4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
53 #include <linux/swapops.h>
54 #include <linux/balloon_compaction.h>
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/vmscan.h>
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
65 /* This context's GFP mask */
68 /* Allocation order */
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup *target_mem_cgroup;
83 /* Scan (total_size >> priority) pages at once */
86 unsigned int may_writepage:1;
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
94 unsigned int hibernation_mode:1;
96 /* One of the zones is ready for compaction */
97 unsigned int compaction_ready:1;
99 /* Incremented by the number of inactive pages that were scanned */
100 unsigned long nr_scanned;
102 /* Number of pages freed so far during a call to shrink_zones() */
103 unsigned long nr_reclaimed;
106 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
108 #ifdef ARCH_HAS_PREFETCH
109 #define prefetch_prev_lru_page(_page, _base, _field) \
111 if ((_page)->lru.prev != _base) { \
114 prev = lru_to_page(&(_page->lru)); \
115 prefetch(&prev->_field); \
119 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122 #ifdef ARCH_HAS_PREFETCHW
123 #define prefetchw_prev_lru_page(_page, _base, _field) \
125 if ((_page)->lru.prev != _base) { \
128 prev = lru_to_page(&(_page->lru)); \
129 prefetchw(&prev->_field); \
133 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
137 * From 0 .. 100. Higher means more swappy.
139 int vm_swappiness = 60;
141 * The total number of pages which are beyond the high watermark within all
144 unsigned long vm_total_pages;
146 static LIST_HEAD(shrinker_list);
147 static DECLARE_RWSEM(shrinker_rwsem);
150 static bool global_reclaim(struct scan_control *sc)
152 return !sc->target_mem_cgroup;
155 static bool global_reclaim(struct scan_control *sc)
161 static 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);
225 kfree(shrinker->nr_deferred);
227 EXPORT_SYMBOL(unregister_shrinker);
229 #define SHRINK_BATCH 128
232 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
233 unsigned long nr_pages_scanned, unsigned long lru_pages)
235 unsigned long freed = 0;
236 unsigned long long delta;
241 int nid = shrinkctl->nid;
242 long batch_size = shrinker->batch ? shrinker->batch
245 freeable = shrinker->count_objects(shrinker, shrinkctl);
250 * copy the current shrinker scan count into a local variable
251 * and zero it so that other concurrent shrinker invocations
252 * don't also do this scanning work.
254 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
257 delta = (4 * nr_pages_scanned) / shrinker->seeks;
259 do_div(delta, lru_pages + 1);
261 if (total_scan < 0) {
263 "shrink_slab: %pF negative objects to delete nr=%ld\n",
264 shrinker->scan_objects, total_scan);
265 total_scan = freeable;
269 * We need to avoid excessive windup on filesystem shrinkers
270 * due to large numbers of GFP_NOFS allocations causing the
271 * shrinkers to return -1 all the time. This results in a large
272 * nr being built up so when a shrink that can do some work
273 * comes along it empties the entire cache due to nr >>>
274 * freeable. This is bad for sustaining a working set in
277 * Hence only allow the shrinker to scan the entire cache when
278 * a large delta change is calculated directly.
280 if (delta < freeable / 4)
281 total_scan = min(total_scan, freeable / 2);
284 * Avoid risking looping forever due to too large nr value:
285 * never try to free more than twice the estimate number of
288 if (total_scan > freeable * 2)
289 total_scan = freeable * 2;
291 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
292 nr_pages_scanned, lru_pages,
293 freeable, delta, total_scan);
296 * Normally, we should not scan less than batch_size objects in one
297 * pass to avoid too frequent shrinker calls, but if the slab has less
298 * than batch_size objects in total and we are really tight on memory,
299 * we will try to reclaim all available objects, otherwise we can end
300 * up failing allocations although there are plenty of reclaimable
301 * objects spread over several slabs with usage less than the
304 * We detect the "tight on memory" situations by looking at the total
305 * number of objects we want to scan (total_scan). If it is greater
306 * than the total number of objects on slab (freeable), we must be
307 * scanning at high prio and therefore should try to reclaim as much as
310 while (total_scan >= batch_size ||
311 total_scan >= freeable) {
313 unsigned long nr_to_scan = min(batch_size, total_scan);
315 shrinkctl->nr_to_scan = nr_to_scan;
316 ret = shrinker->scan_objects(shrinker, shrinkctl);
317 if (ret == SHRINK_STOP)
321 count_vm_events(SLABS_SCANNED, nr_to_scan);
322 total_scan -= nr_to_scan;
328 * move the unused scan count back into the shrinker in a
329 * manner that handles concurrent updates. If we exhausted the
330 * scan, there is no need to do an update.
333 new_nr = atomic_long_add_return(total_scan,
334 &shrinker->nr_deferred[nid]);
336 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
338 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
343 * Call the shrink functions to age shrinkable caches
345 * Here we assume it costs one seek to replace a lru page and that it also
346 * takes a seek to recreate a cache object. With this in mind we age equal
347 * percentages of the lru and ageable caches. This should balance the seeks
348 * generated by these structures.
350 * If the vm encountered mapped pages on the LRU it increase the pressure on
351 * slab to avoid swapping.
353 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
355 * `lru_pages' represents the number of on-LRU pages in all the zones which
356 * are eligible for the caller's allocation attempt. It is used for balancing
357 * slab reclaim versus page reclaim.
359 * Returns the number of slab objects which we shrunk.
361 unsigned long shrink_slab(struct shrink_control *shrinkctl,
362 unsigned long nr_pages_scanned,
363 unsigned long lru_pages)
365 struct shrinker *shrinker;
366 unsigned long freed = 0;
368 if (nr_pages_scanned == 0)
369 nr_pages_scanned = SWAP_CLUSTER_MAX;
371 if (!down_read_trylock(&shrinker_rwsem)) {
373 * If we would return 0, our callers would understand that we
374 * have nothing else to shrink and give up trying. By returning
375 * 1 we keep it going and assume we'll be able to shrink next
382 list_for_each_entry(shrinker, &shrinker_list, list) {
383 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
385 freed += shrink_slab_node(shrinkctl, shrinker,
386 nr_pages_scanned, lru_pages);
390 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
391 if (node_online(shrinkctl->nid))
392 freed += shrink_slab_node(shrinkctl, shrinker,
393 nr_pages_scanned, lru_pages);
397 up_read(&shrinker_rwsem);
403 static inline int is_page_cache_freeable(struct page *page)
406 * A freeable page cache page is referenced only by the caller
407 * that isolated the page, the page cache radix tree and
408 * optional buffer heads at page->private.
410 return page_count(page) - page_has_private(page) == 2;
413 static int may_write_to_queue(struct backing_dev_info *bdi,
414 struct scan_control *sc)
416 if (current->flags & PF_SWAPWRITE)
418 if (!bdi_write_congested(bdi))
420 if (bdi == current->backing_dev_info)
426 * We detected a synchronous write error writing a page out. Probably
427 * -ENOSPC. We need to propagate that into the address_space for a subsequent
428 * fsync(), msync() or close().
430 * The tricky part is that after writepage we cannot touch the mapping: nothing
431 * prevents it from being freed up. But we have a ref on the page and once
432 * that page is locked, the mapping is pinned.
434 * We're allowed to run sleeping lock_page() here because we know the caller has
437 static void handle_write_error(struct address_space *mapping,
438 struct page *page, int error)
441 if (page_mapping(page) == mapping)
442 mapping_set_error(mapping, error);
446 /* possible outcome of pageout() */
448 /* failed to write page out, page is locked */
450 /* move page to the active list, page is locked */
452 /* page has been sent to the disk successfully, page is unlocked */
454 /* page is clean and locked */
459 * pageout is called by shrink_page_list() for each dirty page.
460 * Calls ->writepage().
462 static pageout_t pageout(struct page *page, struct address_space *mapping,
463 struct scan_control *sc)
466 * If the page is dirty, only perform writeback if that write
467 * will be non-blocking. To prevent this allocation from being
468 * stalled by pagecache activity. But note that there may be
469 * stalls if we need to run get_block(). We could test
470 * PagePrivate for that.
472 * If this process is currently in __generic_file_write_iter() against
473 * this page's queue, we can perform writeback even if that
476 * If the page is swapcache, write it back even if that would
477 * block, for some throttling. This happens by accident, because
478 * swap_backing_dev_info is bust: it doesn't reflect the
479 * congestion state of the swapdevs. Easy to fix, if needed.
481 if (!is_page_cache_freeable(page))
485 * Some data journaling orphaned pages can have
486 * page->mapping == NULL while being dirty with clean buffers.
488 if (page_has_private(page)) {
489 if (try_to_free_buffers(page)) {
490 ClearPageDirty(page);
491 pr_info("%s: orphaned page\n", __func__);
497 if (mapping->a_ops->writepage == NULL)
498 return PAGE_ACTIVATE;
499 if (!may_write_to_queue(mapping->backing_dev_info, sc))
502 if (clear_page_dirty_for_io(page)) {
504 struct writeback_control wbc = {
505 .sync_mode = WB_SYNC_NONE,
506 .nr_to_write = SWAP_CLUSTER_MAX,
508 .range_end = LLONG_MAX,
512 SetPageReclaim(page);
513 res = mapping->a_ops->writepage(page, &wbc);
515 handle_write_error(mapping, page, res);
516 if (res == AOP_WRITEPAGE_ACTIVATE) {
517 ClearPageReclaim(page);
518 return PAGE_ACTIVATE;
521 if (!PageWriteback(page)) {
522 /* synchronous write or broken a_ops? */
523 ClearPageReclaim(page);
525 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
526 inc_zone_page_state(page, NR_VMSCAN_WRITE);
534 * Same as remove_mapping, but if the page is removed from the mapping, it
535 * gets returned with a refcount of 0.
537 static int __remove_mapping(struct address_space *mapping, struct page *page,
540 BUG_ON(!PageLocked(page));
541 BUG_ON(mapping != page_mapping(page));
543 spin_lock_irq(&mapping->tree_lock);
545 * The non racy check for a busy page.
547 * Must be careful with the order of the tests. When someone has
548 * a ref to the page, it may be possible that they dirty it then
549 * drop the reference. So if PageDirty is tested before page_count
550 * here, then the following race may occur:
552 * get_user_pages(&page);
553 * [user mapping goes away]
555 * !PageDirty(page) [good]
556 * SetPageDirty(page);
558 * !page_count(page) [good, discard it]
560 * [oops, our write_to data is lost]
562 * Reversing the order of the tests ensures such a situation cannot
563 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
564 * load is not satisfied before that of page->_count.
566 * Note that if SetPageDirty is always performed via set_page_dirty,
567 * and thus under tree_lock, then this ordering is not required.
569 if (!page_freeze_refs(page, 2))
571 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
572 if (unlikely(PageDirty(page))) {
573 page_unfreeze_refs(page, 2);
577 if (PageSwapCache(page)) {
578 swp_entry_t swap = { .val = page_private(page) };
579 __delete_from_swap_cache(page);
580 spin_unlock_irq(&mapping->tree_lock);
581 swapcache_free(swap, page);
583 void (*freepage)(struct page *);
586 freepage = mapping->a_ops->freepage;
588 * Remember a shadow entry for reclaimed file cache in
589 * order to detect refaults, thus thrashing, later on.
591 * But don't store shadows in an address space that is
592 * already exiting. This is not just an optizimation,
593 * inode reclaim needs to empty out the radix tree or
594 * the nodes are lost. Don't plant shadows behind its
597 if (reclaimed && page_is_file_cache(page) &&
598 !mapping_exiting(mapping))
599 shadow = workingset_eviction(mapping, page);
600 __delete_from_page_cache(page, shadow);
601 spin_unlock_irq(&mapping->tree_lock);
602 mem_cgroup_uncharge_cache_page(page);
604 if (freepage != NULL)
611 spin_unlock_irq(&mapping->tree_lock);
616 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
617 * someone else has a ref on the page, abort and return 0. If it was
618 * successfully detached, return 1. Assumes the caller has a single ref on
621 int remove_mapping(struct address_space *mapping, struct page *page)
623 if (__remove_mapping(mapping, page, false)) {
625 * Unfreezing the refcount with 1 rather than 2 effectively
626 * drops the pagecache ref for us without requiring another
629 page_unfreeze_refs(page, 1);
636 * putback_lru_page - put previously isolated page onto appropriate LRU list
637 * @page: page to be put back to appropriate lru list
639 * Add previously isolated @page to appropriate LRU list.
640 * Page may still be unevictable for other reasons.
642 * lru_lock must not be held, interrupts must be enabled.
644 void putback_lru_page(struct page *page)
647 int was_unevictable = PageUnevictable(page);
649 VM_BUG_ON_PAGE(PageLRU(page), page);
652 ClearPageUnevictable(page);
654 if (page_evictable(page)) {
656 * For evictable pages, we can use the cache.
657 * In event of a race, worst case is we end up with an
658 * unevictable page on [in]active list.
659 * We know how to handle that.
661 is_unevictable = false;
665 * Put unevictable pages directly on zone's unevictable
668 is_unevictable = true;
669 add_page_to_unevictable_list(page);
671 * When racing with an mlock or AS_UNEVICTABLE clearing
672 * (page is unlocked) make sure that if the other thread
673 * does not observe our setting of PG_lru and fails
674 * isolation/check_move_unevictable_pages,
675 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
676 * the page back to the evictable list.
678 * The other side is TestClearPageMlocked() or shmem_lock().
684 * page's status can change while we move it among lru. If an evictable
685 * page is on unevictable list, it never be freed. To avoid that,
686 * check after we added it to the list, again.
688 if (is_unevictable && page_evictable(page)) {
689 if (!isolate_lru_page(page)) {
693 /* This means someone else dropped this page from LRU
694 * So, it will be freed or putback to LRU again. There is
695 * nothing to do here.
699 if (was_unevictable && !is_unevictable)
700 count_vm_event(UNEVICTABLE_PGRESCUED);
701 else if (!was_unevictable && is_unevictable)
702 count_vm_event(UNEVICTABLE_PGCULLED);
704 put_page(page); /* drop ref from isolate */
707 enum page_references {
709 PAGEREF_RECLAIM_CLEAN,
714 static enum page_references page_check_references(struct page *page,
715 struct scan_control *sc)
717 int referenced_ptes, referenced_page;
718 unsigned long vm_flags;
720 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
722 referenced_page = TestClearPageReferenced(page);
725 * Mlock lost the isolation race with us. Let try_to_unmap()
726 * move the page to the unevictable list.
728 if (vm_flags & VM_LOCKED)
729 return PAGEREF_RECLAIM;
731 if (referenced_ptes) {
732 if (PageSwapBacked(page))
733 return PAGEREF_ACTIVATE;
735 * All mapped pages start out with page table
736 * references from the instantiating fault, so we need
737 * to look twice if a mapped file page is used more
740 * Mark it and spare it for another trip around the
741 * inactive list. Another page table reference will
742 * lead to its activation.
744 * Note: the mark is set for activated pages as well
745 * so that recently deactivated but used pages are
748 SetPageReferenced(page);
750 if (referenced_page || referenced_ptes > 1)
751 return PAGEREF_ACTIVATE;
754 * Activate file-backed executable pages after first usage.
756 if (vm_flags & VM_EXEC)
757 return PAGEREF_ACTIVATE;
762 /* Reclaim if clean, defer dirty pages to writeback */
763 if (referenced_page && !PageSwapBacked(page))
764 return PAGEREF_RECLAIM_CLEAN;
766 return PAGEREF_RECLAIM;
769 /* Check if a page is dirty or under writeback */
770 static void page_check_dirty_writeback(struct page *page,
771 bool *dirty, bool *writeback)
773 struct address_space *mapping;
776 * Anonymous pages are not handled by flushers and must be written
777 * from reclaim context. Do not stall reclaim based on them
779 if (!page_is_file_cache(page)) {
785 /* By default assume that the page flags are accurate */
786 *dirty = PageDirty(page);
787 *writeback = PageWriteback(page);
789 /* Verify dirty/writeback state if the filesystem supports it */
790 if (!page_has_private(page))
793 mapping = page_mapping(page);
794 if (mapping && mapping->a_ops->is_dirty_writeback)
795 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
799 * shrink_page_list() returns the number of reclaimed pages
801 static unsigned long shrink_page_list(struct list_head *page_list,
803 struct scan_control *sc,
804 enum ttu_flags ttu_flags,
805 unsigned long *ret_nr_dirty,
806 unsigned long *ret_nr_unqueued_dirty,
807 unsigned long *ret_nr_congested,
808 unsigned long *ret_nr_writeback,
809 unsigned long *ret_nr_immediate,
812 LIST_HEAD(ret_pages);
813 LIST_HEAD(free_pages);
815 unsigned long nr_unqueued_dirty = 0;
816 unsigned long nr_dirty = 0;
817 unsigned long nr_congested = 0;
818 unsigned long nr_reclaimed = 0;
819 unsigned long nr_writeback = 0;
820 unsigned long nr_immediate = 0;
824 mem_cgroup_uncharge_start();
825 while (!list_empty(page_list)) {
826 struct address_space *mapping;
829 enum page_references references = PAGEREF_RECLAIM_CLEAN;
830 bool dirty, writeback;
834 page = lru_to_page(page_list);
835 list_del(&page->lru);
837 if (!trylock_page(page))
840 VM_BUG_ON_PAGE(PageActive(page), page);
841 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
845 if (unlikely(!page_evictable(page)))
848 if (!sc->may_unmap && page_mapped(page))
851 /* Double the slab pressure for mapped and swapcache pages */
852 if (page_mapped(page) || PageSwapCache(page))
855 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
856 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
859 * The number of dirty pages determines if a zone is marked
860 * reclaim_congested which affects wait_iff_congested. kswapd
861 * will stall and start writing pages if the tail of the LRU
862 * is all dirty unqueued pages.
864 page_check_dirty_writeback(page, &dirty, &writeback);
865 if (dirty || writeback)
868 if (dirty && !writeback)
872 * Treat this page as congested if the underlying BDI is or if
873 * pages are cycling through the LRU so quickly that the
874 * pages marked for immediate reclaim are making it to the
875 * end of the LRU a second time.
877 mapping = page_mapping(page);
878 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
879 (writeback && PageReclaim(page)))
883 * If a page at the tail of the LRU is under writeback, there
884 * are three cases to consider.
886 * 1) If reclaim is encountering an excessive number of pages
887 * under writeback and this page is both under writeback and
888 * PageReclaim then it indicates that pages are being queued
889 * for IO but are being recycled through the LRU before the
890 * IO can complete. Waiting on the page itself risks an
891 * indefinite stall if it is impossible to writeback the
892 * page due to IO error or disconnected storage so instead
893 * note that the LRU is being scanned too quickly and the
894 * caller can stall after page list has been processed.
896 * 2) Global reclaim encounters a page, memcg encounters a
897 * page that is not marked for immediate reclaim or
898 * the caller does not have __GFP_IO. In this case mark
899 * the page for immediate reclaim and continue scanning.
901 * __GFP_IO is checked because a loop driver thread might
902 * enter reclaim, and deadlock if it waits on a page for
903 * which it is needed to do the write (loop masks off
904 * __GFP_IO|__GFP_FS for this reason); but more thought
905 * would probably show more reasons.
907 * Don't require __GFP_FS, since we're not going into the
908 * FS, just waiting on its writeback completion. Worryingly,
909 * ext4 gfs2 and xfs allocate pages with
910 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
911 * may_enter_fs here is liable to OOM on them.
913 * 3) memcg encounters a page that is not already marked
914 * PageReclaim. memcg does not have any dirty pages
915 * throttling so we could easily OOM just because too many
916 * pages are in writeback and there is nothing else to
917 * reclaim. Wait for the writeback to complete.
919 if (PageWriteback(page)) {
921 if (current_is_kswapd() &&
923 zone_is_reclaim_writeback(zone)) {
928 } else if (global_reclaim(sc) ||
929 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
931 * This is slightly racy - end_page_writeback()
932 * might have just cleared PageReclaim, then
933 * setting PageReclaim here end up interpreted
934 * as PageReadahead - but that does not matter
935 * enough to care. What we do want is for this
936 * page to have PageReclaim set next time memcg
937 * reclaim reaches the tests above, so it will
938 * then wait_on_page_writeback() to avoid OOM;
939 * and it's also appropriate in global reclaim.
941 SetPageReclaim(page);
948 wait_on_page_writeback(page);
953 references = page_check_references(page, sc);
955 switch (references) {
956 case PAGEREF_ACTIVATE:
957 goto activate_locked;
960 case PAGEREF_RECLAIM:
961 case PAGEREF_RECLAIM_CLEAN:
962 ; /* try to reclaim the page below */
966 * Anonymous process memory has backing store?
967 * Try to allocate it some swap space here.
969 if (PageAnon(page) && !PageSwapCache(page)) {
970 if (!(sc->gfp_mask & __GFP_IO))
972 if (!add_to_swap(page, page_list))
973 goto activate_locked;
976 /* Adding to swap updated mapping */
977 mapping = page_mapping(page);
981 * The page is mapped into the page tables of one or more
982 * processes. Try to unmap it here.
984 if (page_mapped(page) && mapping) {
985 switch (try_to_unmap(page, ttu_flags)) {
987 goto activate_locked;
993 ; /* try to free the page below */
997 if (PageDirty(page)) {
999 * Only kswapd can writeback filesystem pages to
1000 * avoid risk of stack overflow but only writeback
1001 * if many dirty pages have been encountered.
1003 if (page_is_file_cache(page) &&
1004 (!current_is_kswapd() ||
1005 !zone_is_reclaim_dirty(zone))) {
1007 * Immediately reclaim when written back.
1008 * Similar in principal to deactivate_page()
1009 * except we already have the page isolated
1010 * and know it's dirty
1012 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1013 SetPageReclaim(page);
1018 if (references == PAGEREF_RECLAIM_CLEAN)
1022 if (!sc->may_writepage)
1025 /* Page is dirty, try to write it out here */
1026 switch (pageout(page, mapping, sc)) {
1030 goto activate_locked;
1032 if (PageWriteback(page))
1034 if (PageDirty(page))
1038 * A synchronous write - probably a ramdisk. Go
1039 * ahead and try to reclaim the page.
1041 if (!trylock_page(page))
1043 if (PageDirty(page) || PageWriteback(page))
1045 mapping = page_mapping(page);
1047 ; /* try to free the page below */
1052 * If the page has buffers, try to free the buffer mappings
1053 * associated with this page. If we succeed we try to free
1056 * We do this even if the page is PageDirty().
1057 * try_to_release_page() does not perform I/O, but it is
1058 * possible for a page to have PageDirty set, but it is actually
1059 * clean (all its buffers are clean). This happens if the
1060 * buffers were written out directly, with submit_bh(). ext3
1061 * will do this, as well as the blockdev mapping.
1062 * try_to_release_page() will discover that cleanness and will
1063 * drop the buffers and mark the page clean - it can be freed.
1065 * Rarely, pages can have buffers and no ->mapping. These are
1066 * the pages which were not successfully invalidated in
1067 * truncate_complete_page(). We try to drop those buffers here
1068 * and if that worked, and the page is no longer mapped into
1069 * process address space (page_count == 1) it can be freed.
1070 * Otherwise, leave the page on the LRU so it is swappable.
1072 if (page_has_private(page)) {
1073 if (!try_to_release_page(page, sc->gfp_mask))
1074 goto activate_locked;
1075 if (!mapping && page_count(page) == 1) {
1077 if (put_page_testzero(page))
1081 * rare race with speculative reference.
1082 * the speculative reference will free
1083 * this page shortly, so we may
1084 * increment nr_reclaimed here (and
1085 * leave it off the LRU).
1093 if (!mapping || !__remove_mapping(mapping, page, true))
1097 * At this point, we have no other references and there is
1098 * no way to pick any more up (removed from LRU, removed
1099 * from pagecache). Can use non-atomic bitops now (and
1100 * we obviously don't have to worry about waking up a process
1101 * waiting on the page lock, because there are no references.
1103 __clear_page_locked(page);
1108 * Is there need to periodically free_page_list? It would
1109 * appear not as the counts should be low
1111 list_add(&page->lru, &free_pages);
1115 if (PageSwapCache(page))
1116 try_to_free_swap(page);
1118 putback_lru_page(page);
1122 /* Not a candidate for swapping, so reclaim swap space. */
1123 if (PageSwapCache(page) && vm_swap_full())
1124 try_to_free_swap(page);
1125 VM_BUG_ON_PAGE(PageActive(page), page);
1126 SetPageActive(page);
1131 list_add(&page->lru, &ret_pages);
1132 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1135 free_hot_cold_page_list(&free_pages, true);
1137 list_splice(&ret_pages, page_list);
1138 count_vm_events(PGACTIVATE, pgactivate);
1139 mem_cgroup_uncharge_end();
1140 *ret_nr_dirty += nr_dirty;
1141 *ret_nr_congested += nr_congested;
1142 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1143 *ret_nr_writeback += nr_writeback;
1144 *ret_nr_immediate += nr_immediate;
1145 return nr_reclaimed;
1148 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1149 struct list_head *page_list)
1151 struct scan_control sc = {
1152 .gfp_mask = GFP_KERNEL,
1153 .priority = DEF_PRIORITY,
1156 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1157 struct page *page, *next;
1158 LIST_HEAD(clean_pages);
1160 list_for_each_entry_safe(page, next, page_list, lru) {
1161 if (page_is_file_cache(page) && !PageDirty(page) &&
1162 !isolated_balloon_page(page)) {
1163 ClearPageActive(page);
1164 list_move(&page->lru, &clean_pages);
1168 ret = shrink_page_list(&clean_pages, zone, &sc,
1169 TTU_UNMAP|TTU_IGNORE_ACCESS,
1170 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1171 list_splice(&clean_pages, page_list);
1172 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1177 * Attempt to remove the specified page from its LRU. Only take this page
1178 * if it is of the appropriate PageActive status. Pages which are being
1179 * freed elsewhere are also ignored.
1181 * page: page to consider
1182 * mode: one of the LRU isolation modes defined above
1184 * returns 0 on success, -ve errno on failure.
1186 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1190 /* Only take pages on the LRU. */
1194 /* Compaction should not handle unevictable pages but CMA can do so */
1195 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1201 * To minimise LRU disruption, the caller can indicate that it only
1202 * wants to isolate pages it will be able to operate on without
1203 * blocking - clean pages for the most part.
1205 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1206 * is used by reclaim when it is cannot write to backing storage
1208 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1209 * that it is possible to migrate without blocking
1211 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1212 /* All the caller can do on PageWriteback is block */
1213 if (PageWriteback(page))
1216 if (PageDirty(page)) {
1217 struct address_space *mapping;
1219 /* ISOLATE_CLEAN means only clean pages */
1220 if (mode & ISOLATE_CLEAN)
1224 * Only pages without mappings or that have a
1225 * ->migratepage callback are possible to migrate
1228 mapping = page_mapping(page);
1229 if (mapping && !mapping->a_ops->migratepage)
1234 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1237 if (likely(get_page_unless_zero(page))) {
1239 * Be careful not to clear PageLRU until after we're
1240 * sure the page is not being freed elsewhere -- the
1241 * page release code relies on it.
1251 * zone->lru_lock is heavily contended. Some of the functions that
1252 * shrink the lists perform better by taking out a batch of pages
1253 * and working on them outside the LRU lock.
1255 * For pagecache intensive workloads, this function is the hottest
1256 * spot in the kernel (apart from copy_*_user functions).
1258 * Appropriate locks must be held before calling this function.
1260 * @nr_to_scan: The number of pages to look through on the list.
1261 * @lruvec: The LRU vector to pull pages from.
1262 * @dst: The temp list to put pages on to.
1263 * @nr_scanned: The number of pages that were scanned.
1264 * @sc: The scan_control struct for this reclaim session
1265 * @mode: One of the LRU isolation modes
1266 * @lru: LRU list id for isolating
1268 * returns how many pages were moved onto *@dst.
1270 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1271 struct lruvec *lruvec, struct list_head *dst,
1272 unsigned long *nr_scanned, struct scan_control *sc,
1273 isolate_mode_t mode, enum lru_list lru)
1275 struct list_head *src = &lruvec->lists[lru];
1276 unsigned long nr_taken = 0;
1279 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1283 page = lru_to_page(src);
1284 prefetchw_prev_lru_page(page, src, flags);
1286 VM_BUG_ON_PAGE(!PageLRU(page), page);
1288 switch (__isolate_lru_page(page, mode)) {
1290 nr_pages = hpage_nr_pages(page);
1291 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1292 list_move(&page->lru, dst);
1293 nr_taken += nr_pages;
1297 /* else it is being freed elsewhere */
1298 list_move(&page->lru, src);
1307 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1308 nr_taken, mode, is_file_lru(lru));
1313 * isolate_lru_page - tries to isolate a page from its LRU list
1314 * @page: page to isolate from its LRU list
1316 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1317 * vmstat statistic corresponding to whatever LRU list the page was on.
1319 * Returns 0 if the page was removed from an LRU list.
1320 * Returns -EBUSY if the page was not on an LRU list.
1322 * The returned page will have PageLRU() cleared. If it was found on
1323 * the active list, it will have PageActive set. If it was found on
1324 * the unevictable list, it will have the PageUnevictable bit set. That flag
1325 * may need to be cleared by the caller before letting the page go.
1327 * The vmstat statistic corresponding to the list on which the page was
1328 * found will be decremented.
1331 * (1) Must be called with an elevated refcount on the page. This is a
1332 * fundamentnal difference from isolate_lru_pages (which is called
1333 * without a stable reference).
1334 * (2) the lru_lock must not be held.
1335 * (3) interrupts must be enabled.
1337 int isolate_lru_page(struct page *page)
1341 VM_BUG_ON_PAGE(!page_count(page), page);
1343 if (PageLRU(page)) {
1344 struct zone *zone = page_zone(page);
1345 struct lruvec *lruvec;
1347 spin_lock_irq(&zone->lru_lock);
1348 lruvec = mem_cgroup_page_lruvec(page, zone);
1349 if (PageLRU(page)) {
1350 int lru = page_lru(page);
1353 del_page_from_lru_list(page, lruvec, lru);
1356 spin_unlock_irq(&zone->lru_lock);
1362 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1363 * then get resheduled. When there are massive number of tasks doing page
1364 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1365 * the LRU list will go small and be scanned faster than necessary, leading to
1366 * unnecessary swapping, thrashing and OOM.
1368 static int too_many_isolated(struct zone *zone, int file,
1369 struct scan_control *sc)
1371 unsigned long inactive, isolated;
1373 if (current_is_kswapd())
1376 if (!global_reclaim(sc))
1380 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1381 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1383 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1384 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1388 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1389 * won't get blocked by normal direct-reclaimers, forming a circular
1392 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1395 return isolated > inactive;
1398 static noinline_for_stack void
1399 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1401 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1402 struct zone *zone = lruvec_zone(lruvec);
1403 LIST_HEAD(pages_to_free);
1406 * Put back any unfreeable pages.
1408 while (!list_empty(page_list)) {
1409 struct page *page = lru_to_page(page_list);
1412 VM_BUG_ON_PAGE(PageLRU(page), page);
1413 list_del(&page->lru);
1414 if (unlikely(!page_evictable(page))) {
1415 spin_unlock_irq(&zone->lru_lock);
1416 putback_lru_page(page);
1417 spin_lock_irq(&zone->lru_lock);
1421 lruvec = mem_cgroup_page_lruvec(page, zone);
1424 lru = page_lru(page);
1425 add_page_to_lru_list(page, lruvec, lru);
1427 if (is_active_lru(lru)) {
1428 int file = is_file_lru(lru);
1429 int numpages = hpage_nr_pages(page);
1430 reclaim_stat->recent_rotated[file] += numpages;
1432 if (put_page_testzero(page)) {
1433 __ClearPageLRU(page);
1434 __ClearPageActive(page);
1435 del_page_from_lru_list(page, lruvec, lru);
1437 if (unlikely(PageCompound(page))) {
1438 spin_unlock_irq(&zone->lru_lock);
1439 (*get_compound_page_dtor(page))(page);
1440 spin_lock_irq(&zone->lru_lock);
1442 list_add(&page->lru, &pages_to_free);
1447 * To save our caller's stack, now use input list for pages to free.
1449 list_splice(&pages_to_free, page_list);
1453 * If a kernel thread (such as nfsd for loop-back mounts) services
1454 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1455 * In that case we should only throttle if the backing device it is
1456 * writing to is congested. In other cases it is safe to throttle.
1458 static int current_may_throttle(void)
1460 return !(current->flags & PF_LESS_THROTTLE) ||
1461 current->backing_dev_info == NULL ||
1462 bdi_write_congested(current->backing_dev_info);
1466 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1467 * of reclaimed pages
1469 static noinline_for_stack unsigned long
1470 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1471 struct scan_control *sc, enum lru_list lru)
1473 LIST_HEAD(page_list);
1474 unsigned long nr_scanned;
1475 unsigned long nr_reclaimed = 0;
1476 unsigned long nr_taken;
1477 unsigned long nr_dirty = 0;
1478 unsigned long nr_congested = 0;
1479 unsigned long nr_unqueued_dirty = 0;
1480 unsigned long nr_writeback = 0;
1481 unsigned long nr_immediate = 0;
1482 isolate_mode_t isolate_mode = 0;
1483 int file = is_file_lru(lru);
1484 struct zone *zone = lruvec_zone(lruvec);
1485 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1487 while (unlikely(too_many_isolated(zone, file, sc))) {
1488 congestion_wait(BLK_RW_ASYNC, HZ/10);
1490 /* We are about to die and free our memory. Return now. */
1491 if (fatal_signal_pending(current))
1492 return SWAP_CLUSTER_MAX;
1498 isolate_mode |= ISOLATE_UNMAPPED;
1499 if (!sc->may_writepage)
1500 isolate_mode |= ISOLATE_CLEAN;
1502 spin_lock_irq(&zone->lru_lock);
1504 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1505 &nr_scanned, sc, isolate_mode, lru);
1507 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1508 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1510 if (global_reclaim(sc)) {
1511 zone->pages_scanned += nr_scanned;
1512 if (current_is_kswapd())
1513 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1515 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1517 spin_unlock_irq(&zone->lru_lock);
1522 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1523 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1524 &nr_writeback, &nr_immediate,
1527 spin_lock_irq(&zone->lru_lock);
1529 reclaim_stat->recent_scanned[file] += nr_taken;
1531 if (global_reclaim(sc)) {
1532 if (current_is_kswapd())
1533 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1536 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1540 putback_inactive_pages(lruvec, &page_list);
1542 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1544 spin_unlock_irq(&zone->lru_lock);
1546 free_hot_cold_page_list(&page_list, true);
1549 * If reclaim is isolating dirty pages under writeback, it implies
1550 * that the long-lived page allocation rate is exceeding the page
1551 * laundering rate. Either the global limits are not being effective
1552 * at throttling processes due to the page distribution throughout
1553 * zones or there is heavy usage of a slow backing device. The
1554 * only option is to throttle from reclaim context which is not ideal
1555 * as there is no guarantee the dirtying process is throttled in the
1556 * same way balance_dirty_pages() manages.
1558 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1559 * of pages under pages flagged for immediate reclaim and stall if any
1560 * are encountered in the nr_immediate check below.
1562 if (nr_writeback && nr_writeback == nr_taken)
1563 zone_set_flag(zone, ZONE_WRITEBACK);
1566 * memcg will stall in page writeback so only consider forcibly
1567 * stalling for global reclaim
1569 if (global_reclaim(sc)) {
1571 * Tag a zone as congested if all the dirty pages scanned were
1572 * backed by a congested BDI and wait_iff_congested will stall.
1574 if (nr_dirty && nr_dirty == nr_congested)
1575 zone_set_flag(zone, ZONE_CONGESTED);
1578 * If dirty pages are scanned that are not queued for IO, it
1579 * implies that flushers are not keeping up. In this case, flag
1580 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1581 * pages from reclaim context.
1583 if (nr_unqueued_dirty == nr_taken)
1584 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1587 * If kswapd scans pages marked marked for immediate
1588 * reclaim and under writeback (nr_immediate), it implies
1589 * that pages are cycling through the LRU faster than
1590 * they are written so also forcibly stall.
1592 if (nr_immediate && current_may_throttle())
1593 congestion_wait(BLK_RW_ASYNC, HZ/10);
1597 * Stall direct reclaim for IO completions if underlying BDIs or zone
1598 * is congested. Allow kswapd to continue until it starts encountering
1599 * unqueued dirty pages or cycling through the LRU too quickly.
1601 if (!sc->hibernation_mode && !current_is_kswapd() &&
1602 current_may_throttle())
1603 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1605 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1607 nr_scanned, nr_reclaimed,
1609 trace_shrink_flags(file));
1610 return nr_reclaimed;
1614 * This moves pages from the active list to the inactive list.
1616 * We move them the other way if the page is referenced by one or more
1617 * processes, from rmap.
1619 * If the pages are mostly unmapped, the processing is fast and it is
1620 * appropriate to hold zone->lru_lock across the whole operation. But if
1621 * the pages are mapped, the processing is slow (page_referenced()) so we
1622 * should drop zone->lru_lock around each page. It's impossible to balance
1623 * this, so instead we remove the pages from the LRU while processing them.
1624 * It is safe to rely on PG_active against the non-LRU pages in here because
1625 * nobody will play with that bit on a non-LRU page.
1627 * The downside is that we have to touch page->_count against each page.
1628 * But we had to alter page->flags anyway.
1631 static void move_active_pages_to_lru(struct lruvec *lruvec,
1632 struct list_head *list,
1633 struct list_head *pages_to_free,
1636 struct zone *zone = lruvec_zone(lruvec);
1637 unsigned long pgmoved = 0;
1641 while (!list_empty(list)) {
1642 page = lru_to_page(list);
1643 lruvec = mem_cgroup_page_lruvec(page, zone);
1645 VM_BUG_ON_PAGE(PageLRU(page), page);
1648 nr_pages = hpage_nr_pages(page);
1649 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1650 list_move(&page->lru, &lruvec->lists[lru]);
1651 pgmoved += nr_pages;
1653 if (put_page_testzero(page)) {
1654 __ClearPageLRU(page);
1655 __ClearPageActive(page);
1656 del_page_from_lru_list(page, lruvec, lru);
1658 if (unlikely(PageCompound(page))) {
1659 spin_unlock_irq(&zone->lru_lock);
1660 (*get_compound_page_dtor(page))(page);
1661 spin_lock_irq(&zone->lru_lock);
1663 list_add(&page->lru, pages_to_free);
1666 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1667 if (!is_active_lru(lru))
1668 __count_vm_events(PGDEACTIVATE, pgmoved);
1671 static void shrink_active_list(unsigned long nr_to_scan,
1672 struct lruvec *lruvec,
1673 struct scan_control *sc,
1676 unsigned long nr_taken;
1677 unsigned long nr_scanned;
1678 unsigned long vm_flags;
1679 LIST_HEAD(l_hold); /* The pages which were snipped off */
1680 LIST_HEAD(l_active);
1681 LIST_HEAD(l_inactive);
1683 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1684 unsigned long nr_rotated = 0;
1685 isolate_mode_t isolate_mode = 0;
1686 int file = is_file_lru(lru);
1687 struct zone *zone = lruvec_zone(lruvec);
1692 isolate_mode |= ISOLATE_UNMAPPED;
1693 if (!sc->may_writepage)
1694 isolate_mode |= ISOLATE_CLEAN;
1696 spin_lock_irq(&zone->lru_lock);
1698 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1699 &nr_scanned, sc, isolate_mode, lru);
1700 if (global_reclaim(sc))
1701 zone->pages_scanned += nr_scanned;
1703 reclaim_stat->recent_scanned[file] += nr_taken;
1705 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1706 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1707 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1708 spin_unlock_irq(&zone->lru_lock);
1710 while (!list_empty(&l_hold)) {
1712 page = lru_to_page(&l_hold);
1713 list_del(&page->lru);
1715 if (unlikely(!page_evictable(page))) {
1716 putback_lru_page(page);
1720 if (unlikely(buffer_heads_over_limit)) {
1721 if (page_has_private(page) && trylock_page(page)) {
1722 if (page_has_private(page))
1723 try_to_release_page(page, 0);
1728 if (page_referenced(page, 0, sc->target_mem_cgroup,
1730 nr_rotated += hpage_nr_pages(page);
1732 * Identify referenced, file-backed active pages and
1733 * give them one more trip around the active list. So
1734 * that executable code get better chances to stay in
1735 * memory under moderate memory pressure. Anon pages
1736 * are not likely to be evicted by use-once streaming
1737 * IO, plus JVM can create lots of anon VM_EXEC pages,
1738 * so we ignore them here.
1740 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1741 list_add(&page->lru, &l_active);
1746 ClearPageActive(page); /* we are de-activating */
1747 list_add(&page->lru, &l_inactive);
1751 * Move pages back to the lru list.
1753 spin_lock_irq(&zone->lru_lock);
1755 * Count referenced pages from currently used mappings as rotated,
1756 * even though only some of them are actually re-activated. This
1757 * helps balance scan pressure between file and anonymous pages in
1760 reclaim_stat->recent_rotated[file] += nr_rotated;
1762 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1763 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1764 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1765 spin_unlock_irq(&zone->lru_lock);
1767 free_hot_cold_page_list(&l_hold, true);
1771 static int inactive_anon_is_low_global(struct zone *zone)
1773 unsigned long active, inactive;
1775 active = zone_page_state(zone, NR_ACTIVE_ANON);
1776 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1778 if (inactive * zone->inactive_ratio < active)
1785 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1786 * @lruvec: LRU vector to check
1788 * Returns true if the zone does not have enough inactive anon pages,
1789 * meaning some active anon pages need to be deactivated.
1791 static int inactive_anon_is_low(struct lruvec *lruvec)
1794 * If we don't have swap space, anonymous page deactivation
1797 if (!total_swap_pages)
1800 if (!mem_cgroup_disabled())
1801 return mem_cgroup_inactive_anon_is_low(lruvec);
1803 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1806 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1813 * inactive_file_is_low - check if file pages need to be deactivated
1814 * @lruvec: LRU vector to check
1816 * When the system is doing streaming IO, memory pressure here
1817 * ensures that active file pages get deactivated, until more
1818 * than half of the file pages are on the inactive list.
1820 * Once we get to that situation, protect the system's working
1821 * set from being evicted by disabling active file page aging.
1823 * This uses a different ratio than the anonymous pages, because
1824 * the page cache uses a use-once replacement algorithm.
1826 static int inactive_file_is_low(struct lruvec *lruvec)
1828 unsigned long inactive;
1829 unsigned long active;
1831 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1832 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1834 return active > inactive;
1837 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1839 if (is_file_lru(lru))
1840 return inactive_file_is_low(lruvec);
1842 return inactive_anon_is_low(lruvec);
1845 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1846 struct lruvec *lruvec, struct scan_control *sc)
1848 if (is_active_lru(lru)) {
1849 if (inactive_list_is_low(lruvec, lru))
1850 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1854 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1865 * Determine how aggressively the anon and file LRU lists should be
1866 * scanned. The relative value of each set of LRU lists is determined
1867 * by looking at the fraction of the pages scanned we did rotate back
1868 * onto the active list instead of evict.
1870 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1871 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1873 static void get_scan_count(struct lruvec *lruvec, int swappiness,
1874 struct scan_control *sc, unsigned long *nr)
1876 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1878 u64 denominator = 0; /* gcc */
1879 struct zone *zone = lruvec_zone(lruvec);
1880 unsigned long anon_prio, file_prio;
1881 enum scan_balance scan_balance;
1882 unsigned long anon, file;
1883 bool force_scan = false;
1884 unsigned long ap, fp;
1890 * If the zone or memcg is small, nr[l] can be 0. This
1891 * results in no scanning on this priority and a potential
1892 * priority drop. Global direct reclaim can go to the next
1893 * zone and tends to have no problems. Global kswapd is for
1894 * zone balancing and it needs to scan a minimum amount. When
1895 * reclaiming for a memcg, a priority drop can cause high
1896 * latencies, so it's better to scan a minimum amount there as
1899 if (current_is_kswapd() && !zone_reclaimable(zone))
1901 if (!global_reclaim(sc))
1904 /* If we have no swap space, do not bother scanning anon pages. */
1905 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1906 scan_balance = SCAN_FILE;
1911 * Global reclaim will swap to prevent OOM even with no
1912 * swappiness, but memcg users want to use this knob to
1913 * disable swapping for individual groups completely when
1914 * using the memory controller's swap limit feature would be
1917 if (!global_reclaim(sc) && !swappiness) {
1918 scan_balance = SCAN_FILE;
1923 * Do not apply any pressure balancing cleverness when the
1924 * system is close to OOM, scan both anon and file equally
1925 * (unless the swappiness setting disagrees with swapping).
1927 if (!sc->priority && swappiness) {
1928 scan_balance = SCAN_EQUAL;
1932 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1933 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1934 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1935 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1938 * Prevent the reclaimer from falling into the cache trap: as
1939 * cache pages start out inactive, every cache fault will tip
1940 * the scan balance towards the file LRU. And as the file LRU
1941 * shrinks, so does the window for rotation from references.
1942 * This means we have a runaway feedback loop where a tiny
1943 * thrashing file LRU becomes infinitely more attractive than
1944 * anon pages. Try to detect this based on file LRU size.
1946 if (global_reclaim(sc)) {
1947 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1949 if (unlikely(file + free <= high_wmark_pages(zone))) {
1950 scan_balance = SCAN_ANON;
1956 * There is enough inactive page cache, do not reclaim
1957 * anything from the anonymous working set right now.
1959 if (!inactive_file_is_low(lruvec)) {
1960 scan_balance = SCAN_FILE;
1964 scan_balance = SCAN_FRACT;
1967 * With swappiness at 100, anonymous and file have the same priority.
1968 * This scanning priority is essentially the inverse of IO cost.
1970 anon_prio = swappiness;
1971 file_prio = 200 - anon_prio;
1974 * OK, so we have swap space and a fair amount of page cache
1975 * pages. We use the recently rotated / recently scanned
1976 * ratios to determine how valuable each cache is.
1978 * Because workloads change over time (and to avoid overflow)
1979 * we keep these statistics as a floating average, which ends
1980 * up weighing recent references more than old ones.
1982 * anon in [0], file in [1]
1984 spin_lock_irq(&zone->lru_lock);
1985 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1986 reclaim_stat->recent_scanned[0] /= 2;
1987 reclaim_stat->recent_rotated[0] /= 2;
1990 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1991 reclaim_stat->recent_scanned[1] /= 2;
1992 reclaim_stat->recent_rotated[1] /= 2;
1996 * The amount of pressure on anon vs file pages is inversely
1997 * proportional to the fraction of recently scanned pages on
1998 * each list that were recently referenced and in active use.
2000 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2001 ap /= reclaim_stat->recent_rotated[0] + 1;
2003 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2004 fp /= reclaim_stat->recent_rotated[1] + 1;
2005 spin_unlock_irq(&zone->lru_lock);
2009 denominator = ap + fp + 1;
2011 some_scanned = false;
2012 /* Only use force_scan on second pass. */
2013 for (pass = 0; !some_scanned && pass < 2; pass++) {
2014 for_each_evictable_lru(lru) {
2015 int file = is_file_lru(lru);
2019 size = get_lru_size(lruvec, lru);
2020 scan = size >> sc->priority;
2022 if (!scan && pass && force_scan)
2023 scan = min(size, SWAP_CLUSTER_MAX);
2025 switch (scan_balance) {
2027 /* Scan lists relative to size */
2031 * Scan types proportional to swappiness and
2032 * their relative recent reclaim efficiency.
2034 scan = div64_u64(scan * fraction[file],
2039 /* Scan one type exclusively */
2040 if ((scan_balance == SCAN_FILE) != file)
2044 /* Look ma, no brain */
2049 * Skip the second pass and don't force_scan,
2050 * if we found something to scan.
2052 some_scanned |= !!scan;
2058 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2060 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2061 struct scan_control *sc)
2063 unsigned long nr[NR_LRU_LISTS];
2064 unsigned long targets[NR_LRU_LISTS];
2065 unsigned long nr_to_scan;
2067 unsigned long nr_reclaimed = 0;
2068 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2069 struct blk_plug plug;
2072 get_scan_count(lruvec, swappiness, sc, nr);
2074 /* Record the original scan target for proportional adjustments later */
2075 memcpy(targets, nr, sizeof(nr));
2078 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2079 * event that can occur when there is little memory pressure e.g.
2080 * multiple streaming readers/writers. Hence, we do not abort scanning
2081 * when the requested number of pages are reclaimed when scanning at
2082 * DEF_PRIORITY on the assumption that the fact we are direct
2083 * reclaiming implies that kswapd is not keeping up and it is best to
2084 * do a batch of work at once. For memcg reclaim one check is made to
2085 * abort proportional reclaim if either the file or anon lru has already
2086 * dropped to zero at the first pass.
2088 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2089 sc->priority == DEF_PRIORITY);
2091 blk_start_plug(&plug);
2092 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2093 nr[LRU_INACTIVE_FILE]) {
2094 unsigned long nr_anon, nr_file, percentage;
2095 unsigned long nr_scanned;
2097 for_each_evictable_lru(lru) {
2099 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2100 nr[lru] -= nr_to_scan;
2102 nr_reclaimed += shrink_list(lru, nr_to_scan,
2107 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2111 * For kswapd and memcg, reclaim at least the number of pages
2112 * requested. Ensure that the anon and file LRUs are scanned
2113 * proportionally what was requested by get_scan_count(). We
2114 * stop reclaiming one LRU and reduce the amount scanning
2115 * proportional to the original scan target.
2117 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2118 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2121 * It's just vindictive to attack the larger once the smaller
2122 * has gone to zero. And given the way we stop scanning the
2123 * smaller below, this makes sure that we only make one nudge
2124 * towards proportionality once we've got nr_to_reclaim.
2126 if (!nr_file || !nr_anon)
2129 if (nr_file > nr_anon) {
2130 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2131 targets[LRU_ACTIVE_ANON] + 1;
2133 percentage = nr_anon * 100 / scan_target;
2135 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2136 targets[LRU_ACTIVE_FILE] + 1;
2138 percentage = nr_file * 100 / scan_target;
2141 /* Stop scanning the smaller of the LRU */
2143 nr[lru + LRU_ACTIVE] = 0;
2146 * Recalculate the other LRU scan count based on its original
2147 * scan target and the percentage scanning already complete
2149 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2150 nr_scanned = targets[lru] - nr[lru];
2151 nr[lru] = targets[lru] * (100 - percentage) / 100;
2152 nr[lru] -= min(nr[lru], nr_scanned);
2155 nr_scanned = targets[lru] - nr[lru];
2156 nr[lru] = targets[lru] * (100 - percentage) / 100;
2157 nr[lru] -= min(nr[lru], nr_scanned);
2159 scan_adjusted = true;
2161 blk_finish_plug(&plug);
2162 sc->nr_reclaimed += nr_reclaimed;
2165 * Even if we did not try to evict anon pages at all, we want to
2166 * rebalance the anon lru active/inactive ratio.
2168 if (inactive_anon_is_low(lruvec))
2169 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2170 sc, LRU_ACTIVE_ANON);
2172 throttle_vm_writeout(sc->gfp_mask);
2175 /* Use reclaim/compaction for costly allocs or under memory pressure */
2176 static bool in_reclaim_compaction(struct scan_control *sc)
2178 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2179 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2180 sc->priority < DEF_PRIORITY - 2))
2187 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2188 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2189 * true if more pages should be reclaimed such that when the page allocator
2190 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2191 * It will give up earlier than that if there is difficulty reclaiming pages.
2193 static inline bool should_continue_reclaim(struct zone *zone,
2194 unsigned long nr_reclaimed,
2195 unsigned long nr_scanned,
2196 struct scan_control *sc)
2198 unsigned long pages_for_compaction;
2199 unsigned long inactive_lru_pages;
2201 /* If not in reclaim/compaction mode, stop */
2202 if (!in_reclaim_compaction(sc))
2205 /* Consider stopping depending on scan and reclaim activity */
2206 if (sc->gfp_mask & __GFP_REPEAT) {
2208 * For __GFP_REPEAT allocations, stop reclaiming if the
2209 * full LRU list has been scanned and we are still failing
2210 * to reclaim pages. This full LRU scan is potentially
2211 * expensive but a __GFP_REPEAT caller really wants to succeed
2213 if (!nr_reclaimed && !nr_scanned)
2217 * For non-__GFP_REPEAT allocations which can presumably
2218 * fail without consequence, stop if we failed to reclaim
2219 * any pages from the last SWAP_CLUSTER_MAX number of
2220 * pages that were scanned. This will return to the
2221 * caller faster at the risk reclaim/compaction and
2222 * the resulting allocation attempt fails
2229 * If we have not reclaimed enough pages for compaction and the
2230 * inactive lists are large enough, continue reclaiming
2232 pages_for_compaction = (2UL << sc->order);
2233 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2234 if (get_nr_swap_pages() > 0)
2235 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2236 if (sc->nr_reclaimed < pages_for_compaction &&
2237 inactive_lru_pages > pages_for_compaction)
2240 /* If compaction would go ahead or the allocation would succeed, stop */
2241 switch (compaction_suitable(zone, sc->order)) {
2242 case COMPACT_PARTIAL:
2243 case COMPACT_CONTINUE:
2250 static bool shrink_zone(struct zone *zone, struct scan_control *sc)
2252 unsigned long nr_reclaimed, nr_scanned;
2253 bool reclaimable = false;
2256 struct mem_cgroup *root = sc->target_mem_cgroup;
2257 struct mem_cgroup_reclaim_cookie reclaim = {
2259 .priority = sc->priority,
2261 struct mem_cgroup *memcg;
2263 nr_reclaimed = sc->nr_reclaimed;
2264 nr_scanned = sc->nr_scanned;
2266 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2268 struct lruvec *lruvec;
2271 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2272 swappiness = mem_cgroup_swappiness(memcg);
2274 shrink_lruvec(lruvec, swappiness, sc);
2277 * Direct reclaim and kswapd have to scan all memory
2278 * cgroups to fulfill the overall scan target for the
2281 * Limit reclaim, on the other hand, only cares about
2282 * nr_to_reclaim pages to be reclaimed and it will
2283 * retry with decreasing priority if one round over the
2284 * whole hierarchy is not sufficient.
2286 if (!global_reclaim(sc) &&
2287 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2288 mem_cgroup_iter_break(root, memcg);
2291 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2294 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2295 sc->nr_scanned - nr_scanned,
2296 sc->nr_reclaimed - nr_reclaimed);
2298 if (sc->nr_reclaimed - nr_reclaimed)
2301 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2302 sc->nr_scanned - nr_scanned, sc));
2307 /* Returns true if compaction should go ahead for a high-order request */
2308 static inline bool compaction_ready(struct zone *zone, int order)
2310 unsigned long balance_gap, watermark;
2314 * Compaction takes time to run and there are potentially other
2315 * callers using the pages just freed. Continue reclaiming until
2316 * there is a buffer of free pages available to give compaction
2317 * a reasonable chance of completing and allocating the page
2319 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2320 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2321 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2322 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2325 * If compaction is deferred, reclaim up to a point where
2326 * compaction will have a chance of success when re-enabled
2328 if (compaction_deferred(zone, order))
2329 return watermark_ok;
2331 /* If compaction is not ready to start, keep reclaiming */
2332 if (!compaction_suitable(zone, order))
2335 return watermark_ok;
2339 * This is the direct reclaim path, for page-allocating processes. We only
2340 * try to reclaim pages from zones which will satisfy the caller's allocation
2343 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2345 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2347 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2348 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2349 * zone defense algorithm.
2351 * If a zone is deemed to be full of pinned pages then just give it a light
2352 * scan then give up on it.
2354 * Returns true if a zone was reclaimable.
2356 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2360 unsigned long nr_soft_reclaimed;
2361 unsigned long nr_soft_scanned;
2362 unsigned long lru_pages = 0;
2363 struct reclaim_state *reclaim_state = current->reclaim_state;
2365 struct shrink_control shrink = {
2366 .gfp_mask = sc->gfp_mask,
2368 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2369 bool reclaimable = false;
2372 * If the number of buffer_heads in the machine exceeds the maximum
2373 * allowed level, force direct reclaim to scan the highmem zone as
2374 * highmem pages could be pinning lowmem pages storing buffer_heads
2376 orig_mask = sc->gfp_mask;
2377 if (buffer_heads_over_limit)
2378 sc->gfp_mask |= __GFP_HIGHMEM;
2380 nodes_clear(shrink.nodes_to_scan);
2382 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2383 gfp_zone(sc->gfp_mask), sc->nodemask) {
2384 if (!populated_zone(zone))
2387 * Take care memory controller reclaiming has small influence
2390 if (global_reclaim(sc)) {
2391 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2394 lru_pages += zone_reclaimable_pages(zone);
2395 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2397 if (sc->priority != DEF_PRIORITY &&
2398 !zone_reclaimable(zone))
2399 continue; /* Let kswapd poll it */
2402 * If we already have plenty of memory free for
2403 * compaction in this zone, don't free any more.
2404 * Even though compaction is invoked for any
2405 * non-zero order, only frequent costly order
2406 * reclamation is disruptive enough to become a
2407 * noticeable problem, like transparent huge
2410 if (IS_ENABLED(CONFIG_COMPACTION) &&
2411 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2412 zonelist_zone_idx(z) <= requested_highidx &&
2413 compaction_ready(zone, sc->order)) {
2414 sc->compaction_ready = true;
2419 * This steals pages from memory cgroups over softlimit
2420 * and returns the number of reclaimed pages and
2421 * scanned pages. This works for global memory pressure
2422 * and balancing, not for a memcg's limit.
2424 nr_soft_scanned = 0;
2425 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2426 sc->order, sc->gfp_mask,
2428 sc->nr_reclaimed += nr_soft_reclaimed;
2429 sc->nr_scanned += nr_soft_scanned;
2430 if (nr_soft_reclaimed)
2432 /* need some check for avoid more shrink_zone() */
2435 if (shrink_zone(zone, sc))
2438 if (global_reclaim(sc) &&
2439 !reclaimable && zone_reclaimable(zone))
2444 * Don't shrink slabs when reclaiming memory from over limit cgroups
2445 * but do shrink slab at least once when aborting reclaim for
2446 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2449 if (global_reclaim(sc)) {
2450 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2451 if (reclaim_state) {
2452 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2453 reclaim_state->reclaimed_slab = 0;
2458 * Restore to original mask to avoid the impact on the caller if we
2459 * promoted it to __GFP_HIGHMEM.
2461 sc->gfp_mask = orig_mask;
2467 * This is the main entry point to direct page reclaim.
2469 * If a full scan of the inactive list fails to free enough memory then we
2470 * are "out of memory" and something needs to be killed.
2472 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2473 * high - the zone may be full of dirty or under-writeback pages, which this
2474 * caller can't do much about. We kick the writeback threads and take explicit
2475 * naps in the hope that some of these pages can be written. But if the
2476 * allocating task holds filesystem locks which prevent writeout this might not
2477 * work, and the allocation attempt will fail.
2479 * returns: 0, if no pages reclaimed
2480 * else, the number of pages reclaimed
2482 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2483 struct scan_control *sc)
2485 unsigned long total_scanned = 0;
2486 unsigned long writeback_threshold;
2487 bool zones_reclaimable;
2489 delayacct_freepages_start();
2491 if (global_reclaim(sc))
2492 count_vm_event(ALLOCSTALL);
2495 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2498 zones_reclaimable = shrink_zones(zonelist, sc);
2500 total_scanned += sc->nr_scanned;
2501 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2504 if (sc->compaction_ready)
2508 * If we're getting trouble reclaiming, start doing
2509 * writepage even in laptop mode.
2511 if (sc->priority < DEF_PRIORITY - 2)
2512 sc->may_writepage = 1;
2515 * Try to write back as many pages as we just scanned. This
2516 * tends to cause slow streaming writers to write data to the
2517 * disk smoothly, at the dirtying rate, which is nice. But
2518 * that's undesirable in laptop mode, where we *want* lumpy
2519 * writeout. So in laptop mode, write out the whole world.
2521 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2522 if (total_scanned > writeback_threshold) {
2523 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2524 WB_REASON_TRY_TO_FREE_PAGES);
2525 sc->may_writepage = 1;
2527 } while (--sc->priority >= 0);
2529 delayacct_freepages_end();
2531 if (sc->nr_reclaimed)
2532 return sc->nr_reclaimed;
2534 /* Aborted reclaim to try compaction? don't OOM, then */
2535 if (sc->compaction_ready)
2538 /* Any of the zones still reclaimable? Don't OOM. */
2539 if (zones_reclaimable)
2545 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2548 unsigned long pfmemalloc_reserve = 0;
2549 unsigned long free_pages = 0;
2553 for (i = 0; i <= ZONE_NORMAL; i++) {
2554 zone = &pgdat->node_zones[i];
2555 if (!populated_zone(zone))
2558 pfmemalloc_reserve += min_wmark_pages(zone);
2559 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2562 /* If there are no reserves (unexpected config) then do not throttle */
2563 if (!pfmemalloc_reserve)
2566 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2568 /* kswapd must be awake if processes are being throttled */
2569 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2570 pgdat->classzone_idx = min(pgdat->classzone_idx,
2571 (enum zone_type)ZONE_NORMAL);
2572 wake_up_interruptible(&pgdat->kswapd_wait);
2579 * Throttle direct reclaimers if backing storage is backed by the network
2580 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2581 * depleted. kswapd will continue to make progress and wake the processes
2582 * when the low watermark is reached.
2584 * Returns true if a fatal signal was delivered during throttling. If this
2585 * happens, the page allocator should not consider triggering the OOM killer.
2587 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2588 nodemask_t *nodemask)
2592 pg_data_t *pgdat = NULL;
2595 * Kernel threads should not be throttled as they may be indirectly
2596 * responsible for cleaning pages necessary for reclaim to make forward
2597 * progress. kjournald for example may enter direct reclaim while
2598 * committing a transaction where throttling it could forcing other
2599 * processes to block on log_wait_commit().
2601 if (current->flags & PF_KTHREAD)
2605 * If a fatal signal is pending, this process should not throttle.
2606 * It should return quickly so it can exit and free its memory
2608 if (fatal_signal_pending(current))
2612 * Check if the pfmemalloc reserves are ok by finding the first node
2613 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2614 * GFP_KERNEL will be required for allocating network buffers when
2615 * swapping over the network so ZONE_HIGHMEM is unusable.
2617 * Throttling is based on the first usable node and throttled processes
2618 * wait on a queue until kswapd makes progress and wakes them. There
2619 * is an affinity then between processes waking up and where reclaim
2620 * progress has been made assuming the process wakes on the same node.
2621 * More importantly, processes running on remote nodes will not compete
2622 * for remote pfmemalloc reserves and processes on different nodes
2623 * should make reasonable progress.
2625 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2626 gfp_mask, nodemask) {
2627 if (zone_idx(zone) > ZONE_NORMAL)
2630 /* Throttle based on the first usable node */
2631 pgdat = zone->zone_pgdat;
2632 if (pfmemalloc_watermark_ok(pgdat))
2637 /* If no zone was usable by the allocation flags then do not throttle */
2641 /* Account for the throttling */
2642 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2645 * If the caller cannot enter the filesystem, it's possible that it
2646 * is due to the caller holding an FS lock or performing a journal
2647 * transaction in the case of a filesystem like ext[3|4]. In this case,
2648 * it is not safe to block on pfmemalloc_wait as kswapd could be
2649 * blocked waiting on the same lock. Instead, throttle for up to a
2650 * second before continuing.
2652 if (!(gfp_mask & __GFP_FS)) {
2653 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2654 pfmemalloc_watermark_ok(pgdat), HZ);
2659 /* Throttle until kswapd wakes the process */
2660 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2661 pfmemalloc_watermark_ok(pgdat));
2664 if (fatal_signal_pending(current))
2671 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2672 gfp_t gfp_mask, nodemask_t *nodemask)
2674 unsigned long nr_reclaimed;
2675 struct scan_control sc = {
2676 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2677 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2679 .nodemask = nodemask,
2680 .priority = DEF_PRIORITY,
2681 .may_writepage = !laptop_mode,
2687 * Do not enter reclaim if fatal signal was delivered while throttled.
2688 * 1 is returned so that the page allocator does not OOM kill at this
2691 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2694 trace_mm_vmscan_direct_reclaim_begin(order,
2698 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2700 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2702 return nr_reclaimed;
2707 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2708 gfp_t gfp_mask, bool noswap,
2710 unsigned long *nr_scanned)
2712 struct scan_control sc = {
2713 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2714 .target_mem_cgroup = memcg,
2715 .may_writepage = !laptop_mode,
2717 .may_swap = !noswap,
2719 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2720 int swappiness = mem_cgroup_swappiness(memcg);
2722 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2723 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2725 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2730 * NOTE: Although we can get the priority field, using it
2731 * here is not a good idea, since it limits the pages we can scan.
2732 * if we don't reclaim here, the shrink_zone from balance_pgdat
2733 * will pick up pages from other mem cgroup's as well. We hack
2734 * the priority and make it zero.
2736 shrink_lruvec(lruvec, swappiness, &sc);
2738 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2740 *nr_scanned = sc.nr_scanned;
2741 return sc.nr_reclaimed;
2744 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2748 struct zonelist *zonelist;
2749 unsigned long nr_reclaimed;
2751 struct scan_control sc = {
2752 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2753 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2754 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2755 .target_mem_cgroup = memcg,
2756 .priority = DEF_PRIORITY,
2757 .may_writepage = !laptop_mode,
2759 .may_swap = !noswap,
2763 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2764 * take care of from where we get pages. So the node where we start the
2765 * scan does not need to be the current node.
2767 nid = mem_cgroup_select_victim_node(memcg);
2769 zonelist = NODE_DATA(nid)->node_zonelists;
2771 trace_mm_vmscan_memcg_reclaim_begin(0,
2775 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2777 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2779 return nr_reclaimed;
2783 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2785 struct mem_cgroup *memcg;
2787 if (!total_swap_pages)
2790 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2792 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2794 if (inactive_anon_is_low(lruvec))
2795 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2796 sc, LRU_ACTIVE_ANON);
2798 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2802 static bool zone_balanced(struct zone *zone, int order,
2803 unsigned long balance_gap, int classzone_idx)
2805 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2806 balance_gap, classzone_idx, 0))
2809 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2810 !compaction_suitable(zone, order))
2817 * pgdat_balanced() is used when checking if a node is balanced.
2819 * For order-0, all zones must be balanced!
2821 * For high-order allocations only zones that meet watermarks and are in a
2822 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2823 * total of balanced pages must be at least 25% of the zones allowed by
2824 * classzone_idx for the node to be considered balanced. Forcing all zones to
2825 * be balanced for high orders can cause excessive reclaim when there are
2827 * The choice of 25% is due to
2828 * o a 16M DMA zone that is balanced will not balance a zone on any
2829 * reasonable sized machine
2830 * o On all other machines, the top zone must be at least a reasonable
2831 * percentage of the middle zones. For example, on 32-bit x86, highmem
2832 * would need to be at least 256M for it to be balance a whole node.
2833 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2834 * to balance a node on its own. These seemed like reasonable ratios.
2836 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2838 unsigned long managed_pages = 0;
2839 unsigned long balanced_pages = 0;
2842 /* Check the watermark levels */
2843 for (i = 0; i <= classzone_idx; i++) {
2844 struct zone *zone = pgdat->node_zones + i;
2846 if (!populated_zone(zone))
2849 managed_pages += zone->managed_pages;
2852 * A special case here:
2854 * balance_pgdat() skips over all_unreclaimable after
2855 * DEF_PRIORITY. Effectively, it considers them balanced so
2856 * they must be considered balanced here as well!
2858 if (!zone_reclaimable(zone)) {
2859 balanced_pages += zone->managed_pages;
2863 if (zone_balanced(zone, order, 0, i))
2864 balanced_pages += zone->managed_pages;
2870 return balanced_pages >= (managed_pages >> 2);
2876 * Prepare kswapd for sleeping. This verifies that there are no processes
2877 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2879 * Returns true if kswapd is ready to sleep
2881 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2884 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2889 * There is a potential race between when kswapd checks its watermarks
2890 * and a process gets throttled. There is also a potential race if
2891 * processes get throttled, kswapd wakes, a large process exits therby
2892 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2893 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2894 * so wake them now if necessary. If necessary, processes will wake
2895 * kswapd and get throttled again
2897 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2898 wake_up(&pgdat->pfmemalloc_wait);
2902 return pgdat_balanced(pgdat, order, classzone_idx);
2906 * kswapd shrinks the zone by the number of pages required to reach
2907 * the high watermark.
2909 * Returns true if kswapd scanned at least the requested number of pages to
2910 * reclaim or if the lack of progress was due to pages under writeback.
2911 * This is used to determine if the scanning priority needs to be raised.
2913 static bool kswapd_shrink_zone(struct zone *zone,
2915 struct scan_control *sc,
2916 unsigned long lru_pages,
2917 unsigned long *nr_attempted)
2919 int testorder = sc->order;
2920 unsigned long balance_gap;
2921 struct reclaim_state *reclaim_state = current->reclaim_state;
2922 struct shrink_control shrink = {
2923 .gfp_mask = sc->gfp_mask,
2925 bool lowmem_pressure;
2927 /* Reclaim above the high watermark. */
2928 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2931 * Kswapd reclaims only single pages with compaction enabled. Trying
2932 * too hard to reclaim until contiguous free pages have become
2933 * available can hurt performance by evicting too much useful data
2934 * from memory. Do not reclaim more than needed for compaction.
2936 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2937 compaction_suitable(zone, sc->order) !=
2942 * We put equal pressure on every zone, unless one zone has way too
2943 * many pages free already. The "too many pages" is defined as the
2944 * high wmark plus a "gap" where the gap is either the low
2945 * watermark or 1% of the zone, whichever is smaller.
2947 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2948 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2951 * If there is no low memory pressure or the zone is balanced then no
2952 * reclaim is necessary
2954 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2955 if (!lowmem_pressure && zone_balanced(zone, testorder,
2956 balance_gap, classzone_idx))
2959 shrink_zone(zone, sc);
2960 nodes_clear(shrink.nodes_to_scan);
2961 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2963 reclaim_state->reclaimed_slab = 0;
2964 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2965 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2967 /* Account for the number of pages attempted to reclaim */
2968 *nr_attempted += sc->nr_to_reclaim;
2970 zone_clear_flag(zone, ZONE_WRITEBACK);
2973 * If a zone reaches its high watermark, consider it to be no longer
2974 * congested. It's possible there are dirty pages backed by congested
2975 * BDIs but as pressure is relieved, speculatively avoid congestion
2978 if (zone_reclaimable(zone) &&
2979 zone_balanced(zone, testorder, 0, classzone_idx)) {
2980 zone_clear_flag(zone, ZONE_CONGESTED);
2981 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2984 return sc->nr_scanned >= sc->nr_to_reclaim;
2988 * For kswapd, balance_pgdat() will work across all this node's zones until
2989 * they are all at high_wmark_pages(zone).
2991 * Returns the final order kswapd was reclaiming at
2993 * There is special handling here for zones which are full of pinned pages.
2994 * This can happen if the pages are all mlocked, or if they are all used by
2995 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2996 * What we do is to detect the case where all pages in the zone have been
2997 * scanned twice and there has been zero successful reclaim. Mark the zone as
2998 * dead and from now on, only perform a short scan. Basically we're polling
2999 * the zone for when the problem goes away.
3001 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3002 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3003 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3004 * lower zones regardless of the number of free pages in the lower zones. This
3005 * interoperates with the page allocator fallback scheme to ensure that aging
3006 * of pages is balanced across the zones.
3008 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3012 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3013 unsigned long nr_soft_reclaimed;
3014 unsigned long nr_soft_scanned;
3015 struct scan_control sc = {
3016 .gfp_mask = GFP_KERNEL,
3018 .priority = DEF_PRIORITY,
3019 .may_writepage = !laptop_mode,
3023 count_vm_event(PAGEOUTRUN);
3026 unsigned long lru_pages = 0;
3027 unsigned long nr_attempted = 0;
3028 bool raise_priority = true;
3029 bool pgdat_needs_compaction = (order > 0);
3031 sc.nr_reclaimed = 0;
3034 * Scan in the highmem->dma direction for the highest
3035 * zone which needs scanning
3037 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3038 struct zone *zone = pgdat->node_zones + i;
3040 if (!populated_zone(zone))
3043 if (sc.priority != DEF_PRIORITY &&
3044 !zone_reclaimable(zone))
3048 * Do some background aging of the anon list, to give
3049 * pages a chance to be referenced before reclaiming.
3051 age_active_anon(zone, &sc);
3054 * If the number of buffer_heads in the machine
3055 * exceeds the maximum allowed level and this node
3056 * has a highmem zone, force kswapd to reclaim from
3057 * it to relieve lowmem pressure.
3059 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3064 if (!zone_balanced(zone, order, 0, 0)) {
3069 * If balanced, clear the dirty and congested
3072 zone_clear_flag(zone, ZONE_CONGESTED);
3073 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3080 for (i = 0; i <= end_zone; i++) {
3081 struct zone *zone = pgdat->node_zones + i;
3083 if (!populated_zone(zone))
3086 lru_pages += zone_reclaimable_pages(zone);
3089 * If any zone is currently balanced then kswapd will
3090 * not call compaction as it is expected that the
3091 * necessary pages are already available.
3093 if (pgdat_needs_compaction &&
3094 zone_watermark_ok(zone, order,
3095 low_wmark_pages(zone),
3097 pgdat_needs_compaction = false;
3101 * If we're getting trouble reclaiming, start doing writepage
3102 * even in laptop mode.
3104 if (sc.priority < DEF_PRIORITY - 2)
3105 sc.may_writepage = 1;
3108 * Now scan the zone in the dma->highmem direction, stopping
3109 * at the last zone which needs scanning.
3111 * We do this because the page allocator works in the opposite
3112 * direction. This prevents the page allocator from allocating
3113 * pages behind kswapd's direction of progress, which would
3114 * cause too much scanning of the lower zones.
3116 for (i = 0; i <= end_zone; i++) {
3117 struct zone *zone = pgdat->node_zones + i;
3119 if (!populated_zone(zone))
3122 if (sc.priority != DEF_PRIORITY &&
3123 !zone_reclaimable(zone))
3128 nr_soft_scanned = 0;
3130 * Call soft limit reclaim before calling shrink_zone.
3132 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3135 sc.nr_reclaimed += nr_soft_reclaimed;
3138 * There should be no need to raise the scanning
3139 * priority if enough pages are already being scanned
3140 * that that high watermark would be met at 100%
3143 if (kswapd_shrink_zone(zone, end_zone, &sc,
3144 lru_pages, &nr_attempted))
3145 raise_priority = false;
3149 * If the low watermark is met there is no need for processes
3150 * to be throttled on pfmemalloc_wait as they should not be
3151 * able to safely make forward progress. Wake them
3153 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3154 pfmemalloc_watermark_ok(pgdat))
3155 wake_up(&pgdat->pfmemalloc_wait);
3158 * Fragmentation may mean that the system cannot be rebalanced
3159 * for high-order allocations in all zones. If twice the
3160 * allocation size has been reclaimed and the zones are still
3161 * not balanced then recheck the watermarks at order-0 to
3162 * prevent kswapd reclaiming excessively. Assume that a
3163 * process requested a high-order can direct reclaim/compact.
3165 if (order && sc.nr_reclaimed >= 2UL << order)
3166 order = sc.order = 0;
3168 /* Check if kswapd should be suspending */
3169 if (try_to_freeze() || kthread_should_stop())
3173 * Compact if necessary and kswapd is reclaiming at least the
3174 * high watermark number of pages as requsted
3176 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3177 compact_pgdat(pgdat, order);
3180 * Raise priority if scanning rate is too low or there was no
3181 * progress in reclaiming pages
3183 if (raise_priority || !sc.nr_reclaimed)
3185 } while (sc.priority >= 1 &&
3186 !pgdat_balanced(pgdat, order, *classzone_idx));
3190 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3191 * makes a decision on the order we were last reclaiming at. However,
3192 * if another caller entered the allocator slow path while kswapd
3193 * was awake, order will remain at the higher level
3195 *classzone_idx = end_zone;
3199 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3204 if (freezing(current) || kthread_should_stop())
3207 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3209 /* Try to sleep for a short interval */
3210 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3211 remaining = schedule_timeout(HZ/10);
3212 finish_wait(&pgdat->kswapd_wait, &wait);
3213 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3217 * After a short sleep, check if it was a premature sleep. If not, then
3218 * go fully to sleep until explicitly woken up.
3220 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3221 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3224 * vmstat counters are not perfectly accurate and the estimated
3225 * value for counters such as NR_FREE_PAGES can deviate from the
3226 * true value by nr_online_cpus * threshold. To avoid the zone
3227 * watermarks being breached while under pressure, we reduce the
3228 * per-cpu vmstat threshold while kswapd is awake and restore
3229 * them before going back to sleep.
3231 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3234 * Compaction records what page blocks it recently failed to
3235 * isolate pages from and skips them in the future scanning.
3236 * When kswapd is going to sleep, it is reasonable to assume
3237 * that pages and compaction may succeed so reset the cache.
3239 reset_isolation_suitable(pgdat);
3241 if (!kthread_should_stop())
3244 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3247 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3249 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3251 finish_wait(&pgdat->kswapd_wait, &wait);
3255 * The background pageout daemon, started as a kernel thread
3256 * from the init process.
3258 * This basically trickles out pages so that we have _some_
3259 * free memory available even if there is no other activity
3260 * that frees anything up. This is needed for things like routing
3261 * etc, where we otherwise might have all activity going on in
3262 * asynchronous contexts that cannot page things out.
3264 * If there are applications that are active memory-allocators
3265 * (most normal use), this basically shouldn't matter.
3267 static int kswapd(void *p)
3269 unsigned long order, new_order;
3270 unsigned balanced_order;
3271 int classzone_idx, new_classzone_idx;
3272 int balanced_classzone_idx;
3273 pg_data_t *pgdat = (pg_data_t*)p;
3274 struct task_struct *tsk = current;
3276 struct reclaim_state reclaim_state = {
3277 .reclaimed_slab = 0,
3279 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3281 lockdep_set_current_reclaim_state(GFP_KERNEL);
3283 if (!cpumask_empty(cpumask))
3284 set_cpus_allowed_ptr(tsk, cpumask);
3285 current->reclaim_state = &reclaim_state;
3288 * Tell the memory management that we're a "memory allocator",
3289 * and that if we need more memory we should get access to it
3290 * regardless (see "__alloc_pages()"). "kswapd" should
3291 * never get caught in the normal page freeing logic.
3293 * (Kswapd normally doesn't need memory anyway, but sometimes
3294 * you need a small amount of memory in order to be able to
3295 * page out something else, and this flag essentially protects
3296 * us from recursively trying to free more memory as we're
3297 * trying to free the first piece of memory in the first place).
3299 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3302 order = new_order = 0;
3304 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3305 balanced_classzone_idx = classzone_idx;
3310 * If the last balance_pgdat was unsuccessful it's unlikely a
3311 * new request of a similar or harder type will succeed soon
3312 * so consider going to sleep on the basis we reclaimed at
3314 if (balanced_classzone_idx >= new_classzone_idx &&
3315 balanced_order == new_order) {
3316 new_order = pgdat->kswapd_max_order;
3317 new_classzone_idx = pgdat->classzone_idx;
3318 pgdat->kswapd_max_order = 0;
3319 pgdat->classzone_idx = pgdat->nr_zones - 1;
3322 if (order < new_order || classzone_idx > new_classzone_idx) {
3324 * Don't sleep if someone wants a larger 'order'
3325 * allocation or has tigher zone constraints
3328 classzone_idx = new_classzone_idx;
3330 kswapd_try_to_sleep(pgdat, balanced_order,
3331 balanced_classzone_idx);
3332 order = pgdat->kswapd_max_order;
3333 classzone_idx = pgdat->classzone_idx;
3335 new_classzone_idx = classzone_idx;
3336 pgdat->kswapd_max_order = 0;
3337 pgdat->classzone_idx = pgdat->nr_zones - 1;
3340 ret = try_to_freeze();
3341 if (kthread_should_stop())
3345 * We can speed up thawing tasks if we don't call balance_pgdat
3346 * after returning from the refrigerator
3349 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3350 balanced_classzone_idx = classzone_idx;
3351 balanced_order = balance_pgdat(pgdat, order,
3352 &balanced_classzone_idx);
3356 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3357 current->reclaim_state = NULL;
3358 lockdep_clear_current_reclaim_state();
3364 * A zone is low on free memory, so wake its kswapd task to service it.
3366 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3370 if (!populated_zone(zone))
3373 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3375 pgdat = zone->zone_pgdat;
3376 if (pgdat->kswapd_max_order < order) {
3377 pgdat->kswapd_max_order = order;
3378 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3380 if (!waitqueue_active(&pgdat->kswapd_wait))
3382 if (zone_balanced(zone, order, 0, 0))
3385 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3386 wake_up_interruptible(&pgdat->kswapd_wait);
3389 #ifdef CONFIG_HIBERNATION
3391 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3394 * Rather than trying to age LRUs the aim is to preserve the overall
3395 * LRU order by reclaiming preferentially
3396 * inactive > active > active referenced > active mapped
3398 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3400 struct reclaim_state reclaim_state;
3401 struct scan_control sc = {
3402 .nr_to_reclaim = nr_to_reclaim,
3403 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3404 .priority = DEF_PRIORITY,
3408 .hibernation_mode = 1,
3410 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3411 struct task_struct *p = current;
3412 unsigned long nr_reclaimed;
3414 p->flags |= PF_MEMALLOC;
3415 lockdep_set_current_reclaim_state(sc.gfp_mask);
3416 reclaim_state.reclaimed_slab = 0;
3417 p->reclaim_state = &reclaim_state;
3419 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3421 p->reclaim_state = NULL;
3422 lockdep_clear_current_reclaim_state();
3423 p->flags &= ~PF_MEMALLOC;
3425 return nr_reclaimed;
3427 #endif /* CONFIG_HIBERNATION */
3429 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3430 not required for correctness. So if the last cpu in a node goes
3431 away, we get changed to run anywhere: as the first one comes back,
3432 restore their cpu bindings. */
3433 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3438 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3439 for_each_node_state(nid, N_MEMORY) {
3440 pg_data_t *pgdat = NODE_DATA(nid);
3441 const struct cpumask *mask;
3443 mask = cpumask_of_node(pgdat->node_id);
3445 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3446 /* One of our CPUs online: restore mask */
3447 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3454 * This kswapd start function will be called by init and node-hot-add.
3455 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3457 int kswapd_run(int nid)
3459 pg_data_t *pgdat = NODE_DATA(nid);
3465 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3466 if (IS_ERR(pgdat->kswapd)) {
3467 /* failure at boot is fatal */
3468 BUG_ON(system_state == SYSTEM_BOOTING);
3469 pr_err("Failed to start kswapd on node %d\n", nid);
3470 ret = PTR_ERR(pgdat->kswapd);
3471 pgdat->kswapd = NULL;
3477 * Called by memory hotplug when all memory in a node is offlined. Caller must
3478 * hold mem_hotplug_begin/end().
3480 void kswapd_stop(int nid)
3482 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3485 kthread_stop(kswapd);
3486 NODE_DATA(nid)->kswapd = NULL;
3490 static int __init kswapd_init(void)
3495 for_each_node_state(nid, N_MEMORY)
3497 hotcpu_notifier(cpu_callback, 0);
3501 module_init(kswapd_init)
3507 * If non-zero call zone_reclaim when the number of free pages falls below
3510 int zone_reclaim_mode __read_mostly;
3512 #define RECLAIM_OFF 0
3513 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3514 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3515 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3518 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3519 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3522 #define ZONE_RECLAIM_PRIORITY 4
3525 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3528 int sysctl_min_unmapped_ratio = 1;
3531 * If the number of slab pages in a zone grows beyond this percentage then
3532 * slab reclaim needs to occur.
3534 int sysctl_min_slab_ratio = 5;
3536 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3538 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3539 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3540 zone_page_state(zone, NR_ACTIVE_FILE);
3543 * It's possible for there to be more file mapped pages than
3544 * accounted for by the pages on the file LRU lists because
3545 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3547 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3550 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3551 static long zone_pagecache_reclaimable(struct zone *zone)
3553 long nr_pagecache_reclaimable;
3557 * If RECLAIM_SWAP is set, then all file pages are considered
3558 * potentially reclaimable. Otherwise, we have to worry about
3559 * pages like swapcache and zone_unmapped_file_pages() provides
3562 if (zone_reclaim_mode & RECLAIM_SWAP)
3563 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3565 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3567 /* If we can't clean pages, remove dirty pages from consideration */
3568 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3569 delta += zone_page_state(zone, NR_FILE_DIRTY);
3571 /* Watch for any possible underflows due to delta */
3572 if (unlikely(delta > nr_pagecache_reclaimable))
3573 delta = nr_pagecache_reclaimable;
3575 return nr_pagecache_reclaimable - delta;
3579 * Try to free up some pages from this zone through reclaim.
3581 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3583 /* Minimum pages needed in order to stay on node */
3584 const unsigned long nr_pages = 1 << order;
3585 struct task_struct *p = current;
3586 struct reclaim_state reclaim_state;
3587 struct scan_control sc = {
3588 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3589 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3591 .priority = ZONE_RECLAIM_PRIORITY,
3592 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3593 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3596 struct shrink_control shrink = {
3597 .gfp_mask = sc.gfp_mask,
3599 unsigned long nr_slab_pages0, nr_slab_pages1;
3603 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3604 * and we also need to be able to write out pages for RECLAIM_WRITE
3607 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3608 lockdep_set_current_reclaim_state(gfp_mask);
3609 reclaim_state.reclaimed_slab = 0;
3610 p->reclaim_state = &reclaim_state;
3612 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3614 * Free memory by calling shrink zone with increasing
3615 * priorities until we have enough memory freed.
3618 shrink_zone(zone, &sc);
3619 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3622 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3623 if (nr_slab_pages0 > zone->min_slab_pages) {
3625 * shrink_slab() does not currently allow us to determine how
3626 * many pages were freed in this zone. So we take the current
3627 * number of slab pages and shake the slab until it is reduced
3628 * by the same nr_pages that we used for reclaiming unmapped
3631 nodes_clear(shrink.nodes_to_scan);
3632 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3634 unsigned long lru_pages = zone_reclaimable_pages(zone);
3636 /* No reclaimable slab or very low memory pressure */
3637 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3640 /* Freed enough memory */
3641 nr_slab_pages1 = zone_page_state(zone,
3642 NR_SLAB_RECLAIMABLE);
3643 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3648 * Update nr_reclaimed by the number of slab pages we
3649 * reclaimed from this zone.
3651 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3652 if (nr_slab_pages1 < nr_slab_pages0)
3653 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3656 p->reclaim_state = NULL;
3657 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3658 lockdep_clear_current_reclaim_state();
3659 return sc.nr_reclaimed >= nr_pages;
3662 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3668 * Zone reclaim reclaims unmapped file backed pages and
3669 * slab pages if we are over the defined limits.
3671 * A small portion of unmapped file backed pages is needed for
3672 * file I/O otherwise pages read by file I/O will be immediately
3673 * thrown out if the zone is overallocated. So we do not reclaim
3674 * if less than a specified percentage of the zone is used by
3675 * unmapped file backed pages.
3677 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3678 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3679 return ZONE_RECLAIM_FULL;
3681 if (!zone_reclaimable(zone))
3682 return ZONE_RECLAIM_FULL;
3685 * Do not scan if the allocation should not be delayed.
3687 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3688 return ZONE_RECLAIM_NOSCAN;
3691 * Only run zone reclaim on the local zone or on zones that do not
3692 * have associated processors. This will favor the local processor
3693 * over remote processors and spread off node memory allocations
3694 * as wide as possible.
3696 node_id = zone_to_nid(zone);
3697 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3698 return ZONE_RECLAIM_NOSCAN;
3700 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3701 return ZONE_RECLAIM_NOSCAN;
3703 ret = __zone_reclaim(zone, gfp_mask, order);
3704 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3707 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3714 * page_evictable - test whether a page is evictable
3715 * @page: the page to test
3717 * Test whether page is evictable--i.e., should be placed on active/inactive
3718 * lists vs unevictable list.
3720 * Reasons page might not be evictable:
3721 * (1) page's mapping marked unevictable
3722 * (2) page is part of an mlocked VMA
3725 int page_evictable(struct page *page)
3727 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3732 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3733 * @pages: array of pages to check
3734 * @nr_pages: number of pages to check
3736 * Checks pages for evictability and moves them to the appropriate lru list.
3738 * This function is only used for SysV IPC SHM_UNLOCK.
3740 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3742 struct lruvec *lruvec;
3743 struct zone *zone = NULL;
3748 for (i = 0; i < nr_pages; i++) {
3749 struct page *page = pages[i];
3750 struct zone *pagezone;
3753 pagezone = page_zone(page);
3754 if (pagezone != zone) {
3756 spin_unlock_irq(&zone->lru_lock);
3758 spin_lock_irq(&zone->lru_lock);
3760 lruvec = mem_cgroup_page_lruvec(page, zone);
3762 if (!PageLRU(page) || !PageUnevictable(page))
3765 if (page_evictable(page)) {
3766 enum lru_list lru = page_lru_base_type(page);
3768 VM_BUG_ON_PAGE(PageActive(page), page);
3769 ClearPageUnevictable(page);
3770 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3771 add_page_to_lru_list(page, lruvec, lru);
3777 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3778 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3779 spin_unlock_irq(&zone->lru_lock);
3782 #endif /* CONFIG_SHMEM */
3784 static void warn_scan_unevictable_pages(void)
3786 printk_once(KERN_WARNING
3787 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3788 "disabled for lack of a legitimate use case. If you have "
3789 "one, please send an email to linux-mm@kvack.org.\n",
3794 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3795 * all nodes' unevictable lists for evictable pages
3797 unsigned long scan_unevictable_pages;
3799 int scan_unevictable_handler(struct ctl_table *table, int write,
3800 void __user *buffer,
3801 size_t *length, loff_t *ppos)
3803 warn_scan_unevictable_pages();
3804 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3805 scan_unevictable_pages = 0;
3811 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3812 * a specified node's per zone unevictable lists for evictable pages.
3815 static ssize_t read_scan_unevictable_node(struct device *dev,
3816 struct device_attribute *attr,
3819 warn_scan_unevictable_pages();
3820 return sprintf(buf, "0\n"); /* always zero; should fit... */
3823 static ssize_t write_scan_unevictable_node(struct device *dev,
3824 struct device_attribute *attr,
3825 const char *buf, size_t count)
3827 warn_scan_unevictable_pages();
3832 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3833 read_scan_unevictable_node,
3834 write_scan_unevictable_node);
3836 int scan_unevictable_register_node(struct node *node)
3838 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3841 void scan_unevictable_unregister_node(struct node *node)
3843 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);