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 /* One of the zones is ready for compaction */
71 /* How many pages shrink_list() should reclaim */
72 unsigned long nr_to_reclaim;
74 unsigned long hibernation_mode;
76 /* This context's GFP mask */
81 /* Can mapped pages be reclaimed? */
84 /* Can pages be swapped as part of reclaim? */
89 /* Scan (total_size >> priority) pages at once */
92 /* anon vs. file LRUs scanning "ratio" */
96 * The memory cgroup that hit its limit and as a result is the
97 * primary target of this reclaim invocation.
99 struct mem_cgroup *target_mem_cgroup;
102 * Nodemask of nodes allowed by the caller. If NULL, all nodes
105 nodemask_t *nodemask;
108 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
110 #ifdef ARCH_HAS_PREFETCH
111 #define prefetch_prev_lru_page(_page, _base, _field) \
113 if ((_page)->lru.prev != _base) { \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
121 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
124 #ifdef ARCH_HAS_PREFETCHW
125 #define prefetchw_prev_lru_page(_page, _base, _field) \
127 if ((_page)->lru.prev != _base) { \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
135 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139 * From 0 .. 100. Higher means more swappy.
141 int vm_swappiness = 60;
142 unsigned long vm_total_pages; /* The total number of pages which the VM controls */
144 static LIST_HEAD(shrinker_list);
145 static DECLARE_RWSEM(shrinker_rwsem);
148 static bool global_reclaim(struct scan_control *sc)
150 return !sc->target_mem_cgroup;
153 static bool global_reclaim(struct scan_control *sc)
159 static unsigned long zone_reclaimable_pages(struct zone *zone)
163 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
164 zone_page_state(zone, NR_INACTIVE_FILE);
166 if (get_nr_swap_pages() > 0)
167 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
168 zone_page_state(zone, NR_INACTIVE_ANON);
173 bool zone_reclaimable(struct zone *zone)
175 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
178 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
180 if (!mem_cgroup_disabled())
181 return mem_cgroup_get_lru_size(lruvec, lru);
183 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
187 * Add a shrinker callback to be called from the vm.
189 int register_shrinker(struct shrinker *shrinker)
191 size_t size = sizeof(*shrinker->nr_deferred);
194 * If we only have one possible node in the system anyway, save
195 * ourselves the trouble and disable NUMA aware behavior. This way we
196 * will save memory and some small loop time later.
198 if (nr_node_ids == 1)
199 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
201 if (shrinker->flags & SHRINKER_NUMA_AWARE)
204 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
205 if (!shrinker->nr_deferred)
208 down_write(&shrinker_rwsem);
209 list_add_tail(&shrinker->list, &shrinker_list);
210 up_write(&shrinker_rwsem);
213 EXPORT_SYMBOL(register_shrinker);
218 void unregister_shrinker(struct shrinker *shrinker)
220 down_write(&shrinker_rwsem);
221 list_del(&shrinker->list);
222 up_write(&shrinker_rwsem);
223 kfree(shrinker->nr_deferred);
225 EXPORT_SYMBOL(unregister_shrinker);
227 #define SHRINK_BATCH 128
230 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
231 unsigned long nr_pages_scanned, unsigned long lru_pages)
233 unsigned long freed = 0;
234 unsigned long long delta;
239 int nid = shrinkctl->nid;
240 long batch_size = shrinker->batch ? shrinker->batch
243 freeable = shrinker->count_objects(shrinker, shrinkctl);
248 * copy the current shrinker scan count into a local variable
249 * and zero it so that other concurrent shrinker invocations
250 * don't also do this scanning work.
252 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
255 delta = (4 * nr_pages_scanned) / shrinker->seeks;
257 do_div(delta, lru_pages + 1);
259 if (total_scan < 0) {
261 "shrink_slab: %pF negative objects to delete nr=%ld\n",
262 shrinker->scan_objects, total_scan);
263 total_scan = freeable;
267 * We need to avoid excessive windup on filesystem shrinkers
268 * due to large numbers of GFP_NOFS allocations causing the
269 * shrinkers to return -1 all the time. This results in a large
270 * nr being built up so when a shrink that can do some work
271 * comes along it empties the entire cache due to nr >>>
272 * freeable. This is bad for sustaining a working set in
275 * Hence only allow the shrinker to scan the entire cache when
276 * a large delta change is calculated directly.
278 if (delta < freeable / 4)
279 total_scan = min(total_scan, freeable / 2);
282 * Avoid risking looping forever due to too large nr value:
283 * never try to free more than twice the estimate number of
286 if (total_scan > freeable * 2)
287 total_scan = freeable * 2;
289 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
290 nr_pages_scanned, lru_pages,
291 freeable, delta, total_scan);
294 * Normally, we should not scan less than batch_size objects in one
295 * pass to avoid too frequent shrinker calls, but if the slab has less
296 * than batch_size objects in total and we are really tight on memory,
297 * we will try to reclaim all available objects, otherwise we can end
298 * up failing allocations although there are plenty of reclaimable
299 * objects spread over several slabs with usage less than the
302 * We detect the "tight on memory" situations by looking at the total
303 * number of objects we want to scan (total_scan). If it is greater
304 * than the total number of objects on slab (freeable), we must be
305 * scanning at high prio and therefore should try to reclaim as much as
308 while (total_scan >= batch_size ||
309 total_scan >= freeable) {
311 unsigned long nr_to_scan = min(batch_size, total_scan);
313 shrinkctl->nr_to_scan = nr_to_scan;
314 ret = shrinker->scan_objects(shrinker, shrinkctl);
315 if (ret == SHRINK_STOP)
319 count_vm_events(SLABS_SCANNED, nr_to_scan);
320 total_scan -= nr_to_scan;
326 * move the unused scan count back into the shrinker in a
327 * manner that handles concurrent updates. If we exhausted the
328 * scan, there is no need to do an update.
331 new_nr = atomic_long_add_return(total_scan,
332 &shrinker->nr_deferred[nid]);
334 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
336 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
341 * Call the shrink functions to age shrinkable caches
343 * Here we assume it costs one seek to replace a lru page and that it also
344 * takes a seek to recreate a cache object. With this in mind we age equal
345 * percentages of the lru and ageable caches. This should balance the seeks
346 * generated by these structures.
348 * If the vm encountered mapped pages on the LRU it increase the pressure on
349 * slab to avoid swapping.
351 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
353 * `lru_pages' represents the number of on-LRU pages in all the zones which
354 * are eligible for the caller's allocation attempt. It is used for balancing
355 * slab reclaim versus page reclaim.
357 * Returns the number of slab objects which we shrunk.
359 unsigned long shrink_slab(struct shrink_control *shrinkctl,
360 unsigned long nr_pages_scanned,
361 unsigned long lru_pages)
363 struct shrinker *shrinker;
364 unsigned long freed = 0;
366 if (nr_pages_scanned == 0)
367 nr_pages_scanned = SWAP_CLUSTER_MAX;
369 if (!down_read_trylock(&shrinker_rwsem)) {
371 * If we would return 0, our callers would understand that we
372 * have nothing else to shrink and give up trying. By returning
373 * 1 we keep it going and assume we'll be able to shrink next
380 list_for_each_entry(shrinker, &shrinker_list, list) {
381 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
383 freed += shrink_slab_node(shrinkctl, shrinker,
384 nr_pages_scanned, lru_pages);
388 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
389 if (node_online(shrinkctl->nid))
390 freed += shrink_slab_node(shrinkctl, shrinker,
391 nr_pages_scanned, lru_pages);
395 up_read(&shrinker_rwsem);
401 static inline int is_page_cache_freeable(struct page *page)
404 * A freeable page cache page is referenced only by the caller
405 * that isolated the page, the page cache radix tree and
406 * optional buffer heads at page->private.
408 return page_count(page) - page_has_private(page) == 2;
411 static int may_write_to_queue(struct backing_dev_info *bdi,
412 struct scan_control *sc)
414 if (current->flags & PF_SWAPWRITE)
416 if (!bdi_write_congested(bdi))
418 if (bdi == current->backing_dev_info)
424 * We detected a synchronous write error writing a page out. Probably
425 * -ENOSPC. We need to propagate that into the address_space for a subsequent
426 * fsync(), msync() or close().
428 * The tricky part is that after writepage we cannot touch the mapping: nothing
429 * prevents it from being freed up. But we have a ref on the page and once
430 * that page is locked, the mapping is pinned.
432 * We're allowed to run sleeping lock_page() here because we know the caller has
435 static void handle_write_error(struct address_space *mapping,
436 struct page *page, int error)
439 if (page_mapping(page) == mapping)
440 mapping_set_error(mapping, error);
444 /* possible outcome of pageout() */
446 /* failed to write page out, page is locked */
448 /* move page to the active list, page is locked */
450 /* page has been sent to the disk successfully, page is unlocked */
452 /* page is clean and locked */
457 * pageout is called by shrink_page_list() for each dirty page.
458 * Calls ->writepage().
460 static pageout_t pageout(struct page *page, struct address_space *mapping,
461 struct scan_control *sc)
464 * If the page is dirty, only perform writeback if that write
465 * will be non-blocking. To prevent this allocation from being
466 * stalled by pagecache activity. But note that there may be
467 * stalls if we need to run get_block(). We could test
468 * PagePrivate for that.
470 * If this process is currently in __generic_file_write_iter() against
471 * this page's queue, we can perform writeback even if that
474 * If the page is swapcache, write it back even if that would
475 * block, for some throttling. This happens by accident, because
476 * swap_backing_dev_info is bust: it doesn't reflect the
477 * congestion state of the swapdevs. Easy to fix, if needed.
479 if (!is_page_cache_freeable(page))
483 * Some data journaling orphaned pages can have
484 * page->mapping == NULL while being dirty with clean buffers.
486 if (page_has_private(page)) {
487 if (try_to_free_buffers(page)) {
488 ClearPageDirty(page);
489 pr_info("%s: orphaned page\n", __func__);
495 if (mapping->a_ops->writepage == NULL)
496 return PAGE_ACTIVATE;
497 if (!may_write_to_queue(mapping->backing_dev_info, sc))
500 if (clear_page_dirty_for_io(page)) {
502 struct writeback_control wbc = {
503 .sync_mode = WB_SYNC_NONE,
504 .nr_to_write = SWAP_CLUSTER_MAX,
506 .range_end = LLONG_MAX,
510 SetPageReclaim(page);
511 res = mapping->a_ops->writepage(page, &wbc);
513 handle_write_error(mapping, page, res);
514 if (res == AOP_WRITEPAGE_ACTIVATE) {
515 ClearPageReclaim(page);
516 return PAGE_ACTIVATE;
519 if (!PageWriteback(page)) {
520 /* synchronous write or broken a_ops? */
521 ClearPageReclaim(page);
523 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
524 inc_zone_page_state(page, NR_VMSCAN_WRITE);
532 * Same as remove_mapping, but if the page is removed from the mapping, it
533 * gets returned with a refcount of 0.
535 static int __remove_mapping(struct address_space *mapping, struct page *page,
538 BUG_ON(!PageLocked(page));
539 BUG_ON(mapping != page_mapping(page));
541 spin_lock_irq(&mapping->tree_lock);
543 * The non racy check for a busy page.
545 * Must be careful with the order of the tests. When someone has
546 * a ref to the page, it may be possible that they dirty it then
547 * drop the reference. So if PageDirty is tested before page_count
548 * here, then the following race may occur:
550 * get_user_pages(&page);
551 * [user mapping goes away]
553 * !PageDirty(page) [good]
554 * SetPageDirty(page);
556 * !page_count(page) [good, discard it]
558 * [oops, our write_to data is lost]
560 * Reversing the order of the tests ensures such a situation cannot
561 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
562 * load is not satisfied before that of page->_count.
564 * Note that if SetPageDirty is always performed via set_page_dirty,
565 * and thus under tree_lock, then this ordering is not required.
567 if (!page_freeze_refs(page, 2))
569 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
570 if (unlikely(PageDirty(page))) {
571 page_unfreeze_refs(page, 2);
575 if (PageSwapCache(page)) {
576 swp_entry_t swap = { .val = page_private(page) };
577 __delete_from_swap_cache(page);
578 spin_unlock_irq(&mapping->tree_lock);
579 swapcache_free(swap, page);
581 void (*freepage)(struct page *);
584 freepage = mapping->a_ops->freepage;
586 * Remember a shadow entry for reclaimed file cache in
587 * order to detect refaults, thus thrashing, later on.
589 * But don't store shadows in an address space that is
590 * already exiting. This is not just an optizimation,
591 * inode reclaim needs to empty out the radix tree or
592 * the nodes are lost. Don't plant shadows behind its
595 if (reclaimed && page_is_file_cache(page) &&
596 !mapping_exiting(mapping))
597 shadow = workingset_eviction(mapping, page);
598 __delete_from_page_cache(page, shadow);
599 spin_unlock_irq(&mapping->tree_lock);
600 mem_cgroup_uncharge_cache_page(page);
602 if (freepage != NULL)
609 spin_unlock_irq(&mapping->tree_lock);
614 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
615 * someone else has a ref on the page, abort and return 0. If it was
616 * successfully detached, return 1. Assumes the caller has a single ref on
619 int remove_mapping(struct address_space *mapping, struct page *page)
621 if (__remove_mapping(mapping, page, false)) {
623 * Unfreezing the refcount with 1 rather than 2 effectively
624 * drops the pagecache ref for us without requiring another
627 page_unfreeze_refs(page, 1);
634 * putback_lru_page - put previously isolated page onto appropriate LRU list
635 * @page: page to be put back to appropriate lru list
637 * Add previously isolated @page to appropriate LRU list.
638 * Page may still be unevictable for other reasons.
640 * lru_lock must not be held, interrupts must be enabled.
642 void putback_lru_page(struct page *page)
645 int was_unevictable = PageUnevictable(page);
647 VM_BUG_ON_PAGE(PageLRU(page), page);
650 ClearPageUnevictable(page);
652 if (page_evictable(page)) {
654 * For evictable pages, we can use the cache.
655 * In event of a race, worst case is we end up with an
656 * unevictable page on [in]active list.
657 * We know how to handle that.
659 is_unevictable = false;
663 * Put unevictable pages directly on zone's unevictable
666 is_unevictable = true;
667 add_page_to_unevictable_list(page);
669 * When racing with an mlock or AS_UNEVICTABLE clearing
670 * (page is unlocked) make sure that if the other thread
671 * does not observe our setting of PG_lru and fails
672 * isolation/check_move_unevictable_pages,
673 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
674 * the page back to the evictable list.
676 * The other side is TestClearPageMlocked() or shmem_lock().
682 * page's status can change while we move it among lru. If an evictable
683 * page is on unevictable list, it never be freed. To avoid that,
684 * check after we added it to the list, again.
686 if (is_unevictable && page_evictable(page)) {
687 if (!isolate_lru_page(page)) {
691 /* This means someone else dropped this page from LRU
692 * So, it will be freed or putback to LRU again. There is
693 * nothing to do here.
697 if (was_unevictable && !is_unevictable)
698 count_vm_event(UNEVICTABLE_PGRESCUED);
699 else if (!was_unevictable && is_unevictable)
700 count_vm_event(UNEVICTABLE_PGCULLED);
702 put_page(page); /* drop ref from isolate */
705 enum page_references {
707 PAGEREF_RECLAIM_CLEAN,
712 static enum page_references page_check_references(struct page *page,
713 struct scan_control *sc)
715 int referenced_ptes, referenced_page;
716 unsigned long vm_flags;
718 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
720 referenced_page = TestClearPageReferenced(page);
723 * Mlock lost the isolation race with us. Let try_to_unmap()
724 * move the page to the unevictable list.
726 if (vm_flags & VM_LOCKED)
727 return PAGEREF_RECLAIM;
729 if (referenced_ptes) {
730 if (PageSwapBacked(page))
731 return PAGEREF_ACTIVATE;
733 * All mapped pages start out with page table
734 * references from the instantiating fault, so we need
735 * to look twice if a mapped file page is used more
738 * Mark it and spare it for another trip around the
739 * inactive list. Another page table reference will
740 * lead to its activation.
742 * Note: the mark is set for activated pages as well
743 * so that recently deactivated but used pages are
746 SetPageReferenced(page);
748 if (referenced_page || referenced_ptes > 1)
749 return PAGEREF_ACTIVATE;
752 * Activate file-backed executable pages after first usage.
754 if (vm_flags & VM_EXEC)
755 return PAGEREF_ACTIVATE;
760 /* Reclaim if clean, defer dirty pages to writeback */
761 if (referenced_page && !PageSwapBacked(page))
762 return PAGEREF_RECLAIM_CLEAN;
764 return PAGEREF_RECLAIM;
767 /* Check if a page is dirty or under writeback */
768 static void page_check_dirty_writeback(struct page *page,
769 bool *dirty, bool *writeback)
771 struct address_space *mapping;
774 * Anonymous pages are not handled by flushers and must be written
775 * from reclaim context. Do not stall reclaim based on them
777 if (!page_is_file_cache(page)) {
783 /* By default assume that the page flags are accurate */
784 *dirty = PageDirty(page);
785 *writeback = PageWriteback(page);
787 /* Verify dirty/writeback state if the filesystem supports it */
788 if (!page_has_private(page))
791 mapping = page_mapping(page);
792 if (mapping && mapping->a_ops->is_dirty_writeback)
793 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
797 * shrink_page_list() returns the number of reclaimed pages
799 static unsigned long shrink_page_list(struct list_head *page_list,
801 struct scan_control *sc,
802 enum ttu_flags ttu_flags,
803 unsigned long *ret_nr_dirty,
804 unsigned long *ret_nr_unqueued_dirty,
805 unsigned long *ret_nr_congested,
806 unsigned long *ret_nr_writeback,
807 unsigned long *ret_nr_immediate,
810 LIST_HEAD(ret_pages);
811 LIST_HEAD(free_pages);
813 unsigned long nr_unqueued_dirty = 0;
814 unsigned long nr_dirty = 0;
815 unsigned long nr_congested = 0;
816 unsigned long nr_reclaimed = 0;
817 unsigned long nr_writeback = 0;
818 unsigned long nr_immediate = 0;
822 mem_cgroup_uncharge_start();
823 while (!list_empty(page_list)) {
824 struct address_space *mapping;
827 enum page_references references = PAGEREF_RECLAIM_CLEAN;
828 bool dirty, writeback;
832 page = lru_to_page(page_list);
833 list_del(&page->lru);
835 if (!trylock_page(page))
838 VM_BUG_ON_PAGE(PageActive(page), page);
839 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
843 if (unlikely(!page_evictable(page)))
846 if (!sc->may_unmap && page_mapped(page))
849 /* Double the slab pressure for mapped and swapcache pages */
850 if (page_mapped(page) || PageSwapCache(page))
853 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
854 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
857 * The number of dirty pages determines if a zone is marked
858 * reclaim_congested which affects wait_iff_congested. kswapd
859 * will stall and start writing pages if the tail of the LRU
860 * is all dirty unqueued pages.
862 page_check_dirty_writeback(page, &dirty, &writeback);
863 if (dirty || writeback)
866 if (dirty && !writeback)
870 * Treat this page as congested if the underlying BDI is or if
871 * pages are cycling through the LRU so quickly that the
872 * pages marked for immediate reclaim are making it to the
873 * end of the LRU a second time.
875 mapping = page_mapping(page);
876 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
877 (writeback && PageReclaim(page)))
881 * If a page at the tail of the LRU is under writeback, there
882 * are three cases to consider.
884 * 1) If reclaim is encountering an excessive number of pages
885 * under writeback and this page is both under writeback and
886 * PageReclaim then it indicates that pages are being queued
887 * for IO but are being recycled through the LRU before the
888 * IO can complete. Waiting on the page itself risks an
889 * indefinite stall if it is impossible to writeback the
890 * page due to IO error or disconnected storage so instead
891 * note that the LRU is being scanned too quickly and the
892 * caller can stall after page list has been processed.
894 * 2) Global reclaim encounters a page, memcg encounters a
895 * page that is not marked for immediate reclaim or
896 * the caller does not have __GFP_IO. In this case mark
897 * the page for immediate reclaim and continue scanning.
899 * __GFP_IO is checked because a loop driver thread might
900 * enter reclaim, and deadlock if it waits on a page for
901 * which it is needed to do the write (loop masks off
902 * __GFP_IO|__GFP_FS for this reason); but more thought
903 * would probably show more reasons.
905 * Don't require __GFP_FS, since we're not going into the
906 * FS, just waiting on its writeback completion. Worryingly,
907 * ext4 gfs2 and xfs allocate pages with
908 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
909 * may_enter_fs here is liable to OOM on them.
911 * 3) memcg encounters a page that is not already marked
912 * PageReclaim. memcg does not have any dirty pages
913 * throttling so we could easily OOM just because too many
914 * pages are in writeback and there is nothing else to
915 * reclaim. Wait for the writeback to complete.
917 if (PageWriteback(page)) {
919 if (current_is_kswapd() &&
921 zone_is_reclaim_writeback(zone)) {
926 } else if (global_reclaim(sc) ||
927 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
929 * This is slightly racy - end_page_writeback()
930 * might have just cleared PageReclaim, then
931 * setting PageReclaim here end up interpreted
932 * as PageReadahead - but that does not matter
933 * enough to care. What we do want is for this
934 * page to have PageReclaim set next time memcg
935 * reclaim reaches the tests above, so it will
936 * then wait_on_page_writeback() to avoid OOM;
937 * and it's also appropriate in global reclaim.
939 SetPageReclaim(page);
946 wait_on_page_writeback(page);
951 references = page_check_references(page, sc);
953 switch (references) {
954 case PAGEREF_ACTIVATE:
955 goto activate_locked;
958 case PAGEREF_RECLAIM:
959 case PAGEREF_RECLAIM_CLEAN:
960 ; /* try to reclaim the page below */
964 * Anonymous process memory has backing store?
965 * Try to allocate it some swap space here.
967 if (PageAnon(page) && !PageSwapCache(page)) {
968 if (!(sc->gfp_mask & __GFP_IO))
970 if (!add_to_swap(page, page_list))
971 goto activate_locked;
974 /* Adding to swap updated mapping */
975 mapping = page_mapping(page);
979 * The page is mapped into the page tables of one or more
980 * processes. Try to unmap it here.
982 if (page_mapped(page) && mapping) {
983 switch (try_to_unmap(page, ttu_flags)) {
985 goto activate_locked;
991 ; /* try to free the page below */
995 if (PageDirty(page)) {
997 * Only kswapd can writeback filesystem pages to
998 * avoid risk of stack overflow but only writeback
999 * if many dirty pages have been encountered.
1001 if (page_is_file_cache(page) &&
1002 (!current_is_kswapd() ||
1003 !zone_is_reclaim_dirty(zone))) {
1005 * Immediately reclaim when written back.
1006 * Similar in principal to deactivate_page()
1007 * except we already have the page isolated
1008 * and know it's dirty
1010 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1011 SetPageReclaim(page);
1016 if (references == PAGEREF_RECLAIM_CLEAN)
1020 if (!sc->may_writepage)
1023 /* Page is dirty, try to write it out here */
1024 switch (pageout(page, mapping, sc)) {
1028 goto activate_locked;
1030 if (PageWriteback(page))
1032 if (PageDirty(page))
1036 * A synchronous write - probably a ramdisk. Go
1037 * ahead and try to reclaim the page.
1039 if (!trylock_page(page))
1041 if (PageDirty(page) || PageWriteback(page))
1043 mapping = page_mapping(page);
1045 ; /* try to free the page below */
1050 * If the page has buffers, try to free the buffer mappings
1051 * associated with this page. If we succeed we try to free
1054 * We do this even if the page is PageDirty().
1055 * try_to_release_page() does not perform I/O, but it is
1056 * possible for a page to have PageDirty set, but it is actually
1057 * clean (all its buffers are clean). This happens if the
1058 * buffers were written out directly, with submit_bh(). ext3
1059 * will do this, as well as the blockdev mapping.
1060 * try_to_release_page() will discover that cleanness and will
1061 * drop the buffers and mark the page clean - it can be freed.
1063 * Rarely, pages can have buffers and no ->mapping. These are
1064 * the pages which were not successfully invalidated in
1065 * truncate_complete_page(). We try to drop those buffers here
1066 * and if that worked, and the page is no longer mapped into
1067 * process address space (page_count == 1) it can be freed.
1068 * Otherwise, leave the page on the LRU so it is swappable.
1070 if (page_has_private(page)) {
1071 if (!try_to_release_page(page, sc->gfp_mask))
1072 goto activate_locked;
1073 if (!mapping && page_count(page) == 1) {
1075 if (put_page_testzero(page))
1079 * rare race with speculative reference.
1080 * the speculative reference will free
1081 * this page shortly, so we may
1082 * increment nr_reclaimed here (and
1083 * leave it off the LRU).
1091 if (!mapping || !__remove_mapping(mapping, page, true))
1095 * At this point, we have no other references and there is
1096 * no way to pick any more up (removed from LRU, removed
1097 * from pagecache). Can use non-atomic bitops now (and
1098 * we obviously don't have to worry about waking up a process
1099 * waiting on the page lock, because there are no references.
1101 __clear_page_locked(page);
1106 * Is there need to periodically free_page_list? It would
1107 * appear not as the counts should be low
1109 list_add(&page->lru, &free_pages);
1113 if (PageSwapCache(page))
1114 try_to_free_swap(page);
1116 putback_lru_page(page);
1120 /* Not a candidate for swapping, so reclaim swap space. */
1121 if (PageSwapCache(page) && vm_swap_full())
1122 try_to_free_swap(page);
1123 VM_BUG_ON_PAGE(PageActive(page), page);
1124 SetPageActive(page);
1129 list_add(&page->lru, &ret_pages);
1130 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1133 free_hot_cold_page_list(&free_pages, true);
1135 list_splice(&ret_pages, page_list);
1136 count_vm_events(PGACTIVATE, pgactivate);
1137 mem_cgroup_uncharge_end();
1138 *ret_nr_dirty += nr_dirty;
1139 *ret_nr_congested += nr_congested;
1140 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1141 *ret_nr_writeback += nr_writeback;
1142 *ret_nr_immediate += nr_immediate;
1143 return nr_reclaimed;
1146 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1147 struct list_head *page_list)
1149 struct scan_control sc = {
1150 .gfp_mask = GFP_KERNEL,
1151 .priority = DEF_PRIORITY,
1154 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1155 struct page *page, *next;
1156 LIST_HEAD(clean_pages);
1158 list_for_each_entry_safe(page, next, page_list, lru) {
1159 if (page_is_file_cache(page) && !PageDirty(page) &&
1160 !isolated_balloon_page(page)) {
1161 ClearPageActive(page);
1162 list_move(&page->lru, &clean_pages);
1166 ret = shrink_page_list(&clean_pages, zone, &sc,
1167 TTU_UNMAP|TTU_IGNORE_ACCESS,
1168 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1169 list_splice(&clean_pages, page_list);
1170 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1175 * Attempt to remove the specified page from its LRU. Only take this page
1176 * if it is of the appropriate PageActive status. Pages which are being
1177 * freed elsewhere are also ignored.
1179 * page: page to consider
1180 * mode: one of the LRU isolation modes defined above
1182 * returns 0 on success, -ve errno on failure.
1184 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1188 /* Only take pages on the LRU. */
1192 /* Compaction should not handle unevictable pages but CMA can do so */
1193 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1199 * To minimise LRU disruption, the caller can indicate that it only
1200 * wants to isolate pages it will be able to operate on without
1201 * blocking - clean pages for the most part.
1203 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1204 * is used by reclaim when it is cannot write to backing storage
1206 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1207 * that it is possible to migrate without blocking
1209 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1210 /* All the caller can do on PageWriteback is block */
1211 if (PageWriteback(page))
1214 if (PageDirty(page)) {
1215 struct address_space *mapping;
1217 /* ISOLATE_CLEAN means only clean pages */
1218 if (mode & ISOLATE_CLEAN)
1222 * Only pages without mappings or that have a
1223 * ->migratepage callback are possible to migrate
1226 mapping = page_mapping(page);
1227 if (mapping && !mapping->a_ops->migratepage)
1232 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1235 if (likely(get_page_unless_zero(page))) {
1237 * Be careful not to clear PageLRU until after we're
1238 * sure the page is not being freed elsewhere -- the
1239 * page release code relies on it.
1249 * zone->lru_lock is heavily contended. Some of the functions that
1250 * shrink the lists perform better by taking out a batch of pages
1251 * and working on them outside the LRU lock.
1253 * For pagecache intensive workloads, this function is the hottest
1254 * spot in the kernel (apart from copy_*_user functions).
1256 * Appropriate locks must be held before calling this function.
1258 * @nr_to_scan: The number of pages to look through on the list.
1259 * @lruvec: The LRU vector to pull pages from.
1260 * @dst: The temp list to put pages on to.
1261 * @nr_scanned: The number of pages that were scanned.
1262 * @sc: The scan_control struct for this reclaim session
1263 * @mode: One of the LRU isolation modes
1264 * @lru: LRU list id for isolating
1266 * returns how many pages were moved onto *@dst.
1268 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1269 struct lruvec *lruvec, struct list_head *dst,
1270 unsigned long *nr_scanned, struct scan_control *sc,
1271 isolate_mode_t mode, enum lru_list lru)
1273 struct list_head *src = &lruvec->lists[lru];
1274 unsigned long nr_taken = 0;
1277 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1281 page = lru_to_page(src);
1282 prefetchw_prev_lru_page(page, src, flags);
1284 VM_BUG_ON_PAGE(!PageLRU(page), page);
1286 switch (__isolate_lru_page(page, mode)) {
1288 nr_pages = hpage_nr_pages(page);
1289 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1290 list_move(&page->lru, dst);
1291 nr_taken += nr_pages;
1295 /* else it is being freed elsewhere */
1296 list_move(&page->lru, src);
1305 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1306 nr_taken, mode, is_file_lru(lru));
1311 * isolate_lru_page - tries to isolate a page from its LRU list
1312 * @page: page to isolate from its LRU list
1314 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1315 * vmstat statistic corresponding to whatever LRU list the page was on.
1317 * Returns 0 if the page was removed from an LRU list.
1318 * Returns -EBUSY if the page was not on an LRU list.
1320 * The returned page will have PageLRU() cleared. If it was found on
1321 * the active list, it will have PageActive set. If it was found on
1322 * the unevictable list, it will have the PageUnevictable bit set. That flag
1323 * may need to be cleared by the caller before letting the page go.
1325 * The vmstat statistic corresponding to the list on which the page was
1326 * found will be decremented.
1329 * (1) Must be called with an elevated refcount on the page. This is a
1330 * fundamentnal difference from isolate_lru_pages (which is called
1331 * without a stable reference).
1332 * (2) the lru_lock must not be held.
1333 * (3) interrupts must be enabled.
1335 int isolate_lru_page(struct page *page)
1339 VM_BUG_ON_PAGE(!page_count(page), page);
1341 if (PageLRU(page)) {
1342 struct zone *zone = page_zone(page);
1343 struct lruvec *lruvec;
1345 spin_lock_irq(&zone->lru_lock);
1346 lruvec = mem_cgroup_page_lruvec(page, zone);
1347 if (PageLRU(page)) {
1348 int lru = page_lru(page);
1351 del_page_from_lru_list(page, lruvec, lru);
1354 spin_unlock_irq(&zone->lru_lock);
1360 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1361 * then get resheduled. When there are massive number of tasks doing page
1362 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1363 * the LRU list will go small and be scanned faster than necessary, leading to
1364 * unnecessary swapping, thrashing and OOM.
1366 static int too_many_isolated(struct zone *zone, int file,
1367 struct scan_control *sc)
1369 unsigned long inactive, isolated;
1371 if (current_is_kswapd())
1374 if (!global_reclaim(sc))
1378 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1379 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1381 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1382 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1386 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1387 * won't get blocked by normal direct-reclaimers, forming a circular
1390 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1393 return isolated > inactive;
1396 static noinline_for_stack void
1397 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1399 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1400 struct zone *zone = lruvec_zone(lruvec);
1401 LIST_HEAD(pages_to_free);
1404 * Put back any unfreeable pages.
1406 while (!list_empty(page_list)) {
1407 struct page *page = lru_to_page(page_list);
1410 VM_BUG_ON_PAGE(PageLRU(page), page);
1411 list_del(&page->lru);
1412 if (unlikely(!page_evictable(page))) {
1413 spin_unlock_irq(&zone->lru_lock);
1414 putback_lru_page(page);
1415 spin_lock_irq(&zone->lru_lock);
1419 lruvec = mem_cgroup_page_lruvec(page, zone);
1422 lru = page_lru(page);
1423 add_page_to_lru_list(page, lruvec, lru);
1425 if (is_active_lru(lru)) {
1426 int file = is_file_lru(lru);
1427 int numpages = hpage_nr_pages(page);
1428 reclaim_stat->recent_rotated[file] += numpages;
1430 if (put_page_testzero(page)) {
1431 __ClearPageLRU(page);
1432 __ClearPageActive(page);
1433 del_page_from_lru_list(page, lruvec, lru);
1435 if (unlikely(PageCompound(page))) {
1436 spin_unlock_irq(&zone->lru_lock);
1437 (*get_compound_page_dtor(page))(page);
1438 spin_lock_irq(&zone->lru_lock);
1440 list_add(&page->lru, &pages_to_free);
1445 * To save our caller's stack, now use input list for pages to free.
1447 list_splice(&pages_to_free, page_list);
1451 * If a kernel thread (such as nfsd for loop-back mounts) services
1452 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1453 * In that case we should only throttle if the backing device it is
1454 * writing to is congested. In other cases it is safe to throttle.
1456 static int current_may_throttle(void)
1458 return !(current->flags & PF_LESS_THROTTLE) ||
1459 current->backing_dev_info == NULL ||
1460 bdi_write_congested(current->backing_dev_info);
1464 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1465 * of reclaimed pages
1467 static noinline_for_stack unsigned long
1468 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1469 struct scan_control *sc, enum lru_list lru)
1471 LIST_HEAD(page_list);
1472 unsigned long nr_scanned;
1473 unsigned long nr_reclaimed = 0;
1474 unsigned long nr_taken;
1475 unsigned long nr_dirty = 0;
1476 unsigned long nr_congested = 0;
1477 unsigned long nr_unqueued_dirty = 0;
1478 unsigned long nr_writeback = 0;
1479 unsigned long nr_immediate = 0;
1480 isolate_mode_t isolate_mode = 0;
1481 int file = is_file_lru(lru);
1482 struct zone *zone = lruvec_zone(lruvec);
1483 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1485 while (unlikely(too_many_isolated(zone, file, sc))) {
1486 congestion_wait(BLK_RW_ASYNC, HZ/10);
1488 /* We are about to die and free our memory. Return now. */
1489 if (fatal_signal_pending(current))
1490 return SWAP_CLUSTER_MAX;
1496 isolate_mode |= ISOLATE_UNMAPPED;
1497 if (!sc->may_writepage)
1498 isolate_mode |= ISOLATE_CLEAN;
1500 spin_lock_irq(&zone->lru_lock);
1502 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1503 &nr_scanned, sc, isolate_mode, lru);
1505 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1506 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1508 if (global_reclaim(sc)) {
1509 zone->pages_scanned += nr_scanned;
1510 if (current_is_kswapd())
1511 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1513 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1515 spin_unlock_irq(&zone->lru_lock);
1520 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1521 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1522 &nr_writeback, &nr_immediate,
1525 spin_lock_irq(&zone->lru_lock);
1527 reclaim_stat->recent_scanned[file] += nr_taken;
1529 if (global_reclaim(sc)) {
1530 if (current_is_kswapd())
1531 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1534 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1538 putback_inactive_pages(lruvec, &page_list);
1540 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1542 spin_unlock_irq(&zone->lru_lock);
1544 free_hot_cold_page_list(&page_list, true);
1547 * If reclaim is isolating dirty pages under writeback, it implies
1548 * that the long-lived page allocation rate is exceeding the page
1549 * laundering rate. Either the global limits are not being effective
1550 * at throttling processes due to the page distribution throughout
1551 * zones or there is heavy usage of a slow backing device. The
1552 * only option is to throttle from reclaim context which is not ideal
1553 * as there is no guarantee the dirtying process is throttled in the
1554 * same way balance_dirty_pages() manages.
1556 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1557 * of pages under pages flagged for immediate reclaim and stall if any
1558 * are encountered in the nr_immediate check below.
1560 if (nr_writeback && nr_writeback == nr_taken)
1561 zone_set_flag(zone, ZONE_WRITEBACK);
1564 * memcg will stall in page writeback so only consider forcibly
1565 * stalling for global reclaim
1567 if (global_reclaim(sc)) {
1569 * Tag a zone as congested if all the dirty pages scanned were
1570 * backed by a congested BDI and wait_iff_congested will stall.
1572 if (nr_dirty && nr_dirty == nr_congested)
1573 zone_set_flag(zone, ZONE_CONGESTED);
1576 * If dirty pages are scanned that are not queued for IO, it
1577 * implies that flushers are not keeping up. In this case, flag
1578 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1579 * pages from reclaim context.
1581 if (nr_unqueued_dirty == nr_taken)
1582 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1585 * If kswapd scans pages marked marked for immediate
1586 * reclaim and under writeback (nr_immediate), it implies
1587 * that pages are cycling through the LRU faster than
1588 * they are written so also forcibly stall.
1590 if (nr_immediate && current_may_throttle())
1591 congestion_wait(BLK_RW_ASYNC, HZ/10);
1595 * Stall direct reclaim for IO completions if underlying BDIs or zone
1596 * is congested. Allow kswapd to continue until it starts encountering
1597 * unqueued dirty pages or cycling through the LRU too quickly.
1599 if (!sc->hibernation_mode && !current_is_kswapd() &&
1600 current_may_throttle())
1601 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1603 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1605 nr_scanned, nr_reclaimed,
1607 trace_shrink_flags(file));
1608 return nr_reclaimed;
1612 * This moves pages from the active list to the inactive list.
1614 * We move them the other way if the page is referenced by one or more
1615 * processes, from rmap.
1617 * If the pages are mostly unmapped, the processing is fast and it is
1618 * appropriate to hold zone->lru_lock across the whole operation. But if
1619 * the pages are mapped, the processing is slow (page_referenced()) so we
1620 * should drop zone->lru_lock around each page. It's impossible to balance
1621 * this, so instead we remove the pages from the LRU while processing them.
1622 * It is safe to rely on PG_active against the non-LRU pages in here because
1623 * nobody will play with that bit on a non-LRU page.
1625 * The downside is that we have to touch page->_count against each page.
1626 * But we had to alter page->flags anyway.
1629 static void move_active_pages_to_lru(struct lruvec *lruvec,
1630 struct list_head *list,
1631 struct list_head *pages_to_free,
1634 struct zone *zone = lruvec_zone(lruvec);
1635 unsigned long pgmoved = 0;
1639 while (!list_empty(list)) {
1640 page = lru_to_page(list);
1641 lruvec = mem_cgroup_page_lruvec(page, zone);
1643 VM_BUG_ON_PAGE(PageLRU(page), page);
1646 nr_pages = hpage_nr_pages(page);
1647 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1648 list_move(&page->lru, &lruvec->lists[lru]);
1649 pgmoved += nr_pages;
1651 if (put_page_testzero(page)) {
1652 __ClearPageLRU(page);
1653 __ClearPageActive(page);
1654 del_page_from_lru_list(page, lruvec, lru);
1656 if (unlikely(PageCompound(page))) {
1657 spin_unlock_irq(&zone->lru_lock);
1658 (*get_compound_page_dtor(page))(page);
1659 spin_lock_irq(&zone->lru_lock);
1661 list_add(&page->lru, pages_to_free);
1664 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1665 if (!is_active_lru(lru))
1666 __count_vm_events(PGDEACTIVATE, pgmoved);
1669 static void shrink_active_list(unsigned long nr_to_scan,
1670 struct lruvec *lruvec,
1671 struct scan_control *sc,
1674 unsigned long nr_taken;
1675 unsigned long nr_scanned;
1676 unsigned long vm_flags;
1677 LIST_HEAD(l_hold); /* The pages which were snipped off */
1678 LIST_HEAD(l_active);
1679 LIST_HEAD(l_inactive);
1681 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1682 unsigned long nr_rotated = 0;
1683 isolate_mode_t isolate_mode = 0;
1684 int file = is_file_lru(lru);
1685 struct zone *zone = lruvec_zone(lruvec);
1690 isolate_mode |= ISOLATE_UNMAPPED;
1691 if (!sc->may_writepage)
1692 isolate_mode |= ISOLATE_CLEAN;
1694 spin_lock_irq(&zone->lru_lock);
1696 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1697 &nr_scanned, sc, isolate_mode, lru);
1698 if (global_reclaim(sc))
1699 zone->pages_scanned += nr_scanned;
1701 reclaim_stat->recent_scanned[file] += nr_taken;
1703 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1704 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1705 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1706 spin_unlock_irq(&zone->lru_lock);
1708 while (!list_empty(&l_hold)) {
1710 page = lru_to_page(&l_hold);
1711 list_del(&page->lru);
1713 if (unlikely(!page_evictable(page))) {
1714 putback_lru_page(page);
1718 if (unlikely(buffer_heads_over_limit)) {
1719 if (page_has_private(page) && trylock_page(page)) {
1720 if (page_has_private(page))
1721 try_to_release_page(page, 0);
1726 if (page_referenced(page, 0, sc->target_mem_cgroup,
1728 nr_rotated += hpage_nr_pages(page);
1730 * Identify referenced, file-backed active pages and
1731 * give them one more trip around the active list. So
1732 * that executable code get better chances to stay in
1733 * memory under moderate memory pressure. Anon pages
1734 * are not likely to be evicted by use-once streaming
1735 * IO, plus JVM can create lots of anon VM_EXEC pages,
1736 * so we ignore them here.
1738 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1739 list_add(&page->lru, &l_active);
1744 ClearPageActive(page); /* we are de-activating */
1745 list_add(&page->lru, &l_inactive);
1749 * Move pages back to the lru list.
1751 spin_lock_irq(&zone->lru_lock);
1753 * Count referenced pages from currently used mappings as rotated,
1754 * even though only some of them are actually re-activated. This
1755 * helps balance scan pressure between file and anonymous pages in
1758 reclaim_stat->recent_rotated[file] += nr_rotated;
1760 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1761 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1762 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1763 spin_unlock_irq(&zone->lru_lock);
1765 free_hot_cold_page_list(&l_hold, true);
1769 static int inactive_anon_is_low_global(struct zone *zone)
1771 unsigned long active, inactive;
1773 active = zone_page_state(zone, NR_ACTIVE_ANON);
1774 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1776 if (inactive * zone->inactive_ratio < active)
1783 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1784 * @lruvec: LRU vector to check
1786 * Returns true if the zone does not have enough inactive anon pages,
1787 * meaning some active anon pages need to be deactivated.
1789 static int inactive_anon_is_low(struct lruvec *lruvec)
1792 * If we don't have swap space, anonymous page deactivation
1795 if (!total_swap_pages)
1798 if (!mem_cgroup_disabled())
1799 return mem_cgroup_inactive_anon_is_low(lruvec);
1801 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1804 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1811 * inactive_file_is_low - check if file pages need to be deactivated
1812 * @lruvec: LRU vector to check
1814 * When the system is doing streaming IO, memory pressure here
1815 * ensures that active file pages get deactivated, until more
1816 * than half of the file pages are on the inactive list.
1818 * Once we get to that situation, protect the system's working
1819 * set from being evicted by disabling active file page aging.
1821 * This uses a different ratio than the anonymous pages, because
1822 * the page cache uses a use-once replacement algorithm.
1824 static int inactive_file_is_low(struct lruvec *lruvec)
1826 unsigned long inactive;
1827 unsigned long active;
1829 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1830 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1832 return active > inactive;
1835 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1837 if (is_file_lru(lru))
1838 return inactive_file_is_low(lruvec);
1840 return inactive_anon_is_low(lruvec);
1843 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1844 struct lruvec *lruvec, struct scan_control *sc)
1846 if (is_active_lru(lru)) {
1847 if (inactive_list_is_low(lruvec, lru))
1848 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1852 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1863 * Determine how aggressively the anon and file LRU lists should be
1864 * scanned. The relative value of each set of LRU lists is determined
1865 * by looking at the fraction of the pages scanned we did rotate back
1866 * onto the active list instead of evict.
1868 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1869 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1871 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1874 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1876 u64 denominator = 0; /* gcc */
1877 struct zone *zone = lruvec_zone(lruvec);
1878 unsigned long anon_prio, file_prio;
1879 enum scan_balance scan_balance;
1880 unsigned long anon, file;
1881 bool force_scan = false;
1882 unsigned long ap, fp;
1888 * If the zone or memcg is small, nr[l] can be 0. This
1889 * results in no scanning on this priority and a potential
1890 * priority drop. Global direct reclaim can go to the next
1891 * zone and tends to have no problems. Global kswapd is for
1892 * zone balancing and it needs to scan a minimum amount. When
1893 * reclaiming for a memcg, a priority drop can cause high
1894 * latencies, so it's better to scan a minimum amount there as
1897 if (current_is_kswapd() && !zone_reclaimable(zone))
1899 if (!global_reclaim(sc))
1902 /* If we have no swap space, do not bother scanning anon pages. */
1903 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1904 scan_balance = SCAN_FILE;
1909 * Global reclaim will swap to prevent OOM even with no
1910 * swappiness, but memcg users want to use this knob to
1911 * disable swapping for individual groups completely when
1912 * using the memory controller's swap limit feature would be
1915 if (!global_reclaim(sc) && !sc->swappiness) {
1916 scan_balance = SCAN_FILE;
1921 * Do not apply any pressure balancing cleverness when the
1922 * system is close to OOM, scan both anon and file equally
1923 * (unless the swappiness setting disagrees with swapping).
1925 if (!sc->priority && sc->swappiness) {
1926 scan_balance = SCAN_EQUAL;
1930 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1931 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1932 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1933 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1936 * Prevent the reclaimer from falling into the cache trap: as
1937 * cache pages start out inactive, every cache fault will tip
1938 * the scan balance towards the file LRU. And as the file LRU
1939 * shrinks, so does the window for rotation from references.
1940 * This means we have a runaway feedback loop where a tiny
1941 * thrashing file LRU becomes infinitely more attractive than
1942 * anon pages. Try to detect this based on file LRU size.
1944 if (global_reclaim(sc)) {
1945 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1947 if (unlikely(file + free <= high_wmark_pages(zone))) {
1948 scan_balance = SCAN_ANON;
1954 * There is enough inactive page cache, do not reclaim
1955 * anything from the anonymous working set right now.
1957 if (!inactive_file_is_low(lruvec)) {
1958 scan_balance = SCAN_FILE;
1962 scan_balance = SCAN_FRACT;
1965 * With swappiness at 100, anonymous and file have the same priority.
1966 * This scanning priority is essentially the inverse of IO cost.
1968 anon_prio = sc->swappiness;
1969 file_prio = 200 - anon_prio;
1972 * OK, so we have swap space and a fair amount of page cache
1973 * pages. We use the recently rotated / recently scanned
1974 * ratios to determine how valuable each cache is.
1976 * Because workloads change over time (and to avoid overflow)
1977 * we keep these statistics as a floating average, which ends
1978 * up weighing recent references more than old ones.
1980 * anon in [0], file in [1]
1982 spin_lock_irq(&zone->lru_lock);
1983 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1984 reclaim_stat->recent_scanned[0] /= 2;
1985 reclaim_stat->recent_rotated[0] /= 2;
1988 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1989 reclaim_stat->recent_scanned[1] /= 2;
1990 reclaim_stat->recent_rotated[1] /= 2;
1994 * The amount of pressure on anon vs file pages is inversely
1995 * proportional to the fraction of recently scanned pages on
1996 * each list that were recently referenced and in active use.
1998 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1999 ap /= reclaim_stat->recent_rotated[0] + 1;
2001 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2002 fp /= reclaim_stat->recent_rotated[1] + 1;
2003 spin_unlock_irq(&zone->lru_lock);
2007 denominator = ap + fp + 1;
2009 some_scanned = false;
2010 /* Only use force_scan on second pass. */
2011 for (pass = 0; !some_scanned && pass < 2; pass++) {
2012 for_each_evictable_lru(lru) {
2013 int file = is_file_lru(lru);
2017 size = get_lru_size(lruvec, lru);
2018 scan = size >> sc->priority;
2020 if (!scan && pass && force_scan)
2021 scan = min(size, SWAP_CLUSTER_MAX);
2023 switch (scan_balance) {
2025 /* Scan lists relative to size */
2029 * Scan types proportional to swappiness and
2030 * their relative recent reclaim efficiency.
2032 scan = div64_u64(scan * fraction[file],
2037 /* Scan one type exclusively */
2038 if ((scan_balance == SCAN_FILE) != file)
2042 /* Look ma, no brain */
2047 * Skip the second pass and don't force_scan,
2048 * if we found something to scan.
2050 some_scanned |= !!scan;
2056 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2058 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2060 unsigned long nr[NR_LRU_LISTS];
2061 unsigned long targets[NR_LRU_LISTS];
2062 unsigned long nr_to_scan;
2064 unsigned long nr_reclaimed = 0;
2065 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2066 struct blk_plug plug;
2069 get_scan_count(lruvec, sc, nr);
2071 /* Record the original scan target for proportional adjustments later */
2072 memcpy(targets, nr, sizeof(nr));
2075 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2076 * event that can occur when there is little memory pressure e.g.
2077 * multiple streaming readers/writers. Hence, we do not abort scanning
2078 * when the requested number of pages are reclaimed when scanning at
2079 * DEF_PRIORITY on the assumption that the fact we are direct
2080 * reclaiming implies that kswapd is not keeping up and it is best to
2081 * do a batch of work at once. For memcg reclaim one check is made to
2082 * abort proportional reclaim if either the file or anon lru has already
2083 * dropped to zero at the first pass.
2085 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2086 sc->priority == DEF_PRIORITY);
2088 blk_start_plug(&plug);
2089 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2090 nr[LRU_INACTIVE_FILE]) {
2091 unsigned long nr_anon, nr_file, percentage;
2092 unsigned long nr_scanned;
2094 for_each_evictable_lru(lru) {
2096 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2097 nr[lru] -= nr_to_scan;
2099 nr_reclaimed += shrink_list(lru, nr_to_scan,
2104 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2108 * For kswapd and memcg, reclaim at least the number of pages
2109 * requested. Ensure that the anon and file LRUs are scanned
2110 * proportionally what was requested by get_scan_count(). We
2111 * stop reclaiming one LRU and reduce the amount scanning
2112 * proportional to the original scan target.
2114 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2115 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2118 * It's just vindictive to attack the larger once the smaller
2119 * has gone to zero. And given the way we stop scanning the
2120 * smaller below, this makes sure that we only make one nudge
2121 * towards proportionality once we've got nr_to_reclaim.
2123 if (!nr_file || !nr_anon)
2126 if (nr_file > nr_anon) {
2127 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2128 targets[LRU_ACTIVE_ANON] + 1;
2130 percentage = nr_anon * 100 / scan_target;
2132 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2133 targets[LRU_ACTIVE_FILE] + 1;
2135 percentage = nr_file * 100 / scan_target;
2138 /* Stop scanning the smaller of the LRU */
2140 nr[lru + LRU_ACTIVE] = 0;
2143 * Recalculate the other LRU scan count based on its original
2144 * scan target and the percentage scanning already complete
2146 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2147 nr_scanned = targets[lru] - nr[lru];
2148 nr[lru] = targets[lru] * (100 - percentage) / 100;
2149 nr[lru] -= min(nr[lru], nr_scanned);
2152 nr_scanned = targets[lru] - nr[lru];
2153 nr[lru] = targets[lru] * (100 - percentage) / 100;
2154 nr[lru] -= min(nr[lru], nr_scanned);
2156 scan_adjusted = true;
2158 blk_finish_plug(&plug);
2159 sc->nr_reclaimed += nr_reclaimed;
2162 * Even if we did not try to evict anon pages at all, we want to
2163 * rebalance the anon lru active/inactive ratio.
2165 if (inactive_anon_is_low(lruvec))
2166 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2167 sc, LRU_ACTIVE_ANON);
2169 throttle_vm_writeout(sc->gfp_mask);
2172 /* Use reclaim/compaction for costly allocs or under memory pressure */
2173 static bool in_reclaim_compaction(struct scan_control *sc)
2175 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2176 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2177 sc->priority < DEF_PRIORITY - 2))
2184 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2185 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2186 * true if more pages should be reclaimed such that when the page allocator
2187 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2188 * It will give up earlier than that if there is difficulty reclaiming pages.
2190 static inline bool should_continue_reclaim(struct zone *zone,
2191 unsigned long nr_reclaimed,
2192 unsigned long nr_scanned,
2193 struct scan_control *sc)
2195 unsigned long pages_for_compaction;
2196 unsigned long inactive_lru_pages;
2198 /* If not in reclaim/compaction mode, stop */
2199 if (!in_reclaim_compaction(sc))
2202 /* Consider stopping depending on scan and reclaim activity */
2203 if (sc->gfp_mask & __GFP_REPEAT) {
2205 * For __GFP_REPEAT allocations, stop reclaiming if the
2206 * full LRU list has been scanned and we are still failing
2207 * to reclaim pages. This full LRU scan is potentially
2208 * expensive but a __GFP_REPEAT caller really wants to succeed
2210 if (!nr_reclaimed && !nr_scanned)
2214 * For non-__GFP_REPEAT allocations which can presumably
2215 * fail without consequence, stop if we failed to reclaim
2216 * any pages from the last SWAP_CLUSTER_MAX number of
2217 * pages that were scanned. This will return to the
2218 * caller faster at the risk reclaim/compaction and
2219 * the resulting allocation attempt fails
2226 * If we have not reclaimed enough pages for compaction and the
2227 * inactive lists are large enough, continue reclaiming
2229 pages_for_compaction = (2UL << sc->order);
2230 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2231 if (get_nr_swap_pages() > 0)
2232 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2233 if (sc->nr_reclaimed < pages_for_compaction &&
2234 inactive_lru_pages > pages_for_compaction)
2237 /* If compaction would go ahead or the allocation would succeed, stop */
2238 switch (compaction_suitable(zone, sc->order)) {
2239 case COMPACT_PARTIAL:
2240 case COMPACT_CONTINUE:
2247 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2249 unsigned long nr_reclaimed, nr_scanned;
2252 struct mem_cgroup *root = sc->target_mem_cgroup;
2253 struct mem_cgroup_reclaim_cookie reclaim = {
2255 .priority = sc->priority,
2257 struct mem_cgroup *memcg;
2259 nr_reclaimed = sc->nr_reclaimed;
2260 nr_scanned = sc->nr_scanned;
2262 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2264 struct lruvec *lruvec;
2266 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2268 sc->swappiness = mem_cgroup_swappiness(memcg);
2269 shrink_lruvec(lruvec, sc);
2272 * Direct reclaim and kswapd have to scan all memory
2273 * cgroups to fulfill the overall scan target for the
2276 * Limit reclaim, on the other hand, only cares about
2277 * nr_to_reclaim pages to be reclaimed and it will
2278 * retry with decreasing priority if one round over the
2279 * whole hierarchy is not sufficient.
2281 if (!global_reclaim(sc) &&
2282 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2283 mem_cgroup_iter_break(root, memcg);
2286 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2289 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2290 sc->nr_scanned - nr_scanned,
2291 sc->nr_reclaimed - nr_reclaimed);
2293 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2294 sc->nr_scanned - nr_scanned, sc));
2297 /* Returns true if compaction should go ahead for a high-order request */
2298 static inline bool compaction_ready(struct zone *zone, int order)
2300 unsigned long balance_gap, watermark;
2304 * Compaction takes time to run and there are potentially other
2305 * callers using the pages just freed. Continue reclaiming until
2306 * there is a buffer of free pages available to give compaction
2307 * a reasonable chance of completing and allocating the page
2309 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2310 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2311 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2312 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2315 * If compaction is deferred, reclaim up to a point where
2316 * compaction will have a chance of success when re-enabled
2318 if (compaction_deferred(zone, order))
2319 return watermark_ok;
2321 /* If compaction is not ready to start, keep reclaiming */
2322 if (!compaction_suitable(zone, order))
2325 return watermark_ok;
2329 * This is the direct reclaim path, for page-allocating processes. We only
2330 * try to reclaim pages from zones which will satisfy the caller's allocation
2333 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2335 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2337 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2338 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2339 * zone defense algorithm.
2341 * If a zone is deemed to be full of pinned pages then just give it a light
2342 * scan then give up on it.
2344 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2348 unsigned long nr_soft_reclaimed;
2349 unsigned long nr_soft_scanned;
2350 unsigned long lru_pages = 0;
2351 struct reclaim_state *reclaim_state = current->reclaim_state;
2353 struct shrink_control shrink = {
2354 .gfp_mask = sc->gfp_mask,
2356 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2359 * If the number of buffer_heads in the machine exceeds the maximum
2360 * allowed level, force direct reclaim to scan the highmem zone as
2361 * highmem pages could be pinning lowmem pages storing buffer_heads
2363 orig_mask = sc->gfp_mask;
2364 if (buffer_heads_over_limit)
2365 sc->gfp_mask |= __GFP_HIGHMEM;
2367 nodes_clear(shrink.nodes_to_scan);
2369 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2370 gfp_zone(sc->gfp_mask), sc->nodemask) {
2371 if (!populated_zone(zone))
2374 * Take care memory controller reclaiming has small influence
2377 if (global_reclaim(sc)) {
2378 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2381 lru_pages += zone_reclaimable_pages(zone);
2382 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2384 if (sc->priority != DEF_PRIORITY &&
2385 !zone_reclaimable(zone))
2386 continue; /* Let kswapd poll it */
2389 * If we already have plenty of memory free for
2390 * compaction in this zone, don't free any more.
2391 * Even though compaction is invoked for any
2392 * non-zero order, only frequent costly order
2393 * reclamation is disruptive enough to become a
2394 * noticeable problem, like transparent huge
2397 if (IS_ENABLED(CONFIG_COMPACTION) &&
2398 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2399 zonelist_zone_idx(z) <= requested_highidx &&
2400 compaction_ready(zone, sc->order)) {
2401 sc->compaction_ready = true;
2406 * This steals pages from memory cgroups over softlimit
2407 * and returns the number of reclaimed pages and
2408 * scanned pages. This works for global memory pressure
2409 * and balancing, not for a memcg's limit.
2411 nr_soft_scanned = 0;
2412 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2413 sc->order, sc->gfp_mask,
2415 sc->nr_reclaimed += nr_soft_reclaimed;
2416 sc->nr_scanned += nr_soft_scanned;
2417 /* need some check for avoid more shrink_zone() */
2420 shrink_zone(zone, sc);
2424 * Don't shrink slabs when reclaiming memory from over limit cgroups
2425 * but do shrink slab at least once when aborting reclaim for
2426 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2429 if (global_reclaim(sc)) {
2430 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2431 if (reclaim_state) {
2432 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2433 reclaim_state->reclaimed_slab = 0;
2438 * Restore to original mask to avoid the impact on the caller if we
2439 * promoted it to __GFP_HIGHMEM.
2441 sc->gfp_mask = orig_mask;
2444 /* All zones in zonelist are unreclaimable? */
2445 static bool all_unreclaimable(struct zonelist *zonelist,
2446 struct scan_control *sc)
2451 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2452 gfp_zone(sc->gfp_mask), sc->nodemask) {
2453 if (!populated_zone(zone))
2455 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2457 if (zone_reclaimable(zone))
2465 * This is the main entry point to direct page reclaim.
2467 * If a full scan of the inactive list fails to free enough memory then we
2468 * are "out of memory" and something needs to be killed.
2470 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2471 * high - the zone may be full of dirty or under-writeback pages, which this
2472 * caller can't do much about. We kick the writeback threads and take explicit
2473 * naps in the hope that some of these pages can be written. But if the
2474 * allocating task holds filesystem locks which prevent writeout this might not
2475 * work, and the allocation attempt will fail.
2477 * returns: 0, if no pages reclaimed
2478 * else, the number of pages reclaimed
2480 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2481 struct scan_control *sc)
2483 unsigned long total_scanned = 0;
2484 unsigned long writeback_threshold;
2486 delayacct_freepages_start();
2488 if (global_reclaim(sc))
2489 count_vm_event(ALLOCSTALL);
2492 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2495 shrink_zones(zonelist, sc);
2497 total_scanned += sc->nr_scanned;
2498 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2501 if (sc->compaction_ready)
2505 * If we're getting trouble reclaiming, start doing
2506 * writepage even in laptop mode.
2508 if (sc->priority < DEF_PRIORITY - 2)
2509 sc->may_writepage = 1;
2512 * Try to write back as many pages as we just scanned. This
2513 * tends to cause slow streaming writers to write data to the
2514 * disk smoothly, at the dirtying rate, which is nice. But
2515 * that's undesirable in laptop mode, where we *want* lumpy
2516 * writeout. So in laptop mode, write out the whole world.
2518 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2519 if (total_scanned > writeback_threshold) {
2520 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2521 WB_REASON_TRY_TO_FREE_PAGES);
2522 sc->may_writepage = 1;
2524 } while (--sc->priority >= 0);
2526 delayacct_freepages_end();
2528 if (sc->nr_reclaimed)
2529 return sc->nr_reclaimed;
2531 /* Aborted reclaim to try compaction? don't OOM, then */
2532 if (sc->compaction_ready)
2535 /* top priority shrink_zones still had more to do? don't OOM, then */
2536 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2542 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2545 unsigned long pfmemalloc_reserve = 0;
2546 unsigned long free_pages = 0;
2550 for (i = 0; i <= ZONE_NORMAL; i++) {
2551 zone = &pgdat->node_zones[i];
2552 if (!populated_zone(zone))
2555 pfmemalloc_reserve += min_wmark_pages(zone);
2556 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2559 /* If there are no reserves (unexpected config) then do not throttle */
2560 if (!pfmemalloc_reserve)
2563 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2565 /* kswapd must be awake if processes are being throttled */
2566 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2567 pgdat->classzone_idx = min(pgdat->classzone_idx,
2568 (enum zone_type)ZONE_NORMAL);
2569 wake_up_interruptible(&pgdat->kswapd_wait);
2576 * Throttle direct reclaimers if backing storage is backed by the network
2577 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2578 * depleted. kswapd will continue to make progress and wake the processes
2579 * when the low watermark is reached.
2581 * Returns true if a fatal signal was delivered during throttling. If this
2582 * happens, the page allocator should not consider triggering the OOM killer.
2584 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2585 nodemask_t *nodemask)
2589 pg_data_t *pgdat = NULL;
2592 * Kernel threads should not be throttled as they may be indirectly
2593 * responsible for cleaning pages necessary for reclaim to make forward
2594 * progress. kjournald for example may enter direct reclaim while
2595 * committing a transaction where throttling it could forcing other
2596 * processes to block on log_wait_commit().
2598 if (current->flags & PF_KTHREAD)
2602 * If a fatal signal is pending, this process should not throttle.
2603 * It should return quickly so it can exit and free its memory
2605 if (fatal_signal_pending(current))
2609 * Check if the pfmemalloc reserves are ok by finding the first node
2610 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2611 * GFP_KERNEL will be required for allocating network buffers when
2612 * swapping over the network so ZONE_HIGHMEM is unusable.
2614 * Throttling is based on the first usable node and throttled processes
2615 * wait on a queue until kswapd makes progress and wakes them. There
2616 * is an affinity then between processes waking up and where reclaim
2617 * progress has been made assuming the process wakes on the same node.
2618 * More importantly, processes running on remote nodes will not compete
2619 * for remote pfmemalloc reserves and processes on different nodes
2620 * should make reasonable progress.
2622 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2623 gfp_mask, nodemask) {
2624 if (zone_idx(zone) > ZONE_NORMAL)
2627 /* Throttle based on the first usable node */
2628 pgdat = zone->zone_pgdat;
2629 if (pfmemalloc_watermark_ok(pgdat))
2634 /* If no zone was usable by the allocation flags then do not throttle */
2638 /* Account for the throttling */
2639 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2642 * If the caller cannot enter the filesystem, it's possible that it
2643 * is due to the caller holding an FS lock or performing a journal
2644 * transaction in the case of a filesystem like ext[3|4]. In this case,
2645 * it is not safe to block on pfmemalloc_wait as kswapd could be
2646 * blocked waiting on the same lock. Instead, throttle for up to a
2647 * second before continuing.
2649 if (!(gfp_mask & __GFP_FS)) {
2650 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2651 pfmemalloc_watermark_ok(pgdat), HZ);
2656 /* Throttle until kswapd wakes the process */
2657 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2658 pfmemalloc_watermark_ok(pgdat));
2661 if (fatal_signal_pending(current))
2668 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2669 gfp_t gfp_mask, nodemask_t *nodemask)
2671 unsigned long nr_reclaimed;
2672 struct scan_control sc = {
2673 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2674 .may_writepage = !laptop_mode,
2675 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2679 .priority = DEF_PRIORITY,
2680 .target_mem_cgroup = NULL,
2681 .nodemask = nodemask,
2685 * Do not enter reclaim if fatal signal was delivered while throttled.
2686 * 1 is returned so that the page allocator does not OOM kill at this
2689 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2692 trace_mm_vmscan_direct_reclaim_begin(order,
2696 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2698 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2700 return nr_reclaimed;
2705 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2706 gfp_t gfp_mask, bool noswap,
2708 unsigned long *nr_scanned)
2710 struct scan_control sc = {
2712 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2713 .may_writepage = !laptop_mode,
2715 .may_swap = !noswap,
2718 .swappiness = mem_cgroup_swappiness(memcg),
2719 .target_mem_cgroup = memcg,
2721 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2723 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2724 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2726 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2731 * NOTE: Although we can get the priority field, using it
2732 * here is not a good idea, since it limits the pages we can scan.
2733 * if we don't reclaim here, the shrink_zone from balance_pgdat
2734 * will pick up pages from other mem cgroup's as well. We hack
2735 * the priority and make it zero.
2737 shrink_lruvec(lruvec, &sc);
2739 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2741 *nr_scanned = sc.nr_scanned;
2742 return sc.nr_reclaimed;
2745 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2749 struct zonelist *zonelist;
2750 unsigned long nr_reclaimed;
2752 struct scan_control sc = {
2753 .may_writepage = !laptop_mode,
2755 .may_swap = !noswap,
2756 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2758 .priority = DEF_PRIORITY,
2759 .target_mem_cgroup = memcg,
2760 .nodemask = NULL, /* we don't care the placement */
2761 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2762 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2766 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2767 * take care of from where we get pages. So the node where we start the
2768 * scan does not need to be the current node.
2770 nid = mem_cgroup_select_victim_node(memcg);
2772 zonelist = NODE_DATA(nid)->node_zonelists;
2774 trace_mm_vmscan_memcg_reclaim_begin(0,
2778 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2780 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2782 return nr_reclaimed;
2786 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2788 struct mem_cgroup *memcg;
2790 if (!total_swap_pages)
2793 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2795 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2797 if (inactive_anon_is_low(lruvec))
2798 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2799 sc, LRU_ACTIVE_ANON);
2801 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2805 static bool zone_balanced(struct zone *zone, int order,
2806 unsigned long balance_gap, int classzone_idx)
2808 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2809 balance_gap, classzone_idx, 0))
2812 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2813 !compaction_suitable(zone, order))
2820 * pgdat_balanced() is used when checking if a node is balanced.
2822 * For order-0, all zones must be balanced!
2824 * For high-order allocations only zones that meet watermarks and are in a
2825 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2826 * total of balanced pages must be at least 25% of the zones allowed by
2827 * classzone_idx for the node to be considered balanced. Forcing all zones to
2828 * be balanced for high orders can cause excessive reclaim when there are
2830 * The choice of 25% is due to
2831 * o a 16M DMA zone that is balanced will not balance a zone on any
2832 * reasonable sized machine
2833 * o On all other machines, the top zone must be at least a reasonable
2834 * percentage of the middle zones. For example, on 32-bit x86, highmem
2835 * would need to be at least 256M for it to be balance a whole node.
2836 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2837 * to balance a node on its own. These seemed like reasonable ratios.
2839 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2841 unsigned long managed_pages = 0;
2842 unsigned long balanced_pages = 0;
2845 /* Check the watermark levels */
2846 for (i = 0; i <= classzone_idx; i++) {
2847 struct zone *zone = pgdat->node_zones + i;
2849 if (!populated_zone(zone))
2852 managed_pages += zone->managed_pages;
2855 * A special case here:
2857 * balance_pgdat() skips over all_unreclaimable after
2858 * DEF_PRIORITY. Effectively, it considers them balanced so
2859 * they must be considered balanced here as well!
2861 if (!zone_reclaimable(zone)) {
2862 balanced_pages += zone->managed_pages;
2866 if (zone_balanced(zone, order, 0, i))
2867 balanced_pages += zone->managed_pages;
2873 return balanced_pages >= (managed_pages >> 2);
2879 * Prepare kswapd for sleeping. This verifies that there are no processes
2880 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2882 * Returns true if kswapd is ready to sleep
2884 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2887 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2892 * There is a potential race between when kswapd checks its watermarks
2893 * and a process gets throttled. There is also a potential race if
2894 * processes get throttled, kswapd wakes, a large process exits therby
2895 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2896 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2897 * so wake them now if necessary. If necessary, processes will wake
2898 * kswapd and get throttled again
2900 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2901 wake_up(&pgdat->pfmemalloc_wait);
2905 return pgdat_balanced(pgdat, order, classzone_idx);
2909 * kswapd shrinks the zone by the number of pages required to reach
2910 * the high watermark.
2912 * Returns true if kswapd scanned at least the requested number of pages to
2913 * reclaim or if the lack of progress was due to pages under writeback.
2914 * This is used to determine if the scanning priority needs to be raised.
2916 static bool kswapd_shrink_zone(struct zone *zone,
2918 struct scan_control *sc,
2919 unsigned long lru_pages,
2920 unsigned long *nr_attempted)
2922 int testorder = sc->order;
2923 unsigned long balance_gap;
2924 struct reclaim_state *reclaim_state = current->reclaim_state;
2925 struct shrink_control shrink = {
2926 .gfp_mask = sc->gfp_mask,
2928 bool lowmem_pressure;
2930 /* Reclaim above the high watermark. */
2931 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2934 * Kswapd reclaims only single pages with compaction enabled. Trying
2935 * too hard to reclaim until contiguous free pages have become
2936 * available can hurt performance by evicting too much useful data
2937 * from memory. Do not reclaim more than needed for compaction.
2939 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2940 compaction_suitable(zone, sc->order) !=
2945 * We put equal pressure on every zone, unless one zone has way too
2946 * many pages free already. The "too many pages" is defined as the
2947 * high wmark plus a "gap" where the gap is either the low
2948 * watermark or 1% of the zone, whichever is smaller.
2950 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2951 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2954 * If there is no low memory pressure or the zone is balanced then no
2955 * reclaim is necessary
2957 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2958 if (!lowmem_pressure && zone_balanced(zone, testorder,
2959 balance_gap, classzone_idx))
2962 shrink_zone(zone, sc);
2963 nodes_clear(shrink.nodes_to_scan);
2964 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2966 reclaim_state->reclaimed_slab = 0;
2967 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2968 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2970 /* Account for the number of pages attempted to reclaim */
2971 *nr_attempted += sc->nr_to_reclaim;
2973 zone_clear_flag(zone, ZONE_WRITEBACK);
2976 * If a zone reaches its high watermark, consider it to be no longer
2977 * congested. It's possible there are dirty pages backed by congested
2978 * BDIs but as pressure is relieved, speculatively avoid congestion
2981 if (zone_reclaimable(zone) &&
2982 zone_balanced(zone, testorder, 0, classzone_idx)) {
2983 zone_clear_flag(zone, ZONE_CONGESTED);
2984 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2987 return sc->nr_scanned >= sc->nr_to_reclaim;
2991 * For kswapd, balance_pgdat() will work across all this node's zones until
2992 * they are all at high_wmark_pages(zone).
2994 * Returns the final order kswapd was reclaiming at
2996 * There is special handling here for zones which are full of pinned pages.
2997 * This can happen if the pages are all mlocked, or if they are all used by
2998 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2999 * What we do is to detect the case where all pages in the zone have been
3000 * scanned twice and there has been zero successful reclaim. Mark the zone as
3001 * dead and from now on, only perform a short scan. Basically we're polling
3002 * the zone for when the problem goes away.
3004 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3005 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3006 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3007 * lower zones regardless of the number of free pages in the lower zones. This
3008 * interoperates with the page allocator fallback scheme to ensure that aging
3009 * of pages is balanced across the zones.
3011 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3015 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3016 unsigned long nr_soft_reclaimed;
3017 unsigned long nr_soft_scanned;
3018 struct scan_control sc = {
3019 .gfp_mask = GFP_KERNEL,
3020 .priority = DEF_PRIORITY,
3023 .may_writepage = !laptop_mode,
3025 .target_mem_cgroup = NULL,
3027 count_vm_event(PAGEOUTRUN);
3030 unsigned long lru_pages = 0;
3031 unsigned long nr_attempted = 0;
3032 bool raise_priority = true;
3033 bool pgdat_needs_compaction = (order > 0);
3035 sc.nr_reclaimed = 0;
3038 * Scan in the highmem->dma direction for the highest
3039 * zone which needs scanning
3041 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3042 struct zone *zone = pgdat->node_zones + i;
3044 if (!populated_zone(zone))
3047 if (sc.priority != DEF_PRIORITY &&
3048 !zone_reclaimable(zone))
3052 * Do some background aging of the anon list, to give
3053 * pages a chance to be referenced before reclaiming.
3055 age_active_anon(zone, &sc);
3058 * If the number of buffer_heads in the machine
3059 * exceeds the maximum allowed level and this node
3060 * has a highmem zone, force kswapd to reclaim from
3061 * it to relieve lowmem pressure.
3063 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3068 if (!zone_balanced(zone, order, 0, 0)) {
3073 * If balanced, clear the dirty and congested
3076 zone_clear_flag(zone, ZONE_CONGESTED);
3077 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3084 for (i = 0; i <= end_zone; i++) {
3085 struct zone *zone = pgdat->node_zones + i;
3087 if (!populated_zone(zone))
3090 lru_pages += zone_reclaimable_pages(zone);
3093 * If any zone is currently balanced then kswapd will
3094 * not call compaction as it is expected that the
3095 * necessary pages are already available.
3097 if (pgdat_needs_compaction &&
3098 zone_watermark_ok(zone, order,
3099 low_wmark_pages(zone),
3101 pgdat_needs_compaction = false;
3105 * If we're getting trouble reclaiming, start doing writepage
3106 * even in laptop mode.
3108 if (sc.priority < DEF_PRIORITY - 2)
3109 sc.may_writepage = 1;
3112 * Now scan the zone in the dma->highmem direction, stopping
3113 * at the last zone which needs scanning.
3115 * We do this because the page allocator works in the opposite
3116 * direction. This prevents the page allocator from allocating
3117 * pages behind kswapd's direction of progress, which would
3118 * cause too much scanning of the lower zones.
3120 for (i = 0; i <= end_zone; i++) {
3121 struct zone *zone = pgdat->node_zones + i;
3123 if (!populated_zone(zone))
3126 if (sc.priority != DEF_PRIORITY &&
3127 !zone_reclaimable(zone))
3132 nr_soft_scanned = 0;
3134 * Call soft limit reclaim before calling shrink_zone.
3136 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3139 sc.nr_reclaimed += nr_soft_reclaimed;
3142 * There should be no need to raise the scanning
3143 * priority if enough pages are already being scanned
3144 * that that high watermark would be met at 100%
3147 if (kswapd_shrink_zone(zone, end_zone, &sc,
3148 lru_pages, &nr_attempted))
3149 raise_priority = false;
3153 * If the low watermark is met there is no need for processes
3154 * to be throttled on pfmemalloc_wait as they should not be
3155 * able to safely make forward progress. Wake them
3157 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3158 pfmemalloc_watermark_ok(pgdat))
3159 wake_up(&pgdat->pfmemalloc_wait);
3162 * Fragmentation may mean that the system cannot be rebalanced
3163 * for high-order allocations in all zones. If twice the
3164 * allocation size has been reclaimed and the zones are still
3165 * not balanced then recheck the watermarks at order-0 to
3166 * prevent kswapd reclaiming excessively. Assume that a
3167 * process requested a high-order can direct reclaim/compact.
3169 if (order && sc.nr_reclaimed >= 2UL << order)
3170 order = sc.order = 0;
3172 /* Check if kswapd should be suspending */
3173 if (try_to_freeze() || kthread_should_stop())
3177 * Compact if necessary and kswapd is reclaiming at least the
3178 * high watermark number of pages as requsted
3180 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3181 compact_pgdat(pgdat, order);
3184 * Raise priority if scanning rate is too low or there was no
3185 * progress in reclaiming pages
3187 if (raise_priority || !sc.nr_reclaimed)
3189 } while (sc.priority >= 1 &&
3190 !pgdat_balanced(pgdat, order, *classzone_idx));
3194 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3195 * makes a decision on the order we were last reclaiming at. However,
3196 * if another caller entered the allocator slow path while kswapd
3197 * was awake, order will remain at the higher level
3199 *classzone_idx = end_zone;
3203 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3208 if (freezing(current) || kthread_should_stop())
3211 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3213 /* Try to sleep for a short interval */
3214 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3215 remaining = schedule_timeout(HZ/10);
3216 finish_wait(&pgdat->kswapd_wait, &wait);
3217 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3221 * After a short sleep, check if it was a premature sleep. If not, then
3222 * go fully to sleep until explicitly woken up.
3224 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3225 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3228 * vmstat counters are not perfectly accurate and the estimated
3229 * value for counters such as NR_FREE_PAGES can deviate from the
3230 * true value by nr_online_cpus * threshold. To avoid the zone
3231 * watermarks being breached while under pressure, we reduce the
3232 * per-cpu vmstat threshold while kswapd is awake and restore
3233 * them before going back to sleep.
3235 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3238 * Compaction records what page blocks it recently failed to
3239 * isolate pages from and skips them in the future scanning.
3240 * When kswapd is going to sleep, it is reasonable to assume
3241 * that pages and compaction may succeed so reset the cache.
3243 reset_isolation_suitable(pgdat);
3245 if (!kthread_should_stop())
3248 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3251 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3253 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3255 finish_wait(&pgdat->kswapd_wait, &wait);
3259 * The background pageout daemon, started as a kernel thread
3260 * from the init process.
3262 * This basically trickles out pages so that we have _some_
3263 * free memory available even if there is no other activity
3264 * that frees anything up. This is needed for things like routing
3265 * etc, where we otherwise might have all activity going on in
3266 * asynchronous contexts that cannot page things out.
3268 * If there are applications that are active memory-allocators
3269 * (most normal use), this basically shouldn't matter.
3271 static int kswapd(void *p)
3273 unsigned long order, new_order;
3274 unsigned balanced_order;
3275 int classzone_idx, new_classzone_idx;
3276 int balanced_classzone_idx;
3277 pg_data_t *pgdat = (pg_data_t*)p;
3278 struct task_struct *tsk = current;
3280 struct reclaim_state reclaim_state = {
3281 .reclaimed_slab = 0,
3283 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3285 lockdep_set_current_reclaim_state(GFP_KERNEL);
3287 if (!cpumask_empty(cpumask))
3288 set_cpus_allowed_ptr(tsk, cpumask);
3289 current->reclaim_state = &reclaim_state;
3292 * Tell the memory management that we're a "memory allocator",
3293 * and that if we need more memory we should get access to it
3294 * regardless (see "__alloc_pages()"). "kswapd" should
3295 * never get caught in the normal page freeing logic.
3297 * (Kswapd normally doesn't need memory anyway, but sometimes
3298 * you need a small amount of memory in order to be able to
3299 * page out something else, and this flag essentially protects
3300 * us from recursively trying to free more memory as we're
3301 * trying to free the first piece of memory in the first place).
3303 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3306 order = new_order = 0;
3308 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3309 balanced_classzone_idx = classzone_idx;
3314 * If the last balance_pgdat was unsuccessful it's unlikely a
3315 * new request of a similar or harder type will succeed soon
3316 * so consider going to sleep on the basis we reclaimed at
3318 if (balanced_classzone_idx >= new_classzone_idx &&
3319 balanced_order == new_order) {
3320 new_order = pgdat->kswapd_max_order;
3321 new_classzone_idx = pgdat->classzone_idx;
3322 pgdat->kswapd_max_order = 0;
3323 pgdat->classzone_idx = pgdat->nr_zones - 1;
3326 if (order < new_order || classzone_idx > new_classzone_idx) {
3328 * Don't sleep if someone wants a larger 'order'
3329 * allocation or has tigher zone constraints
3332 classzone_idx = new_classzone_idx;
3334 kswapd_try_to_sleep(pgdat, balanced_order,
3335 balanced_classzone_idx);
3336 order = pgdat->kswapd_max_order;
3337 classzone_idx = pgdat->classzone_idx;
3339 new_classzone_idx = classzone_idx;
3340 pgdat->kswapd_max_order = 0;
3341 pgdat->classzone_idx = pgdat->nr_zones - 1;
3344 ret = try_to_freeze();
3345 if (kthread_should_stop())
3349 * We can speed up thawing tasks if we don't call balance_pgdat
3350 * after returning from the refrigerator
3353 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3354 balanced_classzone_idx = classzone_idx;
3355 balanced_order = balance_pgdat(pgdat, order,
3356 &balanced_classzone_idx);
3360 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3361 current->reclaim_state = NULL;
3362 lockdep_clear_current_reclaim_state();
3368 * A zone is low on free memory, so wake its kswapd task to service it.
3370 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3374 if (!populated_zone(zone))
3377 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3379 pgdat = zone->zone_pgdat;
3380 if (pgdat->kswapd_max_order < order) {
3381 pgdat->kswapd_max_order = order;
3382 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3384 if (!waitqueue_active(&pgdat->kswapd_wait))
3386 if (zone_balanced(zone, order, 0, 0))
3389 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3390 wake_up_interruptible(&pgdat->kswapd_wait);
3393 #ifdef CONFIG_HIBERNATION
3395 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3398 * Rather than trying to age LRUs the aim is to preserve the overall
3399 * LRU order by reclaiming preferentially
3400 * inactive > active > active referenced > active mapped
3402 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3404 struct reclaim_state reclaim_state;
3405 struct scan_control sc = {
3406 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3410 .nr_to_reclaim = nr_to_reclaim,
3411 .hibernation_mode = 1,
3413 .priority = DEF_PRIORITY,
3415 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3416 struct task_struct *p = current;
3417 unsigned long nr_reclaimed;
3419 p->flags |= PF_MEMALLOC;
3420 lockdep_set_current_reclaim_state(sc.gfp_mask);
3421 reclaim_state.reclaimed_slab = 0;
3422 p->reclaim_state = &reclaim_state;
3424 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3426 p->reclaim_state = NULL;
3427 lockdep_clear_current_reclaim_state();
3428 p->flags &= ~PF_MEMALLOC;
3430 return nr_reclaimed;
3432 #endif /* CONFIG_HIBERNATION */
3434 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3435 not required for correctness. So if the last cpu in a node goes
3436 away, we get changed to run anywhere: as the first one comes back,
3437 restore their cpu bindings. */
3438 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3443 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3444 for_each_node_state(nid, N_MEMORY) {
3445 pg_data_t *pgdat = NODE_DATA(nid);
3446 const struct cpumask *mask;
3448 mask = cpumask_of_node(pgdat->node_id);
3450 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3451 /* One of our CPUs online: restore mask */
3452 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3459 * This kswapd start function will be called by init and node-hot-add.
3460 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3462 int kswapd_run(int nid)
3464 pg_data_t *pgdat = NODE_DATA(nid);
3470 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3471 if (IS_ERR(pgdat->kswapd)) {
3472 /* failure at boot is fatal */
3473 BUG_ON(system_state == SYSTEM_BOOTING);
3474 pr_err("Failed to start kswapd on node %d\n", nid);
3475 ret = PTR_ERR(pgdat->kswapd);
3476 pgdat->kswapd = NULL;
3482 * Called by memory hotplug when all memory in a node is offlined. Caller must
3483 * hold mem_hotplug_begin/end().
3485 void kswapd_stop(int nid)
3487 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3490 kthread_stop(kswapd);
3491 NODE_DATA(nid)->kswapd = NULL;
3495 static int __init kswapd_init(void)
3500 for_each_node_state(nid, N_MEMORY)
3502 hotcpu_notifier(cpu_callback, 0);
3506 module_init(kswapd_init)
3512 * If non-zero call zone_reclaim when the number of free pages falls below
3515 int zone_reclaim_mode __read_mostly;
3517 #define RECLAIM_OFF 0
3518 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3519 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3520 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3523 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3524 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3527 #define ZONE_RECLAIM_PRIORITY 4
3530 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3533 int sysctl_min_unmapped_ratio = 1;
3536 * If the number of slab pages in a zone grows beyond this percentage then
3537 * slab reclaim needs to occur.
3539 int sysctl_min_slab_ratio = 5;
3541 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3543 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3544 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3545 zone_page_state(zone, NR_ACTIVE_FILE);
3548 * It's possible for there to be more file mapped pages than
3549 * accounted for by the pages on the file LRU lists because
3550 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3552 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3555 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3556 static long zone_pagecache_reclaimable(struct zone *zone)
3558 long nr_pagecache_reclaimable;
3562 * If RECLAIM_SWAP is set, then all file pages are considered
3563 * potentially reclaimable. Otherwise, we have to worry about
3564 * pages like swapcache and zone_unmapped_file_pages() provides
3567 if (zone_reclaim_mode & RECLAIM_SWAP)
3568 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3570 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3572 /* If we can't clean pages, remove dirty pages from consideration */
3573 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3574 delta += zone_page_state(zone, NR_FILE_DIRTY);
3576 /* Watch for any possible underflows due to delta */
3577 if (unlikely(delta > nr_pagecache_reclaimable))
3578 delta = nr_pagecache_reclaimable;
3580 return nr_pagecache_reclaimable - delta;
3584 * Try to free up some pages from this zone through reclaim.
3586 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3588 /* Minimum pages needed in order to stay on node */
3589 const unsigned long nr_pages = 1 << order;
3590 struct task_struct *p = current;
3591 struct reclaim_state reclaim_state;
3592 struct scan_control sc = {
3593 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3594 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3596 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3597 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3599 .priority = ZONE_RECLAIM_PRIORITY,
3601 struct shrink_control shrink = {
3602 .gfp_mask = sc.gfp_mask,
3604 unsigned long nr_slab_pages0, nr_slab_pages1;
3608 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3609 * and we also need to be able to write out pages for RECLAIM_WRITE
3612 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3613 lockdep_set_current_reclaim_state(gfp_mask);
3614 reclaim_state.reclaimed_slab = 0;
3615 p->reclaim_state = &reclaim_state;
3617 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3619 * Free memory by calling shrink zone with increasing
3620 * priorities until we have enough memory freed.
3623 shrink_zone(zone, &sc);
3624 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3627 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3628 if (nr_slab_pages0 > zone->min_slab_pages) {
3630 * shrink_slab() does not currently allow us to determine how
3631 * many pages were freed in this zone. So we take the current
3632 * number of slab pages and shake the slab until it is reduced
3633 * by the same nr_pages that we used for reclaiming unmapped
3636 nodes_clear(shrink.nodes_to_scan);
3637 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3639 unsigned long lru_pages = zone_reclaimable_pages(zone);
3641 /* No reclaimable slab or very low memory pressure */
3642 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3645 /* Freed enough memory */
3646 nr_slab_pages1 = zone_page_state(zone,
3647 NR_SLAB_RECLAIMABLE);
3648 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3653 * Update nr_reclaimed by the number of slab pages we
3654 * reclaimed from this zone.
3656 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3657 if (nr_slab_pages1 < nr_slab_pages0)
3658 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3661 p->reclaim_state = NULL;
3662 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3663 lockdep_clear_current_reclaim_state();
3664 return sc.nr_reclaimed >= nr_pages;
3667 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3673 * Zone reclaim reclaims unmapped file backed pages and
3674 * slab pages if we are over the defined limits.
3676 * A small portion of unmapped file backed pages is needed for
3677 * file I/O otherwise pages read by file I/O will be immediately
3678 * thrown out if the zone is overallocated. So we do not reclaim
3679 * if less than a specified percentage of the zone is used by
3680 * unmapped file backed pages.
3682 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3683 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3684 return ZONE_RECLAIM_FULL;
3686 if (!zone_reclaimable(zone))
3687 return ZONE_RECLAIM_FULL;
3690 * Do not scan if the allocation should not be delayed.
3692 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3693 return ZONE_RECLAIM_NOSCAN;
3696 * Only run zone reclaim on the local zone or on zones that do not
3697 * have associated processors. This will favor the local processor
3698 * over remote processors and spread off node memory allocations
3699 * as wide as possible.
3701 node_id = zone_to_nid(zone);
3702 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3703 return ZONE_RECLAIM_NOSCAN;
3705 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3706 return ZONE_RECLAIM_NOSCAN;
3708 ret = __zone_reclaim(zone, gfp_mask, order);
3709 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3712 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3719 * page_evictable - test whether a page is evictable
3720 * @page: the page to test
3722 * Test whether page is evictable--i.e., should be placed on active/inactive
3723 * lists vs unevictable list.
3725 * Reasons page might not be evictable:
3726 * (1) page's mapping marked unevictable
3727 * (2) page is part of an mlocked VMA
3730 int page_evictable(struct page *page)
3732 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3737 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3738 * @pages: array of pages to check
3739 * @nr_pages: number of pages to check
3741 * Checks pages for evictability and moves them to the appropriate lru list.
3743 * This function is only used for SysV IPC SHM_UNLOCK.
3745 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3747 struct lruvec *lruvec;
3748 struct zone *zone = NULL;
3753 for (i = 0; i < nr_pages; i++) {
3754 struct page *page = pages[i];
3755 struct zone *pagezone;
3758 pagezone = page_zone(page);
3759 if (pagezone != zone) {
3761 spin_unlock_irq(&zone->lru_lock);
3763 spin_lock_irq(&zone->lru_lock);
3765 lruvec = mem_cgroup_page_lruvec(page, zone);
3767 if (!PageLRU(page) || !PageUnevictable(page))
3770 if (page_evictable(page)) {
3771 enum lru_list lru = page_lru_base_type(page);
3773 VM_BUG_ON_PAGE(PageActive(page), page);
3774 ClearPageUnevictable(page);
3775 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3776 add_page_to_lru_list(page, lruvec, lru);
3782 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3783 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3784 spin_unlock_irq(&zone->lru_lock);
3787 #endif /* CONFIG_SHMEM */
3789 static void warn_scan_unevictable_pages(void)
3791 printk_once(KERN_WARNING
3792 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3793 "disabled for lack of a legitimate use case. If you have "
3794 "one, please send an email to linux-mm@kvack.org.\n",
3799 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3800 * all nodes' unevictable lists for evictable pages
3802 unsigned long scan_unevictable_pages;
3804 int scan_unevictable_handler(struct ctl_table *table, int write,
3805 void __user *buffer,
3806 size_t *length, loff_t *ppos)
3808 warn_scan_unevictable_pages();
3809 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3810 scan_unevictable_pages = 0;
3816 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3817 * a specified node's per zone unevictable lists for evictable pages.
3820 static ssize_t read_scan_unevictable_node(struct device *dev,
3821 struct device_attribute *attr,
3824 warn_scan_unevictable_pages();
3825 return sprintf(buf, "0\n"); /* always zero; should fit... */
3828 static ssize_t write_scan_unevictable_node(struct device *dev,
3829 struct device_attribute *attr,
3830 const char *buf, size_t count)
3832 warn_scan_unevictable_pages();
3837 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3838 read_scan_unevictable_node,
3839 write_scan_unevictable_node);
3841 int scan_unevictable_register_node(struct node *node)
3843 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3846 void scan_unevictable_unregister_node(struct node *node)
3848 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);