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 /* Incremented by the number of inactive pages that were scanned */
63 unsigned long nr_scanned;
65 /* Number of pages freed so far during a call to shrink_zones() */
66 unsigned long nr_reclaimed;
68 /* How many pages shrink_list() should reclaim */
69 unsigned long nr_to_reclaim;
71 unsigned long hibernation_mode;
73 /* This context's GFP mask */
78 /* Can mapped pages be reclaimed? */
81 /* Can pages be swapped as part of reclaim? */
86 /* Scan (total_size >> priority) pages at once */
89 /* anon vs. file LRUs scanning "ratio" */
93 * The memory cgroup that hit its limit and as a result is the
94 * primary target of this reclaim invocation.
96 struct mem_cgroup *target_mem_cgroup;
99 * Nodemask of nodes allowed by the caller. If NULL, all nodes
102 nodemask_t *nodemask;
105 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
107 #ifdef ARCH_HAS_PREFETCH
108 #define prefetch_prev_lru_page(_page, _base, _field) \
110 if ((_page)->lru.prev != _base) { \
113 prev = lru_to_page(&(_page->lru)); \
114 prefetch(&prev->_field); \
118 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
121 #ifdef ARCH_HAS_PREFETCHW
122 #define prefetchw_prev_lru_page(_page, _base, _field) \
124 if ((_page)->lru.prev != _base) { \
127 prev = lru_to_page(&(_page->lru)); \
128 prefetchw(&prev->_field); \
132 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136 * From 0 .. 100. Higher means more swappy.
138 int vm_swappiness = 60;
139 unsigned long vm_total_pages; /* The total number of pages which the VM controls */
141 static LIST_HEAD(shrinker_list);
142 static DECLARE_RWSEM(shrinker_rwsem);
145 static bool global_reclaim(struct scan_control *sc)
147 return !sc->target_mem_cgroup;
150 static bool global_reclaim(struct scan_control *sc)
156 static unsigned long zone_reclaimable_pages(struct zone *zone)
160 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
161 zone_page_state(zone, NR_INACTIVE_FILE);
163 if (get_nr_swap_pages() > 0)
164 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
165 zone_page_state(zone, NR_INACTIVE_ANON);
170 bool zone_reclaimable(struct zone *zone)
172 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
175 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
177 if (!mem_cgroup_disabled())
178 return mem_cgroup_get_lru_size(lruvec, lru);
180 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
184 * Add a shrinker callback to be called from the vm.
186 int register_shrinker(struct shrinker *shrinker)
188 size_t size = sizeof(*shrinker->nr_deferred);
191 * If we only have one possible node in the system anyway, save
192 * ourselves the trouble and disable NUMA aware behavior. This way we
193 * will save memory and some small loop time later.
195 if (nr_node_ids == 1)
196 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
198 if (shrinker->flags & SHRINKER_NUMA_AWARE)
201 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
202 if (!shrinker->nr_deferred)
205 down_write(&shrinker_rwsem);
206 list_add_tail(&shrinker->list, &shrinker_list);
207 up_write(&shrinker_rwsem);
210 EXPORT_SYMBOL(register_shrinker);
215 void unregister_shrinker(struct shrinker *shrinker)
217 down_write(&shrinker_rwsem);
218 list_del(&shrinker->list);
219 up_write(&shrinker_rwsem);
220 kfree(shrinker->nr_deferred);
222 EXPORT_SYMBOL(unregister_shrinker);
224 #define SHRINK_BATCH 128
227 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
228 unsigned long nr_pages_scanned, unsigned long lru_pages)
230 unsigned long freed = 0;
231 unsigned long long delta;
236 int nid = shrinkctl->nid;
237 long batch_size = shrinker->batch ? shrinker->batch
240 freeable = shrinker->count_objects(shrinker, shrinkctl);
245 * copy the current shrinker scan count into a local variable
246 * and zero it so that other concurrent shrinker invocations
247 * don't also do this scanning work.
249 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
252 delta = (4 * nr_pages_scanned) / shrinker->seeks;
254 do_div(delta, lru_pages + 1);
256 if (total_scan < 0) {
258 "shrink_slab: %pF negative objects to delete nr=%ld\n",
259 shrinker->scan_objects, total_scan);
260 total_scan = freeable;
264 * We need to avoid excessive windup on filesystem shrinkers
265 * due to large numbers of GFP_NOFS allocations causing the
266 * shrinkers to return -1 all the time. This results in a large
267 * nr being built up so when a shrink that can do some work
268 * comes along it empties the entire cache due to nr >>>
269 * freeable. This is bad for sustaining a working set in
272 * Hence only allow the shrinker to scan the entire cache when
273 * a large delta change is calculated directly.
275 if (delta < freeable / 4)
276 total_scan = min(total_scan, freeable / 2);
279 * Avoid risking looping forever due to too large nr value:
280 * never try to free more than twice the estimate number of
283 if (total_scan > freeable * 2)
284 total_scan = freeable * 2;
286 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
287 nr_pages_scanned, lru_pages,
288 freeable, delta, total_scan);
291 * Normally, we should not scan less than batch_size objects in one
292 * pass to avoid too frequent shrinker calls, but if the slab has less
293 * than batch_size objects in total and we are really tight on memory,
294 * we will try to reclaim all available objects, otherwise we can end
295 * up failing allocations although there are plenty of reclaimable
296 * objects spread over several slabs with usage less than the
299 * We detect the "tight on memory" situations by looking at the total
300 * number of objects we want to scan (total_scan). If it is greater
301 * than the total number of objects on slab (freeable), we must be
302 * scanning at high prio and therefore should try to reclaim as much as
305 while (total_scan >= batch_size ||
306 total_scan >= freeable) {
308 unsigned long nr_to_scan = min(batch_size, total_scan);
310 shrinkctl->nr_to_scan = nr_to_scan;
311 ret = shrinker->scan_objects(shrinker, shrinkctl);
312 if (ret == SHRINK_STOP)
316 count_vm_events(SLABS_SCANNED, nr_to_scan);
317 total_scan -= nr_to_scan;
323 * move the unused scan count back into the shrinker in a
324 * manner that handles concurrent updates. If we exhausted the
325 * scan, there is no need to do an update.
328 new_nr = atomic_long_add_return(total_scan,
329 &shrinker->nr_deferred[nid]);
331 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
333 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
338 * Call the shrink functions to age shrinkable caches
340 * Here we assume it costs one seek to replace a lru page and that it also
341 * takes a seek to recreate a cache object. With this in mind we age equal
342 * percentages of the lru and ageable caches. This should balance the seeks
343 * generated by these structures.
345 * If the vm encountered mapped pages on the LRU it increase the pressure on
346 * slab to avoid swapping.
348 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
350 * `lru_pages' represents the number of on-LRU pages in all the zones which
351 * are eligible for the caller's allocation attempt. It is used for balancing
352 * slab reclaim versus page reclaim.
354 * Returns the number of slab objects which we shrunk.
356 unsigned long shrink_slab(struct shrink_control *shrinkctl,
357 unsigned long nr_pages_scanned,
358 unsigned long lru_pages)
360 struct shrinker *shrinker;
361 unsigned long freed = 0;
363 if (nr_pages_scanned == 0)
364 nr_pages_scanned = SWAP_CLUSTER_MAX;
366 if (!down_read_trylock(&shrinker_rwsem)) {
368 * If we would return 0, our callers would understand that we
369 * have nothing else to shrink and give up trying. By returning
370 * 1 we keep it going and assume we'll be able to shrink next
377 list_for_each_entry(shrinker, &shrinker_list, list) {
378 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
380 freed += shrink_slab_node(shrinkctl, shrinker,
381 nr_pages_scanned, lru_pages);
385 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
386 if (node_online(shrinkctl->nid))
387 freed += shrink_slab_node(shrinkctl, shrinker,
388 nr_pages_scanned, lru_pages);
392 up_read(&shrinker_rwsem);
398 static inline int is_page_cache_freeable(struct page *page)
401 * A freeable page cache page is referenced only by the caller
402 * that isolated the page, the page cache radix tree and
403 * optional buffer heads at page->private.
405 return page_count(page) - page_has_private(page) == 2;
408 static int may_write_to_queue(struct backing_dev_info *bdi,
409 struct scan_control *sc)
411 if (current->flags & PF_SWAPWRITE)
413 if (!bdi_write_congested(bdi))
415 if (bdi == current->backing_dev_info)
421 * We detected a synchronous write error writing a page out. Probably
422 * -ENOSPC. We need to propagate that into the address_space for a subsequent
423 * fsync(), msync() or close().
425 * The tricky part is that after writepage we cannot touch the mapping: nothing
426 * prevents it from being freed up. But we have a ref on the page and once
427 * that page is locked, the mapping is pinned.
429 * We're allowed to run sleeping lock_page() here because we know the caller has
432 static void handle_write_error(struct address_space *mapping,
433 struct page *page, int error)
436 if (page_mapping(page) == mapping)
437 mapping_set_error(mapping, error);
441 /* possible outcome of pageout() */
443 /* failed to write page out, page is locked */
445 /* move page to the active list, page is locked */
447 /* page has been sent to the disk successfully, page is unlocked */
449 /* page is clean and locked */
454 * pageout is called by shrink_page_list() for each dirty page.
455 * Calls ->writepage().
457 static pageout_t pageout(struct page *page, struct address_space *mapping,
458 struct scan_control *sc)
461 * If the page is dirty, only perform writeback if that write
462 * will be non-blocking. To prevent this allocation from being
463 * stalled by pagecache activity. But note that there may be
464 * stalls if we need to run get_block(). We could test
465 * PagePrivate for that.
467 * If this process is currently in __generic_file_aio_write() against
468 * this page's queue, we can perform writeback even if that
471 * If the page is swapcache, write it back even if that would
472 * block, for some throttling. This happens by accident, because
473 * swap_backing_dev_info is bust: it doesn't reflect the
474 * congestion state of the swapdevs. Easy to fix, if needed.
476 if (!is_page_cache_freeable(page))
480 * Some data journaling orphaned pages can have
481 * page->mapping == NULL while being dirty with clean buffers.
483 if (page_has_private(page)) {
484 if (try_to_free_buffers(page)) {
485 ClearPageDirty(page);
486 pr_info("%s: orphaned page\n", __func__);
492 if (mapping->a_ops->writepage == NULL)
493 return PAGE_ACTIVATE;
494 if (!may_write_to_queue(mapping->backing_dev_info, sc))
497 if (clear_page_dirty_for_io(page)) {
499 struct writeback_control wbc = {
500 .sync_mode = WB_SYNC_NONE,
501 .nr_to_write = SWAP_CLUSTER_MAX,
503 .range_end = LLONG_MAX,
507 SetPageReclaim(page);
508 res = mapping->a_ops->writepage(page, &wbc);
510 handle_write_error(mapping, page, res);
511 if (res == AOP_WRITEPAGE_ACTIVATE) {
512 ClearPageReclaim(page);
513 return PAGE_ACTIVATE;
516 if (!PageWriteback(page)) {
517 /* synchronous write or broken a_ops? */
518 ClearPageReclaim(page);
520 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
521 inc_zone_page_state(page, NR_VMSCAN_WRITE);
529 * Same as remove_mapping, but if the page is removed from the mapping, it
530 * gets returned with a refcount of 0.
532 static int __remove_mapping(struct address_space *mapping, struct page *page,
535 BUG_ON(!PageLocked(page));
536 BUG_ON(mapping != page_mapping(page));
538 spin_lock_irq(&mapping->tree_lock);
540 * The non racy check for a busy page.
542 * Must be careful with the order of the tests. When someone has
543 * a ref to the page, it may be possible that they dirty it then
544 * drop the reference. So if PageDirty is tested before page_count
545 * here, then the following race may occur:
547 * get_user_pages(&page);
548 * [user mapping goes away]
550 * !PageDirty(page) [good]
551 * SetPageDirty(page);
553 * !page_count(page) [good, discard it]
555 * [oops, our write_to data is lost]
557 * Reversing the order of the tests ensures such a situation cannot
558 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
559 * load is not satisfied before that of page->_count.
561 * Note that if SetPageDirty is always performed via set_page_dirty,
562 * and thus under tree_lock, then this ordering is not required.
564 if (!page_freeze_refs(page, 2))
566 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
567 if (unlikely(PageDirty(page))) {
568 page_unfreeze_refs(page, 2);
572 if (PageSwapCache(page)) {
573 swp_entry_t swap = { .val = page_private(page) };
574 __delete_from_swap_cache(page);
575 spin_unlock_irq(&mapping->tree_lock);
576 swapcache_free(swap, page);
578 void (*freepage)(struct page *);
581 freepage = mapping->a_ops->freepage;
583 * Remember a shadow entry for reclaimed file cache in
584 * order to detect refaults, thus thrashing, later on.
586 * But don't store shadows in an address space that is
587 * already exiting. This is not just an optizimation,
588 * inode reclaim needs to empty out the radix tree or
589 * the nodes are lost. Don't plant shadows behind its
592 if (reclaimed && page_is_file_cache(page) &&
593 !mapping_exiting(mapping))
594 shadow = workingset_eviction(mapping, page);
595 __delete_from_page_cache(page, shadow);
596 spin_unlock_irq(&mapping->tree_lock);
597 mem_cgroup_uncharge_cache_page(page);
599 if (freepage != NULL)
606 spin_unlock_irq(&mapping->tree_lock);
611 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
612 * someone else has a ref on the page, abort and return 0. If it was
613 * successfully detached, return 1. Assumes the caller has a single ref on
616 int remove_mapping(struct address_space *mapping, struct page *page)
618 if (__remove_mapping(mapping, page, false)) {
620 * Unfreezing the refcount with 1 rather than 2 effectively
621 * drops the pagecache ref for us without requiring another
624 page_unfreeze_refs(page, 1);
631 * putback_lru_page - put previously isolated page onto appropriate LRU list
632 * @page: page to be put back to appropriate lru list
634 * Add previously isolated @page to appropriate LRU list.
635 * Page may still be unevictable for other reasons.
637 * lru_lock must not be held, interrupts must be enabled.
639 void putback_lru_page(struct page *page)
642 int was_unevictable = PageUnevictable(page);
644 VM_BUG_ON_PAGE(PageLRU(page), page);
647 ClearPageUnevictable(page);
649 if (page_evictable(page)) {
651 * For evictable pages, we can use the cache.
652 * In event of a race, worst case is we end up with an
653 * unevictable page on [in]active list.
654 * We know how to handle that.
656 is_unevictable = false;
660 * Put unevictable pages directly on zone's unevictable
663 is_unevictable = true;
664 add_page_to_unevictable_list(page);
666 * When racing with an mlock or AS_UNEVICTABLE clearing
667 * (page is unlocked) make sure that if the other thread
668 * does not observe our setting of PG_lru and fails
669 * isolation/check_move_unevictable_pages,
670 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
671 * the page back to the evictable list.
673 * The other side is TestClearPageMlocked() or shmem_lock().
679 * page's status can change while we move it among lru. If an evictable
680 * page is on unevictable list, it never be freed. To avoid that,
681 * check after we added it to the list, again.
683 if (is_unevictable && page_evictable(page)) {
684 if (!isolate_lru_page(page)) {
688 /* This means someone else dropped this page from LRU
689 * So, it will be freed or putback to LRU again. There is
690 * nothing to do here.
694 if (was_unevictable && !is_unevictable)
695 count_vm_event(UNEVICTABLE_PGRESCUED);
696 else if (!was_unevictable && is_unevictable)
697 count_vm_event(UNEVICTABLE_PGCULLED);
699 put_page(page); /* drop ref from isolate */
702 enum page_references {
704 PAGEREF_RECLAIM_CLEAN,
709 static enum page_references page_check_references(struct page *page,
710 struct scan_control *sc)
712 int referenced_ptes, referenced_page;
713 unsigned long vm_flags;
715 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
717 referenced_page = TestClearPageReferenced(page);
720 * Mlock lost the isolation race with us. Let try_to_unmap()
721 * move the page to the unevictable list.
723 if (vm_flags & VM_LOCKED)
724 return PAGEREF_RECLAIM;
726 if (referenced_ptes) {
727 if (PageSwapBacked(page))
728 return PAGEREF_ACTIVATE;
730 * All mapped pages start out with page table
731 * references from the instantiating fault, so we need
732 * to look twice if a mapped file page is used more
735 * Mark it and spare it for another trip around the
736 * inactive list. Another page table reference will
737 * lead to its activation.
739 * Note: the mark is set for activated pages as well
740 * so that recently deactivated but used pages are
743 SetPageReferenced(page);
745 if (referenced_page || referenced_ptes > 1)
746 return PAGEREF_ACTIVATE;
749 * Activate file-backed executable pages after first usage.
751 if (vm_flags & VM_EXEC)
752 return PAGEREF_ACTIVATE;
757 /* Reclaim if clean, defer dirty pages to writeback */
758 if (referenced_page && !PageSwapBacked(page))
759 return PAGEREF_RECLAIM_CLEAN;
761 return PAGEREF_RECLAIM;
764 /* Check if a page is dirty or under writeback */
765 static void page_check_dirty_writeback(struct page *page,
766 bool *dirty, bool *writeback)
768 struct address_space *mapping;
771 * Anonymous pages are not handled by flushers and must be written
772 * from reclaim context. Do not stall reclaim based on them
774 if (!page_is_file_cache(page)) {
780 /* By default assume that the page flags are accurate */
781 *dirty = PageDirty(page);
782 *writeback = PageWriteback(page);
784 /* Verify dirty/writeback state if the filesystem supports it */
785 if (!page_has_private(page))
788 mapping = page_mapping(page);
789 if (mapping && mapping->a_ops->is_dirty_writeback)
790 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
794 * shrink_page_list() returns the number of reclaimed pages
796 static unsigned long shrink_page_list(struct list_head *page_list,
798 struct scan_control *sc,
799 enum ttu_flags ttu_flags,
800 unsigned long *ret_nr_dirty,
801 unsigned long *ret_nr_unqueued_dirty,
802 unsigned long *ret_nr_congested,
803 unsigned long *ret_nr_writeback,
804 unsigned long *ret_nr_immediate,
807 LIST_HEAD(ret_pages);
808 LIST_HEAD(free_pages);
810 unsigned long nr_unqueued_dirty = 0;
811 unsigned long nr_dirty = 0;
812 unsigned long nr_congested = 0;
813 unsigned long nr_reclaimed = 0;
814 unsigned long nr_writeback = 0;
815 unsigned long nr_immediate = 0;
819 mem_cgroup_uncharge_start();
820 while (!list_empty(page_list)) {
821 struct address_space *mapping;
824 enum page_references references = PAGEREF_RECLAIM_CLEAN;
825 bool dirty, writeback;
829 page = lru_to_page(page_list);
830 list_del(&page->lru);
832 if (!trylock_page(page))
835 VM_BUG_ON_PAGE(PageActive(page), page);
836 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
840 if (unlikely(!page_evictable(page)))
843 if (!sc->may_unmap && page_mapped(page))
846 /* Double the slab pressure for mapped and swapcache pages */
847 if (page_mapped(page) || PageSwapCache(page))
850 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
851 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
854 * The number of dirty pages determines if a zone is marked
855 * reclaim_congested which affects wait_iff_congested. kswapd
856 * will stall and start writing pages if the tail of the LRU
857 * is all dirty unqueued pages.
859 page_check_dirty_writeback(page, &dirty, &writeback);
860 if (dirty || writeback)
863 if (dirty && !writeback)
867 * Treat this page as congested if the underlying BDI is or if
868 * pages are cycling through the LRU so quickly that the
869 * pages marked for immediate reclaim are making it to the
870 * end of the LRU a second time.
872 mapping = page_mapping(page);
873 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
874 (writeback && PageReclaim(page)))
878 * If a page at the tail of the LRU is under writeback, there
879 * are three cases to consider.
881 * 1) If reclaim is encountering an excessive number of pages
882 * under writeback and this page is both under writeback and
883 * PageReclaim then it indicates that pages are being queued
884 * for IO but are being recycled through the LRU before the
885 * IO can complete. Waiting on the page itself risks an
886 * indefinite stall if it is impossible to writeback the
887 * page due to IO error or disconnected storage so instead
888 * note that the LRU is being scanned too quickly and the
889 * caller can stall after page list has been processed.
891 * 2) Global reclaim encounters a page, memcg encounters a
892 * page that is not marked for immediate reclaim or
893 * the caller does not have __GFP_IO. In this case mark
894 * the page for immediate reclaim and continue scanning.
896 * __GFP_IO is checked because a loop driver thread might
897 * enter reclaim, and deadlock if it waits on a page for
898 * which it is needed to do the write (loop masks off
899 * __GFP_IO|__GFP_FS for this reason); but more thought
900 * would probably show more reasons.
902 * Don't require __GFP_FS, since we're not going into the
903 * FS, just waiting on its writeback completion. Worryingly,
904 * ext4 gfs2 and xfs allocate pages with
905 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
906 * may_enter_fs here is liable to OOM on them.
908 * 3) memcg encounters a page that is not already marked
909 * PageReclaim. memcg does not have any dirty pages
910 * throttling so we could easily OOM just because too many
911 * pages are in writeback and there is nothing else to
912 * reclaim. Wait for the writeback to complete.
914 if (PageWriteback(page)) {
916 if (current_is_kswapd() &&
918 zone_is_reclaim_writeback(zone)) {
923 } else if (global_reclaim(sc) ||
924 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
926 * This is slightly racy - end_page_writeback()
927 * might have just cleared PageReclaim, then
928 * setting PageReclaim here end up interpreted
929 * as PageReadahead - but that does not matter
930 * enough to care. What we do want is for this
931 * page to have PageReclaim set next time memcg
932 * reclaim reaches the tests above, so it will
933 * then wait_on_page_writeback() to avoid OOM;
934 * and it's also appropriate in global reclaim.
936 SetPageReclaim(page);
943 wait_on_page_writeback(page);
948 references = page_check_references(page, sc);
950 switch (references) {
951 case PAGEREF_ACTIVATE:
952 goto activate_locked;
955 case PAGEREF_RECLAIM:
956 case PAGEREF_RECLAIM_CLEAN:
957 ; /* try to reclaim the page below */
961 * Anonymous process memory has backing store?
962 * Try to allocate it some swap space here.
964 if (PageAnon(page) && !PageSwapCache(page)) {
965 if (!(sc->gfp_mask & __GFP_IO))
967 if (!add_to_swap(page, page_list))
968 goto activate_locked;
971 /* Adding to swap updated mapping */
972 mapping = page_mapping(page);
976 * The page is mapped into the page tables of one or more
977 * processes. Try to unmap it here.
979 if (page_mapped(page) && mapping) {
980 switch (try_to_unmap(page, ttu_flags)) {
982 goto activate_locked;
988 ; /* try to free the page below */
992 if (PageDirty(page)) {
994 * Only kswapd can writeback filesystem pages to
995 * avoid risk of stack overflow but only writeback
996 * if many dirty pages have been encountered.
998 if (page_is_file_cache(page) &&
999 (!current_is_kswapd() ||
1000 !zone_is_reclaim_dirty(zone))) {
1002 * Immediately reclaim when written back.
1003 * Similar in principal to deactivate_page()
1004 * except we already have the page isolated
1005 * and know it's dirty
1007 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1008 SetPageReclaim(page);
1013 if (references == PAGEREF_RECLAIM_CLEAN)
1017 if (!sc->may_writepage)
1020 /* Page is dirty, try to write it out here */
1021 switch (pageout(page, mapping, sc)) {
1025 goto activate_locked;
1027 if (PageWriteback(page))
1029 if (PageDirty(page))
1033 * A synchronous write - probably a ramdisk. Go
1034 * ahead and try to reclaim the page.
1036 if (!trylock_page(page))
1038 if (PageDirty(page) || PageWriteback(page))
1040 mapping = page_mapping(page);
1042 ; /* try to free the page below */
1047 * If the page has buffers, try to free the buffer mappings
1048 * associated with this page. If we succeed we try to free
1051 * We do this even if the page is PageDirty().
1052 * try_to_release_page() does not perform I/O, but it is
1053 * possible for a page to have PageDirty set, but it is actually
1054 * clean (all its buffers are clean). This happens if the
1055 * buffers were written out directly, with submit_bh(). ext3
1056 * will do this, as well as the blockdev mapping.
1057 * try_to_release_page() will discover that cleanness and will
1058 * drop the buffers and mark the page clean - it can be freed.
1060 * Rarely, pages can have buffers and no ->mapping. These are
1061 * the pages which were not successfully invalidated in
1062 * truncate_complete_page(). We try to drop those buffers here
1063 * and if that worked, and the page is no longer mapped into
1064 * process address space (page_count == 1) it can be freed.
1065 * Otherwise, leave the page on the LRU so it is swappable.
1067 if (page_has_private(page)) {
1068 if (!try_to_release_page(page, sc->gfp_mask))
1069 goto activate_locked;
1070 if (!mapping && page_count(page) == 1) {
1072 if (put_page_testzero(page))
1076 * rare race with speculative reference.
1077 * the speculative reference will free
1078 * this page shortly, so we may
1079 * increment nr_reclaimed here (and
1080 * leave it off the LRU).
1088 if (!mapping || !__remove_mapping(mapping, page, true))
1092 * At this point, we have no other references and there is
1093 * no way to pick any more up (removed from LRU, removed
1094 * from pagecache). Can use non-atomic bitops now (and
1095 * we obviously don't have to worry about waking up a process
1096 * waiting on the page lock, because there are no references.
1098 __clear_page_locked(page);
1103 * Is there need to periodically free_page_list? It would
1104 * appear not as the counts should be low
1106 list_add(&page->lru, &free_pages);
1110 if (PageSwapCache(page))
1111 try_to_free_swap(page);
1113 putback_lru_page(page);
1117 /* Not a candidate for swapping, so reclaim swap space. */
1118 if (PageSwapCache(page) && vm_swap_full())
1119 try_to_free_swap(page);
1120 VM_BUG_ON_PAGE(PageActive(page), page);
1121 SetPageActive(page);
1126 list_add(&page->lru, &ret_pages);
1127 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1130 free_hot_cold_page_list(&free_pages, true);
1132 list_splice(&ret_pages, page_list);
1133 count_vm_events(PGACTIVATE, pgactivate);
1134 mem_cgroup_uncharge_end();
1135 *ret_nr_dirty += nr_dirty;
1136 *ret_nr_congested += nr_congested;
1137 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1138 *ret_nr_writeback += nr_writeback;
1139 *ret_nr_immediate += nr_immediate;
1140 return nr_reclaimed;
1143 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1144 struct list_head *page_list)
1146 struct scan_control sc = {
1147 .gfp_mask = GFP_KERNEL,
1148 .priority = DEF_PRIORITY,
1151 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1152 struct page *page, *next;
1153 LIST_HEAD(clean_pages);
1155 list_for_each_entry_safe(page, next, page_list, lru) {
1156 if (page_is_file_cache(page) && !PageDirty(page) &&
1157 !isolated_balloon_page(page)) {
1158 ClearPageActive(page);
1159 list_move(&page->lru, &clean_pages);
1163 ret = shrink_page_list(&clean_pages, zone, &sc,
1164 TTU_UNMAP|TTU_IGNORE_ACCESS,
1165 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1166 list_splice(&clean_pages, page_list);
1167 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1172 * Attempt to remove the specified page from its LRU. Only take this page
1173 * if it is of the appropriate PageActive status. Pages which are being
1174 * freed elsewhere are also ignored.
1176 * page: page to consider
1177 * mode: one of the LRU isolation modes defined above
1179 * returns 0 on success, -ve errno on failure.
1181 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1185 /* Only take pages on the LRU. */
1189 /* Compaction should not handle unevictable pages but CMA can do so */
1190 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1196 * To minimise LRU disruption, the caller can indicate that it only
1197 * wants to isolate pages it will be able to operate on without
1198 * blocking - clean pages for the most part.
1200 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1201 * is used by reclaim when it is cannot write to backing storage
1203 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1204 * that it is possible to migrate without blocking
1206 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1207 /* All the caller can do on PageWriteback is block */
1208 if (PageWriteback(page))
1211 if (PageDirty(page)) {
1212 struct address_space *mapping;
1214 /* ISOLATE_CLEAN means only clean pages */
1215 if (mode & ISOLATE_CLEAN)
1219 * Only pages without mappings or that have a
1220 * ->migratepage callback are possible to migrate
1223 mapping = page_mapping(page);
1224 if (mapping && !mapping->a_ops->migratepage)
1229 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1232 if (likely(get_page_unless_zero(page))) {
1234 * Be careful not to clear PageLRU until after we're
1235 * sure the page is not being freed elsewhere -- the
1236 * page release code relies on it.
1246 * zone->lru_lock is heavily contended. Some of the functions that
1247 * shrink the lists perform better by taking out a batch of pages
1248 * and working on them outside the LRU lock.
1250 * For pagecache intensive workloads, this function is the hottest
1251 * spot in the kernel (apart from copy_*_user functions).
1253 * Appropriate locks must be held before calling this function.
1255 * @nr_to_scan: The number of pages to look through on the list.
1256 * @lruvec: The LRU vector to pull pages from.
1257 * @dst: The temp list to put pages on to.
1258 * @nr_scanned: The number of pages that were scanned.
1259 * @sc: The scan_control struct for this reclaim session
1260 * @mode: One of the LRU isolation modes
1261 * @lru: LRU list id for isolating
1263 * returns how many pages were moved onto *@dst.
1265 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1266 struct lruvec *lruvec, struct list_head *dst,
1267 unsigned long *nr_scanned, struct scan_control *sc,
1268 isolate_mode_t mode, enum lru_list lru)
1270 struct list_head *src = &lruvec->lists[lru];
1271 unsigned long nr_taken = 0;
1274 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1278 page = lru_to_page(src);
1279 prefetchw_prev_lru_page(page, src, flags);
1281 VM_BUG_ON_PAGE(!PageLRU(page), page);
1283 switch (__isolate_lru_page(page, mode)) {
1285 nr_pages = hpage_nr_pages(page);
1286 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1287 list_move(&page->lru, dst);
1288 nr_taken += nr_pages;
1292 /* else it is being freed elsewhere */
1293 list_move(&page->lru, src);
1302 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1303 nr_taken, mode, is_file_lru(lru));
1308 * isolate_lru_page - tries to isolate a page from its LRU list
1309 * @page: page to isolate from its LRU list
1311 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1312 * vmstat statistic corresponding to whatever LRU list the page was on.
1314 * Returns 0 if the page was removed from an LRU list.
1315 * Returns -EBUSY if the page was not on an LRU list.
1317 * The returned page will have PageLRU() cleared. If it was found on
1318 * the active list, it will have PageActive set. If it was found on
1319 * the unevictable list, it will have the PageUnevictable bit set. That flag
1320 * may need to be cleared by the caller before letting the page go.
1322 * The vmstat statistic corresponding to the list on which the page was
1323 * found will be decremented.
1326 * (1) Must be called with an elevated refcount on the page. This is a
1327 * fundamentnal difference from isolate_lru_pages (which is called
1328 * without a stable reference).
1329 * (2) the lru_lock must not be held.
1330 * (3) interrupts must be enabled.
1332 int isolate_lru_page(struct page *page)
1336 VM_BUG_ON_PAGE(!page_count(page), page);
1338 if (PageLRU(page)) {
1339 struct zone *zone = page_zone(page);
1340 struct lruvec *lruvec;
1342 spin_lock_irq(&zone->lru_lock);
1343 lruvec = mem_cgroup_page_lruvec(page, zone);
1344 if (PageLRU(page)) {
1345 int lru = page_lru(page);
1348 del_page_from_lru_list(page, lruvec, lru);
1351 spin_unlock_irq(&zone->lru_lock);
1357 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1358 * then get resheduled. When there are massive number of tasks doing page
1359 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1360 * the LRU list will go small and be scanned faster than necessary, leading to
1361 * unnecessary swapping, thrashing and OOM.
1363 static int too_many_isolated(struct zone *zone, int file,
1364 struct scan_control *sc)
1366 unsigned long inactive, isolated;
1368 if (current_is_kswapd())
1371 if (!global_reclaim(sc))
1375 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1376 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1378 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1379 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1383 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1384 * won't get blocked by normal direct-reclaimers, forming a circular
1387 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1390 return isolated > inactive;
1393 static noinline_for_stack void
1394 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1396 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1397 struct zone *zone = lruvec_zone(lruvec);
1398 LIST_HEAD(pages_to_free);
1401 * Put back any unfreeable pages.
1403 while (!list_empty(page_list)) {
1404 struct page *page = lru_to_page(page_list);
1407 VM_BUG_ON_PAGE(PageLRU(page), page);
1408 list_del(&page->lru);
1409 if (unlikely(!page_evictable(page))) {
1410 spin_unlock_irq(&zone->lru_lock);
1411 putback_lru_page(page);
1412 spin_lock_irq(&zone->lru_lock);
1416 lruvec = mem_cgroup_page_lruvec(page, zone);
1419 lru = page_lru(page);
1420 add_page_to_lru_list(page, lruvec, lru);
1422 if (is_active_lru(lru)) {
1423 int file = is_file_lru(lru);
1424 int numpages = hpage_nr_pages(page);
1425 reclaim_stat->recent_rotated[file] += numpages;
1427 if (put_page_testzero(page)) {
1428 __ClearPageLRU(page);
1429 __ClearPageActive(page);
1430 del_page_from_lru_list(page, lruvec, lru);
1432 if (unlikely(PageCompound(page))) {
1433 spin_unlock_irq(&zone->lru_lock);
1434 (*get_compound_page_dtor(page))(page);
1435 spin_lock_irq(&zone->lru_lock);
1437 list_add(&page->lru, &pages_to_free);
1442 * To save our caller's stack, now use input list for pages to free.
1444 list_splice(&pages_to_free, page_list);
1448 * If a kernel thread (such as nfsd for loop-back mounts) services
1449 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1450 * In that case we should only throttle if the backing device it is
1451 * writing to is congested. In other cases it is safe to throttle.
1453 static int current_may_throttle(void)
1455 return !(current->flags & PF_LESS_THROTTLE) ||
1456 current->backing_dev_info == NULL ||
1457 bdi_write_congested(current->backing_dev_info);
1461 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1462 * of reclaimed pages
1464 static noinline_for_stack unsigned long
1465 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1466 struct scan_control *sc, enum lru_list lru)
1468 LIST_HEAD(page_list);
1469 unsigned long nr_scanned;
1470 unsigned long nr_reclaimed = 0;
1471 unsigned long nr_taken;
1472 unsigned long nr_dirty = 0;
1473 unsigned long nr_congested = 0;
1474 unsigned long nr_unqueued_dirty = 0;
1475 unsigned long nr_writeback = 0;
1476 unsigned long nr_immediate = 0;
1477 isolate_mode_t isolate_mode = 0;
1478 int file = is_file_lru(lru);
1479 struct zone *zone = lruvec_zone(lruvec);
1480 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1482 while (unlikely(too_many_isolated(zone, file, sc))) {
1483 congestion_wait(BLK_RW_ASYNC, HZ/10);
1485 /* We are about to die and free our memory. Return now. */
1486 if (fatal_signal_pending(current))
1487 return SWAP_CLUSTER_MAX;
1493 isolate_mode |= ISOLATE_UNMAPPED;
1494 if (!sc->may_writepage)
1495 isolate_mode |= ISOLATE_CLEAN;
1497 spin_lock_irq(&zone->lru_lock);
1499 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1500 &nr_scanned, sc, isolate_mode, lru);
1502 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1503 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1505 if (global_reclaim(sc)) {
1506 zone->pages_scanned += nr_scanned;
1507 if (current_is_kswapd())
1508 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1510 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1512 spin_unlock_irq(&zone->lru_lock);
1517 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1518 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1519 &nr_writeback, &nr_immediate,
1522 spin_lock_irq(&zone->lru_lock);
1524 reclaim_stat->recent_scanned[file] += nr_taken;
1526 if (global_reclaim(sc)) {
1527 if (current_is_kswapd())
1528 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1531 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1535 putback_inactive_pages(lruvec, &page_list);
1537 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1539 spin_unlock_irq(&zone->lru_lock);
1541 free_hot_cold_page_list(&page_list, true);
1544 * If reclaim is isolating dirty pages under writeback, it implies
1545 * that the long-lived page allocation rate is exceeding the page
1546 * laundering rate. Either the global limits are not being effective
1547 * at throttling processes due to the page distribution throughout
1548 * zones or there is heavy usage of a slow backing device. The
1549 * only option is to throttle from reclaim context which is not ideal
1550 * as there is no guarantee the dirtying process is throttled in the
1551 * same way balance_dirty_pages() manages.
1553 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1554 * of pages under pages flagged for immediate reclaim and stall if any
1555 * are encountered in the nr_immediate check below.
1557 if (nr_writeback && nr_writeback == nr_taken)
1558 zone_set_flag(zone, ZONE_WRITEBACK);
1561 * memcg will stall in page writeback so only consider forcibly
1562 * stalling for global reclaim
1564 if (global_reclaim(sc)) {
1566 * Tag a zone as congested if all the dirty pages scanned were
1567 * backed by a congested BDI and wait_iff_congested will stall.
1569 if (nr_dirty && nr_dirty == nr_congested)
1570 zone_set_flag(zone, ZONE_CONGESTED);
1573 * If dirty pages are scanned that are not queued for IO, it
1574 * implies that flushers are not keeping up. In this case, flag
1575 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1576 * pages from reclaim context. It will forcibly stall in the
1579 if (nr_unqueued_dirty == nr_taken)
1580 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1583 * In addition, if kswapd scans pages marked marked for
1584 * immediate reclaim and under writeback (nr_immediate), it
1585 * implies that pages are cycling through the LRU faster than
1586 * they are written so also forcibly stall.
1588 if ((nr_unqueued_dirty == nr_taken || nr_immediate) &&
1589 current_may_throttle())
1590 congestion_wait(BLK_RW_ASYNC, HZ/10);
1594 * Stall direct reclaim for IO completions if underlying BDIs or zone
1595 * is congested. Allow kswapd to continue until it starts encountering
1596 * unqueued dirty pages or cycling through the LRU too quickly.
1598 if (!sc->hibernation_mode && !current_is_kswapd() &&
1599 current_may_throttle())
1600 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1602 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1604 nr_scanned, nr_reclaimed,
1606 trace_shrink_flags(file));
1607 return nr_reclaimed;
1611 * This moves pages from the active list to the inactive list.
1613 * We move them the other way if the page is referenced by one or more
1614 * processes, from rmap.
1616 * If the pages are mostly unmapped, the processing is fast and it is
1617 * appropriate to hold zone->lru_lock across the whole operation. But if
1618 * the pages are mapped, the processing is slow (page_referenced()) so we
1619 * should drop zone->lru_lock around each page. It's impossible to balance
1620 * this, so instead we remove the pages from the LRU while processing them.
1621 * It is safe to rely on PG_active against the non-LRU pages in here because
1622 * nobody will play with that bit on a non-LRU page.
1624 * The downside is that we have to touch page->_count against each page.
1625 * But we had to alter page->flags anyway.
1628 static void move_active_pages_to_lru(struct lruvec *lruvec,
1629 struct list_head *list,
1630 struct list_head *pages_to_free,
1633 struct zone *zone = lruvec_zone(lruvec);
1634 unsigned long pgmoved = 0;
1638 while (!list_empty(list)) {
1639 page = lru_to_page(list);
1640 lruvec = mem_cgroup_page_lruvec(page, zone);
1642 VM_BUG_ON_PAGE(PageLRU(page), page);
1645 nr_pages = hpage_nr_pages(page);
1646 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1647 list_move(&page->lru, &lruvec->lists[lru]);
1648 pgmoved += nr_pages;
1650 if (put_page_testzero(page)) {
1651 __ClearPageLRU(page);
1652 __ClearPageActive(page);
1653 del_page_from_lru_list(page, lruvec, lru);
1655 if (unlikely(PageCompound(page))) {
1656 spin_unlock_irq(&zone->lru_lock);
1657 (*get_compound_page_dtor(page))(page);
1658 spin_lock_irq(&zone->lru_lock);
1660 list_add(&page->lru, pages_to_free);
1663 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1664 if (!is_active_lru(lru))
1665 __count_vm_events(PGDEACTIVATE, pgmoved);
1668 static void shrink_active_list(unsigned long nr_to_scan,
1669 struct lruvec *lruvec,
1670 struct scan_control *sc,
1673 unsigned long nr_taken;
1674 unsigned long nr_scanned;
1675 unsigned long vm_flags;
1676 LIST_HEAD(l_hold); /* The pages which were snipped off */
1677 LIST_HEAD(l_active);
1678 LIST_HEAD(l_inactive);
1680 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1681 unsigned long nr_rotated = 0;
1682 isolate_mode_t isolate_mode = 0;
1683 int file = is_file_lru(lru);
1684 struct zone *zone = lruvec_zone(lruvec);
1689 isolate_mode |= ISOLATE_UNMAPPED;
1690 if (!sc->may_writepage)
1691 isolate_mode |= ISOLATE_CLEAN;
1693 spin_lock_irq(&zone->lru_lock);
1695 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1696 &nr_scanned, sc, isolate_mode, lru);
1697 if (global_reclaim(sc))
1698 zone->pages_scanned += nr_scanned;
1700 reclaim_stat->recent_scanned[file] += nr_taken;
1702 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1703 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1704 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1705 spin_unlock_irq(&zone->lru_lock);
1707 while (!list_empty(&l_hold)) {
1709 page = lru_to_page(&l_hold);
1710 list_del(&page->lru);
1712 if (unlikely(!page_evictable(page))) {
1713 putback_lru_page(page);
1717 if (unlikely(buffer_heads_over_limit)) {
1718 if (page_has_private(page) && trylock_page(page)) {
1719 if (page_has_private(page))
1720 try_to_release_page(page, 0);
1725 if (page_referenced(page, 0, sc->target_mem_cgroup,
1727 nr_rotated += hpage_nr_pages(page);
1729 * Identify referenced, file-backed active pages and
1730 * give them one more trip around the active list. So
1731 * that executable code get better chances to stay in
1732 * memory under moderate memory pressure. Anon pages
1733 * are not likely to be evicted by use-once streaming
1734 * IO, plus JVM can create lots of anon VM_EXEC pages,
1735 * so we ignore them here.
1737 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1738 list_add(&page->lru, &l_active);
1743 ClearPageActive(page); /* we are de-activating */
1744 list_add(&page->lru, &l_inactive);
1748 * Move pages back to the lru list.
1750 spin_lock_irq(&zone->lru_lock);
1752 * Count referenced pages from currently used mappings as rotated,
1753 * even though only some of them are actually re-activated. This
1754 * helps balance scan pressure between file and anonymous pages in
1757 reclaim_stat->recent_rotated[file] += nr_rotated;
1759 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1760 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1761 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1762 spin_unlock_irq(&zone->lru_lock);
1764 free_hot_cold_page_list(&l_hold, true);
1768 static int inactive_anon_is_low_global(struct zone *zone)
1770 unsigned long active, inactive;
1772 active = zone_page_state(zone, NR_ACTIVE_ANON);
1773 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1775 if (inactive * zone->inactive_ratio < active)
1782 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1783 * @lruvec: LRU vector to check
1785 * Returns true if the zone does not have enough inactive anon pages,
1786 * meaning some active anon pages need to be deactivated.
1788 static int inactive_anon_is_low(struct lruvec *lruvec)
1791 * If we don't have swap space, anonymous page deactivation
1794 if (!total_swap_pages)
1797 if (!mem_cgroup_disabled())
1798 return mem_cgroup_inactive_anon_is_low(lruvec);
1800 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1803 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1810 * inactive_file_is_low - check if file pages need to be deactivated
1811 * @lruvec: LRU vector to check
1813 * When the system is doing streaming IO, memory pressure here
1814 * ensures that active file pages get deactivated, until more
1815 * than half of the file pages are on the inactive list.
1817 * Once we get to that situation, protect the system's working
1818 * set from being evicted by disabling active file page aging.
1820 * This uses a different ratio than the anonymous pages, because
1821 * the page cache uses a use-once replacement algorithm.
1823 static int inactive_file_is_low(struct lruvec *lruvec)
1825 unsigned long inactive;
1826 unsigned long active;
1828 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1829 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1831 return active > inactive;
1834 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1836 if (is_file_lru(lru))
1837 return inactive_file_is_low(lruvec);
1839 return inactive_anon_is_low(lruvec);
1842 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1843 struct lruvec *lruvec, struct scan_control *sc)
1845 if (is_active_lru(lru)) {
1846 if (inactive_list_is_low(lruvec, lru))
1847 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1851 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1862 * Determine how aggressively the anon and file LRU lists should be
1863 * scanned. The relative value of each set of LRU lists is determined
1864 * by looking at the fraction of the pages scanned we did rotate back
1865 * onto the active list instead of evict.
1867 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1868 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1870 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1873 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1875 u64 denominator = 0; /* gcc */
1876 struct zone *zone = lruvec_zone(lruvec);
1877 unsigned long anon_prio, file_prio;
1878 enum scan_balance scan_balance;
1879 unsigned long anon, file;
1880 bool force_scan = false;
1881 unsigned long ap, fp;
1887 * If the zone or memcg is small, nr[l] can be 0. This
1888 * results in no scanning on this priority and a potential
1889 * priority drop. Global direct reclaim can go to the next
1890 * zone and tends to have no problems. Global kswapd is for
1891 * zone balancing and it needs to scan a minimum amount. When
1892 * reclaiming for a memcg, a priority drop can cause high
1893 * latencies, so it's better to scan a minimum amount there as
1896 if (current_is_kswapd() && !zone_reclaimable(zone))
1898 if (!global_reclaim(sc))
1901 /* If we have no swap space, do not bother scanning anon pages. */
1902 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1903 scan_balance = SCAN_FILE;
1908 * Global reclaim will swap to prevent OOM even with no
1909 * swappiness, but memcg users want to use this knob to
1910 * disable swapping for individual groups completely when
1911 * using the memory controller's swap limit feature would be
1914 if (!global_reclaim(sc) && !sc->swappiness) {
1915 scan_balance = SCAN_FILE;
1920 * Do not apply any pressure balancing cleverness when the
1921 * system is close to OOM, scan both anon and file equally
1922 * (unless the swappiness setting disagrees with swapping).
1924 if (!sc->priority && sc->swappiness) {
1925 scan_balance = SCAN_EQUAL;
1929 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1930 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1931 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1932 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1935 * Prevent the reclaimer from falling into the cache trap: as
1936 * cache pages start out inactive, every cache fault will tip
1937 * the scan balance towards the file LRU. And as the file LRU
1938 * shrinks, so does the window for rotation from references.
1939 * This means we have a runaway feedback loop where a tiny
1940 * thrashing file LRU becomes infinitely more attractive than
1941 * anon pages. Try to detect this based on file LRU size.
1943 if (global_reclaim(sc)) {
1944 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1946 if (unlikely(file + free <= high_wmark_pages(zone))) {
1947 scan_balance = SCAN_ANON;
1953 * There is enough inactive page cache, do not reclaim
1954 * anything from the anonymous working set right now.
1956 if (!inactive_file_is_low(lruvec)) {
1957 scan_balance = SCAN_FILE;
1961 scan_balance = SCAN_FRACT;
1964 * With swappiness at 100, anonymous and file have the same priority.
1965 * This scanning priority is essentially the inverse of IO cost.
1967 anon_prio = sc->swappiness;
1968 file_prio = 200 - anon_prio;
1971 * OK, so we have swap space and a fair amount of page cache
1972 * pages. We use the recently rotated / recently scanned
1973 * ratios to determine how valuable each cache is.
1975 * Because workloads change over time (and to avoid overflow)
1976 * we keep these statistics as a floating average, which ends
1977 * up weighing recent references more than old ones.
1979 * anon in [0], file in [1]
1981 spin_lock_irq(&zone->lru_lock);
1982 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1983 reclaim_stat->recent_scanned[0] /= 2;
1984 reclaim_stat->recent_rotated[0] /= 2;
1987 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1988 reclaim_stat->recent_scanned[1] /= 2;
1989 reclaim_stat->recent_rotated[1] /= 2;
1993 * The amount of pressure on anon vs file pages is inversely
1994 * proportional to the fraction of recently scanned pages on
1995 * each list that were recently referenced and in active use.
1997 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1998 ap /= reclaim_stat->recent_rotated[0] + 1;
2000 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2001 fp /= reclaim_stat->recent_rotated[1] + 1;
2002 spin_unlock_irq(&zone->lru_lock);
2006 denominator = ap + fp + 1;
2008 some_scanned = false;
2009 /* Only use force_scan on second pass. */
2010 for (pass = 0; !some_scanned && pass < 2; pass++) {
2011 for_each_evictable_lru(lru) {
2012 int file = is_file_lru(lru);
2016 size = get_lru_size(lruvec, lru);
2017 scan = size >> sc->priority;
2019 if (!scan && pass && force_scan)
2020 scan = min(size, SWAP_CLUSTER_MAX);
2022 switch (scan_balance) {
2024 /* Scan lists relative to size */
2028 * Scan types proportional to swappiness and
2029 * their relative recent reclaim efficiency.
2031 scan = div64_u64(scan * fraction[file],
2036 /* Scan one type exclusively */
2037 if ((scan_balance == SCAN_FILE) != file)
2041 /* Look ma, no brain */
2046 * Skip the second pass and don't force_scan,
2047 * if we found something to scan.
2049 some_scanned |= !!scan;
2055 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2057 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2059 unsigned long nr[NR_LRU_LISTS];
2060 unsigned long targets[NR_LRU_LISTS];
2061 unsigned long nr_to_scan;
2063 unsigned long nr_reclaimed = 0;
2064 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2065 struct blk_plug plug;
2068 get_scan_count(lruvec, sc, nr);
2070 /* Record the original scan target for proportional adjustments later */
2071 memcpy(targets, nr, sizeof(nr));
2074 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2075 * event that can occur when there is little memory pressure e.g.
2076 * multiple streaming readers/writers. Hence, we do not abort scanning
2077 * when the requested number of pages are reclaimed when scanning at
2078 * DEF_PRIORITY on the assumption that the fact we are direct
2079 * reclaiming implies that kswapd is not keeping up and it is best to
2080 * do a batch of work at once. For memcg reclaim one check is made to
2081 * abort proportional reclaim if either the file or anon lru has already
2082 * dropped to zero at the first pass.
2084 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2085 sc->priority == DEF_PRIORITY);
2087 blk_start_plug(&plug);
2088 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2089 nr[LRU_INACTIVE_FILE]) {
2090 unsigned long nr_anon, nr_file, percentage;
2091 unsigned long nr_scanned;
2093 for_each_evictable_lru(lru) {
2095 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2096 nr[lru] -= nr_to_scan;
2098 nr_reclaimed += shrink_list(lru, nr_to_scan,
2103 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2107 * For kswapd and memcg, reclaim at least the number of pages
2108 * requested. Ensure that the anon and file LRUs are scanned
2109 * proportionally what was requested by get_scan_count(). We
2110 * stop reclaiming one LRU and reduce the amount scanning
2111 * proportional to the original scan target.
2113 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2114 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2117 * It's just vindictive to attack the larger once the smaller
2118 * has gone to zero. And given the way we stop scanning the
2119 * smaller below, this makes sure that we only make one nudge
2120 * towards proportionality once we've got nr_to_reclaim.
2122 if (!nr_file || !nr_anon)
2125 if (nr_file > nr_anon) {
2126 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2127 targets[LRU_ACTIVE_ANON] + 1;
2129 percentage = nr_anon * 100 / scan_target;
2131 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2132 targets[LRU_ACTIVE_FILE] + 1;
2134 percentage = nr_file * 100 / scan_target;
2137 /* Stop scanning the smaller of the LRU */
2139 nr[lru + LRU_ACTIVE] = 0;
2142 * Recalculate the other LRU scan count based on its original
2143 * scan target and the percentage scanning already complete
2145 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2146 nr_scanned = targets[lru] - nr[lru];
2147 nr[lru] = targets[lru] * (100 - percentage) / 100;
2148 nr[lru] -= min(nr[lru], nr_scanned);
2151 nr_scanned = targets[lru] - nr[lru];
2152 nr[lru] = targets[lru] * (100 - percentage) / 100;
2153 nr[lru] -= min(nr[lru], nr_scanned);
2155 scan_adjusted = true;
2157 blk_finish_plug(&plug);
2158 sc->nr_reclaimed += nr_reclaimed;
2161 * Even if we did not try to evict anon pages at all, we want to
2162 * rebalance the anon lru active/inactive ratio.
2164 if (inactive_anon_is_low(lruvec))
2165 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2166 sc, LRU_ACTIVE_ANON);
2168 throttle_vm_writeout(sc->gfp_mask);
2171 /* Use reclaim/compaction for costly allocs or under memory pressure */
2172 static bool in_reclaim_compaction(struct scan_control *sc)
2174 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2175 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2176 sc->priority < DEF_PRIORITY - 2))
2183 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2184 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2185 * true if more pages should be reclaimed such that when the page allocator
2186 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2187 * It will give up earlier than that if there is difficulty reclaiming pages.
2189 static inline bool should_continue_reclaim(struct zone *zone,
2190 unsigned long nr_reclaimed,
2191 unsigned long nr_scanned,
2192 struct scan_control *sc)
2194 unsigned long pages_for_compaction;
2195 unsigned long inactive_lru_pages;
2197 /* If not in reclaim/compaction mode, stop */
2198 if (!in_reclaim_compaction(sc))
2201 /* Consider stopping depending on scan and reclaim activity */
2202 if (sc->gfp_mask & __GFP_REPEAT) {
2204 * For __GFP_REPEAT allocations, stop reclaiming if the
2205 * full LRU list has been scanned and we are still failing
2206 * to reclaim pages. This full LRU scan is potentially
2207 * expensive but a __GFP_REPEAT caller really wants to succeed
2209 if (!nr_reclaimed && !nr_scanned)
2213 * For non-__GFP_REPEAT allocations which can presumably
2214 * fail without consequence, stop if we failed to reclaim
2215 * any pages from the last SWAP_CLUSTER_MAX number of
2216 * pages that were scanned. This will return to the
2217 * caller faster at the risk reclaim/compaction and
2218 * the resulting allocation attempt fails
2225 * If we have not reclaimed enough pages for compaction and the
2226 * inactive lists are large enough, continue reclaiming
2228 pages_for_compaction = (2UL << sc->order);
2229 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2230 if (get_nr_swap_pages() > 0)
2231 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2232 if (sc->nr_reclaimed < pages_for_compaction &&
2233 inactive_lru_pages > pages_for_compaction)
2236 /* If compaction would go ahead or the allocation would succeed, stop */
2237 switch (compaction_suitable(zone, sc->order)) {
2238 case COMPACT_PARTIAL:
2239 case COMPACT_CONTINUE:
2246 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2248 unsigned long nr_reclaimed, nr_scanned;
2251 struct mem_cgroup *root = sc->target_mem_cgroup;
2252 struct mem_cgroup_reclaim_cookie reclaim = {
2254 .priority = sc->priority,
2256 struct mem_cgroup *memcg;
2258 nr_reclaimed = sc->nr_reclaimed;
2259 nr_scanned = sc->nr_scanned;
2261 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2263 struct lruvec *lruvec;
2265 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2267 sc->swappiness = mem_cgroup_swappiness(memcg);
2268 shrink_lruvec(lruvec, sc);
2271 * Direct reclaim and kswapd have to scan all memory
2272 * cgroups to fulfill the overall scan target for the
2275 * Limit reclaim, on the other hand, only cares about
2276 * nr_to_reclaim pages to be reclaimed and it will
2277 * retry with decreasing priority if one round over the
2278 * whole hierarchy is not sufficient.
2280 if (!global_reclaim(sc) &&
2281 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2282 mem_cgroup_iter_break(root, memcg);
2285 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2288 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2289 sc->nr_scanned - nr_scanned,
2290 sc->nr_reclaimed - nr_reclaimed);
2292 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2293 sc->nr_scanned - nr_scanned, sc));
2296 /* Returns true if compaction should go ahead for a high-order request */
2297 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2299 unsigned long balance_gap, watermark;
2302 /* Do not consider compaction for orders reclaim is meant to satisfy */
2303 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2307 * Compaction takes time to run and there are potentially other
2308 * callers using the pages just freed. Continue reclaiming until
2309 * there is a buffer of free pages available to give compaction
2310 * a reasonable chance of completing and allocating the page
2312 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2313 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2314 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2315 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2318 * If compaction is deferred, reclaim up to a point where
2319 * compaction will have a chance of success when re-enabled
2321 if (compaction_deferred(zone, sc->order))
2322 return watermark_ok;
2324 /* If compaction is not ready to start, keep reclaiming */
2325 if (!compaction_suitable(zone, sc->order))
2328 return watermark_ok;
2332 * This is the direct reclaim path, for page-allocating processes. We only
2333 * try to reclaim pages from zones which will satisfy the caller's allocation
2336 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2338 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2340 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2341 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2342 * zone defense algorithm.
2344 * If a zone is deemed to be full of pinned pages then just give it a light
2345 * scan then give up on it.
2347 * This function returns true if a zone is being reclaimed for a costly
2348 * high-order allocation and compaction is ready to begin. This indicates to
2349 * the caller that it should consider retrying the allocation instead of
2352 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2356 unsigned long nr_soft_reclaimed;
2357 unsigned long nr_soft_scanned;
2358 unsigned long lru_pages = 0;
2359 bool aborted_reclaim = false;
2360 struct reclaim_state *reclaim_state = current->reclaim_state;
2362 struct shrink_control shrink = {
2363 .gfp_mask = sc->gfp_mask,
2365 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2368 * If the number of buffer_heads in the machine exceeds the maximum
2369 * allowed level, force direct reclaim to scan the highmem zone as
2370 * highmem pages could be pinning lowmem pages storing buffer_heads
2372 orig_mask = sc->gfp_mask;
2373 if (buffer_heads_over_limit)
2374 sc->gfp_mask |= __GFP_HIGHMEM;
2376 nodes_clear(shrink.nodes_to_scan);
2378 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2379 gfp_zone(sc->gfp_mask), sc->nodemask) {
2380 if (!populated_zone(zone))
2383 * Take care memory controller reclaiming has small influence
2386 if (global_reclaim(sc)) {
2387 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2390 lru_pages += zone_reclaimable_pages(zone);
2391 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2393 if (sc->priority != DEF_PRIORITY &&
2394 !zone_reclaimable(zone))
2395 continue; /* Let kswapd poll it */
2396 if (IS_ENABLED(CONFIG_COMPACTION)) {
2398 * If we already have plenty of memory free for
2399 * compaction in this zone, don't free any more.
2400 * Even though compaction is invoked for any
2401 * non-zero order, only frequent costly order
2402 * reclamation is disruptive enough to become a
2403 * noticeable problem, like transparent huge
2406 if ((zonelist_zone_idx(z) <= requested_highidx)
2407 && compaction_ready(zone, sc)) {
2408 aborted_reclaim = true;
2413 * This steals pages from memory cgroups over softlimit
2414 * and returns the number of reclaimed pages and
2415 * scanned pages. This works for global memory pressure
2416 * and balancing, not for a memcg's limit.
2418 nr_soft_scanned = 0;
2419 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2420 sc->order, sc->gfp_mask,
2422 sc->nr_reclaimed += nr_soft_reclaimed;
2423 sc->nr_scanned += nr_soft_scanned;
2424 /* need some check for avoid more shrink_zone() */
2427 shrink_zone(zone, sc);
2431 * Don't shrink slabs when reclaiming memory from over limit cgroups
2432 * but do shrink slab at least once when aborting reclaim for
2433 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2436 if (global_reclaim(sc)) {
2437 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2438 if (reclaim_state) {
2439 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2440 reclaim_state->reclaimed_slab = 0;
2445 * Restore to original mask to avoid the impact on the caller if we
2446 * promoted it to __GFP_HIGHMEM.
2448 sc->gfp_mask = orig_mask;
2450 return aborted_reclaim;
2453 /* All zones in zonelist are unreclaimable? */
2454 static bool all_unreclaimable(struct zonelist *zonelist,
2455 struct scan_control *sc)
2460 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2461 gfp_zone(sc->gfp_mask), sc->nodemask) {
2462 if (!populated_zone(zone))
2464 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2466 if (zone_reclaimable(zone))
2474 * This is the main entry point to direct page reclaim.
2476 * If a full scan of the inactive list fails to free enough memory then we
2477 * are "out of memory" and something needs to be killed.
2479 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2480 * high - the zone may be full of dirty or under-writeback pages, which this
2481 * caller can't do much about. We kick the writeback threads and take explicit
2482 * naps in the hope that some of these pages can be written. But if the
2483 * allocating task holds filesystem locks which prevent writeout this might not
2484 * work, and the allocation attempt will fail.
2486 * returns: 0, if no pages reclaimed
2487 * else, the number of pages reclaimed
2489 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2490 struct scan_control *sc)
2492 unsigned long total_scanned = 0;
2493 unsigned long writeback_threshold;
2494 bool aborted_reclaim;
2496 delayacct_freepages_start();
2498 if (global_reclaim(sc))
2499 count_vm_event(ALLOCSTALL);
2502 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2505 aborted_reclaim = shrink_zones(zonelist, sc);
2507 total_scanned += sc->nr_scanned;
2508 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2512 * If we're getting trouble reclaiming, start doing
2513 * writepage even in laptop mode.
2515 if (sc->priority < DEF_PRIORITY - 2)
2516 sc->may_writepage = 1;
2519 * Try to write back as many pages as we just scanned. This
2520 * tends to cause slow streaming writers to write data to the
2521 * disk smoothly, at the dirtying rate, which is nice. But
2522 * that's undesirable in laptop mode, where we *want* lumpy
2523 * writeout. So in laptop mode, write out the whole world.
2525 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2526 if (total_scanned > writeback_threshold) {
2527 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2528 WB_REASON_TRY_TO_FREE_PAGES);
2529 sc->may_writepage = 1;
2531 } while (--sc->priority >= 0 && !aborted_reclaim);
2534 delayacct_freepages_end();
2536 if (sc->nr_reclaimed)
2537 return sc->nr_reclaimed;
2540 * As hibernation is going on, kswapd is freezed so that it can't mark
2541 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2544 if (oom_killer_disabled)
2547 /* Aborted reclaim to try compaction? don't OOM, then */
2548 if (aborted_reclaim)
2551 /* top priority shrink_zones still had more to do? don't OOM, then */
2552 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2558 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2561 unsigned long pfmemalloc_reserve = 0;
2562 unsigned long free_pages = 0;
2566 for (i = 0; i <= ZONE_NORMAL; i++) {
2567 zone = &pgdat->node_zones[i];
2568 if (!populated_zone(zone))
2571 pfmemalloc_reserve += min_wmark_pages(zone);
2572 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2575 /* If there are no reserves (unexpected config) then do not throttle */
2576 if (!pfmemalloc_reserve)
2579 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2581 /* kswapd must be awake if processes are being throttled */
2582 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2583 pgdat->classzone_idx = min(pgdat->classzone_idx,
2584 (enum zone_type)ZONE_NORMAL);
2585 wake_up_interruptible(&pgdat->kswapd_wait);
2592 * Throttle direct reclaimers if backing storage is backed by the network
2593 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2594 * depleted. kswapd will continue to make progress and wake the processes
2595 * when the low watermark is reached.
2597 * Returns true if a fatal signal was delivered during throttling. If this
2598 * happens, the page allocator should not consider triggering the OOM killer.
2600 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2601 nodemask_t *nodemask)
2605 pg_data_t *pgdat = NULL;
2608 * Kernel threads should not be throttled as they may be indirectly
2609 * responsible for cleaning pages necessary for reclaim to make forward
2610 * progress. kjournald for example may enter direct reclaim while
2611 * committing a transaction where throttling it could forcing other
2612 * processes to block on log_wait_commit().
2614 if (current->flags & PF_KTHREAD)
2618 * If a fatal signal is pending, this process should not throttle.
2619 * It should return quickly so it can exit and free its memory
2621 if (fatal_signal_pending(current))
2625 * Check if the pfmemalloc reserves are ok by finding the first node
2626 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2627 * GFP_KERNEL will be required for allocating network buffers when
2628 * swapping over the network so ZONE_HIGHMEM is unusable.
2630 * Throttling is based on the first usable node and throttled processes
2631 * wait on a queue until kswapd makes progress and wakes them. There
2632 * is an affinity then between processes waking up and where reclaim
2633 * progress has been made assuming the process wakes on the same node.
2634 * More importantly, processes running on remote nodes will not compete
2635 * for remote pfmemalloc reserves and processes on different nodes
2636 * should make reasonable progress.
2638 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2639 gfp_mask, nodemask) {
2640 if (zone_idx(zone) > ZONE_NORMAL)
2643 /* Throttle based on the first usable node */
2644 pgdat = zone->zone_pgdat;
2645 if (pfmemalloc_watermark_ok(pgdat))
2650 /* If no zone was usable by the allocation flags then do not throttle */
2654 /* Account for the throttling */
2655 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2658 * If the caller cannot enter the filesystem, it's possible that it
2659 * is due to the caller holding an FS lock or performing a journal
2660 * transaction in the case of a filesystem like ext[3|4]. In this case,
2661 * it is not safe to block on pfmemalloc_wait as kswapd could be
2662 * blocked waiting on the same lock. Instead, throttle for up to a
2663 * second before continuing.
2665 if (!(gfp_mask & __GFP_FS)) {
2666 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2667 pfmemalloc_watermark_ok(pgdat), HZ);
2672 /* Throttle until kswapd wakes the process */
2673 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2674 pfmemalloc_watermark_ok(pgdat));
2677 if (fatal_signal_pending(current))
2684 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2685 gfp_t gfp_mask, nodemask_t *nodemask)
2687 unsigned long nr_reclaimed;
2688 struct scan_control sc = {
2689 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2690 .may_writepage = !laptop_mode,
2691 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2695 .priority = DEF_PRIORITY,
2696 .target_mem_cgroup = NULL,
2697 .nodemask = nodemask,
2701 * Do not enter reclaim if fatal signal was delivered while throttled.
2702 * 1 is returned so that the page allocator does not OOM kill at this
2705 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2708 trace_mm_vmscan_direct_reclaim_begin(order,
2712 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2714 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2716 return nr_reclaimed;
2721 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2722 gfp_t gfp_mask, bool noswap,
2724 unsigned long *nr_scanned)
2726 struct scan_control sc = {
2728 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2729 .may_writepage = !laptop_mode,
2731 .may_swap = !noswap,
2734 .swappiness = mem_cgroup_swappiness(memcg),
2735 .target_mem_cgroup = memcg,
2737 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2739 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2740 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2742 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2747 * NOTE: Although we can get the priority field, using it
2748 * here is not a good idea, since it limits the pages we can scan.
2749 * if we don't reclaim here, the shrink_zone from balance_pgdat
2750 * will pick up pages from other mem cgroup's as well. We hack
2751 * the priority and make it zero.
2753 shrink_lruvec(lruvec, &sc);
2755 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2757 *nr_scanned = sc.nr_scanned;
2758 return sc.nr_reclaimed;
2761 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2765 struct zonelist *zonelist;
2766 unsigned long nr_reclaimed;
2768 struct scan_control sc = {
2769 .may_writepage = !laptop_mode,
2771 .may_swap = !noswap,
2772 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2774 .priority = DEF_PRIORITY,
2775 .target_mem_cgroup = memcg,
2776 .nodemask = NULL, /* we don't care the placement */
2777 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2778 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2782 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2783 * take care of from where we get pages. So the node where we start the
2784 * scan does not need to be the current node.
2786 nid = mem_cgroup_select_victim_node(memcg);
2788 zonelist = NODE_DATA(nid)->node_zonelists;
2790 trace_mm_vmscan_memcg_reclaim_begin(0,
2794 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2796 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2798 return nr_reclaimed;
2802 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2804 struct mem_cgroup *memcg;
2806 if (!total_swap_pages)
2809 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2811 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2813 if (inactive_anon_is_low(lruvec))
2814 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2815 sc, LRU_ACTIVE_ANON);
2817 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2821 static bool zone_balanced(struct zone *zone, int order,
2822 unsigned long balance_gap, int classzone_idx)
2824 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2825 balance_gap, classzone_idx, 0))
2828 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2829 !compaction_suitable(zone, order))
2836 * pgdat_balanced() is used when checking if a node is balanced.
2838 * For order-0, all zones must be balanced!
2840 * For high-order allocations only zones that meet watermarks and are in a
2841 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2842 * total of balanced pages must be at least 25% of the zones allowed by
2843 * classzone_idx for the node to be considered balanced. Forcing all zones to
2844 * be balanced for high orders can cause excessive reclaim when there are
2846 * The choice of 25% is due to
2847 * o a 16M DMA zone that is balanced will not balance a zone on any
2848 * reasonable sized machine
2849 * o On all other machines, the top zone must be at least a reasonable
2850 * percentage of the middle zones. For example, on 32-bit x86, highmem
2851 * would need to be at least 256M for it to be balance a whole node.
2852 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2853 * to balance a node on its own. These seemed like reasonable ratios.
2855 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2857 unsigned long managed_pages = 0;
2858 unsigned long balanced_pages = 0;
2861 /* Check the watermark levels */
2862 for (i = 0; i <= classzone_idx; i++) {
2863 struct zone *zone = pgdat->node_zones + i;
2865 if (!populated_zone(zone))
2868 managed_pages += zone->managed_pages;
2871 * A special case here:
2873 * balance_pgdat() skips over all_unreclaimable after
2874 * DEF_PRIORITY. Effectively, it considers them balanced so
2875 * they must be considered balanced here as well!
2877 if (!zone_reclaimable(zone)) {
2878 balanced_pages += zone->managed_pages;
2882 if (zone_balanced(zone, order, 0, i))
2883 balanced_pages += zone->managed_pages;
2889 return balanced_pages >= (managed_pages >> 2);
2895 * Prepare kswapd for sleeping. This verifies that there are no processes
2896 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2898 * Returns true if kswapd is ready to sleep
2900 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2903 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2908 * There is a potential race between when kswapd checks its watermarks
2909 * and a process gets throttled. There is also a potential race if
2910 * processes get throttled, kswapd wakes, a large process exits therby
2911 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2912 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2913 * so wake them now if necessary. If necessary, processes will wake
2914 * kswapd and get throttled again
2916 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2917 wake_up(&pgdat->pfmemalloc_wait);
2921 return pgdat_balanced(pgdat, order, classzone_idx);
2925 * kswapd shrinks the zone by the number of pages required to reach
2926 * the high watermark.
2928 * Returns true if kswapd scanned at least the requested number of pages to
2929 * reclaim or if the lack of progress was due to pages under writeback.
2930 * This is used to determine if the scanning priority needs to be raised.
2932 static bool kswapd_shrink_zone(struct zone *zone,
2934 struct scan_control *sc,
2935 unsigned long lru_pages,
2936 unsigned long *nr_attempted)
2938 int testorder = sc->order;
2939 unsigned long balance_gap;
2940 struct reclaim_state *reclaim_state = current->reclaim_state;
2941 struct shrink_control shrink = {
2942 .gfp_mask = sc->gfp_mask,
2944 bool lowmem_pressure;
2946 /* Reclaim above the high watermark. */
2947 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2950 * Kswapd reclaims only single pages with compaction enabled. Trying
2951 * too hard to reclaim until contiguous free pages have become
2952 * available can hurt performance by evicting too much useful data
2953 * from memory. Do not reclaim more than needed for compaction.
2955 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2956 compaction_suitable(zone, sc->order) !=
2961 * We put equal pressure on every zone, unless one zone has way too
2962 * many pages free already. The "too many pages" is defined as the
2963 * high wmark plus a "gap" where the gap is either the low
2964 * watermark or 1% of the zone, whichever is smaller.
2966 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2967 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2970 * If there is no low memory pressure or the zone is balanced then no
2971 * reclaim is necessary
2973 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2974 if (!lowmem_pressure && zone_balanced(zone, testorder,
2975 balance_gap, classzone_idx))
2978 shrink_zone(zone, sc);
2979 nodes_clear(shrink.nodes_to_scan);
2980 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2982 reclaim_state->reclaimed_slab = 0;
2983 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2984 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2986 /* Account for the number of pages attempted to reclaim */
2987 *nr_attempted += sc->nr_to_reclaim;
2989 zone_clear_flag(zone, ZONE_WRITEBACK);
2992 * If a zone reaches its high watermark, consider it to be no longer
2993 * congested. It's possible there are dirty pages backed by congested
2994 * BDIs but as pressure is relieved, speculatively avoid congestion
2997 if (zone_reclaimable(zone) &&
2998 zone_balanced(zone, testorder, 0, classzone_idx)) {
2999 zone_clear_flag(zone, ZONE_CONGESTED);
3000 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3003 return sc->nr_scanned >= sc->nr_to_reclaim;
3007 * For kswapd, balance_pgdat() will work across all this node's zones until
3008 * they are all at high_wmark_pages(zone).
3010 * Returns the final order kswapd was reclaiming at
3012 * There is special handling here for zones which are full of pinned pages.
3013 * This can happen if the pages are all mlocked, or if they are all used by
3014 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3015 * What we do is to detect the case where all pages in the zone have been
3016 * scanned twice and there has been zero successful reclaim. Mark the zone as
3017 * dead and from now on, only perform a short scan. Basically we're polling
3018 * the zone for when the problem goes away.
3020 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3021 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3022 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3023 * lower zones regardless of the number of free pages in the lower zones. This
3024 * interoperates with the page allocator fallback scheme to ensure that aging
3025 * of pages is balanced across the zones.
3027 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3031 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3032 unsigned long nr_soft_reclaimed;
3033 unsigned long nr_soft_scanned;
3034 struct scan_control sc = {
3035 .gfp_mask = GFP_KERNEL,
3036 .priority = DEF_PRIORITY,
3039 .may_writepage = !laptop_mode,
3041 .target_mem_cgroup = NULL,
3043 count_vm_event(PAGEOUTRUN);
3046 unsigned long lru_pages = 0;
3047 unsigned long nr_attempted = 0;
3048 bool raise_priority = true;
3049 bool pgdat_needs_compaction = (order > 0);
3051 sc.nr_reclaimed = 0;
3054 * Scan in the highmem->dma direction for the highest
3055 * zone which needs scanning
3057 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3058 struct zone *zone = pgdat->node_zones + i;
3060 if (!populated_zone(zone))
3063 if (sc.priority != DEF_PRIORITY &&
3064 !zone_reclaimable(zone))
3068 * Do some background aging of the anon list, to give
3069 * pages a chance to be referenced before reclaiming.
3071 age_active_anon(zone, &sc);
3074 * If the number of buffer_heads in the machine
3075 * exceeds the maximum allowed level and this node
3076 * has a highmem zone, force kswapd to reclaim from
3077 * it to relieve lowmem pressure.
3079 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3084 if (!zone_balanced(zone, order, 0, 0)) {
3089 * If balanced, clear the dirty and congested
3092 zone_clear_flag(zone, ZONE_CONGESTED);
3093 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3100 for (i = 0; i <= end_zone; i++) {
3101 struct zone *zone = pgdat->node_zones + i;
3103 if (!populated_zone(zone))
3106 lru_pages += zone_reclaimable_pages(zone);
3109 * If any zone is currently balanced then kswapd will
3110 * not call compaction as it is expected that the
3111 * necessary pages are already available.
3113 if (pgdat_needs_compaction &&
3114 zone_watermark_ok(zone, order,
3115 low_wmark_pages(zone),
3117 pgdat_needs_compaction = false;
3121 * If we're getting trouble reclaiming, start doing writepage
3122 * even in laptop mode.
3124 if (sc.priority < DEF_PRIORITY - 2)
3125 sc.may_writepage = 1;
3128 * Now scan the zone in the dma->highmem direction, stopping
3129 * at the last zone which needs scanning.
3131 * We do this because the page allocator works in the opposite
3132 * direction. This prevents the page allocator from allocating
3133 * pages behind kswapd's direction of progress, which would
3134 * cause too much scanning of the lower zones.
3136 for (i = 0; i <= end_zone; i++) {
3137 struct zone *zone = pgdat->node_zones + i;
3139 if (!populated_zone(zone))
3142 if (sc.priority != DEF_PRIORITY &&
3143 !zone_reclaimable(zone))
3148 nr_soft_scanned = 0;
3150 * Call soft limit reclaim before calling shrink_zone.
3152 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3155 sc.nr_reclaimed += nr_soft_reclaimed;
3158 * There should be no need to raise the scanning
3159 * priority if enough pages are already being scanned
3160 * that that high watermark would be met at 100%
3163 if (kswapd_shrink_zone(zone, end_zone, &sc,
3164 lru_pages, &nr_attempted))
3165 raise_priority = false;
3169 * If the low watermark is met there is no need for processes
3170 * to be throttled on pfmemalloc_wait as they should not be
3171 * able to safely make forward progress. Wake them
3173 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3174 pfmemalloc_watermark_ok(pgdat))
3175 wake_up(&pgdat->pfmemalloc_wait);
3178 * Fragmentation may mean that the system cannot be rebalanced
3179 * for high-order allocations in all zones. If twice the
3180 * allocation size has been reclaimed and the zones are still
3181 * not balanced then recheck the watermarks at order-0 to
3182 * prevent kswapd reclaiming excessively. Assume that a
3183 * process requested a high-order can direct reclaim/compact.
3185 if (order && sc.nr_reclaimed >= 2UL << order)
3186 order = sc.order = 0;
3188 /* Check if kswapd should be suspending */
3189 if (try_to_freeze() || kthread_should_stop())
3193 * Compact if necessary and kswapd is reclaiming at least the
3194 * high watermark number of pages as requsted
3196 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3197 compact_pgdat(pgdat, order);
3200 * Raise priority if scanning rate is too low or there was no
3201 * progress in reclaiming pages
3203 if (raise_priority || !sc.nr_reclaimed)
3205 } while (sc.priority >= 1 &&
3206 !pgdat_balanced(pgdat, order, *classzone_idx));
3210 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3211 * makes a decision on the order we were last reclaiming at. However,
3212 * if another caller entered the allocator slow path while kswapd
3213 * was awake, order will remain at the higher level
3215 *classzone_idx = end_zone;
3219 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3224 if (freezing(current) || kthread_should_stop())
3227 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3229 /* Try to sleep for a short interval */
3230 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3231 remaining = schedule_timeout(HZ/10);
3232 finish_wait(&pgdat->kswapd_wait, &wait);
3233 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3237 * After a short sleep, check if it was a premature sleep. If not, then
3238 * go fully to sleep until explicitly woken up.
3240 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3241 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3244 * vmstat counters are not perfectly accurate and the estimated
3245 * value for counters such as NR_FREE_PAGES can deviate from the
3246 * true value by nr_online_cpus * threshold. To avoid the zone
3247 * watermarks being breached while under pressure, we reduce the
3248 * per-cpu vmstat threshold while kswapd is awake and restore
3249 * them before going back to sleep.
3251 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3254 * Compaction records what page blocks it recently failed to
3255 * isolate pages from and skips them in the future scanning.
3256 * When kswapd is going to sleep, it is reasonable to assume
3257 * that pages and compaction may succeed so reset the cache.
3259 reset_isolation_suitable(pgdat);
3261 if (!kthread_should_stop())
3264 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3267 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3269 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3271 finish_wait(&pgdat->kswapd_wait, &wait);
3275 * The background pageout daemon, started as a kernel thread
3276 * from the init process.
3278 * This basically trickles out pages so that we have _some_
3279 * free memory available even if there is no other activity
3280 * that frees anything up. This is needed for things like routing
3281 * etc, where we otherwise might have all activity going on in
3282 * asynchronous contexts that cannot page things out.
3284 * If there are applications that are active memory-allocators
3285 * (most normal use), this basically shouldn't matter.
3287 static int kswapd(void *p)
3289 unsigned long order, new_order;
3290 unsigned balanced_order;
3291 int classzone_idx, new_classzone_idx;
3292 int balanced_classzone_idx;
3293 pg_data_t *pgdat = (pg_data_t*)p;
3294 struct task_struct *tsk = current;
3296 struct reclaim_state reclaim_state = {
3297 .reclaimed_slab = 0,
3299 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3301 lockdep_set_current_reclaim_state(GFP_KERNEL);
3303 if (!cpumask_empty(cpumask))
3304 set_cpus_allowed_ptr(tsk, cpumask);
3305 current->reclaim_state = &reclaim_state;
3308 * Tell the memory management that we're a "memory allocator",
3309 * and that if we need more memory we should get access to it
3310 * regardless (see "__alloc_pages()"). "kswapd" should
3311 * never get caught in the normal page freeing logic.
3313 * (Kswapd normally doesn't need memory anyway, but sometimes
3314 * you need a small amount of memory in order to be able to
3315 * page out something else, and this flag essentially protects
3316 * us from recursively trying to free more memory as we're
3317 * trying to free the first piece of memory in the first place).
3319 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3322 order = new_order = 0;
3324 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3325 balanced_classzone_idx = classzone_idx;
3330 * If the last balance_pgdat was unsuccessful it's unlikely a
3331 * new request of a similar or harder type will succeed soon
3332 * so consider going to sleep on the basis we reclaimed at
3334 if (balanced_classzone_idx >= new_classzone_idx &&
3335 balanced_order == new_order) {
3336 new_order = pgdat->kswapd_max_order;
3337 new_classzone_idx = pgdat->classzone_idx;
3338 pgdat->kswapd_max_order = 0;
3339 pgdat->classzone_idx = pgdat->nr_zones - 1;
3342 if (order < new_order || classzone_idx > new_classzone_idx) {
3344 * Don't sleep if someone wants a larger 'order'
3345 * allocation or has tigher zone constraints
3348 classzone_idx = new_classzone_idx;
3350 kswapd_try_to_sleep(pgdat, balanced_order,
3351 balanced_classzone_idx);
3352 order = pgdat->kswapd_max_order;
3353 classzone_idx = pgdat->classzone_idx;
3355 new_classzone_idx = classzone_idx;
3356 pgdat->kswapd_max_order = 0;
3357 pgdat->classzone_idx = pgdat->nr_zones - 1;
3360 ret = try_to_freeze();
3361 if (kthread_should_stop())
3365 * We can speed up thawing tasks if we don't call balance_pgdat
3366 * after returning from the refrigerator
3369 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3370 balanced_classzone_idx = classzone_idx;
3371 balanced_order = balance_pgdat(pgdat, order,
3372 &balanced_classzone_idx);
3376 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3377 current->reclaim_state = NULL;
3378 lockdep_clear_current_reclaim_state();
3384 * A zone is low on free memory, so wake its kswapd task to service it.
3386 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3390 if (!populated_zone(zone))
3393 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3395 pgdat = zone->zone_pgdat;
3396 if (pgdat->kswapd_max_order < order) {
3397 pgdat->kswapd_max_order = order;
3398 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3400 if (!waitqueue_active(&pgdat->kswapd_wait))
3402 if (zone_balanced(zone, order, 0, 0))
3405 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3406 wake_up_interruptible(&pgdat->kswapd_wait);
3409 #ifdef CONFIG_HIBERNATION
3411 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3414 * Rather than trying to age LRUs the aim is to preserve the overall
3415 * LRU order by reclaiming preferentially
3416 * inactive > active > active referenced > active mapped
3418 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3420 struct reclaim_state reclaim_state;
3421 struct scan_control sc = {
3422 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3426 .nr_to_reclaim = nr_to_reclaim,
3427 .hibernation_mode = 1,
3429 .priority = DEF_PRIORITY,
3431 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3432 struct task_struct *p = current;
3433 unsigned long nr_reclaimed;
3435 p->flags |= PF_MEMALLOC;
3436 lockdep_set_current_reclaim_state(sc.gfp_mask);
3437 reclaim_state.reclaimed_slab = 0;
3438 p->reclaim_state = &reclaim_state;
3440 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3442 p->reclaim_state = NULL;
3443 lockdep_clear_current_reclaim_state();
3444 p->flags &= ~PF_MEMALLOC;
3446 return nr_reclaimed;
3448 #endif /* CONFIG_HIBERNATION */
3450 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3451 not required for correctness. So if the last cpu in a node goes
3452 away, we get changed to run anywhere: as the first one comes back,
3453 restore their cpu bindings. */
3454 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3459 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3460 for_each_node_state(nid, N_MEMORY) {
3461 pg_data_t *pgdat = NODE_DATA(nid);
3462 const struct cpumask *mask;
3464 mask = cpumask_of_node(pgdat->node_id);
3466 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3467 /* One of our CPUs online: restore mask */
3468 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3475 * This kswapd start function will be called by init and node-hot-add.
3476 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3478 int kswapd_run(int nid)
3480 pg_data_t *pgdat = NODE_DATA(nid);
3486 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3487 if (IS_ERR(pgdat->kswapd)) {
3488 /* failure at boot is fatal */
3489 BUG_ON(system_state == SYSTEM_BOOTING);
3490 pr_err("Failed to start kswapd on node %d\n", nid);
3491 ret = PTR_ERR(pgdat->kswapd);
3492 pgdat->kswapd = NULL;
3498 * Called by memory hotplug when all memory in a node is offlined. Caller must
3499 * hold mem_hotplug_begin/end().
3501 void kswapd_stop(int nid)
3503 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3506 kthread_stop(kswapd);
3507 NODE_DATA(nid)->kswapd = NULL;
3511 static int __init kswapd_init(void)
3516 for_each_node_state(nid, N_MEMORY)
3518 hotcpu_notifier(cpu_callback, 0);
3522 module_init(kswapd_init)
3528 * If non-zero call zone_reclaim when the number of free pages falls below
3531 int zone_reclaim_mode __read_mostly;
3533 #define RECLAIM_OFF 0
3534 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3535 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3536 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3539 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3540 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3543 #define ZONE_RECLAIM_PRIORITY 4
3546 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3549 int sysctl_min_unmapped_ratio = 1;
3552 * If the number of slab pages in a zone grows beyond this percentage then
3553 * slab reclaim needs to occur.
3555 int sysctl_min_slab_ratio = 5;
3557 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3559 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3560 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3561 zone_page_state(zone, NR_ACTIVE_FILE);
3564 * It's possible for there to be more file mapped pages than
3565 * accounted for by the pages on the file LRU lists because
3566 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3568 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3571 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3572 static long zone_pagecache_reclaimable(struct zone *zone)
3574 long nr_pagecache_reclaimable;
3578 * If RECLAIM_SWAP is set, then all file pages are considered
3579 * potentially reclaimable. Otherwise, we have to worry about
3580 * pages like swapcache and zone_unmapped_file_pages() provides
3583 if (zone_reclaim_mode & RECLAIM_SWAP)
3584 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3586 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3588 /* If we can't clean pages, remove dirty pages from consideration */
3589 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3590 delta += zone_page_state(zone, NR_FILE_DIRTY);
3592 /* Watch for any possible underflows due to delta */
3593 if (unlikely(delta > nr_pagecache_reclaimable))
3594 delta = nr_pagecache_reclaimable;
3596 return nr_pagecache_reclaimable - delta;
3600 * Try to free up some pages from this zone through reclaim.
3602 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3604 /* Minimum pages needed in order to stay on node */
3605 const unsigned long nr_pages = 1 << order;
3606 struct task_struct *p = current;
3607 struct reclaim_state reclaim_state;
3608 struct scan_control sc = {
3609 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3610 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3612 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3613 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3615 .priority = ZONE_RECLAIM_PRIORITY,
3617 struct shrink_control shrink = {
3618 .gfp_mask = sc.gfp_mask,
3620 unsigned long nr_slab_pages0, nr_slab_pages1;
3624 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3625 * and we also need to be able to write out pages for RECLAIM_WRITE
3628 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3629 lockdep_set_current_reclaim_state(gfp_mask);
3630 reclaim_state.reclaimed_slab = 0;
3631 p->reclaim_state = &reclaim_state;
3633 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3635 * Free memory by calling shrink zone with increasing
3636 * priorities until we have enough memory freed.
3639 shrink_zone(zone, &sc);
3640 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3643 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3644 if (nr_slab_pages0 > zone->min_slab_pages) {
3646 * shrink_slab() does not currently allow us to determine how
3647 * many pages were freed in this zone. So we take the current
3648 * number of slab pages and shake the slab until it is reduced
3649 * by the same nr_pages that we used for reclaiming unmapped
3652 nodes_clear(shrink.nodes_to_scan);
3653 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3655 unsigned long lru_pages = zone_reclaimable_pages(zone);
3657 /* No reclaimable slab or very low memory pressure */
3658 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3661 /* Freed enough memory */
3662 nr_slab_pages1 = zone_page_state(zone,
3663 NR_SLAB_RECLAIMABLE);
3664 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3669 * Update nr_reclaimed by the number of slab pages we
3670 * reclaimed from this zone.
3672 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3673 if (nr_slab_pages1 < nr_slab_pages0)
3674 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3677 p->reclaim_state = NULL;
3678 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3679 lockdep_clear_current_reclaim_state();
3680 return sc.nr_reclaimed >= nr_pages;
3683 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3689 * Zone reclaim reclaims unmapped file backed pages and
3690 * slab pages if we are over the defined limits.
3692 * A small portion of unmapped file backed pages is needed for
3693 * file I/O otherwise pages read by file I/O will be immediately
3694 * thrown out if the zone is overallocated. So we do not reclaim
3695 * if less than a specified percentage of the zone is used by
3696 * unmapped file backed pages.
3698 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3699 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3700 return ZONE_RECLAIM_FULL;
3702 if (!zone_reclaimable(zone))
3703 return ZONE_RECLAIM_FULL;
3706 * Do not scan if the allocation should not be delayed.
3708 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3709 return ZONE_RECLAIM_NOSCAN;
3712 * Only run zone reclaim on the local zone or on zones that do not
3713 * have associated processors. This will favor the local processor
3714 * over remote processors and spread off node memory allocations
3715 * as wide as possible.
3717 node_id = zone_to_nid(zone);
3718 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3719 return ZONE_RECLAIM_NOSCAN;
3721 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3722 return ZONE_RECLAIM_NOSCAN;
3724 ret = __zone_reclaim(zone, gfp_mask, order);
3725 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3728 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3735 * page_evictable - test whether a page is evictable
3736 * @page: the page to test
3738 * Test whether page is evictable--i.e., should be placed on active/inactive
3739 * lists vs unevictable list.
3741 * Reasons page might not be evictable:
3742 * (1) page's mapping marked unevictable
3743 * (2) page is part of an mlocked VMA
3746 int page_evictable(struct page *page)
3748 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3753 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3754 * @pages: array of pages to check
3755 * @nr_pages: number of pages to check
3757 * Checks pages for evictability and moves them to the appropriate lru list.
3759 * This function is only used for SysV IPC SHM_UNLOCK.
3761 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3763 struct lruvec *lruvec;
3764 struct zone *zone = NULL;
3769 for (i = 0; i < nr_pages; i++) {
3770 struct page *page = pages[i];
3771 struct zone *pagezone;
3774 pagezone = page_zone(page);
3775 if (pagezone != zone) {
3777 spin_unlock_irq(&zone->lru_lock);
3779 spin_lock_irq(&zone->lru_lock);
3781 lruvec = mem_cgroup_page_lruvec(page, zone);
3783 if (!PageLRU(page) || !PageUnevictable(page))
3786 if (page_evictable(page)) {
3787 enum lru_list lru = page_lru_base_type(page);
3789 VM_BUG_ON_PAGE(PageActive(page), page);
3790 ClearPageUnevictable(page);
3791 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3792 add_page_to_lru_list(page, lruvec, lru);
3798 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3799 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3800 spin_unlock_irq(&zone->lru_lock);
3803 #endif /* CONFIG_SHMEM */
3805 static void warn_scan_unevictable_pages(void)
3807 printk_once(KERN_WARNING
3808 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3809 "disabled for lack of a legitimate use case. If you have "
3810 "one, please send an email to linux-mm@kvack.org.\n",
3815 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3816 * all nodes' unevictable lists for evictable pages
3818 unsigned long scan_unevictable_pages;
3820 int scan_unevictable_handler(struct ctl_table *table, int write,
3821 void __user *buffer,
3822 size_t *length, loff_t *ppos)
3824 warn_scan_unevictable_pages();
3825 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3826 scan_unevictable_pages = 0;
3832 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3833 * a specified node's per zone unevictable lists for evictable pages.
3836 static ssize_t read_scan_unevictable_node(struct device *dev,
3837 struct device_attribute *attr,
3840 warn_scan_unevictable_pages();
3841 return sprintf(buf, "0\n"); /* always zero; should fit... */
3844 static ssize_t write_scan_unevictable_node(struct device *dev,
3845 struct device_attribute *attr,
3846 const char *buf, size_t count)
3848 warn_scan_unevictable_pages();
3853 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3854 read_scan_unevictable_node,
3855 write_scan_unevictable_node);
3857 int scan_unevictable_register_node(struct node *node)
3859 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3862 void scan_unevictable_unregister_node(struct node *node)
3864 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);