclaimed;
3370 #endif /* CONFIG_HIBERNATION */
3372 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3373 not required for correctness. So if the last cpu in a node goes
3374 away, we get changed to run anywhere: as the first one comes back,
3375 restore their cpu bindings. */
3376 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3381 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3382 for_each_node_state(nid, N_MEMORY) {
3383 pg_data_t *pgdat = NODE_DATA(nid);
3384 const struct cpumask *mask;
3386 mask = cpumask_of_node(pgdat->node_id);
3388 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3389 /* One of our CPUs online: restore mask */
3390 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3397 * This kswapd start function will be called by init and node-hot-add.
3398 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3400 int kswapd_run(int nid)
3402 pg_data_t *pgdat = NODE_DATA(nid);
3408 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3409 if (IS_ERR(pgdat->kswapd)) {
3410 /* failure at boot is fatal */
3411 BUG_ON(system_state == SYSTEM_BOOTING);
3412 pr_err("Failed to start kswapd on node %d\n", nid);
3413 ret = PTR_ERR(pgdat->kswapd);
3414 pgdat->kswapd = NULL;
3420 * Called by memory hotplug when all memory in a node is offlined. Caller must
3421 * hold lock_memory_hotplug().
3423 void kswapd_stop(int nid)
3425 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3428 kthread_stop(kswapd);
3429 NODE_DATA(nid)->kswapd = NULL;
3433 static int __init kswapd_init(void)
3438 for_each_node_state(nid, N_MEMORY)
3440 hotcpu_notifier(cpu_callback, 0);
3444 module_init(kswapd_init)
3450 * If non-zero call zone_reclaim when the number of free pages falls below
3453 int zone_reclaim_mode __read_mostly;
3455 #define RECLAIM_OFF 0
3456 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3457 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3458 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3461 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3462 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3465 #define ZONE_RECLAIM_PRIORITY 4
3468 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3471 int sysctl_min_unmapped_ratio = 1;
3474 * If the number of slab pages in a zone grows beyond this percentage then
3475 * slab reclaim needs to occur.
3477 int sysctl_min_slab_ratio = 5;
3479 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3481 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3482 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3483 zone_page_state(zone, NR_ACTIVE_FILE);
3486 * It's possible for there to be more file mapped pages than
3487 * accounted for by the pages on the file LRU lists because
3488 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3490 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3493 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3494 static long zone_pagecache_reclaimable(struct zone *zone)
3496 long nr_pagecache_reclaimable;
3500 * If RECLAIM_SWAP is set, then all file pages are considered
3501 * potentially reclaimable. Otherwise, we have to worry about
3502 * pages like swapcache and zone_unmapped_file_pages() provides
3505 if (zone_reclaim_mode & RECLAIM_SWAP)
3506 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3508 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3510 /* If we can't clean pages, remove dirty pages from consideration */
3511 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3512 delta += zone_page_state(zone, NR_FILE_DIRTY);
3514 /* Watch for any possible underflows due to delta */
3515 if (unlikely(delta > nr_pagecache_reclaimable))
3516 delta = nr_pagecache_reclaimable;
3518 return nr_pagecache_reclaimable - delta;
3522 * Try to free up some pages from this zone through reclaim.
3524 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3526 /* Minimum pages needed in order to stay on node */
3527 const unsigned long nr_pages = 1 << order;
3528 struct task_struct *p = current;
3529 struct reclaim_state reclaim_state;
3530 struct scan_control sc = {
3531 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3532 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3534 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3535 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3537 .priority = ZONE_RECLAIM_PRIORITY,
3539 struct shrink_control shrink = {
3540 .gfp_mask = sc.gfp_mask,
3542 unsigned long nr_slab_pages0, nr_slab_pages1;
3546 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3547 * and we also need to be able to write out pages for RECLAIM_WRITE
3550 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3551 lockdep_set_current_reclaim_state(gfp_mask);
3552 reclaim_state.reclaimed_slab = 0;
3553 p->reclaim_state = &reclaim_state;
3555 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3557 * Free memory by calling shrink zone with increasing
3558 * priorities until we have enough memory freed.
3561 shrink_zone(zone, &sc);
3562 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3565 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3566 if (nr_slab_pages0 > zone->min_slab_pages) {
3568 * shrink_slab() does not currently allow us to determine how
3569 * many pages were freed in this zone. So we take the current
3570 * number of slab pages and shake the slab until it is reduced
3571 * by the same nr_pages that we used for reclaiming unmapped
3574 nodes_clear(shrink.nodes_to_scan);
3575 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3577 unsigned long lru_pages = zone_reclaimable_pages(zone);
3579 /* No reclaimable slab or very low memory pressure */
3580 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3583 /* Freed enough memory */
3584 nr_slab_pages1 = zone_page_state(zone,
3585 NR_SLAB_RECLAIMABLE);
3586 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3591 * Update nr_reclaimed by the number of slab pages we
3592 * reclaimed from this zone.
3594 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3595 if (nr_slab_pages1 < nr_slab_pages0)
3596 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3599 p->reclaim_state = NULL;
3600 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3601 lockdep_clear_current_reclaim_state();
3602 return sc.nr_reclaimed >= nr_pages;
3605 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3611 * Zone reclaim reclaims unmapped file backed pages and
3612 * slab pages if we are over the defined limits.
3614 * A small portion of unmapped file backed pages is needed for
3615 * file I/O otherwise pages read by file I/O will be immediately
3616 * thrown out if the zone is overallocated. So we do not reclaim
3617 * if less than a specified percentage of the zone is used by
3618 * unmapped file backed pages.
3620 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3621 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3622 return ZONE_RECLAIM_FULL;
3624 if (!zone_reclaimable(zone))
3625 return ZONE_RECLAIM_FULL;
3628 * Do not scan if the allocation should not be delayed.
3630 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3631 return ZONE_RECLAIM_NOSCAN;
3634 * Only run zone reclaim on the local zone or on zones that do not
3635 * have associated processors. This will favor the local processor
3636 * over remote processors and spread off node memory allocations
3637 * as wide as possible.
3639 node_id = zone_to_nid(zone);
3640 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3641 return ZONE_RECLAIM_NOSCAN;
3643 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3644 return ZONE_RECLAIM_NOSCAN;
3646 ret = __zone_reclaim(zone, gfp_mask, order);
3647 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3650 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3657 * page_evictable - test whether a page is evictable
3658 * @page: the page to test
3660 * Test whether page is evictable--i.e., should be placed on active/inactive
3661 * lists vs unevictable list.
3663 * Reasons page might not be evictable:
3664 * (1) page's mapping marked unevictable
3665 * (2) page is part of an mlocked VMA
3668 int page_evictable(struct page *page)
3670 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3675 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3676 * @pages: array of pages to check
3677 * @nr_pages: number of pages to check
3679 * Checks pages for evictability and moves them to the appropriate lru list.
3681 * This function is only used for SysV IPC SHM_UNLOCK.
3683 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3685 struct lruvec *lruvec;
3686 struct zone *zone = NULL;
3691 for (i = 0; i < nr_pages; i++) {
3692 struct page *page = pages[i];
3693 struct zone *pagezone;
3696 pagezone = page_zone(page);
3697 if (pagezone != zone) {
3699 spin_unlock_irq(&zone->lru_lock);
3701 spin_lock_irq(&zone->lru_lock);
3703 lruvec = mem_cgroup_page_lruvec(page, zone);
3705 if (!PageLRU(page) || !PageUnevictable(page))
3708 if (page_evictable(page)) {
3709 enum lru_list lru = page_lru_base_type(page);
3711 VM_BUG_ON(PageActive(page));
3712 ClearPageUnevictable(page);
3713 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3714 add_page_to_lru_list(page, lruvec, lru);
3720 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3721 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3722 spin_unlock_irq(&zone->lru_lock);
3725 #endif /* CONFIG_SHMEM */
3727 static void warn_scan_unevictable_pages(void)
3729 printk_once(KERN_WARNING
3730 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3731 "disabled for lack of a legitimate use case. If you have "
3732 "one, please send an email to linux-mm@kvack.org.\n",
3737 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3738 * all nodes' unevictable lists for evictable pages
3740 unsigned long scan_unevictable_pages;
3742 int scan_unevictable_handler(struct ctl_table *table, int write,
3743 void __user *buffer,
3744 size_t *length, loff_t *ppos)
3746 warn_scan_unevictable_pages();
3747 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3748 scan_unevictable_pages = 0;
3754 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3755 * a specified node's per zone unevictable lists for evictable pages.
3758 static ssize_t read_scan_unevictable_node(struct device *dev,
3759 struct device_attribute *attr,
3762 warn_scan_unevictable_pages();
3763 return sprintf(buf, "0\n"); /* always zero; should fit... */
3766 static ssize_t write_scan_unevictable_node(struct device *dev,
3767 struct device_attribute *attr,
3768 const char *buf, size_t count)
3770 warn_scan_unevictable_pages();
3775 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3776 read_scan_unevictable_node,
3777 write_scan_unevictable_node);
3779 int scan_unevictable_register_node(struct node *node)
3781 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3784 void scan_unevictable_unregister_node(struct node *node)
3786 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);