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 struct mem_cgroup *root = sc->target_mem_cgroup;
146 return !mem_cgroup_disabled() &&
147 mem_cgroup_soft_reclaim_eligible(root, root) != SKIP_TREE;
150 static bool global_reclaim(struct scan_control *sc)
155 static bool mem_cgroup_should_soft_reclaim(struct scan_control *sc)
161 unsigned long zone_reclaimable_pages(struct zone *zone)
165 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
166 zone_page_state(zone, NR_INACTIVE_FILE);
168 if (get_nr_swap_pages() > 0)
169 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
170 zone_page_state(zone, NR_INACTIVE_ANON);
175 bool zone_reclaimable(struct zone *zone)
177 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
180 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
182 if (!mem_cgroup_disabled())
183 return mem_cgroup_get_lru_size(lruvec, lru);
185 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
189 * Add a shrinker callback to be called from the vm.
191 int register_shrinker(struct shrinker *shrinker)
193 size_t size = sizeof(*shrinker->nr_deferred);
196 * If we only have one possible node in the system anyway, save
197 * ourselves the trouble and disable NUMA aware behavior. This way we
198 * will save memory and some small loop time later.
200 if (nr_node_ids == 1)
201 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
203 if (shrinker->flags & SHRINKER_NUMA_AWARE)
206 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
207 if (!shrinker->nr_deferred)
210 down_write(&shrinker_rwsem);
211 list_add_tail(&shrinker->list, &shrinker_list);
212 up_write(&shrinker_rwsem);
215 EXPORT_SYMBOL(register_shrinker);
220 void unregister_shrinker(struct shrinker *shrinker)
222 down_write(&shrinker_rwsem);
223 list_del(&shrinker->list);
224 up_write(&shrinker_rwsem);
226 EXPORT_SYMBOL(unregister_shrinker);
228 #define SHRINK_BATCH 128
231 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
232 unsigned long nr_pages_scanned, unsigned long lru_pages)
234 unsigned long freed = 0;
235 unsigned long long delta;
240 int nid = shrinkctl->nid;
241 long batch_size = shrinker->batch ? shrinker->batch
244 max_pass = shrinker->count_objects(shrinker, shrinkctl);
249 * copy the current shrinker scan count into a local variable
250 * and zero it so that other concurrent shrinker invocations
251 * don't also do this scanning work.
253 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
256 delta = (4 * nr_pages_scanned) / shrinker->seeks;
258 do_div(delta, lru_pages + 1);
260 if (total_scan < 0) {
262 "shrink_slab: %pF negative objects to delete nr=%ld\n",
263 shrinker->scan_objects, total_scan);
264 total_scan = max_pass;
268 * We need to avoid excessive windup on filesystem shrinkers
269 * due to large numbers of GFP_NOFS allocations causing the
270 * shrinkers to return -1 all the time. This results in a large
271 * nr being built up so when a shrink that can do some work
272 * comes along it empties the entire cache due to nr >>>
273 * max_pass. This is bad for sustaining a working set in
276 * Hence only allow the shrinker to scan the entire cache when
277 * a large delta change is calculated directly.
279 if (delta < max_pass / 4)
280 total_scan = min(total_scan, max_pass / 2);
283 * Avoid risking looping forever due to too large nr value:
284 * never try to free more than twice the estimate number of
287 if (total_scan > max_pass * 2)
288 total_scan = max_pass * 2;
290 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
291 nr_pages_scanned, lru_pages,
292 max_pass, delta, total_scan);
294 while (total_scan >= batch_size) {
297 shrinkctl->nr_to_scan = batch_size;
298 ret = shrinker->scan_objects(shrinker, shrinkctl);
299 if (ret == SHRINK_STOP)
303 count_vm_events(SLABS_SCANNED, batch_size);
304 total_scan -= batch_size;
310 * move the unused scan count back into the shrinker in a
311 * manner that handles concurrent updates. If we exhausted the
312 * scan, there is no need to do an update.
315 new_nr = atomic_long_add_return(total_scan,
316 &shrinker->nr_deferred[nid]);
318 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
320 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
325 * Call the shrink functions to age shrinkable caches
327 * Here we assume it costs one seek to replace a lru page and that it also
328 * takes a seek to recreate a cache object. With this in mind we age equal
329 * percentages of the lru and ageable caches. This should balance the seeks
330 * generated by these structures.
332 * If the vm encountered mapped pages on the LRU it increase the pressure on
333 * slab to avoid swapping.
335 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
337 * `lru_pages' represents the number of on-LRU pages in all the zones which
338 * are eligible for the caller's allocation attempt. It is used for balancing
339 * slab reclaim versus page reclaim.
341 * Returns the number of slab objects which we shrunk.
343 unsigned long shrink_slab(struct shrink_control *shrinkctl,
344 unsigned long nr_pages_scanned,
345 unsigned long lru_pages)
347 struct shrinker *shrinker;
348 unsigned long freed = 0;
350 if (nr_pages_scanned == 0)
351 nr_pages_scanned = SWAP_CLUSTER_MAX;
353 if (!down_read_trylock(&shrinker_rwsem)) {
355 * If we would return 0, our callers would understand that we
356 * have nothing else to shrink and give up trying. By returning
357 * 1 we keep it going and assume we'll be able to shrink next
364 list_for_each_entry(shrinker, &shrinker_list, list) {
365 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
366 if (!node_online(shrinkctl->nid))
369 if (!(shrinker->flags & SHRINKER_NUMA_AWARE) &&
370 (shrinkctl->nid != 0))
373 freed += shrink_slab_node(shrinkctl, shrinker,
374 nr_pages_scanned, lru_pages);
378 up_read(&shrinker_rwsem);
384 static inline int is_page_cache_freeable(struct page *page)
387 * A freeable page cache page is referenced only by the caller
388 * that isolated the page, the page cache radix tree and
389 * optional buffer heads at page->private.
391 return page_count(page) - page_has_private(page) == 2;
394 static int may_write_to_queue(struct backing_dev_info *bdi,
395 struct scan_control *sc)
397 if (current->flags & PF_SWAPWRITE)
399 if (!bdi_write_congested(bdi))
401 if (bdi == current->backing_dev_info)
407 * We detected a synchronous write error writing a page out. Probably
408 * -ENOSPC. We need to propagate that into the address_space for a subsequent
409 * fsync(), msync() or close().
411 * The tricky part is that after writepage we cannot touch the mapping: nothing
412 * prevents it from being freed up. But we have a ref on the page and once
413 * that page is locked, the mapping is pinned.
415 * We're allowed to run sleeping lock_page() here because we know the caller has
418 static void handle_write_error(struct address_space *mapping,
419 struct page *page, int error)
422 if (page_mapping(page) == mapping)
423 mapping_set_error(mapping, error);
427 /* possible outcome of pageout() */
429 /* failed to write page out, page is locked */
431 /* move page to the active list, page is locked */
433 /* page has been sent to the disk successfully, page is unlocked */
435 /* page is clean and locked */
440 * pageout is called by shrink_page_list() for each dirty page.
441 * Calls ->writepage().
443 static pageout_t pageout(struct page *page, struct address_space *mapping,
444 struct scan_control *sc)
447 * If the page is dirty, only perform writeback if that write
448 * will be non-blocking. To prevent this allocation from being
449 * stalled by pagecache activity. But note that there may be
450 * stalls if we need to run get_block(). We could test
451 * PagePrivate for that.
453 * If this process is currently in __generic_file_aio_write() against
454 * this page's queue, we can perform writeback even if that
457 * If the page is swapcache, write it back even if that would
458 * block, for some throttling. This happens by accident, because
459 * swap_backing_dev_info is bust: it doesn't reflect the
460 * congestion state of the swapdevs. Easy to fix, if needed.
462 if (!is_page_cache_freeable(page))
466 * Some data journaling orphaned pages can have
467 * page->mapping == NULL while being dirty with clean buffers.
469 if (page_has_private(page)) {
470 if (try_to_free_buffers(page)) {
471 ClearPageDirty(page);
472 printk("%s: orphaned page\n", __func__);
478 if (mapping->a_ops->writepage == NULL)
479 return PAGE_ACTIVATE;
480 if (!may_write_to_queue(mapping->backing_dev_info, sc))
483 if (clear_page_dirty_for_io(page)) {
485 struct writeback_control wbc = {
486 .sync_mode = WB_SYNC_NONE,
487 .nr_to_write = SWAP_CLUSTER_MAX,
489 .range_end = LLONG_MAX,
493 SetPageReclaim(page);
494 res = mapping->a_ops->writepage(page, &wbc);
496 handle_write_error(mapping, page, res);
497 if (res == AOP_WRITEPAGE_ACTIVATE) {
498 ClearPageReclaim(page);
499 return PAGE_ACTIVATE;
502 if (!PageWriteback(page)) {
503 /* synchronous write or broken a_ops? */
504 ClearPageReclaim(page);
506 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
507 inc_zone_page_state(page, NR_VMSCAN_WRITE);
515 * Same as remove_mapping, but if the page is removed from the mapping, it
516 * gets returned with a refcount of 0.
518 static int __remove_mapping(struct address_space *mapping, struct page *page)
520 BUG_ON(!PageLocked(page));
521 BUG_ON(mapping != page_mapping(page));
523 spin_lock_irq(&mapping->tree_lock);
525 * The non racy check for a busy page.
527 * Must be careful with the order of the tests. When someone has
528 * a ref to the page, it may be possible that they dirty it then
529 * drop the reference. So if PageDirty is tested before page_count
530 * here, then the following race may occur:
532 * get_user_pages(&page);
533 * [user mapping goes away]
535 * !PageDirty(page) [good]
536 * SetPageDirty(page);
538 * !page_count(page) [good, discard it]
540 * [oops, our write_to data is lost]
542 * Reversing the order of the tests ensures such a situation cannot
543 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
544 * load is not satisfied before that of page->_count.
546 * Note that if SetPageDirty is always performed via set_page_dirty,
547 * and thus under tree_lock, then this ordering is not required.
549 if (!page_freeze_refs(page, 2))
551 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
552 if (unlikely(PageDirty(page))) {
553 page_unfreeze_refs(page, 2);
557 if (PageSwapCache(page)) {
558 swp_entry_t swap = { .val = page_private(page) };
559 __delete_from_swap_cache(page);
560 spin_unlock_irq(&mapping->tree_lock);
561 swapcache_free(swap, page);
563 void (*freepage)(struct page *);
565 freepage = mapping->a_ops->freepage;
567 __delete_from_page_cache(page);
568 spin_unlock_irq(&mapping->tree_lock);
569 mem_cgroup_uncharge_cache_page(page);
571 if (freepage != NULL)
578 spin_unlock_irq(&mapping->tree_lock);
583 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
584 * someone else has a ref on the page, abort and return 0. If it was
585 * successfully detached, return 1. Assumes the caller has a single ref on
588 int remove_mapping(struct address_space *mapping, struct page *page)
590 if (__remove_mapping(mapping, page)) {
592 * Unfreezing the refcount with 1 rather than 2 effectively
593 * drops the pagecache ref for us without requiring another
596 page_unfreeze_refs(page, 1);
603 * putback_lru_page - put previously isolated page onto appropriate LRU list
604 * @page: page to be put back to appropriate lru list
606 * Add previously isolated @page to appropriate LRU list.
607 * Page may still be unevictable for other reasons.
609 * lru_lock must not be held, interrupts must be enabled.
611 void putback_lru_page(struct page *page)
614 int was_unevictable = PageUnevictable(page);
616 VM_BUG_ON(PageLRU(page));
619 ClearPageUnevictable(page);
621 if (page_evictable(page)) {
623 * For evictable pages, we can use the cache.
624 * In event of a race, worst case is we end up with an
625 * unevictable page on [in]active list.
626 * We know how to handle that.
628 is_unevictable = false;
632 * Put unevictable pages directly on zone's unevictable
635 is_unevictable = true;
636 add_page_to_unevictable_list(page);
638 * When racing with an mlock or AS_UNEVICTABLE clearing
639 * (page is unlocked) make sure that if the other thread
640 * does not observe our setting of PG_lru and fails
641 * isolation/check_move_unevictable_pages,
642 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
643 * the page back to the evictable list.
645 * The other side is TestClearPageMlocked() or shmem_lock().
651 * page's status can change while we move it among lru. If an evictable
652 * page is on unevictable list, it never be freed. To avoid that,
653 * check after we added it to the list, again.
655 if (is_unevictable && page_evictable(page)) {
656 if (!isolate_lru_page(page)) {
660 /* This means someone else dropped this page from LRU
661 * So, it will be freed or putback to LRU again. There is
662 * nothing to do here.
666 if (was_unevictable && !is_unevictable)
667 count_vm_event(UNEVICTABLE_PGRESCUED);
668 else if (!was_unevictable && is_unevictable)
669 count_vm_event(UNEVICTABLE_PGCULLED);
671 put_page(page); /* drop ref from isolate */
674 enum page_references {
676 PAGEREF_RECLAIM_CLEAN,
681 static enum page_references page_check_references(struct page *page,
682 struct scan_control *sc)
684 int referenced_ptes, referenced_page;
685 unsigned long vm_flags;
687 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
689 referenced_page = TestClearPageReferenced(page);
692 * Mlock lost the isolation race with us. Let try_to_unmap()
693 * move the page to the unevictable list.
695 if (vm_flags & VM_LOCKED)
696 return PAGEREF_RECLAIM;
698 if (referenced_ptes) {
699 if (PageSwapBacked(page))
700 return PAGEREF_ACTIVATE;
702 * All mapped pages start out with page table
703 * references from the instantiating fault, so we need
704 * to look twice if a mapped file page is used more
707 * Mark it and spare it for another trip around the
708 * inactive list. Another page table reference will
709 * lead to its activation.
711 * Note: the mark is set for activated pages as well
712 * so that recently deactivated but used pages are
715 SetPageReferenced(page);
717 if (referenced_page || referenced_ptes > 1)
718 return PAGEREF_ACTIVATE;
721 * Activate file-backed executable pages after first usage.
723 if (vm_flags & VM_EXEC)
724 return PAGEREF_ACTIVATE;
729 /* Reclaim if clean, defer dirty pages to writeback */
730 if (referenced_page && !PageSwapBacked(page))
731 return PAGEREF_RECLAIM_CLEAN;
733 return PAGEREF_RECLAIM;
736 /* Check if a page is dirty or under writeback */
737 static void page_check_dirty_writeback(struct page *page,
738 bool *dirty, bool *writeback)
740 struct address_space *mapping;
743 * Anonymous pages are not handled by flushers and must be written
744 * from reclaim context. Do not stall reclaim based on them
746 if (!page_is_file_cache(page)) {
752 /* By default assume that the page flags are accurate */
753 *dirty = PageDirty(page);
754 *writeback = PageWriteback(page);
756 /* Verify dirty/writeback state if the filesystem supports it */
757 if (!page_has_private(page))
760 mapping = page_mapping(page);
761 if (mapping && mapping->a_ops->is_dirty_writeback)
762 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
766 * shrink_page_list() returns the number of reclaimed pages
768 static unsigned long shrink_page_list(struct list_head *page_list,
770 struct scan_control *sc,
771 enum ttu_flags ttu_flags,
772 unsigned long *ret_nr_dirty,
773 unsigned long *ret_nr_unqueued_dirty,
774 unsigned long *ret_nr_congested,
775 unsigned long *ret_nr_writeback,
776 unsigned long *ret_nr_immediate,
779 LIST_HEAD(ret_pages);
780 LIST_HEAD(free_pages);
782 unsigned long nr_unqueued_dirty = 0;
783 unsigned long nr_dirty = 0;
784 unsigned long nr_congested = 0;
785 unsigned long nr_reclaimed = 0;
786 unsigned long nr_writeback = 0;
787 unsigned long nr_immediate = 0;
791 mem_cgroup_uncharge_start();
792 while (!list_empty(page_list)) {
793 struct address_space *mapping;
796 enum page_references references = PAGEREF_RECLAIM_CLEAN;
797 bool dirty, writeback;
801 page = lru_to_page(page_list);
802 list_del(&page->lru);
804 if (!trylock_page(page))
807 VM_BUG_ON(PageActive(page));
808 VM_BUG_ON(page_zone(page) != zone);
812 if (unlikely(!page_evictable(page)))
815 if (!sc->may_unmap && page_mapped(page))
818 /* Double the slab pressure for mapped and swapcache pages */
819 if (page_mapped(page) || PageSwapCache(page))
822 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
823 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
826 * The number of dirty pages determines if a zone is marked
827 * reclaim_congested which affects wait_iff_congested. kswapd
828 * will stall and start writing pages if the tail of the LRU
829 * is all dirty unqueued pages.
831 page_check_dirty_writeback(page, &dirty, &writeback);
832 if (dirty || writeback)
835 if (dirty && !writeback)
839 * Treat this page as congested if the underlying BDI is or if
840 * pages are cycling through the LRU so quickly that the
841 * pages marked for immediate reclaim are making it to the
842 * end of the LRU a second time.
844 mapping = page_mapping(page);
845 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
846 (writeback && PageReclaim(page)))
850 * If a page at the tail of the LRU is under writeback, there
851 * are three cases to consider.
853 * 1) If reclaim is encountering an excessive number of pages
854 * under writeback and this page is both under writeback and
855 * PageReclaim then it indicates that pages are being queued
856 * for IO but are being recycled through the LRU before the
857 * IO can complete. Waiting on the page itself risks an
858 * indefinite stall if it is impossible to writeback the
859 * page due to IO error or disconnected storage so instead
860 * note that the LRU is being scanned too quickly and the
861 * caller can stall after page list has been processed.
863 * 2) Global reclaim encounters a page, memcg encounters a
864 * page that is not marked for immediate reclaim or
865 * the caller does not have __GFP_IO. In this case mark
866 * the page for immediate reclaim and continue scanning.
868 * __GFP_IO is checked because a loop driver thread might
869 * enter reclaim, and deadlock if it waits on a page for
870 * which it is needed to do the write (loop masks off
871 * __GFP_IO|__GFP_FS for this reason); but more thought
872 * would probably show more reasons.
874 * Don't require __GFP_FS, since we're not going into the
875 * FS, just waiting on its writeback completion. Worryingly,
876 * ext4 gfs2 and xfs allocate pages with
877 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
878 * may_enter_fs here is liable to OOM on them.
880 * 3) memcg encounters a page that is not already marked
881 * PageReclaim. memcg does not have any dirty pages
882 * throttling so we could easily OOM just because too many
883 * pages are in writeback and there is nothing else to
884 * reclaim. Wait for the writeback to complete.
886 if (PageWriteback(page)) {
888 if (current_is_kswapd() &&
890 zone_is_reclaim_writeback(zone)) {
895 } else if (global_reclaim(sc) ||
896 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
898 * This is slightly racy - end_page_writeback()
899 * might have just cleared PageReclaim, then
900 * setting PageReclaim here end up interpreted
901 * as PageReadahead - but that does not matter
902 * enough to care. What we do want is for this
903 * page to have PageReclaim set next time memcg
904 * reclaim reaches the tests above, so it will
905 * then wait_on_page_writeback() to avoid OOM;
906 * and it's also appropriate in global reclaim.
908 SetPageReclaim(page);
915 wait_on_page_writeback(page);
920 references = page_check_references(page, sc);
922 switch (references) {
923 case PAGEREF_ACTIVATE:
924 goto activate_locked;
927 case PAGEREF_RECLAIM:
928 case PAGEREF_RECLAIM_CLEAN:
929 ; /* try to reclaim the page below */
933 * Anonymous process memory has backing store?
934 * Try to allocate it some swap space here.
936 if (PageAnon(page) && !PageSwapCache(page)) {
937 if (!(sc->gfp_mask & __GFP_IO))
939 if (!add_to_swap(page, page_list))
940 goto activate_locked;
943 /* Adding to swap updated mapping */
944 mapping = page_mapping(page);
948 * The page is mapped into the page tables of one or more
949 * processes. Try to unmap it here.
951 if (page_mapped(page) && mapping) {
952 switch (try_to_unmap(page, ttu_flags)) {
954 goto activate_locked;
960 ; /* try to free the page below */
964 if (PageDirty(page)) {
966 * Only kswapd can writeback filesystem pages to
967 * avoid risk of stack overflow but only writeback
968 * if many dirty pages have been encountered.
970 if (page_is_file_cache(page) &&
971 (!current_is_kswapd() ||
972 !zone_is_reclaim_dirty(zone))) {
974 * Immediately reclaim when written back.
975 * Similar in principal to deactivate_page()
976 * except we already have the page isolated
977 * and know it's dirty
979 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
980 SetPageReclaim(page);
985 if (references == PAGEREF_RECLAIM_CLEAN)
989 if (!sc->may_writepage)
992 /* Page is dirty, try to write it out here */
993 switch (pageout(page, mapping, sc)) {
997 goto activate_locked;
999 if (PageWriteback(page))
1001 if (PageDirty(page))
1005 * A synchronous write - probably a ramdisk. Go
1006 * ahead and try to reclaim the page.
1008 if (!trylock_page(page))
1010 if (PageDirty(page) || PageWriteback(page))
1012 mapping = page_mapping(page);
1014 ; /* try to free the page below */
1019 * If the page has buffers, try to free the buffer mappings
1020 * associated with this page. If we succeed we try to free
1023 * We do this even if the page is PageDirty().
1024 * try_to_release_page() does not perform I/O, but it is
1025 * possible for a page to have PageDirty set, but it is actually
1026 * clean (all its buffers are clean). This happens if the
1027 * buffers were written out directly, with submit_bh(). ext3
1028 * will do this, as well as the blockdev mapping.
1029 * try_to_release_page() will discover that cleanness and will
1030 * drop the buffers and mark the page clean - it can be freed.
1032 * Rarely, pages can have buffers and no ->mapping. These are
1033 * the pages which were not successfully invalidated in
1034 * truncate_complete_page(). We try to drop those buffers here
1035 * and if that worked, and the page is no longer mapped into
1036 * process address space (page_count == 1) it can be freed.
1037 * Otherwise, leave the page on the LRU so it is swappable.
1039 if (page_has_private(page)) {
1040 if (!try_to_release_page(page, sc->gfp_mask))
1041 goto activate_locked;
1042 if (!mapping && page_count(page) == 1) {
1044 if (put_page_testzero(page))
1048 * rare race with speculative reference.
1049 * the speculative reference will free
1050 * this page shortly, so we may
1051 * increment nr_reclaimed here (and
1052 * leave it off the LRU).
1060 if (!mapping || !__remove_mapping(mapping, page))
1064 * At this point, we have no other references and there is
1065 * no way to pick any more up (removed from LRU, removed
1066 * from pagecache). Can use non-atomic bitops now (and
1067 * we obviously don't have to worry about waking up a process
1068 * waiting on the page lock, because there are no references.
1070 __clear_page_locked(page);
1075 * Is there need to periodically free_page_list? It would
1076 * appear not as the counts should be low
1078 list_add(&page->lru, &free_pages);
1082 if (PageSwapCache(page))
1083 try_to_free_swap(page);
1085 putback_lru_page(page);
1089 /* Not a candidate for swapping, so reclaim swap space. */
1090 if (PageSwapCache(page) && vm_swap_full())
1091 try_to_free_swap(page);
1092 VM_BUG_ON(PageActive(page));
1093 SetPageActive(page);
1098 list_add(&page->lru, &ret_pages);
1099 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1102 free_hot_cold_page_list(&free_pages, 1);
1104 list_splice(&ret_pages, page_list);
1105 count_vm_events(PGACTIVATE, pgactivate);
1106 mem_cgroup_uncharge_end();
1107 *ret_nr_dirty += nr_dirty;
1108 *ret_nr_congested += nr_congested;
1109 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1110 *ret_nr_writeback += nr_writeback;
1111 *ret_nr_immediate += nr_immediate;
1112 return nr_reclaimed;
1115 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1116 struct list_head *page_list)
1118 struct scan_control sc = {
1119 .gfp_mask = GFP_KERNEL,
1120 .priority = DEF_PRIORITY,
1123 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1124 struct page *page, *next;
1125 LIST_HEAD(clean_pages);
1127 list_for_each_entry_safe(page, next, page_list, lru) {
1128 if (page_is_file_cache(page) && !PageDirty(page)) {
1129 ClearPageActive(page);
1130 list_move(&page->lru, &clean_pages);
1134 ret = shrink_page_list(&clean_pages, zone, &sc,
1135 TTU_UNMAP|TTU_IGNORE_ACCESS,
1136 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1137 list_splice(&clean_pages, page_list);
1138 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1143 * Attempt to remove the specified page from its LRU. Only take this page
1144 * if it is of the appropriate PageActive status. Pages which are being
1145 * freed elsewhere are also ignored.
1147 * page: page to consider
1148 * mode: one of the LRU isolation modes defined above
1150 * returns 0 on success, -ve errno on failure.
1152 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1156 /* Only take pages on the LRU. */
1160 /* Compaction should not handle unevictable pages but CMA can do so */
1161 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1167 * To minimise LRU disruption, the caller can indicate that it only
1168 * wants to isolate pages it will be able to operate on without
1169 * blocking - clean pages for the most part.
1171 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1172 * is used by reclaim when it is cannot write to backing storage
1174 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1175 * that it is possible to migrate without blocking
1177 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1178 /* All the caller can do on PageWriteback is block */
1179 if (PageWriteback(page))
1182 if (PageDirty(page)) {
1183 struct address_space *mapping;
1185 /* ISOLATE_CLEAN means only clean pages */
1186 if (mode & ISOLATE_CLEAN)
1190 * Only pages without mappings or that have a
1191 * ->migratepage callback are possible to migrate
1194 mapping = page_mapping(page);
1195 if (mapping && !mapping->a_ops->migratepage)
1200 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1203 if (likely(get_page_unless_zero(page))) {
1205 * Be careful not to clear PageLRU until after we're
1206 * sure the page is not being freed elsewhere -- the
1207 * page release code relies on it.
1217 * zone->lru_lock is heavily contended. Some of the functions that
1218 * shrink the lists perform better by taking out a batch of pages
1219 * and working on them outside the LRU lock.
1221 * For pagecache intensive workloads, this function is the hottest
1222 * spot in the kernel (apart from copy_*_user functions).
1224 * Appropriate locks must be held before calling this function.
1226 * @nr_to_scan: The number of pages to look through on the list.
1227 * @lruvec: The LRU vector to pull pages from.
1228 * @dst: The temp list to put pages on to.
1229 * @nr_scanned: The number of pages that were scanned.
1230 * @sc: The scan_control struct for this reclaim session
1231 * @mode: One of the LRU isolation modes
1232 * @lru: LRU list id for isolating
1234 * returns how many pages were moved onto *@dst.
1236 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1237 struct lruvec *lruvec, struct list_head *dst,
1238 unsigned long *nr_scanned, struct scan_control *sc,
1239 isolate_mode_t mode, enum lru_list lru)
1241 struct list_head *src = &lruvec->lists[lru];
1242 unsigned long nr_taken = 0;
1245 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1249 page = lru_to_page(src);
1250 prefetchw_prev_lru_page(page, src, flags);
1252 VM_BUG_ON(!PageLRU(page));
1254 switch (__isolate_lru_page(page, mode)) {
1256 nr_pages = hpage_nr_pages(page);
1257 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1258 list_move(&page->lru, dst);
1259 nr_taken += nr_pages;
1263 /* else it is being freed elsewhere */
1264 list_move(&page->lru, src);
1273 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1274 nr_taken, mode, is_file_lru(lru));
1279 * isolate_lru_page - tries to isolate a page from its LRU list
1280 * @page: page to isolate from its LRU list
1282 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1283 * vmstat statistic corresponding to whatever LRU list the page was on.
1285 * Returns 0 if the page was removed from an LRU list.
1286 * Returns -EBUSY if the page was not on an LRU list.
1288 * The returned page will have PageLRU() cleared. If it was found on
1289 * the active list, it will have PageActive set. If it was found on
1290 * the unevictable list, it will have the PageUnevictable bit set. That flag
1291 * may need to be cleared by the caller before letting the page go.
1293 * The vmstat statistic corresponding to the list on which the page was
1294 * found will be decremented.
1297 * (1) Must be called with an elevated refcount on the page. This is a
1298 * fundamentnal difference from isolate_lru_pages (which is called
1299 * without a stable reference).
1300 * (2) the lru_lock must not be held.
1301 * (3) interrupts must be enabled.
1303 int isolate_lru_page(struct page *page)
1307 VM_BUG_ON(!page_count(page));
1309 if (PageLRU(page)) {
1310 struct zone *zone = page_zone(page);
1311 struct lruvec *lruvec;
1313 spin_lock_irq(&zone->lru_lock);
1314 lruvec = mem_cgroup_page_lruvec(page, zone);
1315 if (PageLRU(page)) {
1316 int lru = page_lru(page);
1319 del_page_from_lru_list(page, lruvec, lru);
1322 spin_unlock_irq(&zone->lru_lock);
1328 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1329 * then get resheduled. When there are massive number of tasks doing page
1330 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1331 * the LRU list will go small and be scanned faster than necessary, leading to
1332 * unnecessary swapping, thrashing and OOM.
1334 static int too_many_isolated(struct zone *zone, int file,
1335 struct scan_control *sc)
1337 unsigned long inactive, isolated;
1339 if (current_is_kswapd())
1342 if (!global_reclaim(sc))
1346 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1347 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1349 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1350 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1354 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1355 * won't get blocked by normal direct-reclaimers, forming a circular
1358 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1361 return isolated > inactive;
1364 static noinline_for_stack void
1365 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1367 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1368 struct zone *zone = lruvec_zone(lruvec);
1369 LIST_HEAD(pages_to_free);
1372 * Put back any unfreeable pages.
1374 while (!list_empty(page_list)) {
1375 struct page *page = lru_to_page(page_list);
1378 VM_BUG_ON(PageLRU(page));
1379 list_del(&page->lru);
1380 if (unlikely(!page_evictable(page))) {
1381 spin_unlock_irq(&zone->lru_lock);
1382 putback_lru_page(page);
1383 spin_lock_irq(&zone->lru_lock);
1387 lruvec = mem_cgroup_page_lruvec(page, zone);
1390 lru = page_lru(page);
1391 add_page_to_lru_list(page, lruvec, lru);
1393 if (is_active_lru(lru)) {
1394 int file = is_file_lru(lru);
1395 int numpages = hpage_nr_pages(page);
1396 reclaim_stat->recent_rotated[file] += numpages;
1398 if (put_page_testzero(page)) {
1399 __ClearPageLRU(page);
1400 __ClearPageActive(page);
1401 del_page_from_lru_list(page, lruvec, lru);
1403 if (unlikely(PageCompound(page))) {
1404 spin_unlock_irq(&zone->lru_lock);
1405 (*get_compound_page_dtor(page))(page);
1406 spin_lock_irq(&zone->lru_lock);
1408 list_add(&page->lru, &pages_to_free);
1413 * To save our caller's stack, now use input list for pages to free.
1415 list_splice(&pages_to_free, page_list);
1419 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1420 * of reclaimed pages
1422 static noinline_for_stack unsigned long
1423 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1424 struct scan_control *sc, enum lru_list lru)
1426 LIST_HEAD(page_list);
1427 unsigned long nr_scanned;
1428 unsigned long nr_reclaimed = 0;
1429 unsigned long nr_taken;
1430 unsigned long nr_dirty = 0;
1431 unsigned long nr_congested = 0;
1432 unsigned long nr_unqueued_dirty = 0;
1433 unsigned long nr_writeback = 0;
1434 unsigned long nr_immediate = 0;
1435 isolate_mode_t isolate_mode = 0;
1436 int file = is_file_lru(lru);
1437 struct zone *zone = lruvec_zone(lruvec);
1438 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1440 while (unlikely(too_many_isolated(zone, file, sc))) {
1441 congestion_wait(BLK_RW_ASYNC, HZ/10);
1443 /* We are about to die and free our memory. Return now. */
1444 if (fatal_signal_pending(current))
1445 return SWAP_CLUSTER_MAX;
1451 isolate_mode |= ISOLATE_UNMAPPED;
1452 if (!sc->may_writepage)
1453 isolate_mode |= ISOLATE_CLEAN;
1455 spin_lock_irq(&zone->lru_lock);
1457 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1458 &nr_scanned, sc, isolate_mode, lru);
1460 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1461 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1463 if (global_reclaim(sc)) {
1464 zone->pages_scanned += nr_scanned;
1465 if (current_is_kswapd())
1466 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1468 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1470 spin_unlock_irq(&zone->lru_lock);
1475 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1476 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1477 &nr_writeback, &nr_immediate,
1480 spin_lock_irq(&zone->lru_lock);
1482 reclaim_stat->recent_scanned[file] += nr_taken;
1484 if (global_reclaim(sc)) {
1485 if (current_is_kswapd())
1486 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1489 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1493 putback_inactive_pages(lruvec, &page_list);
1495 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1497 spin_unlock_irq(&zone->lru_lock);
1499 free_hot_cold_page_list(&page_list, 1);
1502 * If reclaim is isolating dirty pages under writeback, it implies
1503 * that the long-lived page allocation rate is exceeding the page
1504 * laundering rate. Either the global limits are not being effective
1505 * at throttling processes due to the page distribution throughout
1506 * zones or there is heavy usage of a slow backing device. The
1507 * only option is to throttle from reclaim context which is not ideal
1508 * as there is no guarantee the dirtying process is throttled in the
1509 * same way balance_dirty_pages() manages.
1511 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1512 * of pages under pages flagged for immediate reclaim and stall if any
1513 * are encountered in the nr_immediate check below.
1515 if (nr_writeback && nr_writeback == nr_taken)
1516 zone_set_flag(zone, ZONE_WRITEBACK);
1519 * memcg will stall in page writeback so only consider forcibly
1520 * stalling for global reclaim
1522 if (global_reclaim(sc)) {
1524 * Tag a zone as congested if all the dirty pages scanned were
1525 * backed by a congested BDI and wait_iff_congested will stall.
1527 if (nr_dirty && nr_dirty == nr_congested)
1528 zone_set_flag(zone, ZONE_CONGESTED);
1531 * If dirty pages are scanned that are not queued for IO, it
1532 * implies that flushers are not keeping up. In this case, flag
1533 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1534 * pages from reclaim context. It will forcibly stall in the
1537 if (nr_unqueued_dirty == nr_taken)
1538 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1541 * In addition, if kswapd scans pages marked marked for
1542 * immediate reclaim and under writeback (nr_immediate), it
1543 * implies that pages are cycling through the LRU faster than
1544 * they are written so also forcibly stall.
1546 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1547 congestion_wait(BLK_RW_ASYNC, HZ/10);
1551 * Stall direct reclaim for IO completions if underlying BDIs or zone
1552 * is congested. Allow kswapd to continue until it starts encountering
1553 * unqueued dirty pages or cycling through the LRU too quickly.
1555 if (!sc->hibernation_mode && !current_is_kswapd())
1556 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1558 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1560 nr_scanned, nr_reclaimed,
1562 trace_shrink_flags(file));
1563 return nr_reclaimed;
1567 * This moves pages from the active list to the inactive list.
1569 * We move them the other way if the page is referenced by one or more
1570 * processes, from rmap.
1572 * If the pages are mostly unmapped, the processing is fast and it is
1573 * appropriate to hold zone->lru_lock across the whole operation. But if
1574 * the pages are mapped, the processing is slow (page_referenced()) so we
1575 * should drop zone->lru_lock around each page. It's impossible to balance
1576 * this, so instead we remove the pages from the LRU while processing them.
1577 * It is safe to rely on PG_active against the non-LRU pages in here because
1578 * nobody will play with that bit on a non-LRU page.
1580 * The downside is that we have to touch page->_count against each page.
1581 * But we had to alter page->flags anyway.
1584 static void move_active_pages_to_lru(struct lruvec *lruvec,
1585 struct list_head *list,
1586 struct list_head *pages_to_free,
1589 struct zone *zone = lruvec_zone(lruvec);
1590 unsigned long pgmoved = 0;
1594 while (!list_empty(list)) {
1595 page = lru_to_page(list);
1596 lruvec = mem_cgroup_page_lruvec(page, zone);
1598 VM_BUG_ON(PageLRU(page));
1601 nr_pages = hpage_nr_pages(page);
1602 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1603 list_move(&page->lru, &lruvec->lists[lru]);
1604 pgmoved += nr_pages;
1606 if (put_page_testzero(page)) {
1607 __ClearPageLRU(page);
1608 __ClearPageActive(page);
1609 del_page_from_lru_list(page, lruvec, lru);
1611 if (unlikely(PageCompound(page))) {
1612 spin_unlock_irq(&zone->lru_lock);
1613 (*get_compound_page_dtor(page))(page);
1614 spin_lock_irq(&zone->lru_lock);
1616 list_add(&page->lru, pages_to_free);
1619 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1620 if (!is_active_lru(lru))
1621 __count_vm_events(PGDEACTIVATE, pgmoved);
1624 static void shrink_active_list(unsigned long nr_to_scan,
1625 struct lruvec *lruvec,
1626 struct scan_control *sc,
1629 unsigned long nr_taken;
1630 unsigned long nr_scanned;
1631 unsigned long vm_flags;
1632 LIST_HEAD(l_hold); /* The pages which were snipped off */
1633 LIST_HEAD(l_active);
1634 LIST_HEAD(l_inactive);
1636 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1637 unsigned long nr_rotated = 0;
1638 isolate_mode_t isolate_mode = 0;
1639 int file = is_file_lru(lru);
1640 struct zone *zone = lruvec_zone(lruvec);
1645 isolate_mode |= ISOLATE_UNMAPPED;
1646 if (!sc->may_writepage)
1647 isolate_mode |= ISOLATE_CLEAN;
1649 spin_lock_irq(&zone->lru_lock);
1651 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1652 &nr_scanned, sc, isolate_mode, lru);
1653 if (global_reclaim(sc))
1654 zone->pages_scanned += nr_scanned;
1656 reclaim_stat->recent_scanned[file] += nr_taken;
1658 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1659 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1660 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1661 spin_unlock_irq(&zone->lru_lock);
1663 while (!list_empty(&l_hold)) {
1665 page = lru_to_page(&l_hold);
1666 list_del(&page->lru);
1668 if (unlikely(!page_evictable(page))) {
1669 putback_lru_page(page);
1673 if (unlikely(buffer_heads_over_limit)) {
1674 if (page_has_private(page) && trylock_page(page)) {
1675 if (page_has_private(page))
1676 try_to_release_page(page, 0);
1681 if (page_referenced(page, 0, sc->target_mem_cgroup,
1683 nr_rotated += hpage_nr_pages(page);
1685 * Identify referenced, file-backed active pages and
1686 * give them one more trip around the active list. So
1687 * that executable code get better chances to stay in
1688 * memory under moderate memory pressure. Anon pages
1689 * are not likely to be evicted by use-once streaming
1690 * IO, plus JVM can create lots of anon VM_EXEC pages,
1691 * so we ignore them here.
1693 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1694 list_add(&page->lru, &l_active);
1699 ClearPageActive(page); /* we are de-activating */
1700 list_add(&page->lru, &l_inactive);
1704 * Move pages back to the lru list.
1706 spin_lock_irq(&zone->lru_lock);
1708 * Count referenced pages from currently used mappings as rotated,
1709 * even though only some of them are actually re-activated. This
1710 * helps balance scan pressure between file and anonymous pages in
1713 reclaim_stat->recent_rotated[file] += nr_rotated;
1715 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1716 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1717 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1718 spin_unlock_irq(&zone->lru_lock);
1720 free_hot_cold_page_list(&l_hold, 1);
1724 static int inactive_anon_is_low_global(struct zone *zone)
1726 unsigned long active, inactive;
1728 active = zone_page_state(zone, NR_ACTIVE_ANON);
1729 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1731 if (inactive * zone->inactive_ratio < active)
1738 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1739 * @lruvec: LRU vector to check
1741 * Returns true if the zone does not have enough inactive anon pages,
1742 * meaning some active anon pages need to be deactivated.
1744 static int inactive_anon_is_low(struct lruvec *lruvec)
1747 * If we don't have swap space, anonymous page deactivation
1750 if (!total_swap_pages)
1753 if (!mem_cgroup_disabled())
1754 return mem_cgroup_inactive_anon_is_low(lruvec);
1756 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1759 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1766 * inactive_file_is_low - check if file pages need to be deactivated
1767 * @lruvec: LRU vector to check
1769 * When the system is doing streaming IO, memory pressure here
1770 * ensures that active file pages get deactivated, until more
1771 * than half of the file pages are on the inactive list.
1773 * Once we get to that situation, protect the system's working
1774 * set from being evicted by disabling active file page aging.
1776 * This uses a different ratio than the anonymous pages, because
1777 * the page cache uses a use-once replacement algorithm.
1779 static int inactive_file_is_low(struct lruvec *lruvec)
1781 unsigned long inactive;
1782 unsigned long active;
1784 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1785 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1787 return active > inactive;
1790 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1792 if (is_file_lru(lru))
1793 return inactive_file_is_low(lruvec);
1795 return inactive_anon_is_low(lruvec);
1798 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1799 struct lruvec *lruvec, struct scan_control *sc)
1801 if (is_active_lru(lru)) {
1802 if (inactive_list_is_low(lruvec, lru))
1803 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1807 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1810 static int vmscan_swappiness(struct scan_control *sc)
1812 if (global_reclaim(sc))
1813 return vm_swappiness;
1814 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1825 * Determine how aggressively the anon and file LRU lists should be
1826 * scanned. The relative value of each set of LRU lists is determined
1827 * by looking at the fraction of the pages scanned we did rotate back
1828 * onto the active list instead of evict.
1830 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1831 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1833 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1836 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1838 u64 denominator = 0; /* gcc */
1839 struct zone *zone = lruvec_zone(lruvec);
1840 unsigned long anon_prio, file_prio;
1841 enum scan_balance scan_balance;
1842 unsigned long anon, file, free;
1843 bool force_scan = false;
1844 unsigned long ap, fp;
1848 * If the zone or memcg is small, nr[l] can be 0. This
1849 * results in no scanning on this priority and a potential
1850 * priority drop. Global direct reclaim can go to the next
1851 * zone and tends to have no problems. Global kswapd is for
1852 * zone balancing and it needs to scan a minimum amount. When
1853 * reclaiming for a memcg, a priority drop can cause high
1854 * latencies, so it's better to scan a minimum amount there as
1857 if (current_is_kswapd() && !zone_reclaimable(zone))
1859 if (!global_reclaim(sc))
1862 /* If we have no swap space, do not bother scanning anon pages. */
1863 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1864 scan_balance = SCAN_FILE;
1869 * Global reclaim will swap to prevent OOM even with no
1870 * swappiness, but memcg users want to use this knob to
1871 * disable swapping for individual groups completely when
1872 * using the memory controller's swap limit feature would be
1875 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1876 scan_balance = SCAN_FILE;
1881 * Do not apply any pressure balancing cleverness when the
1882 * system is close to OOM, scan both anon and file equally
1883 * (unless the swappiness setting disagrees with swapping).
1885 if (!sc->priority && vmscan_swappiness(sc)) {
1886 scan_balance = SCAN_EQUAL;
1890 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1891 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1892 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1893 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1896 * If it's foreseeable that reclaiming the file cache won't be
1897 * enough to get the zone back into a desirable shape, we have
1898 * to swap. Better start now and leave the - probably heavily
1899 * thrashing - remaining file pages alone.
1901 if (global_reclaim(sc)) {
1902 free = zone_page_state(zone, NR_FREE_PAGES);
1903 if (unlikely(file + free <= high_wmark_pages(zone))) {
1904 scan_balance = SCAN_ANON;
1910 * There is enough inactive page cache, do not reclaim
1911 * anything from the anonymous working set right now.
1913 if (!inactive_file_is_low(lruvec)) {
1914 scan_balance = SCAN_FILE;
1918 scan_balance = SCAN_FRACT;
1921 * With swappiness at 100, anonymous and file have the same priority.
1922 * This scanning priority is essentially the inverse of IO cost.
1924 anon_prio = vmscan_swappiness(sc);
1925 file_prio = 200 - anon_prio;
1928 * OK, so we have swap space and a fair amount of page cache
1929 * pages. We use the recently rotated / recently scanned
1930 * ratios to determine how valuable each cache is.
1932 * Because workloads change over time (and to avoid overflow)
1933 * we keep these statistics as a floating average, which ends
1934 * up weighing recent references more than old ones.
1936 * anon in [0], file in [1]
1938 spin_lock_irq(&zone->lru_lock);
1939 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1940 reclaim_stat->recent_scanned[0] /= 2;
1941 reclaim_stat->recent_rotated[0] /= 2;
1944 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1945 reclaim_stat->recent_scanned[1] /= 2;
1946 reclaim_stat->recent_rotated[1] /= 2;
1950 * The amount of pressure on anon vs file pages is inversely
1951 * proportional to the fraction of recently scanned pages on
1952 * each list that were recently referenced and in active use.
1954 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1955 ap /= reclaim_stat->recent_rotated[0] + 1;
1957 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1958 fp /= reclaim_stat->recent_rotated[1] + 1;
1959 spin_unlock_irq(&zone->lru_lock);
1963 denominator = ap + fp + 1;
1965 for_each_evictable_lru(lru) {
1966 int file = is_file_lru(lru);
1970 size = get_lru_size(lruvec, lru);
1971 scan = size >> sc->priority;
1973 if (!scan && force_scan)
1974 scan = min(size, SWAP_CLUSTER_MAX);
1976 switch (scan_balance) {
1978 /* Scan lists relative to size */
1982 * Scan types proportional to swappiness and
1983 * their relative recent reclaim efficiency.
1985 scan = div64_u64(scan * fraction[file], denominator);
1989 /* Scan one type exclusively */
1990 if ((scan_balance == SCAN_FILE) != file)
1994 /* Look ma, no brain */
2002 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2004 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2006 unsigned long nr[NR_LRU_LISTS];
2007 unsigned long targets[NR_LRU_LISTS];
2008 unsigned long nr_to_scan;
2010 unsigned long nr_reclaimed = 0;
2011 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2012 struct blk_plug plug;
2013 bool scan_adjusted = false;
2015 get_scan_count(lruvec, sc, nr);
2017 /* Record the original scan target for proportional adjustments later */
2018 memcpy(targets, nr, sizeof(nr));
2020 blk_start_plug(&plug);
2021 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2022 nr[LRU_INACTIVE_FILE]) {
2023 unsigned long nr_anon, nr_file, percentage;
2024 unsigned long nr_scanned;
2026 for_each_evictable_lru(lru) {
2028 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2029 nr[lru] -= nr_to_scan;
2031 nr_reclaimed += shrink_list(lru, nr_to_scan,
2036 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2040 * For global direct reclaim, reclaim only the number of pages
2041 * requested. Less care is taken to scan proportionally as it
2042 * is more important to minimise direct reclaim stall latency
2043 * than it is to properly age the LRU lists.
2045 if (global_reclaim(sc) && !current_is_kswapd())
2049 * For kswapd and memcg, reclaim at least the number of pages
2050 * requested. Ensure that the anon and file LRUs shrink
2051 * proportionally what was requested by get_scan_count(). We
2052 * stop reclaiming one LRU and reduce the amount scanning
2053 * proportional to the original scan target.
2055 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2056 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2058 if (nr_file > nr_anon) {
2059 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2060 targets[LRU_ACTIVE_ANON] + 1;
2062 percentage = nr_anon * 100 / scan_target;
2064 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2065 targets[LRU_ACTIVE_FILE] + 1;
2067 percentage = nr_file * 100 / scan_target;
2070 /* Stop scanning the smaller of the LRU */
2072 nr[lru + LRU_ACTIVE] = 0;
2075 * Recalculate the other LRU scan count based on its original
2076 * scan target and the percentage scanning already complete
2078 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2079 nr_scanned = targets[lru] - nr[lru];
2080 nr[lru] = targets[lru] * (100 - percentage) / 100;
2081 nr[lru] -= min(nr[lru], nr_scanned);
2084 nr_scanned = targets[lru] - nr[lru];
2085 nr[lru] = targets[lru] * (100 - percentage) / 100;
2086 nr[lru] -= min(nr[lru], nr_scanned);
2088 scan_adjusted = true;
2090 blk_finish_plug(&plug);
2091 sc->nr_reclaimed += nr_reclaimed;
2094 * Even if we did not try to evict anon pages at all, we want to
2095 * rebalance the anon lru active/inactive ratio.
2097 if (inactive_anon_is_low(lruvec))
2098 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2099 sc, LRU_ACTIVE_ANON);
2101 throttle_vm_writeout(sc->gfp_mask);
2104 /* Use reclaim/compaction for costly allocs or under memory pressure */
2105 static bool in_reclaim_compaction(struct scan_control *sc)
2107 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2108 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2109 sc->priority < DEF_PRIORITY - 2))
2116 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2117 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2118 * true if more pages should be reclaimed such that when the page allocator
2119 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2120 * It will give up earlier than that if there is difficulty reclaiming pages.
2122 static inline bool should_continue_reclaim(struct zone *zone,
2123 unsigned long nr_reclaimed,
2124 unsigned long nr_scanned,
2125 struct scan_control *sc)
2127 unsigned long pages_for_compaction;
2128 unsigned long inactive_lru_pages;
2130 /* If not in reclaim/compaction mode, stop */
2131 if (!in_reclaim_compaction(sc))
2134 /* Consider stopping depending on scan and reclaim activity */
2135 if (sc->gfp_mask & __GFP_REPEAT) {
2137 * For __GFP_REPEAT allocations, stop reclaiming if the
2138 * full LRU list has been scanned and we are still failing
2139 * to reclaim pages. This full LRU scan is potentially
2140 * expensive but a __GFP_REPEAT caller really wants to succeed
2142 if (!nr_reclaimed && !nr_scanned)
2146 * For non-__GFP_REPEAT allocations which can presumably
2147 * fail without consequence, stop if we failed to reclaim
2148 * any pages from the last SWAP_CLUSTER_MAX number of
2149 * pages that were scanned. This will return to the
2150 * caller faster at the risk reclaim/compaction and
2151 * the resulting allocation attempt fails
2158 * If we have not reclaimed enough pages for compaction and the
2159 * inactive lists are large enough, continue reclaiming
2161 pages_for_compaction = (2UL << sc->order);
2162 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2163 if (get_nr_swap_pages() > 0)
2164 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2165 if (sc->nr_reclaimed < pages_for_compaction &&
2166 inactive_lru_pages > pages_for_compaction)
2169 /* If compaction would go ahead or the allocation would succeed, stop */
2170 switch (compaction_suitable(zone, sc->order)) {
2171 case COMPACT_PARTIAL:
2172 case COMPACT_CONTINUE:
2180 __shrink_zone(struct zone *zone, struct scan_control *sc, bool soft_reclaim)
2182 unsigned long nr_reclaimed, nr_scanned;
2185 struct mem_cgroup *root = sc->target_mem_cgroup;
2186 struct mem_cgroup_reclaim_cookie reclaim = {
2188 .priority = sc->priority,
2190 struct mem_cgroup *memcg = NULL;
2191 mem_cgroup_iter_filter filter = (soft_reclaim) ?
2192 mem_cgroup_soft_reclaim_eligible : NULL;
2194 nr_reclaimed = sc->nr_reclaimed;
2195 nr_scanned = sc->nr_scanned;
2197 while ((memcg = mem_cgroup_iter_cond(root, memcg, &reclaim, filter))) {
2198 struct lruvec *lruvec;
2200 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2202 shrink_lruvec(lruvec, sc);
2205 * Direct reclaim and kswapd have to scan all memory
2206 * cgroups to fulfill the overall scan target for the
2209 * Limit reclaim, on the other hand, only cares about
2210 * nr_to_reclaim pages to be reclaimed and it will
2211 * retry with decreasing priority if one round over the
2212 * whole hierarchy is not sufficient.
2214 if (!global_reclaim(sc) &&
2215 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2216 mem_cgroup_iter_break(root, memcg);
2221 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2222 sc->nr_scanned - nr_scanned,
2223 sc->nr_reclaimed - nr_reclaimed);
2225 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2226 sc->nr_scanned - nr_scanned, sc));
2230 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2232 bool do_soft_reclaim = mem_cgroup_should_soft_reclaim(sc);
2233 unsigned long nr_scanned = sc->nr_scanned;
2235 __shrink_zone(zone, sc, do_soft_reclaim);
2238 * No group is over the soft limit or those that are do not have
2239 * pages in the zone we are reclaiming so we have to reclaim everybody
2241 if (do_soft_reclaim && (sc->nr_scanned == nr_scanned)) {
2242 __shrink_zone(zone, sc, false);
2247 /* Returns true if compaction should go ahead for a high-order request */
2248 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2250 unsigned long balance_gap, watermark;
2253 /* Do not consider compaction for orders reclaim is meant to satisfy */
2254 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2258 * Compaction takes time to run and there are potentially other
2259 * callers using the pages just freed. Continue reclaiming until
2260 * there is a buffer of free pages available to give compaction
2261 * a reasonable chance of completing and allocating the page
2263 balance_gap = min(low_wmark_pages(zone),
2264 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2265 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2266 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2267 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2270 * If compaction is deferred, reclaim up to a point where
2271 * compaction will have a chance of success when re-enabled
2273 if (compaction_deferred(zone, sc->order))
2274 return watermark_ok;
2276 /* If compaction is not ready to start, keep reclaiming */
2277 if (!compaction_suitable(zone, sc->order))
2280 return watermark_ok;
2284 * This is the direct reclaim path, for page-allocating processes. We only
2285 * try to reclaim pages from zones which will satisfy the caller's allocation
2288 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2290 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2292 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2293 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2294 * zone defense algorithm.
2296 * If a zone is deemed to be full of pinned pages then just give it a light
2297 * scan then give up on it.
2299 * This function returns true if a zone is being reclaimed for a costly
2300 * high-order allocation and compaction is ready to begin. This indicates to
2301 * the caller that it should consider retrying the allocation instead of
2304 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2308 bool aborted_reclaim = false;
2311 * If the number of buffer_heads in the machine exceeds the maximum
2312 * allowed level, force direct reclaim to scan the highmem zone as
2313 * highmem pages could be pinning lowmem pages storing buffer_heads
2315 if (buffer_heads_over_limit)
2316 sc->gfp_mask |= __GFP_HIGHMEM;
2318 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2319 gfp_zone(sc->gfp_mask), sc->nodemask) {
2320 if (!populated_zone(zone))
2323 * Take care memory controller reclaiming has small influence
2326 if (global_reclaim(sc)) {
2327 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2329 if (sc->priority != DEF_PRIORITY &&
2330 !zone_reclaimable(zone))
2331 continue; /* Let kswapd poll it */
2332 if (IS_ENABLED(CONFIG_COMPACTION)) {
2334 * If we already have plenty of memory free for
2335 * compaction in this zone, don't free any more.
2336 * Even though compaction is invoked for any
2337 * non-zero order, only frequent costly order
2338 * reclamation is disruptive enough to become a
2339 * noticeable problem, like transparent huge
2342 if (compaction_ready(zone, sc)) {
2343 aborted_reclaim = true;
2347 /* need some check for avoid more shrink_zone() */
2350 shrink_zone(zone, sc);
2353 return aborted_reclaim;
2356 /* All zones in zonelist are unreclaimable? */
2357 static bool all_unreclaimable(struct zonelist *zonelist,
2358 struct scan_control *sc)
2363 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2364 gfp_zone(sc->gfp_mask), sc->nodemask) {
2365 if (!populated_zone(zone))
2367 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2369 if (zone_reclaimable(zone))
2377 * This is the main entry point to direct page reclaim.
2379 * If a full scan of the inactive list fails to free enough memory then we
2380 * are "out of memory" and something needs to be killed.
2382 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2383 * high - the zone may be full of dirty or under-writeback pages, which this
2384 * caller can't do much about. We kick the writeback threads and take explicit
2385 * naps in the hope that some of these pages can be written. But if the
2386 * allocating task holds filesystem locks which prevent writeout this might not
2387 * work, and the allocation attempt will fail.
2389 * returns: 0, if no pages reclaimed
2390 * else, the number of pages reclaimed
2392 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2393 struct scan_control *sc,
2394 struct shrink_control *shrink)
2396 unsigned long total_scanned = 0;
2397 struct reclaim_state *reclaim_state = current->reclaim_state;
2400 unsigned long writeback_threshold;
2401 bool aborted_reclaim;
2403 delayacct_freepages_start();
2405 if (global_reclaim(sc))
2406 count_vm_event(ALLOCSTALL);
2409 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2412 aborted_reclaim = shrink_zones(zonelist, sc);
2415 * Don't shrink slabs when reclaiming memory from over limit
2416 * cgroups but do shrink slab at least once when aborting
2417 * reclaim for compaction to avoid unevenly scanning file/anon
2418 * LRU pages over slab pages.
2420 if (global_reclaim(sc)) {
2421 unsigned long lru_pages = 0;
2423 nodes_clear(shrink->nodes_to_scan);
2424 for_each_zone_zonelist(zone, z, zonelist,
2425 gfp_zone(sc->gfp_mask)) {
2426 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2429 lru_pages += zone_reclaimable_pages(zone);
2430 node_set(zone_to_nid(zone),
2431 shrink->nodes_to_scan);
2434 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2435 if (reclaim_state) {
2436 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2437 reclaim_state->reclaimed_slab = 0;
2440 total_scanned += sc->nr_scanned;
2441 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2445 * If we're getting trouble reclaiming, start doing
2446 * writepage even in laptop mode.
2448 if (sc->priority < DEF_PRIORITY - 2)
2449 sc->may_writepage = 1;
2452 * Try to write back as many pages as we just scanned. This
2453 * tends to cause slow streaming writers to write data to the
2454 * disk smoothly, at the dirtying rate, which is nice. But
2455 * that's undesirable in laptop mode, where we *want* lumpy
2456 * writeout. So in laptop mode, write out the whole world.
2458 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2459 if (total_scanned > writeback_threshold) {
2460 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2461 WB_REASON_TRY_TO_FREE_PAGES);
2462 sc->may_writepage = 1;
2464 } while (--sc->priority >= 0 && !aborted_reclaim);
2467 delayacct_freepages_end();
2469 if (sc->nr_reclaimed)
2470 return sc->nr_reclaimed;
2473 * As hibernation is going on, kswapd is freezed so that it can't mark
2474 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2477 if (oom_killer_disabled)
2480 /* Aborted reclaim to try compaction? don't OOM, then */
2481 if (aborted_reclaim)
2484 /* top priority shrink_zones still had more to do? don't OOM, then */
2485 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2491 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2494 unsigned long pfmemalloc_reserve = 0;
2495 unsigned long free_pages = 0;
2499 for (i = 0; i <= ZONE_NORMAL; i++) {
2500 zone = &pgdat->node_zones[i];
2501 pfmemalloc_reserve += min_wmark_pages(zone);
2502 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2505 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2507 /* kswapd must be awake if processes are being throttled */
2508 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2509 pgdat->classzone_idx = min(pgdat->classzone_idx,
2510 (enum zone_type)ZONE_NORMAL);
2511 wake_up_interruptible(&pgdat->kswapd_wait);
2518 * Throttle direct reclaimers if backing storage is backed by the network
2519 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2520 * depleted. kswapd will continue to make progress and wake the processes
2521 * when the low watermark is reached.
2523 * Returns true if a fatal signal was delivered during throttling. If this
2524 * happens, the page allocator should not consider triggering the OOM killer.
2526 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2527 nodemask_t *nodemask)
2530 int high_zoneidx = gfp_zone(gfp_mask);
2534 * Kernel threads should not be throttled as they may be indirectly
2535 * responsible for cleaning pages necessary for reclaim to make forward
2536 * progress. kjournald for example may enter direct reclaim while
2537 * committing a transaction where throttling it could forcing other
2538 * processes to block on log_wait_commit().
2540 if (current->flags & PF_KTHREAD)
2544 * If a fatal signal is pending, this process should not throttle.
2545 * It should return quickly so it can exit and free its memory
2547 if (fatal_signal_pending(current))
2550 /* Check if the pfmemalloc reserves are ok */
2551 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2552 pgdat = zone->zone_pgdat;
2553 if (pfmemalloc_watermark_ok(pgdat))
2556 /* Account for the throttling */
2557 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2560 * If the caller cannot enter the filesystem, it's possible that it
2561 * is due to the caller holding an FS lock or performing a journal
2562 * transaction in the case of a filesystem like ext[3|4]. In this case,
2563 * it is not safe to block on pfmemalloc_wait as kswapd could be
2564 * blocked waiting on the same lock. Instead, throttle for up to a
2565 * second before continuing.
2567 if (!(gfp_mask & __GFP_FS)) {
2568 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2569 pfmemalloc_watermark_ok(pgdat), HZ);
2574 /* Throttle until kswapd wakes the process */
2575 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2576 pfmemalloc_watermark_ok(pgdat));
2579 if (fatal_signal_pending(current))
2586 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2587 gfp_t gfp_mask, nodemask_t *nodemask)
2589 unsigned long nr_reclaimed;
2590 struct scan_control sc = {
2591 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2592 .may_writepage = !laptop_mode,
2593 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2597 .priority = DEF_PRIORITY,
2598 .target_mem_cgroup = NULL,
2599 .nodemask = nodemask,
2601 struct shrink_control shrink = {
2602 .gfp_mask = sc.gfp_mask,
2606 * Do not enter reclaim if fatal signal was delivered while throttled.
2607 * 1 is returned so that the page allocator does not OOM kill at this
2610 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2613 trace_mm_vmscan_direct_reclaim_begin(order,
2617 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2619 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2621 return nr_reclaimed;
2626 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2627 gfp_t gfp_mask, bool noswap,
2629 unsigned long *nr_scanned)
2631 struct scan_control sc = {
2633 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2634 .may_writepage = !laptop_mode,
2636 .may_swap = !noswap,
2639 .target_mem_cgroup = memcg,
2641 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2643 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2644 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2646 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2651 * NOTE: Although we can get the priority field, using it
2652 * here is not a good idea, since it limits the pages we can scan.
2653 * if we don't reclaim here, the shrink_zone from balance_pgdat
2654 * will pick up pages from other mem cgroup's as well. We hack
2655 * the priority and make it zero.
2657 shrink_lruvec(lruvec, &sc);
2659 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2661 *nr_scanned = sc.nr_scanned;
2662 return sc.nr_reclaimed;
2665 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2669 struct zonelist *zonelist;
2670 unsigned long nr_reclaimed;
2672 struct scan_control sc = {
2673 .may_writepage = !laptop_mode,
2675 .may_swap = !noswap,
2676 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2678 .priority = DEF_PRIORITY,
2679 .target_mem_cgroup = memcg,
2680 .nodemask = NULL, /* we don't care the placement */
2681 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2682 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2684 struct shrink_control shrink = {
2685 .gfp_mask = sc.gfp_mask,
2689 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2690 * take care of from where we get pages. So the node where we start the
2691 * scan does not need to be the current node.
2693 nid = mem_cgroup_select_victim_node(memcg);
2695 zonelist = NODE_DATA(nid)->node_zonelists;
2697 trace_mm_vmscan_memcg_reclaim_begin(0,
2701 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2703 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2705 return nr_reclaimed;
2709 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2711 struct mem_cgroup *memcg;
2713 if (!total_swap_pages)
2716 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2718 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2720 if (inactive_anon_is_low(lruvec))
2721 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2722 sc, LRU_ACTIVE_ANON);
2724 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2728 static bool zone_balanced(struct zone *zone, int order,
2729 unsigned long balance_gap, int classzone_idx)
2731 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2732 balance_gap, classzone_idx, 0))
2735 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2736 !compaction_suitable(zone, order))
2743 * pgdat_balanced() is used when checking if a node is balanced.
2745 * For order-0, all zones must be balanced!
2747 * For high-order allocations only zones that meet watermarks and are in a
2748 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2749 * total of balanced pages must be at least 25% of the zones allowed by
2750 * classzone_idx for the node to be considered balanced. Forcing all zones to
2751 * be balanced for high orders can cause excessive reclaim when there are
2753 * The choice of 25% is due to
2754 * o a 16M DMA zone that is balanced will not balance a zone on any
2755 * reasonable sized machine
2756 * o On all other machines, the top zone must be at least a reasonable
2757 * percentage of the middle zones. For example, on 32-bit x86, highmem
2758 * would need to be at least 256M for it to be balance a whole node.
2759 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2760 * to balance a node on its own. These seemed like reasonable ratios.
2762 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2764 unsigned long managed_pages = 0;
2765 unsigned long balanced_pages = 0;
2768 /* Check the watermark levels */
2769 for (i = 0; i <= classzone_idx; i++) {
2770 struct zone *zone = pgdat->node_zones + i;
2772 if (!populated_zone(zone))
2775 managed_pages += zone->managed_pages;
2778 * A special case here:
2780 * balance_pgdat() skips over all_unreclaimable after
2781 * DEF_PRIORITY. Effectively, it considers them balanced so
2782 * they must be considered balanced here as well!
2784 if (!zone_reclaimable(zone)) {
2785 balanced_pages += zone->managed_pages;
2789 if (zone_balanced(zone, order, 0, i))
2790 balanced_pages += zone->managed_pages;
2796 return balanced_pages >= (managed_pages >> 2);
2802 * Prepare kswapd for sleeping. This verifies that there are no processes
2803 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2805 * Returns true if kswapd is ready to sleep
2807 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2810 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2815 * There is a potential race between when kswapd checks its watermarks
2816 * and a process gets throttled. There is also a potential race if
2817 * processes get throttled, kswapd wakes, a large process exits therby
2818 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2819 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2820 * so wake them now if necessary. If necessary, processes will wake
2821 * kswapd and get throttled again
2823 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2824 wake_up(&pgdat->pfmemalloc_wait);
2828 return pgdat_balanced(pgdat, order, classzone_idx);
2832 * kswapd shrinks the zone by the number of pages required to reach
2833 * the high watermark.
2835 * Returns true if kswapd scanned at least the requested number of pages to
2836 * reclaim or if the lack of progress was due to pages under writeback.
2837 * This is used to determine if the scanning priority needs to be raised.
2839 static bool kswapd_shrink_zone(struct zone *zone,
2841 struct scan_control *sc,
2842 unsigned long lru_pages,
2843 unsigned long *nr_attempted)
2845 int testorder = sc->order;
2846 unsigned long balance_gap;
2847 struct reclaim_state *reclaim_state = current->reclaim_state;
2848 struct shrink_control shrink = {
2849 .gfp_mask = sc->gfp_mask,
2851 bool lowmem_pressure;
2853 /* Reclaim above the high watermark. */
2854 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2857 * Kswapd reclaims only single pages with compaction enabled. Trying
2858 * too hard to reclaim until contiguous free pages have become
2859 * available can hurt performance by evicting too much useful data
2860 * from memory. Do not reclaim more than needed for compaction.
2862 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2863 compaction_suitable(zone, sc->order) !=
2868 * We put equal pressure on every zone, unless one zone has way too
2869 * many pages free already. The "too many pages" is defined as the
2870 * high wmark plus a "gap" where the gap is either the low
2871 * watermark or 1% of the zone, whichever is smaller.
2873 balance_gap = min(low_wmark_pages(zone),
2874 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2875 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2878 * If there is no low memory pressure or the zone is balanced then no
2879 * reclaim is necessary
2881 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2882 if (!lowmem_pressure && zone_balanced(zone, testorder,
2883 balance_gap, classzone_idx))
2886 shrink_zone(zone, sc);
2887 nodes_clear(shrink.nodes_to_scan);
2888 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2890 reclaim_state->reclaimed_slab = 0;
2891 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2892 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2894 /* Account for the number of pages attempted to reclaim */
2895 *nr_attempted += sc->nr_to_reclaim;
2897 zone_clear_flag(zone, ZONE_WRITEBACK);
2900 * If a zone reaches its high watermark, consider it to be no longer
2901 * congested. It's possible there are dirty pages backed by congested
2902 * BDIs but as pressure is relieved, speculatively avoid congestion
2905 if (zone_reclaimable(zone) &&
2906 zone_balanced(zone, testorder, 0, classzone_idx)) {
2907 zone_clear_flag(zone, ZONE_CONGESTED);
2908 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2911 return sc->nr_scanned >= sc->nr_to_reclaim;
2915 * For kswapd, balance_pgdat() will work across all this node's zones until
2916 * they are all at high_wmark_pages(zone).
2918 * Returns the final order kswapd was reclaiming at
2920 * There is special handling here for zones which are full of pinned pages.
2921 * This can happen if the pages are all mlocked, or if they are all used by
2922 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2923 * What we do is to detect the case where all pages in the zone have been
2924 * scanned twice and there has been zero successful reclaim. Mark the zone as
2925 * dead and from now on, only perform a short scan. Basically we're polling
2926 * the zone for when the problem goes away.
2928 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2929 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2930 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2931 * lower zones regardless of the number of free pages in the lower zones. This
2932 * interoperates with the page allocator fallback scheme to ensure that aging
2933 * of pages is balanced across the zones.
2935 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2939 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2940 struct scan_control sc = {
2941 .gfp_mask = GFP_KERNEL,
2942 .priority = DEF_PRIORITY,
2945 .may_writepage = !laptop_mode,
2947 .target_mem_cgroup = NULL,
2949 count_vm_event(PAGEOUTRUN);
2952 unsigned long lru_pages = 0;
2953 unsigned long nr_attempted = 0;
2954 bool raise_priority = true;
2955 bool pgdat_needs_compaction = (order > 0);
2957 sc.nr_reclaimed = 0;
2960 * Scan in the highmem->dma direction for the highest
2961 * zone which needs scanning
2963 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2964 struct zone *zone = pgdat->node_zones + i;
2966 if (!populated_zone(zone))
2969 if (sc.priority != DEF_PRIORITY &&
2970 !zone_reclaimable(zone))
2974 * Do some background aging of the anon list, to give
2975 * pages a chance to be referenced before reclaiming.
2977 age_active_anon(zone, &sc);
2980 * If the number of buffer_heads in the machine
2981 * exceeds the maximum allowed level and this node
2982 * has a highmem zone, force kswapd to reclaim from
2983 * it to relieve lowmem pressure.
2985 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2990 if (!zone_balanced(zone, order, 0, 0)) {
2995 * If balanced, clear the dirty and congested
2998 zone_clear_flag(zone, ZONE_CONGESTED);
2999 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3006 for (i = 0; i <= end_zone; i++) {
3007 struct zone *zone = pgdat->node_zones + i;
3009 if (!populated_zone(zone))
3012 lru_pages += zone_reclaimable_pages(zone);
3015 * If any zone is currently balanced then kswapd will
3016 * not call compaction as it is expected that the
3017 * necessary pages are already available.
3019 if (pgdat_needs_compaction &&
3020 zone_watermark_ok(zone, order,
3021 low_wmark_pages(zone),
3023 pgdat_needs_compaction = false;
3027 * If we're getting trouble reclaiming, start doing writepage
3028 * even in laptop mode.
3030 if (sc.priority < DEF_PRIORITY - 2)
3031 sc.may_writepage = 1;
3034 * Now scan the zone in the dma->highmem direction, stopping
3035 * at the last zone which needs scanning.
3037 * We do this because the page allocator works in the opposite
3038 * direction. This prevents the page allocator from allocating
3039 * pages behind kswapd's direction of progress, which would
3040 * cause too much scanning of the lower zones.
3042 for (i = 0; i <= end_zone; i++) {
3043 struct zone *zone = pgdat->node_zones + i;
3045 if (!populated_zone(zone))
3048 if (sc.priority != DEF_PRIORITY &&
3049 !zone_reclaimable(zone))
3055 * There should be no need to raise the scanning
3056 * priority if enough pages are already being scanned
3057 * that that high watermark would be met at 100%
3060 if (kswapd_shrink_zone(zone, end_zone, &sc,
3061 lru_pages, &nr_attempted))
3062 raise_priority = false;
3066 * If the low watermark is met there is no need for processes
3067 * to be throttled on pfmemalloc_wait as they should not be
3068 * able to safely make forward progress. Wake them
3070 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3071 pfmemalloc_watermark_ok(pgdat))
3072 wake_up(&pgdat->pfmemalloc_wait);
3075 * Fragmentation may mean that the system cannot be rebalanced
3076 * for high-order allocations in all zones. If twice the
3077 * allocation size has been reclaimed and the zones are still
3078 * not balanced then recheck the watermarks at order-0 to
3079 * prevent kswapd reclaiming excessively. Assume that a
3080 * process requested a high-order can direct reclaim/compact.
3082 if (order && sc.nr_reclaimed >= 2UL << order)
3083 order = sc.order = 0;
3085 /* Check if kswapd should be suspending */
3086 if (try_to_freeze() || kthread_should_stop())
3090 * Compact if necessary and kswapd is reclaiming at least the
3091 * high watermark number of pages as requsted
3093 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3094 compact_pgdat(pgdat, order);
3097 * Raise priority if scanning rate is too low or there was no
3098 * progress in reclaiming pages
3100 if (raise_priority || !sc.nr_reclaimed)
3102 } while (sc.priority >= 1 &&
3103 !pgdat_balanced(pgdat, order, *classzone_idx));
3107 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3108 * makes a decision on the order we were last reclaiming at. However,
3109 * if another caller entered the allocator slow path while kswapd
3110 * was awake, order will remain at the higher level
3112 *classzone_idx = end_zone;
3116 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3121 if (freezing(current) || kthread_should_stop())
3124 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3126 /* Try to sleep for a short interval */
3127 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3128 remaining = schedule_timeout(HZ/10);
3129 finish_wait(&pgdat->kswapd_wait, &wait);
3130 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3134 * After a short sleep, check if it was a premature sleep. If not, then
3135 * go fully to sleep until explicitly woken up.
3137 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3138 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3141 * vmstat counters are not perfectly accurate and the estimated
3142 * value for counters such as NR_FREE_PAGES can deviate from the
3143 * true value by nr_online_cpus * threshold. To avoid the zone
3144 * watermarks being breached while under pressure, we reduce the
3145 * per-cpu vmstat threshold while kswapd is awake and restore
3146 * them before going back to sleep.
3148 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3151 * Compaction records what page blocks it recently failed to
3152 * isolate pages from and skips them in the future scanning.
3153 * When kswapd is going to sleep, it is reasonable to assume
3154 * that pages and compaction may succeed so reset the cache.
3156 reset_isolation_suitable(pgdat);
3158 if (!kthread_should_stop())
3161 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3164 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3166 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3168 finish_wait(&pgdat->kswapd_wait, &wait);
3172 * The background pageout daemon, started as a kernel thread
3173 * from the init process.
3175 * This basically trickles out pages so that we have _some_
3176 * free memory available even if there is no other activity
3177 * that frees anything up. This is needed for things like routing
3178 * etc, where we otherwise might have all activity going on in
3179 * asynchronous contexts that cannot page things out.
3181 * If there are applications that are active memory-allocators
3182 * (most normal use), this basically shouldn't matter.
3184 static int kswapd(void *p)
3186 unsigned long order, new_order;
3187 unsigned balanced_order;
3188 int classzone_idx, new_classzone_idx;
3189 int balanced_classzone_idx;
3190 pg_data_t *pgdat = (pg_data_t*)p;
3191 struct task_struct *tsk = current;
3193 struct reclaim_state reclaim_state = {
3194 .reclaimed_slab = 0,
3196 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3198 lockdep_set_current_reclaim_state(GFP_KERNEL);
3200 if (!cpumask_empty(cpumask))
3201 set_cpus_allowed_ptr(tsk, cpumask);
3202 current->reclaim_state = &reclaim_state;
3205 * Tell the memory management that we're a "memory allocator",
3206 * and that if we need more memory we should get access to it
3207 * regardless (see "__alloc_pages()"). "kswapd" should
3208 * never get caught in the normal page freeing logic.
3210 * (Kswapd normally doesn't need memory anyway, but sometimes
3211 * you need a small amount of memory in order to be able to
3212 * page out something else, and this flag essentially protects
3213 * us from recursively trying to free more memory as we're
3214 * trying to free the first piece of memory in the first place).
3216 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3219 order = new_order = 0;
3221 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3222 balanced_classzone_idx = classzone_idx;
3227 * If the last balance_pgdat was unsuccessful it's unlikely a
3228 * new request of a similar or harder type will succeed soon
3229 * so consider going to sleep on the basis we reclaimed at
3231 if (balanced_classzone_idx >= new_classzone_idx &&
3232 balanced_order == new_order) {
3233 new_order = pgdat->kswapd_max_order;
3234 new_classzone_idx = pgdat->classzone_idx;
3235 pgdat->kswapd_max_order = 0;
3236 pgdat->classzone_idx = pgdat->nr_zones - 1;
3239 if (order < new_order || classzone_idx > new_classzone_idx) {
3241 * Don't sleep if someone wants a larger 'order'
3242 * allocation or has tigher zone constraints
3245 classzone_idx = new_classzone_idx;
3247 kswapd_try_to_sleep(pgdat, balanced_order,
3248 balanced_classzone_idx);
3249 order = pgdat->kswapd_max_order;
3250 classzone_idx = pgdat->classzone_idx;
3252 new_classzone_idx = classzone_idx;
3253 pgdat->kswapd_max_order = 0;
3254 pgdat->classzone_idx = pgdat->nr_zones - 1;
3257 ret = try_to_freeze();
3258 if (kthread_should_stop())
3262 * We can speed up thawing tasks if we don't call balance_pgdat
3263 * after returning from the refrigerator
3266 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3267 balanced_classzone_idx = classzone_idx;
3268 balanced_order = balance_pgdat(pgdat, order,
3269 &balanced_classzone_idx);
3273 current->reclaim_state = NULL;
3278 * A zone is low on free memory, so wake its kswapd task to service it.
3280 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3284 if (!populated_zone(zone))
3287 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3289 pgdat = zone->zone_pgdat;
3290 if (pgdat->kswapd_max_order < order) {
3291 pgdat->kswapd_max_order = order;
3292 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3294 if (!waitqueue_active(&pgdat->kswapd_wait))
3296 if (zone_balanced(zone, order, 0, 0))
3299 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3300 wake_up_interruptible(&pgdat->kswapd_wait);
3304 * The reclaimable count would be mostly accurate.
3305 * The less reclaimable pages may be
3306 * - mlocked pages, which will be moved to unevictable list when encountered
3307 * - mapped pages, which may require several travels to be reclaimed
3308 * - dirty pages, which is not "instantly" reclaimable
3310 unsigned long global_reclaimable_pages(void)
3314 nr = global_page_state(NR_ACTIVE_FILE) +
3315 global_page_state(NR_INACTIVE_FILE);
3317 if (get_nr_swap_pages() > 0)
3318 nr += global_page_state(NR_ACTIVE_ANON) +
3319 global_page_state(NR_INACTIVE_ANON);
3324 #ifdef CONFIG_HIBERNATION
3326 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3329 * Rather than trying to age LRUs the aim is to preserve the overall
3330 * LRU order by reclaiming preferentially
3331 * inactive > active > active referenced > active mapped
3333 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3335 struct reclaim_state reclaim_state;
3336 struct scan_control sc = {
3337 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3341 .nr_to_reclaim = nr_to_reclaim,
3342 .hibernation_mode = 1,
3344 .priority = DEF_PRIORITY,
3346 struct shrink_control shrink = {
3347 .gfp_mask = sc.gfp_mask,
3349 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3350 struct task_struct *p = current;
3351 unsigned long nr_reclaimed;
3353 p->flags |= PF_MEMALLOC;
3354 lockdep_set_current_reclaim_state(sc.gfp_mask);
3355 reclaim_state.reclaimed_slab = 0;
3356 p->reclaim_state = &reclaim_state;
3358 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3360 p->reclaim_state = NULL;
3361 lockdep_clear_current_reclaim_state();
3362 p->flags &= ~PF_MEMALLOC;
3364 return nr_reclaimed;
3366 #endif /* CONFIG_HIBERNATION */
3368 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3369 not required for correctness. So if the last cpu in a node goes
3370 away, we get changed to run anywhere: as the first one comes back,
3371 restore their cpu bindings. */
3372 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3377 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3378 for_each_node_state(nid, N_MEMORY) {
3379 pg_data_t *pgdat = NODE_DATA(nid);
3380 const struct cpumask *mask;
3382 mask = cpumask_of_node(pgdat->node_id);
3384 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3385 /* One of our CPUs online: restore mask */
3386 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3393 * This kswapd start function will be called by init and node-hot-add.
3394 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3396 int kswapd_run(int nid)
3398 pg_data_t *pgdat = NODE_DATA(nid);
3404 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3405 if (IS_ERR(pgdat->kswapd)) {
3406 /* failure at boot is fatal */
3407 BUG_ON(system_state == SYSTEM_BOOTING);
3408 pr_err("Failed to start kswapd on node %d\n", nid);
3409 ret = PTR_ERR(pgdat->kswapd);
3410 pgdat->kswapd = NULL;
3416 * Called by memory hotplug when all memory in a node is offlined. Caller must
3417 * hold lock_memory_hotplug().
3419 void kswapd_stop(int nid)
3421 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3424 kthread_stop(kswapd);
3425 NODE_DATA(nid)->kswapd = NULL;
3429 static int __init kswapd_init(void)
3434 for_each_node_state(nid, N_MEMORY)
3436 hotcpu_notifier(cpu_callback, 0);
3440 module_init(kswapd_init)
3446 * If non-zero call zone_reclaim when the number of free pages falls below
3449 int zone_reclaim_mode __read_mostly;
3451 #define RECLAIM_OFF 0
3452 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3453 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3454 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3457 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3458 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3461 #define ZONE_RECLAIM_PRIORITY 4
3464 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3467 int sysctl_min_unmapped_ratio = 1;
3470 * If the number of slab pages in a zone grows beyond this percentage then
3471 * slab reclaim needs to occur.
3473 int sysctl_min_slab_ratio = 5;
3475 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3477 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3478 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3479 zone_page_state(zone, NR_ACTIVE_FILE);
3482 * It's possible for there to be more file mapped pages than
3483 * accounted for by the pages on the file LRU lists because
3484 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3486 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3489 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3490 static long zone_pagecache_reclaimable(struct zone *zone)
3492 long nr_pagecache_reclaimable;
3496 * If RECLAIM_SWAP is set, then all file pages are considered
3497 * potentially reclaimable. Otherwise, we have to worry about
3498 * pages like swapcache and zone_unmapped_file_pages() provides
3501 if (zone_reclaim_mode & RECLAIM_SWAP)
3502 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3504 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3506 /* If we can't clean pages, remove dirty pages from consideration */
3507 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3508 delta += zone_page_state(zone, NR_FILE_DIRTY);
3510 /* Watch for any possible underflows due to delta */
3511 if (unlikely(delta > nr_pagecache_reclaimable))
3512 delta = nr_pagecache_reclaimable;
3514 return nr_pagecache_reclaimable - delta;
3518 * Try to free up some pages from this zone through reclaim.
3520 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3522 /* Minimum pages needed in order to stay on node */
3523 const unsigned long nr_pages = 1 << order;
3524 struct task_struct *p = current;
3525 struct reclaim_state reclaim_state;
3526 struct scan_control sc = {
3527 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3528 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3530 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3531 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3533 .priority = ZONE_RECLAIM_PRIORITY,
3535 struct shrink_control shrink = {
3536 .gfp_mask = sc.gfp_mask,
3538 unsigned long nr_slab_pages0, nr_slab_pages1;
3542 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3543 * and we also need to be able to write out pages for RECLAIM_WRITE
3546 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3547 lockdep_set_current_reclaim_state(gfp_mask);
3548 reclaim_state.reclaimed_slab = 0;
3549 p->reclaim_state = &reclaim_state;
3551 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3553 * Free memory by calling shrink zone with increasing
3554 * priorities until we have enough memory freed.
3557 shrink_zone(zone, &sc);
3558 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3561 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3562 if (nr_slab_pages0 > zone->min_slab_pages) {
3564 * shrink_slab() does not currently allow us to determine how
3565 * many pages were freed in this zone. So we take the current
3566 * number of slab pages and shake the slab until it is reduced
3567 * by the same nr_pages that we used for reclaiming unmapped
3570 nodes_clear(shrink.nodes_to_scan);
3571 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3573 unsigned long lru_pages = zone_reclaimable_pages(zone);
3575 /* No reclaimable slab or very low memory pressure */
3576 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3579 /* Freed enough memory */
3580 nr_slab_pages1 = zone_page_state(zone,
3581 NR_SLAB_RECLAIMABLE);
3582 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3587 * Update nr_reclaimed by the number of slab pages we
3588 * reclaimed from this zone.
3590 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3591 if (nr_slab_pages1 < nr_slab_pages0)
3592 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3595 p->reclaim_state = NULL;
3596 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3597 lockdep_clear_current_reclaim_state();
3598 return sc.nr_reclaimed >= nr_pages;
3601 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3607 * Zone reclaim reclaims unmapped file backed pages and
3608 * slab pages if we are over the defined limits.
3610 * A small portion of unmapped file backed pages is needed for
3611 * file I/O otherwise pages read by file I/O will be immediately
3612 * thrown out if the zone is overallocated. So we do not reclaim
3613 * if less than a specified percentage of the zone is used by
3614 * unmapped file backed pages.
3616 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3617 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3618 return ZONE_RECLAIM_FULL;
3620 if (!zone_reclaimable(zone))
3621 return ZONE_RECLAIM_FULL;
3624 * Do not scan if the allocation should not be delayed.
3626 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3627 return ZONE_RECLAIM_NOSCAN;
3630 * Only run zone reclaim on the local zone or on zones that do not
3631 * have associated processors. This will favor the local processor
3632 * over remote processors and spread off node memory allocations
3633 * as wide as possible.
3635 node_id = zone_to_nid(zone);
3636 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3637 return ZONE_RECLAIM_NOSCAN;
3639 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3640 return ZONE_RECLAIM_NOSCAN;
3642 ret = __zone_reclaim(zone, gfp_mask, order);
3643 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3646 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3653 * page_evictable - test whether a page is evictable
3654 * @page: the page to test
3656 * Test whether page is evictable--i.e., should be placed on active/inactive
3657 * lists vs unevictable list.
3659 * Reasons page might not be evictable:
3660 * (1) page's mapping marked unevictable
3661 * (2) page is part of an mlocked VMA
3664 int page_evictable(struct page *page)
3666 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3671 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3672 * @pages: array of pages to check
3673 * @nr_pages: number of pages to check
3675 * Checks pages for evictability and moves them to the appropriate lru list.
3677 * This function is only used for SysV IPC SHM_UNLOCK.
3679 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3681 struct lruvec *lruvec;
3682 struct zone *zone = NULL;
3687 for (i = 0; i < nr_pages; i++) {
3688 struct page *page = pages[i];
3689 struct zone *pagezone;
3692 pagezone = page_zone(page);
3693 if (pagezone != zone) {
3695 spin_unlock_irq(&zone->lru_lock);
3697 spin_lock_irq(&zone->lru_lock);
3699 lruvec = mem_cgroup_page_lruvec(page, zone);
3701 if (!PageLRU(page) || !PageUnevictable(page))
3704 if (page_evictable(page)) {
3705 enum lru_list lru = page_lru_base_type(page);
3707 VM_BUG_ON(PageActive(page));
3708 ClearPageUnevictable(page);
3709 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3710 add_page_to_lru_list(page, lruvec, lru);
3716 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3717 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3718 spin_unlock_irq(&zone->lru_lock);
3721 #endif /* CONFIG_SHMEM */
3723 static void warn_scan_unevictable_pages(void)
3725 printk_once(KERN_WARNING
3726 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3727 "disabled for lack of a legitimate use case. If you have "
3728 "one, please send an email to linux-mm@kvack.org.\n",
3733 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3734 * all nodes' unevictable lists for evictable pages
3736 unsigned long scan_unevictable_pages;
3738 int scan_unevictable_handler(struct ctl_table *table, int write,
3739 void __user *buffer,
3740 size_t *length, loff_t *ppos)
3742 warn_scan_unevictable_pages();
3743 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3744 scan_unevictable_pages = 0;
3750 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3751 * a specified node's per zone unevictable lists for evictable pages.
3754 static ssize_t read_scan_unevictable_node(struct device *dev,
3755 struct device_attribute *attr,
3758 warn_scan_unevictable_pages();
3759 return sprintf(buf, "0\n"); /* always zero; should fit... */
3762 static ssize_t write_scan_unevictable_node(struct device *dev,
3763 struct device_attribute *attr,
3764 const char *buf, size_t count)
3766 warn_scan_unevictable_pages();
3771 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3772 read_scan_unevictable_node,
3773 write_scan_unevictable_node);
3775 int scan_unevictable_register_node(struct node *node)
3777 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3780 void scan_unevictable_unregister_node(struct node *node)
3782 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);