4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29 #include <linux/mm_inline.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned;
61 /* Number of pages freed so far during a call to shrink_zones() */
62 unsigned long nr_reclaimed;
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim;
67 unsigned long hibernation_mode;
69 /* This context's GFP mask */
74 /* Can mapped pages be reclaimed? */
77 /* Can pages be swapped as part of reclaim? */
82 /* Scan (total_size >> priority) pages at once */
86 * The memory cgroup that hit its limit and as a result is the
87 * primary target of this reclaim invocation.
89 struct mem_cgroup *target_mem_cgroup;
92 * Nodemask of nodes allowed by the caller. If NULL, all nodes
98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100 #ifdef ARCH_HAS_PREFETCH
101 #define prefetch_prev_lru_page(_page, _base, _field) \
103 if ((_page)->lru.prev != _base) { \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
114 #ifdef ARCH_HAS_PREFETCHW
115 #define prefetchw_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
129 * From 0 .. 100. Higher means more swappy.
131 int vm_swappiness = 60;
132 unsigned long vm_total_pages; /* The total number of pages which the VM controls */
134 static LIST_HEAD(shrinker_list);
135 static DECLARE_RWSEM(shrinker_rwsem);
138 static bool global_reclaim(struct scan_control *sc)
140 return !sc->target_mem_cgroup;
143 static bool mem_cgroup_should_soft_reclaim(struct scan_control *sc)
145 return !mem_cgroup_disabled();
148 static bool global_reclaim(struct scan_control *sc)
153 static bool mem_cgroup_should_soft_reclaim(struct scan_control *sc)
159 unsigned long zone_reclaimable_pages(struct zone *zone)
163 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
164 zone_page_state(zone, NR_INACTIVE_FILE);
166 if (get_nr_swap_pages() > 0)
167 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
168 zone_page_state(zone, NR_INACTIVE_ANON);
173 bool zone_reclaimable(struct zone *zone)
175 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
178 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
180 if (!mem_cgroup_disabled())
181 return mem_cgroup_get_lru_size(lruvec, lru);
183 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
187 * Add a shrinker callback to be called from the vm
189 void register_shrinker(struct shrinker *shrinker)
191 atomic_long_set(&shrinker->nr_in_batch, 0);
192 down_write(&shrinker_rwsem);
193 list_add_tail(&shrinker->list, &shrinker_list);
194 up_write(&shrinker_rwsem);
196 EXPORT_SYMBOL(register_shrinker);
201 void unregister_shrinker(struct shrinker *shrinker)
203 down_write(&shrinker_rwsem);
204 list_del(&shrinker->list);
205 up_write(&shrinker_rwsem);
207 EXPORT_SYMBOL(unregister_shrinker);
209 static inline int do_shrinker_shrink(struct shrinker *shrinker,
210 struct shrink_control *sc,
211 unsigned long nr_to_scan)
213 sc->nr_to_scan = nr_to_scan;
214 return (*shrinker->shrink)(shrinker, sc);
217 #define SHRINK_BATCH 128
219 * Call the shrink functions to age shrinkable caches
221 * Here we assume it costs one seek to replace a lru page and that it also
222 * takes a seek to recreate a cache object. With this in mind we age equal
223 * percentages of the lru and ageable caches. This should balance the seeks
224 * generated by these structures.
226 * If the vm encountered mapped pages on the LRU it increase the pressure on
227 * slab to avoid swapping.
229 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
231 * `lru_pages' represents the number of on-LRU pages in all the zones which
232 * are eligible for the caller's allocation attempt. It is used for balancing
233 * slab reclaim versus page reclaim.
235 * Returns the number of slab objects which we shrunk.
237 unsigned long shrink_slab(struct shrink_control *shrink,
238 unsigned long nr_pages_scanned,
239 unsigned long lru_pages)
241 struct shrinker *shrinker;
242 unsigned long ret = 0;
244 if (nr_pages_scanned == 0)
245 nr_pages_scanned = SWAP_CLUSTER_MAX;
247 if (!down_read_trylock(&shrinker_rwsem)) {
248 /* Assume we'll be able to shrink next time */
253 list_for_each_entry(shrinker, &shrinker_list, list) {
254 unsigned long long delta;
260 long batch_size = shrinker->batch ? shrinker->batch
263 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
268 * copy the current shrinker scan count into a local variable
269 * and zero it so that other concurrent shrinker invocations
270 * don't also do this scanning work.
272 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
275 delta = (4 * nr_pages_scanned) / shrinker->seeks;
277 do_div(delta, lru_pages + 1);
279 if (total_scan < 0) {
280 printk(KERN_ERR "shrink_slab: %pF negative objects to "
282 shrinker->shrink, total_scan);
283 total_scan = max_pass;
287 * We need to avoid excessive windup on filesystem shrinkers
288 * due to large numbers of GFP_NOFS allocations causing the
289 * shrinkers to return -1 all the time. This results in a large
290 * nr being built up so when a shrink that can do some work
291 * comes along it empties the entire cache due to nr >>>
292 * max_pass. This is bad for sustaining a working set in
295 * Hence only allow the shrinker to scan the entire cache when
296 * a large delta change is calculated directly.
298 if (delta < max_pass / 4)
299 total_scan = min(total_scan, max_pass / 2);
302 * Avoid risking looping forever due to too large nr value:
303 * never try to free more than twice the estimate number of
306 if (total_scan > max_pass * 2)
307 total_scan = max_pass * 2;
309 trace_mm_shrink_slab_start(shrinker, shrink, nr,
310 nr_pages_scanned, lru_pages,
311 max_pass, delta, total_scan);
313 while (total_scan >= batch_size) {
316 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
317 shrink_ret = do_shrinker_shrink(shrinker, shrink,
319 if (shrink_ret == -1)
321 if (shrink_ret < nr_before)
322 ret += nr_before - shrink_ret;
323 count_vm_events(SLABS_SCANNED, batch_size);
324 total_scan -= batch_size;
330 * move the unused scan count back into the shrinker in a
331 * manner that handles concurrent updates. If we exhausted the
332 * scan, there is no need to do an update.
335 new_nr = atomic_long_add_return(total_scan,
336 &shrinker->nr_in_batch);
338 new_nr = atomic_long_read(&shrinker->nr_in_batch);
340 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
342 up_read(&shrinker_rwsem);
348 static inline int is_page_cache_freeable(struct page *page)
351 * A freeable page cache page is referenced only by the caller
352 * that isolated the page, the page cache radix tree and
353 * optional buffer heads at page->private.
355 return page_count(page) - page_has_private(page) == 2;
358 static int may_write_to_queue(struct backing_dev_info *bdi,
359 struct scan_control *sc)
361 if (current->flags & PF_SWAPWRITE)
363 if (!bdi_write_congested(bdi))
365 if (bdi == current->backing_dev_info)
371 * We detected a synchronous write error writing a page out. Probably
372 * -ENOSPC. We need to propagate that into the address_space for a subsequent
373 * fsync(), msync() or close().
375 * The tricky part is that after writepage we cannot touch the mapping: nothing
376 * prevents it from being freed up. But we have a ref on the page and once
377 * that page is locked, the mapping is pinned.
379 * We're allowed to run sleeping lock_page() here because we know the caller has
382 static void handle_write_error(struct address_space *mapping,
383 struct page *page, int error)
386 if (page_mapping(page) == mapping)
387 mapping_set_error(mapping, error);
391 /* possible outcome of pageout() */
393 /* failed to write page out, page is locked */
395 /* move page to the active list, page is locked */
397 /* page has been sent to the disk successfully, page is unlocked */
399 /* page is clean and locked */
404 * pageout is called by shrink_page_list() for each dirty page.
405 * Calls ->writepage().
407 static pageout_t pageout(struct page *page, struct address_space *mapping,
408 struct scan_control *sc)
411 * If the page is dirty, only perform writeback if that write
412 * will be non-blocking. To prevent this allocation from being
413 * stalled by pagecache activity. But note that there may be
414 * stalls if we need to run get_block(). We could test
415 * PagePrivate for that.
417 * If this process is currently in __generic_file_aio_write() against
418 * this page's queue, we can perform writeback even if that
421 * If the page is swapcache, write it back even if that would
422 * block, for some throttling. This happens by accident, because
423 * swap_backing_dev_info is bust: it doesn't reflect the
424 * congestion state of the swapdevs. Easy to fix, if needed.
426 if (!is_page_cache_freeable(page))
430 * Some data journaling orphaned pages can have
431 * page->mapping == NULL while being dirty with clean buffers.
433 if (page_has_private(page)) {
434 if (try_to_free_buffers(page)) {
435 ClearPageDirty(page);
436 printk("%s: orphaned page\n", __func__);
442 if (mapping->a_ops->writepage == NULL)
443 return PAGE_ACTIVATE;
444 if (!may_write_to_queue(mapping->backing_dev_info, sc))
447 if (clear_page_dirty_for_io(page)) {
449 struct writeback_control wbc = {
450 .sync_mode = WB_SYNC_NONE,
451 .nr_to_write = SWAP_CLUSTER_MAX,
453 .range_end = LLONG_MAX,
457 SetPageReclaim(page);
458 res = mapping->a_ops->writepage(page, &wbc);
460 handle_write_error(mapping, page, res);
461 if (res == AOP_WRITEPAGE_ACTIVATE) {
462 ClearPageReclaim(page);
463 return PAGE_ACTIVATE;
466 if (!PageWriteback(page)) {
467 /* synchronous write or broken a_ops? */
468 ClearPageReclaim(page);
470 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
471 inc_zone_page_state(page, NR_VMSCAN_WRITE);
479 * Same as remove_mapping, but if the page is removed from the mapping, it
480 * gets returned with a refcount of 0.
482 static int __remove_mapping(struct address_space *mapping, struct page *page)
484 BUG_ON(!PageLocked(page));
485 BUG_ON(mapping != page_mapping(page));
487 spin_lock_irq(&mapping->tree_lock);
489 * The non racy check for a busy page.
491 * Must be careful with the order of the tests. When someone has
492 * a ref to the page, it may be possible that they dirty it then
493 * drop the reference. So if PageDirty is tested before page_count
494 * here, then the following race may occur:
496 * get_user_pages(&page);
497 * [user mapping goes away]
499 * !PageDirty(page) [good]
500 * SetPageDirty(page);
502 * !page_count(page) [good, discard it]
504 * [oops, our write_to data is lost]
506 * Reversing the order of the tests ensures such a situation cannot
507 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
508 * load is not satisfied before that of page->_count.
510 * Note that if SetPageDirty is always performed via set_page_dirty,
511 * and thus under tree_lock, then this ordering is not required.
513 if (!page_freeze_refs(page, 2))
515 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
516 if (unlikely(PageDirty(page))) {
517 page_unfreeze_refs(page, 2);
521 if (PageSwapCache(page)) {
522 swp_entry_t swap = { .val = page_private(page) };
523 __delete_from_swap_cache(page);
524 spin_unlock_irq(&mapping->tree_lock);
525 swapcache_free(swap, page);
527 void (*freepage)(struct page *);
529 freepage = mapping->a_ops->freepage;
531 __delete_from_page_cache(page);
532 spin_unlock_irq(&mapping->tree_lock);
533 mem_cgroup_uncharge_cache_page(page);
535 if (freepage != NULL)
542 spin_unlock_irq(&mapping->tree_lock);
547 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
548 * someone else has a ref on the page, abort and return 0. If it was
549 * successfully detached, return 1. Assumes the caller has a single ref on
552 int remove_mapping(struct address_space *mapping, struct page *page)
554 if (__remove_mapping(mapping, page)) {
556 * Unfreezing the refcount with 1 rather than 2 effectively
557 * drops the pagecache ref for us without requiring another
560 page_unfreeze_refs(page, 1);
567 * putback_lru_page - put previously isolated page onto appropriate LRU list
568 * @page: page to be put back to appropriate lru list
570 * Add previously isolated @page to appropriate LRU list.
571 * Page may still be unevictable for other reasons.
573 * lru_lock must not be held, interrupts must be enabled.
575 void putback_lru_page(struct page *page)
578 int was_unevictable = PageUnevictable(page);
580 VM_BUG_ON(PageLRU(page));
583 ClearPageUnevictable(page);
585 if (page_evictable(page)) {
587 * For evictable pages, we can use the cache.
588 * In event of a race, worst case is we end up with an
589 * unevictable page on [in]active list.
590 * We know how to handle that.
592 is_unevictable = false;
596 * Put unevictable pages directly on zone's unevictable
599 is_unevictable = true;
600 add_page_to_unevictable_list(page);
602 * When racing with an mlock or AS_UNEVICTABLE clearing
603 * (page is unlocked) make sure that if the other thread
604 * does not observe our setting of PG_lru and fails
605 * isolation/check_move_unevictable_pages,
606 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
607 * the page back to the evictable list.
609 * The other side is TestClearPageMlocked() or shmem_lock().
615 * page's status can change while we move it among lru. If an evictable
616 * page is on unevictable list, it never be freed. To avoid that,
617 * check after we added it to the list, again.
619 if (is_unevictable && page_evictable(page)) {
620 if (!isolate_lru_page(page)) {
624 /* This means someone else dropped this page from LRU
625 * So, it will be freed or putback to LRU again. There is
626 * nothing to do here.
630 if (was_unevictable && !is_unevictable)
631 count_vm_event(UNEVICTABLE_PGRESCUED);
632 else if (!was_unevictable && is_unevictable)
633 count_vm_event(UNEVICTABLE_PGCULLED);
635 put_page(page); /* drop ref from isolate */
638 enum page_references {
640 PAGEREF_RECLAIM_CLEAN,
645 static enum page_references page_check_references(struct page *page,
646 struct scan_control *sc)
648 int referenced_ptes, referenced_page;
649 unsigned long vm_flags;
651 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
653 referenced_page = TestClearPageReferenced(page);
656 * Mlock lost the isolation race with us. Let try_to_unmap()
657 * move the page to the unevictable list.
659 if (vm_flags & VM_LOCKED)
660 return PAGEREF_RECLAIM;
662 if (referenced_ptes) {
663 if (PageSwapBacked(page))
664 return PAGEREF_ACTIVATE;
666 * All mapped pages start out with page table
667 * references from the instantiating fault, so we need
668 * to look twice if a mapped file page is used more
671 * Mark it and spare it for another trip around the
672 * inactive list. Another page table reference will
673 * lead to its activation.
675 * Note: the mark is set for activated pages as well
676 * so that recently deactivated but used pages are
679 SetPageReferenced(page);
681 if (referenced_page || referenced_ptes > 1)
682 return PAGEREF_ACTIVATE;
685 * Activate file-backed executable pages after first usage.
687 if (vm_flags & VM_EXEC)
688 return PAGEREF_ACTIVATE;
693 /* Reclaim if clean, defer dirty pages to writeback */
694 if (referenced_page && !PageSwapBacked(page))
695 return PAGEREF_RECLAIM_CLEAN;
697 return PAGEREF_RECLAIM;
700 /* Check if a page is dirty or under writeback */
701 static void page_check_dirty_writeback(struct page *page,
702 bool *dirty, bool *writeback)
704 struct address_space *mapping;
707 * Anonymous pages are not handled by flushers and must be written
708 * from reclaim context. Do not stall reclaim based on them
710 if (!page_is_file_cache(page)) {
716 /* By default assume that the page flags are accurate */
717 *dirty = PageDirty(page);
718 *writeback = PageWriteback(page);
720 /* Verify dirty/writeback state if the filesystem supports it */
721 if (!page_has_private(page))
724 mapping = page_mapping(page);
725 if (mapping && mapping->a_ops->is_dirty_writeback)
726 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
730 * shrink_page_list() returns the number of reclaimed pages
732 static unsigned long shrink_page_list(struct list_head *page_list,
734 struct scan_control *sc,
735 enum ttu_flags ttu_flags,
736 unsigned long *ret_nr_dirty,
737 unsigned long *ret_nr_unqueued_dirty,
738 unsigned long *ret_nr_congested,
739 unsigned long *ret_nr_writeback,
740 unsigned long *ret_nr_immediate,
743 LIST_HEAD(ret_pages);
744 LIST_HEAD(free_pages);
746 unsigned long nr_unqueued_dirty = 0;
747 unsigned long nr_dirty = 0;
748 unsigned long nr_congested = 0;
749 unsigned long nr_reclaimed = 0;
750 unsigned long nr_writeback = 0;
751 unsigned long nr_immediate = 0;
755 mem_cgroup_uncharge_start();
756 while (!list_empty(page_list)) {
757 struct address_space *mapping;
760 enum page_references references = PAGEREF_RECLAIM_CLEAN;
761 bool dirty, writeback;
765 page = lru_to_page(page_list);
766 list_del(&page->lru);
768 if (!trylock_page(page))
771 VM_BUG_ON(PageActive(page));
772 VM_BUG_ON(page_zone(page) != zone);
776 if (unlikely(!page_evictable(page)))
779 if (!sc->may_unmap && page_mapped(page))
782 /* Double the slab pressure for mapped and swapcache pages */
783 if (page_mapped(page) || PageSwapCache(page))
786 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
787 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
790 * The number of dirty pages determines if a zone is marked
791 * reclaim_congested which affects wait_iff_congested. kswapd
792 * will stall and start writing pages if the tail of the LRU
793 * is all dirty unqueued pages.
795 page_check_dirty_writeback(page, &dirty, &writeback);
796 if (dirty || writeback)
799 if (dirty && !writeback)
803 * Treat this page as congested if the underlying BDI is or if
804 * pages are cycling through the LRU so quickly that the
805 * pages marked for immediate reclaim are making it to the
806 * end of the LRU a second time.
808 mapping = page_mapping(page);
809 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
810 (writeback && PageReclaim(page)))
814 * If a page at the tail of the LRU is under writeback, there
815 * are three cases to consider.
817 * 1) If reclaim is encountering an excessive number of pages
818 * under writeback and this page is both under writeback and
819 * PageReclaim then it indicates that pages are being queued
820 * for IO but are being recycled through the LRU before the
821 * IO can complete. Waiting on the page itself risks an
822 * indefinite stall if it is impossible to writeback the
823 * page due to IO error or disconnected storage so instead
824 * note that the LRU is being scanned too quickly and the
825 * caller can stall after page list has been processed.
827 * 2) Global reclaim encounters a page, memcg encounters a
828 * page that is not marked for immediate reclaim or
829 * the caller does not have __GFP_IO. In this case mark
830 * the page for immediate reclaim and continue scanning.
832 * __GFP_IO is checked because a loop driver thread might
833 * enter reclaim, and deadlock if it waits on a page for
834 * which it is needed to do the write (loop masks off
835 * __GFP_IO|__GFP_FS for this reason); but more thought
836 * would probably show more reasons.
838 * Don't require __GFP_FS, since we're not going into the
839 * FS, just waiting on its writeback completion. Worryingly,
840 * ext4 gfs2 and xfs allocate pages with
841 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
842 * may_enter_fs here is liable to OOM on them.
844 * 3) memcg encounters a page that is not already marked
845 * PageReclaim. memcg does not have any dirty pages
846 * throttling so we could easily OOM just because too many
847 * pages are in writeback and there is nothing else to
848 * reclaim. Wait for the writeback to complete.
850 if (PageWriteback(page)) {
852 if (current_is_kswapd() &&
854 zone_is_reclaim_writeback(zone)) {
859 } else if (global_reclaim(sc) ||
860 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
862 * This is slightly racy - end_page_writeback()
863 * might have just cleared PageReclaim, then
864 * setting PageReclaim here end up interpreted
865 * as PageReadahead - but that does not matter
866 * enough to care. What we do want is for this
867 * page to have PageReclaim set next time memcg
868 * reclaim reaches the tests above, so it will
869 * then wait_on_page_writeback() to avoid OOM;
870 * and it's also appropriate in global reclaim.
872 SetPageReclaim(page);
879 wait_on_page_writeback(page);
884 references = page_check_references(page, sc);
886 switch (references) {
887 case PAGEREF_ACTIVATE:
888 goto activate_locked;
891 case PAGEREF_RECLAIM:
892 case PAGEREF_RECLAIM_CLEAN:
893 ; /* try to reclaim the page below */
897 * Anonymous process memory has backing store?
898 * Try to allocate it some swap space here.
900 if (PageAnon(page) && !PageSwapCache(page)) {
901 if (!(sc->gfp_mask & __GFP_IO))
903 if (!add_to_swap(page, page_list))
904 goto activate_locked;
907 /* Adding to swap updated mapping */
908 mapping = page_mapping(page);
912 * The page is mapped into the page tables of one or more
913 * processes. Try to unmap it here.
915 if (page_mapped(page) && mapping) {
916 switch (try_to_unmap(page, ttu_flags)) {
918 goto activate_locked;
924 ; /* try to free the page below */
928 if (PageDirty(page)) {
930 * Only kswapd can writeback filesystem pages to
931 * avoid risk of stack overflow but only writeback
932 * if many dirty pages have been encountered.
934 if (page_is_file_cache(page) &&
935 (!current_is_kswapd() ||
936 !zone_is_reclaim_dirty(zone))) {
938 * Immediately reclaim when written back.
939 * Similar in principal to deactivate_page()
940 * except we already have the page isolated
941 * and know it's dirty
943 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
944 SetPageReclaim(page);
949 if (references == PAGEREF_RECLAIM_CLEAN)
953 if (!sc->may_writepage)
956 /* Page is dirty, try to write it out here */
957 switch (pageout(page, mapping, sc)) {
961 goto activate_locked;
963 if (PageWriteback(page))
969 * A synchronous write - probably a ramdisk. Go
970 * ahead and try to reclaim the page.
972 if (!trylock_page(page))
974 if (PageDirty(page) || PageWriteback(page))
976 mapping = page_mapping(page);
978 ; /* try to free the page below */
983 * If the page has buffers, try to free the buffer mappings
984 * associated with this page. If we succeed we try to free
987 * We do this even if the page is PageDirty().
988 * try_to_release_page() does not perform I/O, but it is
989 * possible for a page to have PageDirty set, but it is actually
990 * clean (all its buffers are clean). This happens if the
991 * buffers were written out directly, with submit_bh(). ext3
992 * will do this, as well as the blockdev mapping.
993 * try_to_release_page() will discover that cleanness and will
994 * drop the buffers and mark the page clean - it can be freed.
996 * Rarely, pages can have buffers and no ->mapping. These are
997 * the pages which were not successfully invalidated in
998 * truncate_complete_page(). We try to drop those buffers here
999 * and if that worked, and the page is no longer mapped into
1000 * process address space (page_count == 1) it can be freed.
1001 * Otherwise, leave the page on the LRU so it is swappable.
1003 if (page_has_private(page)) {
1004 if (!try_to_release_page(page, sc->gfp_mask))
1005 goto activate_locked;
1006 if (!mapping && page_count(page) == 1) {
1008 if (put_page_testzero(page))
1012 * rare race with speculative reference.
1013 * the speculative reference will free
1014 * this page shortly, so we may
1015 * increment nr_reclaimed here (and
1016 * leave it off the LRU).
1024 if (!mapping || !__remove_mapping(mapping, page))
1028 * At this point, we have no other references and there is
1029 * no way to pick any more up (removed from LRU, removed
1030 * from pagecache). Can use non-atomic bitops now (and
1031 * we obviously don't have to worry about waking up a process
1032 * waiting on the page lock, because there are no references.
1034 __clear_page_locked(page);
1039 * Is there need to periodically free_page_list? It would
1040 * appear not as the counts should be low
1042 list_add(&page->lru, &free_pages);
1046 if (PageSwapCache(page))
1047 try_to_free_swap(page);
1049 putback_lru_page(page);
1053 /* Not a candidate for swapping, so reclaim swap space. */
1054 if (PageSwapCache(page) && vm_swap_full())
1055 try_to_free_swap(page);
1056 VM_BUG_ON(PageActive(page));
1057 SetPageActive(page);
1062 list_add(&page->lru, &ret_pages);
1063 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1066 free_hot_cold_page_list(&free_pages, 1);
1068 list_splice(&ret_pages, page_list);
1069 count_vm_events(PGACTIVATE, pgactivate);
1070 mem_cgroup_uncharge_end();
1071 *ret_nr_dirty += nr_dirty;
1072 *ret_nr_congested += nr_congested;
1073 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1074 *ret_nr_writeback += nr_writeback;
1075 *ret_nr_immediate += nr_immediate;
1076 return nr_reclaimed;
1079 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1080 struct list_head *page_list)
1082 struct scan_control sc = {
1083 .gfp_mask = GFP_KERNEL,
1084 .priority = DEF_PRIORITY,
1087 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1088 struct page *page, *next;
1089 LIST_HEAD(clean_pages);
1091 list_for_each_entry_safe(page, next, page_list, lru) {
1092 if (page_is_file_cache(page) && !PageDirty(page)) {
1093 ClearPageActive(page);
1094 list_move(&page->lru, &clean_pages);
1098 ret = shrink_page_list(&clean_pages, zone, &sc,
1099 TTU_UNMAP|TTU_IGNORE_ACCESS,
1100 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1101 list_splice(&clean_pages, page_list);
1102 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1107 * Attempt to remove the specified page from its LRU. Only take this page
1108 * if it is of the appropriate PageActive status. Pages which are being
1109 * freed elsewhere are also ignored.
1111 * page: page to consider
1112 * mode: one of the LRU isolation modes defined above
1114 * returns 0 on success, -ve errno on failure.
1116 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1120 /* Only take pages on the LRU. */
1124 /* Compaction should not handle unevictable pages but CMA can do so */
1125 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1131 * To minimise LRU disruption, the caller can indicate that it only
1132 * wants to isolate pages it will be able to operate on without
1133 * blocking - clean pages for the most part.
1135 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1136 * is used by reclaim when it is cannot write to backing storage
1138 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1139 * that it is possible to migrate without blocking
1141 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1142 /* All the caller can do on PageWriteback is block */
1143 if (PageWriteback(page))
1146 if (PageDirty(page)) {
1147 struct address_space *mapping;
1149 /* ISOLATE_CLEAN means only clean pages */
1150 if (mode & ISOLATE_CLEAN)
1154 * Only pages without mappings or that have a
1155 * ->migratepage callback are possible to migrate
1158 mapping = page_mapping(page);
1159 if (mapping && !mapping->a_ops->migratepage)
1164 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1167 if (likely(get_page_unless_zero(page))) {
1169 * Be careful not to clear PageLRU until after we're
1170 * sure the page is not being freed elsewhere -- the
1171 * page release code relies on it.
1181 * zone->lru_lock is heavily contended. Some of the functions that
1182 * shrink the lists perform better by taking out a batch of pages
1183 * and working on them outside the LRU lock.
1185 * For pagecache intensive workloads, this function is the hottest
1186 * spot in the kernel (apart from copy_*_user functions).
1188 * Appropriate locks must be held before calling this function.
1190 * @nr_to_scan: The number of pages to look through on the list.
1191 * @lruvec: The LRU vector to pull pages from.
1192 * @dst: The temp list to put pages on to.
1193 * @nr_scanned: The number of pages that were scanned.
1194 * @sc: The scan_control struct for this reclaim session
1195 * @mode: One of the LRU isolation modes
1196 * @lru: LRU list id for isolating
1198 * returns how many pages were moved onto *@dst.
1200 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1201 struct lruvec *lruvec, struct list_head *dst,
1202 unsigned long *nr_scanned, struct scan_control *sc,
1203 isolate_mode_t mode, enum lru_list lru)
1205 struct list_head *src = &lruvec->lists[lru];
1206 unsigned long nr_taken = 0;
1209 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1213 page = lru_to_page(src);
1214 prefetchw_prev_lru_page(page, src, flags);
1216 VM_BUG_ON(!PageLRU(page));
1218 switch (__isolate_lru_page(page, mode)) {
1220 nr_pages = hpage_nr_pages(page);
1221 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1222 list_move(&page->lru, dst);
1223 nr_taken += nr_pages;
1227 /* else it is being freed elsewhere */
1228 list_move(&page->lru, src);
1237 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1238 nr_taken, mode, is_file_lru(lru));
1243 * isolate_lru_page - tries to isolate a page from its LRU list
1244 * @page: page to isolate from its LRU list
1246 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1247 * vmstat statistic corresponding to whatever LRU list the page was on.
1249 * Returns 0 if the page was removed from an LRU list.
1250 * Returns -EBUSY if the page was not on an LRU list.
1252 * The returned page will have PageLRU() cleared. If it was found on
1253 * the active list, it will have PageActive set. If it was found on
1254 * the unevictable list, it will have the PageUnevictable bit set. That flag
1255 * may need to be cleared by the caller before letting the page go.
1257 * The vmstat statistic corresponding to the list on which the page was
1258 * found will be decremented.
1261 * (1) Must be called with an elevated refcount on the page. This is a
1262 * fundamentnal difference from isolate_lru_pages (which is called
1263 * without a stable reference).
1264 * (2) the lru_lock must not be held.
1265 * (3) interrupts must be enabled.
1267 int isolate_lru_page(struct page *page)
1271 VM_BUG_ON(!page_count(page));
1273 if (PageLRU(page)) {
1274 struct zone *zone = page_zone(page);
1275 struct lruvec *lruvec;
1277 spin_lock_irq(&zone->lru_lock);
1278 lruvec = mem_cgroup_page_lruvec(page, zone);
1279 if (PageLRU(page)) {
1280 int lru = page_lru(page);
1283 del_page_from_lru_list(page, lruvec, lru);
1286 spin_unlock_irq(&zone->lru_lock);
1292 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1293 * then get resheduled. When there are massive number of tasks doing page
1294 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1295 * the LRU list will go small and be scanned faster than necessary, leading to
1296 * unnecessary swapping, thrashing and OOM.
1298 static int too_many_isolated(struct zone *zone, int file,
1299 struct scan_control *sc)
1301 unsigned long inactive, isolated;
1303 if (current_is_kswapd())
1306 if (!global_reclaim(sc))
1310 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1311 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1313 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1314 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1318 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1319 * won't get blocked by normal direct-reclaimers, forming a circular
1322 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1325 return isolated > inactive;
1328 static noinline_for_stack void
1329 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1331 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1332 struct zone *zone = lruvec_zone(lruvec);
1333 LIST_HEAD(pages_to_free);
1336 * Put back any unfreeable pages.
1338 while (!list_empty(page_list)) {
1339 struct page *page = lru_to_page(page_list);
1342 VM_BUG_ON(PageLRU(page));
1343 list_del(&page->lru);
1344 if (unlikely(!page_evictable(page))) {
1345 spin_unlock_irq(&zone->lru_lock);
1346 putback_lru_page(page);
1347 spin_lock_irq(&zone->lru_lock);
1351 lruvec = mem_cgroup_page_lruvec(page, zone);
1354 lru = page_lru(page);
1355 add_page_to_lru_list(page, lruvec, lru);
1357 if (is_active_lru(lru)) {
1358 int file = is_file_lru(lru);
1359 int numpages = hpage_nr_pages(page);
1360 reclaim_stat->recent_rotated[file] += numpages;
1362 if (put_page_testzero(page)) {
1363 __ClearPageLRU(page);
1364 __ClearPageActive(page);
1365 del_page_from_lru_list(page, lruvec, lru);
1367 if (unlikely(PageCompound(page))) {
1368 spin_unlock_irq(&zone->lru_lock);
1369 (*get_compound_page_dtor(page))(page);
1370 spin_lock_irq(&zone->lru_lock);
1372 list_add(&page->lru, &pages_to_free);
1377 * To save our caller's stack, now use input list for pages to free.
1379 list_splice(&pages_to_free, page_list);
1383 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1384 * of reclaimed pages
1386 static noinline_for_stack unsigned long
1387 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1388 struct scan_control *sc, enum lru_list lru)
1390 LIST_HEAD(page_list);
1391 unsigned long nr_scanned;
1392 unsigned long nr_reclaimed = 0;
1393 unsigned long nr_taken;
1394 unsigned long nr_dirty = 0;
1395 unsigned long nr_congested = 0;
1396 unsigned long nr_unqueued_dirty = 0;
1397 unsigned long nr_writeback = 0;
1398 unsigned long nr_immediate = 0;
1399 isolate_mode_t isolate_mode = 0;
1400 int file = is_file_lru(lru);
1401 struct zone *zone = lruvec_zone(lruvec);
1402 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1404 while (unlikely(too_many_isolated(zone, file, sc))) {
1405 congestion_wait(BLK_RW_ASYNC, HZ/10);
1407 /* We are about to die and free our memory. Return now. */
1408 if (fatal_signal_pending(current))
1409 return SWAP_CLUSTER_MAX;
1415 isolate_mode |= ISOLATE_UNMAPPED;
1416 if (!sc->may_writepage)
1417 isolate_mode |= ISOLATE_CLEAN;
1419 spin_lock_irq(&zone->lru_lock);
1421 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1422 &nr_scanned, sc, isolate_mode, lru);
1424 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1425 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1427 if (global_reclaim(sc)) {
1428 zone->pages_scanned += nr_scanned;
1429 if (current_is_kswapd())
1430 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1432 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1434 spin_unlock_irq(&zone->lru_lock);
1439 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1440 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1441 &nr_writeback, &nr_immediate,
1444 spin_lock_irq(&zone->lru_lock);
1446 reclaim_stat->recent_scanned[file] += nr_taken;
1448 if (global_reclaim(sc)) {
1449 if (current_is_kswapd())
1450 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1453 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1457 putback_inactive_pages(lruvec, &page_list);
1459 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1461 spin_unlock_irq(&zone->lru_lock);
1463 free_hot_cold_page_list(&page_list, 1);
1466 * If reclaim is isolating dirty pages under writeback, it implies
1467 * that the long-lived page allocation rate is exceeding the page
1468 * laundering rate. Either the global limits are not being effective
1469 * at throttling processes due to the page distribution throughout
1470 * zones or there is heavy usage of a slow backing device. The
1471 * only option is to throttle from reclaim context which is not ideal
1472 * as there is no guarantee the dirtying process is throttled in the
1473 * same way balance_dirty_pages() manages.
1475 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1476 * of pages under pages flagged for immediate reclaim and stall if any
1477 * are encountered in the nr_immediate check below.
1479 if (nr_writeback && nr_writeback == nr_taken)
1480 zone_set_flag(zone, ZONE_WRITEBACK);
1483 * memcg will stall in page writeback so only consider forcibly
1484 * stalling for global reclaim
1486 if (global_reclaim(sc)) {
1488 * Tag a zone as congested if all the dirty pages scanned were
1489 * backed by a congested BDI and wait_iff_congested will stall.
1491 if (nr_dirty && nr_dirty == nr_congested)
1492 zone_set_flag(zone, ZONE_CONGESTED);
1495 * If dirty pages are scanned that are not queued for IO, it
1496 * implies that flushers are not keeping up. In this case, flag
1497 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1498 * pages from reclaim context. It will forcibly stall in the
1501 if (nr_unqueued_dirty == nr_taken)
1502 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1505 * In addition, if kswapd scans pages marked marked for
1506 * immediate reclaim and under writeback (nr_immediate), it
1507 * implies that pages are cycling through the LRU faster than
1508 * they are written so also forcibly stall.
1510 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1511 congestion_wait(BLK_RW_ASYNC, HZ/10);
1515 * Stall direct reclaim for IO completions if underlying BDIs or zone
1516 * is congested. Allow kswapd to continue until it starts encountering
1517 * unqueued dirty pages or cycling through the LRU too quickly.
1519 if (!sc->hibernation_mode && !current_is_kswapd())
1520 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1522 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1524 nr_scanned, nr_reclaimed,
1526 trace_shrink_flags(file));
1527 return nr_reclaimed;
1531 * This moves pages from the active list to the inactive list.
1533 * We move them the other way if the page is referenced by one or more
1534 * processes, from rmap.
1536 * If the pages are mostly unmapped, the processing is fast and it is
1537 * appropriate to hold zone->lru_lock across the whole operation. But if
1538 * the pages are mapped, the processing is slow (page_referenced()) so we
1539 * should drop zone->lru_lock around each page. It's impossible to balance
1540 * this, so instead we remove the pages from the LRU while processing them.
1541 * It is safe to rely on PG_active against the non-LRU pages in here because
1542 * nobody will play with that bit on a non-LRU page.
1544 * The downside is that we have to touch page->_count against each page.
1545 * But we had to alter page->flags anyway.
1548 static void move_active_pages_to_lru(struct lruvec *lruvec,
1549 struct list_head *list,
1550 struct list_head *pages_to_free,
1553 struct zone *zone = lruvec_zone(lruvec);
1554 unsigned long pgmoved = 0;
1558 while (!list_empty(list)) {
1559 page = lru_to_page(list);
1560 lruvec = mem_cgroup_page_lruvec(page, zone);
1562 VM_BUG_ON(PageLRU(page));
1565 nr_pages = hpage_nr_pages(page);
1566 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1567 list_move(&page->lru, &lruvec->lists[lru]);
1568 pgmoved += nr_pages;
1570 if (put_page_testzero(page)) {
1571 __ClearPageLRU(page);
1572 __ClearPageActive(page);
1573 del_page_from_lru_list(page, lruvec, lru);
1575 if (unlikely(PageCompound(page))) {
1576 spin_unlock_irq(&zone->lru_lock);
1577 (*get_compound_page_dtor(page))(page);
1578 spin_lock_irq(&zone->lru_lock);
1580 list_add(&page->lru, pages_to_free);
1583 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1584 if (!is_active_lru(lru))
1585 __count_vm_events(PGDEACTIVATE, pgmoved);
1588 static void shrink_active_list(unsigned long nr_to_scan,
1589 struct lruvec *lruvec,
1590 struct scan_control *sc,
1593 unsigned long nr_taken;
1594 unsigned long nr_scanned;
1595 unsigned long vm_flags;
1596 LIST_HEAD(l_hold); /* The pages which were snipped off */
1597 LIST_HEAD(l_active);
1598 LIST_HEAD(l_inactive);
1600 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1601 unsigned long nr_rotated = 0;
1602 isolate_mode_t isolate_mode = 0;
1603 int file = is_file_lru(lru);
1604 struct zone *zone = lruvec_zone(lruvec);
1609 isolate_mode |= ISOLATE_UNMAPPED;
1610 if (!sc->may_writepage)
1611 isolate_mode |= ISOLATE_CLEAN;
1613 spin_lock_irq(&zone->lru_lock);
1615 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1616 &nr_scanned, sc, isolate_mode, lru);
1617 if (global_reclaim(sc))
1618 zone->pages_scanned += nr_scanned;
1620 reclaim_stat->recent_scanned[file] += nr_taken;
1622 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1623 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1624 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1625 spin_unlock_irq(&zone->lru_lock);
1627 while (!list_empty(&l_hold)) {
1629 page = lru_to_page(&l_hold);
1630 list_del(&page->lru);
1632 if (unlikely(!page_evictable(page))) {
1633 putback_lru_page(page);
1637 if (unlikely(buffer_heads_over_limit)) {
1638 if (page_has_private(page) && trylock_page(page)) {
1639 if (page_has_private(page))
1640 try_to_release_page(page, 0);
1645 if (page_referenced(page, 0, sc->target_mem_cgroup,
1647 nr_rotated += hpage_nr_pages(page);
1649 * Identify referenced, file-backed active pages and
1650 * give them one more trip around the active list. So
1651 * that executable code get better chances to stay in
1652 * memory under moderate memory pressure. Anon pages
1653 * are not likely to be evicted by use-once streaming
1654 * IO, plus JVM can create lots of anon VM_EXEC pages,
1655 * so we ignore them here.
1657 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1658 list_add(&page->lru, &l_active);
1663 ClearPageActive(page); /* we are de-activating */
1664 list_add(&page->lru, &l_inactive);
1668 * Move pages back to the lru list.
1670 spin_lock_irq(&zone->lru_lock);
1672 * Count referenced pages from currently used mappings as rotated,
1673 * even though only some of them are actually re-activated. This
1674 * helps balance scan pressure between file and anonymous pages in
1677 reclaim_stat->recent_rotated[file] += nr_rotated;
1679 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1680 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1681 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1682 spin_unlock_irq(&zone->lru_lock);
1684 free_hot_cold_page_list(&l_hold, 1);
1688 static int inactive_anon_is_low_global(struct zone *zone)
1690 unsigned long active, inactive;
1692 active = zone_page_state(zone, NR_ACTIVE_ANON);
1693 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1695 if (inactive * zone->inactive_ratio < active)
1702 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1703 * @lruvec: LRU vector to check
1705 * Returns true if the zone does not have enough inactive anon pages,
1706 * meaning some active anon pages need to be deactivated.
1708 static int inactive_anon_is_low(struct lruvec *lruvec)
1711 * If we don't have swap space, anonymous page deactivation
1714 if (!total_swap_pages)
1717 if (!mem_cgroup_disabled())
1718 return mem_cgroup_inactive_anon_is_low(lruvec);
1720 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1723 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1730 * inactive_file_is_low - check if file pages need to be deactivated
1731 * @lruvec: LRU vector to check
1733 * When the system is doing streaming IO, memory pressure here
1734 * ensures that active file pages get deactivated, until more
1735 * than half of the file pages are on the inactive list.
1737 * Once we get to that situation, protect the system's working
1738 * set from being evicted by disabling active file page aging.
1740 * This uses a different ratio than the anonymous pages, because
1741 * the page cache uses a use-once replacement algorithm.
1743 static int inactive_file_is_low(struct lruvec *lruvec)
1745 unsigned long inactive;
1746 unsigned long active;
1748 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1749 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1751 return active > inactive;
1754 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1756 if (is_file_lru(lru))
1757 return inactive_file_is_low(lruvec);
1759 return inactive_anon_is_low(lruvec);
1762 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1763 struct lruvec *lruvec, struct scan_control *sc)
1765 if (is_active_lru(lru)) {
1766 if (inactive_list_is_low(lruvec, lru))
1767 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1771 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1774 static int vmscan_swappiness(struct scan_control *sc)
1776 if (global_reclaim(sc))
1777 return vm_swappiness;
1778 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1789 * Determine how aggressively the anon and file LRU lists should be
1790 * scanned. The relative value of each set of LRU lists is determined
1791 * by looking at the fraction of the pages scanned we did rotate back
1792 * onto the active list instead of evict.
1794 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1795 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1797 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1800 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1802 u64 denominator = 0; /* gcc */
1803 struct zone *zone = lruvec_zone(lruvec);
1804 unsigned long anon_prio, file_prio;
1805 enum scan_balance scan_balance;
1806 unsigned long anon, file, free;
1807 bool force_scan = false;
1808 unsigned long ap, fp;
1812 * If the zone or memcg is small, nr[l] can be 0. This
1813 * results in no scanning on this priority and a potential
1814 * priority drop. Global direct reclaim can go to the next
1815 * zone and tends to have no problems. Global kswapd is for
1816 * zone balancing and it needs to scan a minimum amount. When
1817 * reclaiming for a memcg, a priority drop can cause high
1818 * latencies, so it's better to scan a minimum amount there as
1821 if (current_is_kswapd() && !zone_reclaimable(zone))
1823 if (!global_reclaim(sc))
1826 /* If we have no swap space, do not bother scanning anon pages. */
1827 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1828 scan_balance = SCAN_FILE;
1833 * Global reclaim will swap to prevent OOM even with no
1834 * swappiness, but memcg users want to use this knob to
1835 * disable swapping for individual groups completely when
1836 * using the memory controller's swap limit feature would be
1839 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1840 scan_balance = SCAN_FILE;
1845 * Do not apply any pressure balancing cleverness when the
1846 * system is close to OOM, scan both anon and file equally
1847 * (unless the swappiness setting disagrees with swapping).
1849 if (!sc->priority && vmscan_swappiness(sc)) {
1850 scan_balance = SCAN_EQUAL;
1854 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1855 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1856 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1857 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1860 * If it's foreseeable that reclaiming the file cache won't be
1861 * enough to get the zone back into a desirable shape, we have
1862 * to swap. Better start now and leave the - probably heavily
1863 * thrashing - remaining file pages alone.
1865 if (global_reclaim(sc)) {
1866 free = zone_page_state(zone, NR_FREE_PAGES);
1867 if (unlikely(file + free <= high_wmark_pages(zone))) {
1868 scan_balance = SCAN_ANON;
1874 * There is enough inactive page cache, do not reclaim
1875 * anything from the anonymous working set right now.
1877 if (!inactive_file_is_low(lruvec)) {
1878 scan_balance = SCAN_FILE;
1882 scan_balance = SCAN_FRACT;
1885 * With swappiness at 100, anonymous and file have the same priority.
1886 * This scanning priority is essentially the inverse of IO cost.
1888 anon_prio = vmscan_swappiness(sc);
1889 file_prio = 200 - anon_prio;
1892 * OK, so we have swap space and a fair amount of page cache
1893 * pages. We use the recently rotated / recently scanned
1894 * ratios to determine how valuable each cache is.
1896 * Because workloads change over time (and to avoid overflow)
1897 * we keep these statistics as a floating average, which ends
1898 * up weighing recent references more than old ones.
1900 * anon in [0], file in [1]
1902 spin_lock_irq(&zone->lru_lock);
1903 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1904 reclaim_stat->recent_scanned[0] /= 2;
1905 reclaim_stat->recent_rotated[0] /= 2;
1908 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1909 reclaim_stat->recent_scanned[1] /= 2;
1910 reclaim_stat->recent_rotated[1] /= 2;
1914 * The amount of pressure on anon vs file pages is inversely
1915 * proportional to the fraction of recently scanned pages on
1916 * each list that were recently referenced and in active use.
1918 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1919 ap /= reclaim_stat->recent_rotated[0] + 1;
1921 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1922 fp /= reclaim_stat->recent_rotated[1] + 1;
1923 spin_unlock_irq(&zone->lru_lock);
1927 denominator = ap + fp + 1;
1929 for_each_evictable_lru(lru) {
1930 int file = is_file_lru(lru);
1934 size = get_lru_size(lruvec, lru);
1935 scan = size >> sc->priority;
1937 if (!scan && force_scan)
1938 scan = min(size, SWAP_CLUSTER_MAX);
1940 switch (scan_balance) {
1942 /* Scan lists relative to size */
1946 * Scan types proportional to swappiness and
1947 * their relative recent reclaim efficiency.
1949 scan = div64_u64(scan * fraction[file], denominator);
1953 /* Scan one type exclusively */
1954 if ((scan_balance == SCAN_FILE) != file)
1958 /* Look ma, no brain */
1966 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1968 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1970 unsigned long nr[NR_LRU_LISTS];
1971 unsigned long targets[NR_LRU_LISTS];
1972 unsigned long nr_to_scan;
1974 unsigned long nr_reclaimed = 0;
1975 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1976 struct blk_plug plug;
1977 bool scan_adjusted = false;
1979 get_scan_count(lruvec, sc, nr);
1981 /* Record the original scan target for proportional adjustments later */
1982 memcpy(targets, nr, sizeof(nr));
1984 blk_start_plug(&plug);
1985 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1986 nr[LRU_INACTIVE_FILE]) {
1987 unsigned long nr_anon, nr_file, percentage;
1988 unsigned long nr_scanned;
1990 for_each_evictable_lru(lru) {
1992 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1993 nr[lru] -= nr_to_scan;
1995 nr_reclaimed += shrink_list(lru, nr_to_scan,
2000 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2004 * For global direct reclaim, reclaim only the number of pages
2005 * requested. Less care is taken to scan proportionally as it
2006 * is more important to minimise direct reclaim stall latency
2007 * than it is to properly age the LRU lists.
2009 if (global_reclaim(sc) && !current_is_kswapd())
2013 * For kswapd and memcg, reclaim at least the number of pages
2014 * requested. Ensure that the anon and file LRUs shrink
2015 * proportionally what was requested by get_scan_count(). We
2016 * stop reclaiming one LRU and reduce the amount scanning
2017 * proportional to the original scan target.
2019 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2020 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2022 if (nr_file > nr_anon) {
2023 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2024 targets[LRU_ACTIVE_ANON] + 1;
2026 percentage = nr_anon * 100 / scan_target;
2028 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2029 targets[LRU_ACTIVE_FILE] + 1;
2031 percentage = nr_file * 100 / scan_target;
2034 /* Stop scanning the smaller of the LRU */
2036 nr[lru + LRU_ACTIVE] = 0;
2039 * Recalculate the other LRU scan count based on its original
2040 * scan target and the percentage scanning already complete
2042 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2043 nr_scanned = targets[lru] - nr[lru];
2044 nr[lru] = targets[lru] * (100 - percentage) / 100;
2045 nr[lru] -= min(nr[lru], nr_scanned);
2048 nr_scanned = targets[lru] - nr[lru];
2049 nr[lru] = targets[lru] * (100 - percentage) / 100;
2050 nr[lru] -= min(nr[lru], nr_scanned);
2052 scan_adjusted = true;
2054 blk_finish_plug(&plug);
2055 sc->nr_reclaimed += nr_reclaimed;
2058 * Even if we did not try to evict anon pages at all, we want to
2059 * rebalance the anon lru active/inactive ratio.
2061 if (inactive_anon_is_low(lruvec))
2062 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2063 sc, LRU_ACTIVE_ANON);
2065 throttle_vm_writeout(sc->gfp_mask);
2068 /* Use reclaim/compaction for costly allocs or under memory pressure */
2069 static bool in_reclaim_compaction(struct scan_control *sc)
2071 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2072 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2073 sc->priority < DEF_PRIORITY - 2))
2080 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2081 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2082 * true if more pages should be reclaimed such that when the page allocator
2083 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2084 * It will give up earlier than that if there is difficulty reclaiming pages.
2086 static inline bool should_continue_reclaim(struct zone *zone,
2087 unsigned long nr_reclaimed,
2088 unsigned long nr_scanned,
2089 struct scan_control *sc)
2091 unsigned long pages_for_compaction;
2092 unsigned long inactive_lru_pages;
2094 /* If not in reclaim/compaction mode, stop */
2095 if (!in_reclaim_compaction(sc))
2098 /* Consider stopping depending on scan and reclaim activity */
2099 if (sc->gfp_mask & __GFP_REPEAT) {
2101 * For __GFP_REPEAT allocations, stop reclaiming if the
2102 * full LRU list has been scanned and we are still failing
2103 * to reclaim pages. This full LRU scan is potentially
2104 * expensive but a __GFP_REPEAT caller really wants to succeed
2106 if (!nr_reclaimed && !nr_scanned)
2110 * For non-__GFP_REPEAT allocations which can presumably
2111 * fail without consequence, stop if we failed to reclaim
2112 * any pages from the last SWAP_CLUSTER_MAX number of
2113 * pages that were scanned. This will return to the
2114 * caller faster at the risk reclaim/compaction and
2115 * the resulting allocation attempt fails
2122 * If we have not reclaimed enough pages for compaction and the
2123 * inactive lists are large enough, continue reclaiming
2125 pages_for_compaction = (2UL << sc->order);
2126 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2127 if (get_nr_swap_pages() > 0)
2128 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2129 if (sc->nr_reclaimed < pages_for_compaction &&
2130 inactive_lru_pages > pages_for_compaction)
2133 /* If compaction would go ahead or the allocation would succeed, stop */
2134 switch (compaction_suitable(zone, sc->order)) {
2135 case COMPACT_PARTIAL:
2136 case COMPACT_CONTINUE:
2144 __shrink_zone(struct zone *zone, struct scan_control *sc, bool soft_reclaim)
2146 unsigned long nr_reclaimed, nr_scanned;
2149 struct mem_cgroup *root = sc->target_mem_cgroup;
2150 struct mem_cgroup_reclaim_cookie reclaim = {
2152 .priority = sc->priority,
2154 struct mem_cgroup *memcg = NULL;
2155 mem_cgroup_iter_filter filter = (soft_reclaim) ?
2156 mem_cgroup_soft_reclaim_eligible : NULL;
2158 nr_reclaimed = sc->nr_reclaimed;
2159 nr_scanned = sc->nr_scanned;
2161 while ((memcg = mem_cgroup_iter_cond(root, memcg, &reclaim, filter))) {
2162 struct lruvec *lruvec;
2164 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2166 shrink_lruvec(lruvec, sc);
2169 * Direct reclaim and kswapd have to scan all memory
2170 * cgroups to fulfill the overall scan target for the
2173 * Limit reclaim, on the other hand, only cares about
2174 * nr_to_reclaim pages to be reclaimed and it will
2175 * retry with decreasing priority if one round over the
2176 * whole hierarchy is not sufficient.
2178 if (!global_reclaim(sc) &&
2179 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2180 mem_cgroup_iter_break(root, memcg);
2185 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2186 sc->nr_scanned - nr_scanned,
2187 sc->nr_reclaimed - nr_reclaimed);
2189 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2190 sc->nr_scanned - nr_scanned, sc));
2194 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2196 bool do_soft_reclaim = mem_cgroup_should_soft_reclaim(sc);
2197 unsigned long nr_scanned = sc->nr_scanned;
2199 __shrink_zone(zone, sc, do_soft_reclaim);
2202 * No group is over the soft limit or those that are do not have
2203 * pages in the zone we are reclaiming so we have to reclaim everybody
2205 if (do_soft_reclaim && (sc->nr_scanned == nr_scanned)) {
2206 __shrink_zone(zone, sc, false);
2211 /* Returns true if compaction should go ahead for a high-order request */
2212 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2214 unsigned long balance_gap, watermark;
2217 /* Do not consider compaction for orders reclaim is meant to satisfy */
2218 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2222 * Compaction takes time to run and there are potentially other
2223 * callers using the pages just freed. Continue reclaiming until
2224 * there is a buffer of free pages available to give compaction
2225 * a reasonable chance of completing and allocating the page
2227 balance_gap = min(low_wmark_pages(zone),
2228 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2229 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2230 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2231 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2234 * If compaction is deferred, reclaim up to a point where
2235 * compaction will have a chance of success when re-enabled
2237 if (compaction_deferred(zone, sc->order))
2238 return watermark_ok;
2240 /* If compaction is not ready to start, keep reclaiming */
2241 if (!compaction_suitable(zone, sc->order))
2244 return watermark_ok;
2248 * This is the direct reclaim path, for page-allocating processes. We only
2249 * try to reclaim pages from zones which will satisfy the caller's allocation
2252 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2254 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2256 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2257 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2258 * zone defense algorithm.
2260 * If a zone is deemed to be full of pinned pages then just give it a light
2261 * scan then give up on it.
2263 * This function returns true if a zone is being reclaimed for a costly
2264 * high-order allocation and compaction is ready to begin. This indicates to
2265 * the caller that it should consider retrying the allocation instead of
2268 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2272 bool aborted_reclaim = false;
2275 * If the number of buffer_heads in the machine exceeds the maximum
2276 * allowed level, force direct reclaim to scan the highmem zone as
2277 * highmem pages could be pinning lowmem pages storing buffer_heads
2279 if (buffer_heads_over_limit)
2280 sc->gfp_mask |= __GFP_HIGHMEM;
2282 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2283 gfp_zone(sc->gfp_mask), sc->nodemask) {
2284 if (!populated_zone(zone))
2287 * Take care memory controller reclaiming has small influence
2290 if (global_reclaim(sc)) {
2291 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2293 if (sc->priority != DEF_PRIORITY &&
2294 !zone_reclaimable(zone))
2295 continue; /* Let kswapd poll it */
2296 if (IS_ENABLED(CONFIG_COMPACTION)) {
2298 * If we already have plenty of memory free for
2299 * compaction in this zone, don't free any more.
2300 * Even though compaction is invoked for any
2301 * non-zero order, only frequent costly order
2302 * reclamation is disruptive enough to become a
2303 * noticeable problem, like transparent huge
2306 if (compaction_ready(zone, sc)) {
2307 aborted_reclaim = true;
2311 /* need some check for avoid more shrink_zone() */
2314 shrink_zone(zone, sc);
2317 return aborted_reclaim;
2320 /* All zones in zonelist are unreclaimable? */
2321 static bool all_unreclaimable(struct zonelist *zonelist,
2322 struct scan_control *sc)
2327 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2328 gfp_zone(sc->gfp_mask), sc->nodemask) {
2329 if (!populated_zone(zone))
2331 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2333 if (zone_reclaimable(zone))
2341 * This is the main entry point to direct page reclaim.
2343 * If a full scan of the inactive list fails to free enough memory then we
2344 * are "out of memory" and something needs to be killed.
2346 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2347 * high - the zone may be full of dirty or under-writeback pages, which this
2348 * caller can't do much about. We kick the writeback threads and take explicit
2349 * naps in the hope that some of these pages can be written. But if the
2350 * allocating task holds filesystem locks which prevent writeout this might not
2351 * work, and the allocation attempt will fail.
2353 * returns: 0, if no pages reclaimed
2354 * else, the number of pages reclaimed
2356 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2357 struct scan_control *sc,
2358 struct shrink_control *shrink)
2360 unsigned long total_scanned = 0;
2361 struct reclaim_state *reclaim_state = current->reclaim_state;
2364 unsigned long writeback_threshold;
2365 bool aborted_reclaim;
2367 delayacct_freepages_start();
2369 if (global_reclaim(sc))
2370 count_vm_event(ALLOCSTALL);
2373 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2376 aborted_reclaim = shrink_zones(zonelist, sc);
2379 * Don't shrink slabs when reclaiming memory from over limit
2380 * cgroups but do shrink slab at least once when aborting
2381 * reclaim for compaction to avoid unevenly scanning file/anon
2382 * LRU pages over slab pages.
2384 if (global_reclaim(sc)) {
2385 unsigned long lru_pages = 0;
2386 for_each_zone_zonelist(zone, z, zonelist,
2387 gfp_zone(sc->gfp_mask)) {
2388 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2391 lru_pages += zone_reclaimable_pages(zone);
2394 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2395 if (reclaim_state) {
2396 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2397 reclaim_state->reclaimed_slab = 0;
2400 total_scanned += sc->nr_scanned;
2401 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2405 * If we're getting trouble reclaiming, start doing
2406 * writepage even in laptop mode.
2408 if (sc->priority < DEF_PRIORITY - 2)
2409 sc->may_writepage = 1;
2412 * Try to write back as many pages as we just scanned. This
2413 * tends to cause slow streaming writers to write data to the
2414 * disk smoothly, at the dirtying rate, which is nice. But
2415 * that's undesirable in laptop mode, where we *want* lumpy
2416 * writeout. So in laptop mode, write out the whole world.
2418 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2419 if (total_scanned > writeback_threshold) {
2420 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2421 WB_REASON_TRY_TO_FREE_PAGES);
2422 sc->may_writepage = 1;
2424 } while (--sc->priority >= 0 && !aborted_reclaim);
2427 delayacct_freepages_end();
2429 if (sc->nr_reclaimed)
2430 return sc->nr_reclaimed;
2433 * As hibernation is going on, kswapd is freezed so that it can't mark
2434 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2437 if (oom_killer_disabled)
2440 /* Aborted reclaim to try compaction? don't OOM, then */
2441 if (aborted_reclaim)
2444 /* top priority shrink_zones still had more to do? don't OOM, then */
2445 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2451 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2454 unsigned long pfmemalloc_reserve = 0;
2455 unsigned long free_pages = 0;
2459 for (i = 0; i <= ZONE_NORMAL; i++) {
2460 zone = &pgdat->node_zones[i];
2461 pfmemalloc_reserve += min_wmark_pages(zone);
2462 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2465 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2467 /* kswapd must be awake if processes are being throttled */
2468 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2469 pgdat->classzone_idx = min(pgdat->classzone_idx,
2470 (enum zone_type)ZONE_NORMAL);
2471 wake_up_interruptible(&pgdat->kswapd_wait);
2478 * Throttle direct reclaimers if backing storage is backed by the network
2479 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2480 * depleted. kswapd will continue to make progress and wake the processes
2481 * when the low watermark is reached.
2483 * Returns true if a fatal signal was delivered during throttling. If this
2484 * happens, the page allocator should not consider triggering the OOM killer.
2486 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2487 nodemask_t *nodemask)
2490 int high_zoneidx = gfp_zone(gfp_mask);
2494 * Kernel threads should not be throttled as they may be indirectly
2495 * responsible for cleaning pages necessary for reclaim to make forward
2496 * progress. kjournald for example may enter direct reclaim while
2497 * committing a transaction where throttling it could forcing other
2498 * processes to block on log_wait_commit().
2500 if (current->flags & PF_KTHREAD)
2504 * If a fatal signal is pending, this process should not throttle.
2505 * It should return quickly so it can exit and free its memory
2507 if (fatal_signal_pending(current))
2510 /* Check if the pfmemalloc reserves are ok */
2511 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2512 pgdat = zone->zone_pgdat;
2513 if (pfmemalloc_watermark_ok(pgdat))
2516 /* Account for the throttling */
2517 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2520 * If the caller cannot enter the filesystem, it's possible that it
2521 * is due to the caller holding an FS lock or performing a journal
2522 * transaction in the case of a filesystem like ext[3|4]. In this case,
2523 * it is not safe to block on pfmemalloc_wait as kswapd could be
2524 * blocked waiting on the same lock. Instead, throttle for up to a
2525 * second before continuing.
2527 if (!(gfp_mask & __GFP_FS)) {
2528 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2529 pfmemalloc_watermark_ok(pgdat), HZ);
2534 /* Throttle until kswapd wakes the process */
2535 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2536 pfmemalloc_watermark_ok(pgdat));
2539 if (fatal_signal_pending(current))
2546 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2547 gfp_t gfp_mask, nodemask_t *nodemask)
2549 unsigned long nr_reclaimed;
2550 struct scan_control sc = {
2551 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2552 .may_writepage = !laptop_mode,
2553 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2557 .priority = DEF_PRIORITY,
2558 .target_mem_cgroup = NULL,
2559 .nodemask = nodemask,
2561 struct shrink_control shrink = {
2562 .gfp_mask = sc.gfp_mask,
2566 * Do not enter reclaim if fatal signal was delivered while throttled.
2567 * 1 is returned so that the page allocator does not OOM kill at this
2570 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2573 trace_mm_vmscan_direct_reclaim_begin(order,
2577 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2579 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2581 return nr_reclaimed;
2586 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2587 gfp_t gfp_mask, bool noswap,
2589 unsigned long *nr_scanned)
2591 struct scan_control sc = {
2593 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2594 .may_writepage = !laptop_mode,
2596 .may_swap = !noswap,
2599 .target_mem_cgroup = memcg,
2601 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2603 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2604 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2606 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2611 * NOTE: Although we can get the priority field, using it
2612 * here is not a good idea, since it limits the pages we can scan.
2613 * if we don't reclaim here, the shrink_zone from balance_pgdat
2614 * will pick up pages from other mem cgroup's as well. We hack
2615 * the priority and make it zero.
2617 shrink_lruvec(lruvec, &sc);
2619 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2621 *nr_scanned = sc.nr_scanned;
2622 return sc.nr_reclaimed;
2625 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2629 struct zonelist *zonelist;
2630 unsigned long nr_reclaimed;
2632 struct scan_control sc = {
2633 .may_writepage = !laptop_mode,
2635 .may_swap = !noswap,
2636 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2638 .priority = DEF_PRIORITY,
2639 .target_mem_cgroup = memcg,
2640 .nodemask = NULL, /* we don't care the placement */
2641 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2642 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2644 struct shrink_control shrink = {
2645 .gfp_mask = sc.gfp_mask,
2649 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2650 * take care of from where we get pages. So the node where we start the
2651 * scan does not need to be the current node.
2653 nid = mem_cgroup_select_victim_node(memcg);
2655 zonelist = NODE_DATA(nid)->node_zonelists;
2657 trace_mm_vmscan_memcg_reclaim_begin(0,
2661 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2663 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2665 return nr_reclaimed;
2669 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2671 struct mem_cgroup *memcg;
2673 if (!total_swap_pages)
2676 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2678 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2680 if (inactive_anon_is_low(lruvec))
2681 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2682 sc, LRU_ACTIVE_ANON);
2684 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2688 static bool zone_balanced(struct zone *zone, int order,
2689 unsigned long balance_gap, int classzone_idx)
2691 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2692 balance_gap, classzone_idx, 0))
2695 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2696 !compaction_suitable(zone, order))
2703 * pgdat_balanced() is used when checking if a node is balanced.
2705 * For order-0, all zones must be balanced!
2707 * For high-order allocations only zones that meet watermarks and are in a
2708 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2709 * total of balanced pages must be at least 25% of the zones allowed by
2710 * classzone_idx for the node to be considered balanced. Forcing all zones to
2711 * be balanced for high orders can cause excessive reclaim when there are
2713 * The choice of 25% is due to
2714 * o a 16M DMA zone that is balanced will not balance a zone on any
2715 * reasonable sized machine
2716 * o On all other machines, the top zone must be at least a reasonable
2717 * percentage of the middle zones. For example, on 32-bit x86, highmem
2718 * would need to be at least 256M for it to be balance a whole node.
2719 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2720 * to balance a node on its own. These seemed like reasonable ratios.
2722 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2724 unsigned long managed_pages = 0;
2725 unsigned long balanced_pages = 0;
2728 /* Check the watermark levels */
2729 for (i = 0; i <= classzone_idx; i++) {
2730 struct zone *zone = pgdat->node_zones + i;
2732 if (!populated_zone(zone))
2735 managed_pages += zone->managed_pages;
2738 * A special case here:
2740 * balance_pgdat() skips over all_unreclaimable after
2741 * DEF_PRIORITY. Effectively, it considers them balanced so
2742 * they must be considered balanced here as well!
2744 if (!zone_reclaimable(zone)) {
2745 balanced_pages += zone->managed_pages;
2749 if (zone_balanced(zone, order, 0, i))
2750 balanced_pages += zone->managed_pages;
2756 return balanced_pages >= (managed_pages >> 2);
2762 * Prepare kswapd for sleeping. This verifies that there are no processes
2763 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2765 * Returns true if kswapd is ready to sleep
2767 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2770 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2775 * There is a potential race between when kswapd checks its watermarks
2776 * and a process gets throttled. There is also a potential race if
2777 * processes get throttled, kswapd wakes, a large process exits therby
2778 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2779 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2780 * so wake them now if necessary. If necessary, processes will wake
2781 * kswapd and get throttled again
2783 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2784 wake_up(&pgdat->pfmemalloc_wait);
2788 return pgdat_balanced(pgdat, order, classzone_idx);
2792 * kswapd shrinks the zone by the number of pages required to reach
2793 * the high watermark.
2795 * Returns true if kswapd scanned at least the requested number of pages to
2796 * reclaim or if the lack of progress was due to pages under writeback.
2797 * This is used to determine if the scanning priority needs to be raised.
2799 static bool kswapd_shrink_zone(struct zone *zone,
2801 struct scan_control *sc,
2802 unsigned long lru_pages,
2803 unsigned long *nr_attempted)
2805 int testorder = sc->order;
2806 unsigned long balance_gap;
2807 struct reclaim_state *reclaim_state = current->reclaim_state;
2808 struct shrink_control shrink = {
2809 .gfp_mask = sc->gfp_mask,
2811 bool lowmem_pressure;
2813 /* Reclaim above the high watermark. */
2814 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2817 * Kswapd reclaims only single pages with compaction enabled. Trying
2818 * too hard to reclaim until contiguous free pages have become
2819 * available can hurt performance by evicting too much useful data
2820 * from memory. Do not reclaim more than needed for compaction.
2822 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2823 compaction_suitable(zone, sc->order) !=
2828 * We put equal pressure on every zone, unless one zone has way too
2829 * many pages free already. The "too many pages" is defined as the
2830 * high wmark plus a "gap" where the gap is either the low
2831 * watermark or 1% of the zone, whichever is smaller.
2833 balance_gap = min(low_wmark_pages(zone),
2834 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2835 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2838 * If there is no low memory pressure or the zone is balanced then no
2839 * reclaim is necessary
2841 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2842 if (!lowmem_pressure && zone_balanced(zone, testorder,
2843 balance_gap, classzone_idx))
2846 shrink_zone(zone, sc);
2848 reclaim_state->reclaimed_slab = 0;
2849 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2850 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2852 /* Account for the number of pages attempted to reclaim */
2853 *nr_attempted += sc->nr_to_reclaim;
2855 zone_clear_flag(zone, ZONE_WRITEBACK);
2858 * If a zone reaches its high watermark, consider it to be no longer
2859 * congested. It's possible there are dirty pages backed by congested
2860 * BDIs but as pressure is relieved, speculatively avoid congestion
2863 if (zone_reclaimable(zone) &&
2864 zone_balanced(zone, testorder, 0, classzone_idx)) {
2865 zone_clear_flag(zone, ZONE_CONGESTED);
2866 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2869 return sc->nr_scanned >= sc->nr_to_reclaim;
2873 * For kswapd, balance_pgdat() will work across all this node's zones until
2874 * they are all at high_wmark_pages(zone).
2876 * Returns the final order kswapd was reclaiming at
2878 * There is special handling here for zones which are full of pinned pages.
2879 * This can happen if the pages are all mlocked, or if they are all used by
2880 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2881 * What we do is to detect the case where all pages in the zone have been
2882 * scanned twice and there has been zero successful reclaim. Mark the zone as
2883 * dead and from now on, only perform a short scan. Basically we're polling
2884 * the zone for when the problem goes away.
2886 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2887 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2888 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2889 * lower zones regardless of the number of free pages in the lower zones. This
2890 * interoperates with the page allocator fallback scheme to ensure that aging
2891 * of pages is balanced across the zones.
2893 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2897 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2898 struct scan_control sc = {
2899 .gfp_mask = GFP_KERNEL,
2900 .priority = DEF_PRIORITY,
2903 .may_writepage = !laptop_mode,
2905 .target_mem_cgroup = NULL,
2907 count_vm_event(PAGEOUTRUN);
2910 unsigned long lru_pages = 0;
2911 unsigned long nr_attempted = 0;
2912 bool raise_priority = true;
2913 bool pgdat_needs_compaction = (order > 0);
2915 sc.nr_reclaimed = 0;
2918 * Scan in the highmem->dma direction for the highest
2919 * zone which needs scanning
2921 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2922 struct zone *zone = pgdat->node_zones + i;
2924 if (!populated_zone(zone))
2927 if (sc.priority != DEF_PRIORITY &&
2928 !zone_reclaimable(zone))
2932 * Do some background aging of the anon list, to give
2933 * pages a chance to be referenced before reclaiming.
2935 age_active_anon(zone, &sc);
2938 * If the number of buffer_heads in the machine
2939 * exceeds the maximum allowed level and this node
2940 * has a highmem zone, force kswapd to reclaim from
2941 * it to relieve lowmem pressure.
2943 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2948 if (!zone_balanced(zone, order, 0, 0)) {
2953 * If balanced, clear the dirty and congested
2956 zone_clear_flag(zone, ZONE_CONGESTED);
2957 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2964 for (i = 0; i <= end_zone; i++) {
2965 struct zone *zone = pgdat->node_zones + i;
2967 if (!populated_zone(zone))
2970 lru_pages += zone_reclaimable_pages(zone);
2973 * If any zone is currently balanced then kswapd will
2974 * not call compaction as it is expected that the
2975 * necessary pages are already available.
2977 if (pgdat_needs_compaction &&
2978 zone_watermark_ok(zone, order,
2979 low_wmark_pages(zone),
2981 pgdat_needs_compaction = false;
2985 * If we're getting trouble reclaiming, start doing writepage
2986 * even in laptop mode.
2988 if (sc.priority < DEF_PRIORITY - 2)
2989 sc.may_writepage = 1;
2992 * Now scan the zone in the dma->highmem direction, stopping
2993 * at the last zone which needs scanning.
2995 * We do this because the page allocator works in the opposite
2996 * direction. This prevents the page allocator from allocating
2997 * pages behind kswapd's direction of progress, which would
2998 * cause too much scanning of the lower zones.
3000 for (i = 0; i <= end_zone; i++) {
3001 struct zone *zone = pgdat->node_zones + i;
3003 if (!populated_zone(zone))
3006 if (sc.priority != DEF_PRIORITY &&
3007 !zone_reclaimable(zone))
3013 * There should be no need to raise the scanning
3014 * priority if enough pages are already being scanned
3015 * that that high watermark would be met at 100%
3018 if (kswapd_shrink_zone(zone, end_zone, &sc,
3019 lru_pages, &nr_attempted))
3020 raise_priority = false;
3024 * If the low watermark is met there is no need for processes
3025 * to be throttled on pfmemalloc_wait as they should not be
3026 * able to safely make forward progress. Wake them
3028 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3029 pfmemalloc_watermark_ok(pgdat))
3030 wake_up(&pgdat->pfmemalloc_wait);
3033 * Fragmentation may mean that the system cannot be rebalanced
3034 * for high-order allocations in all zones. If twice the
3035 * allocation size has been reclaimed and the zones are still
3036 * not balanced then recheck the watermarks at order-0 to
3037 * prevent kswapd reclaiming excessively. Assume that a
3038 * process requested a high-order can direct reclaim/compact.
3040 if (order && sc.nr_reclaimed >= 2UL << order)
3041 order = sc.order = 0;
3043 /* Check if kswapd should be suspending */
3044 if (try_to_freeze() || kthread_should_stop())
3048 * Compact if necessary and kswapd is reclaiming at least the
3049 * high watermark number of pages as requsted
3051 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3052 compact_pgdat(pgdat, order);
3055 * Raise priority if scanning rate is too low or there was no
3056 * progress in reclaiming pages
3058 if (raise_priority || !sc.nr_reclaimed)
3060 } while (sc.priority >= 1 &&
3061 !pgdat_balanced(pgdat, order, *classzone_idx));
3065 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3066 * makes a decision on the order we were last reclaiming at. However,
3067 * if another caller entered the allocator slow path while kswapd
3068 * was awake, order will remain at the higher level
3070 *classzone_idx = end_zone;
3074 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3079 if (freezing(current) || kthread_should_stop())
3082 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3084 /* Try to sleep for a short interval */
3085 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3086 remaining = schedule_timeout(HZ/10);
3087 finish_wait(&pgdat->kswapd_wait, &wait);
3088 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3092 * After a short sleep, check if it was a premature sleep. If not, then
3093 * go fully to sleep until explicitly woken up.
3095 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3096 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3099 * vmstat counters are not perfectly accurate and the estimated
3100 * value for counters such as NR_FREE_PAGES can deviate from the
3101 * true value by nr_online_cpus * threshold. To avoid the zone
3102 * watermarks being breached while under pressure, we reduce the
3103 * per-cpu vmstat threshold while kswapd is awake and restore
3104 * them before going back to sleep.
3106 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3109 * Compaction records what page blocks it recently failed to
3110 * isolate pages from and skips them in the future scanning.
3111 * When kswapd is going to sleep, it is reasonable to assume
3112 * that pages and compaction may succeed so reset the cache.
3114 reset_isolation_suitable(pgdat);
3116 if (!kthread_should_stop())
3119 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3122 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3124 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3126 finish_wait(&pgdat->kswapd_wait, &wait);
3130 * The background pageout daemon, started as a kernel thread
3131 * from the init process.
3133 * This basically trickles out pages so that we have _some_
3134 * free memory available even if there is no other activity
3135 * that frees anything up. This is needed for things like routing
3136 * etc, where we otherwise might have all activity going on in
3137 * asynchronous contexts that cannot page things out.
3139 * If there are applications that are active memory-allocators
3140 * (most normal use), this basically shouldn't matter.
3142 static int kswapd(void *p)
3144 unsigned long order, new_order;
3145 unsigned balanced_order;
3146 int classzone_idx, new_classzone_idx;
3147 int balanced_classzone_idx;
3148 pg_data_t *pgdat = (pg_data_t*)p;
3149 struct task_struct *tsk = current;
3151 struct reclaim_state reclaim_state = {
3152 .reclaimed_slab = 0,
3154 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3156 lockdep_set_current_reclaim_state(GFP_KERNEL);
3158 if (!cpumask_empty(cpumask))
3159 set_cpus_allowed_ptr(tsk, cpumask);
3160 current->reclaim_state = &reclaim_state;
3163 * Tell the memory management that we're a "memory allocator",
3164 * and that if we need more memory we should get access to it
3165 * regardless (see "__alloc_pages()"). "kswapd" should
3166 * never get caught in the normal page freeing logic.
3168 * (Kswapd normally doesn't need memory anyway, but sometimes
3169 * you need a small amount of memory in order to be able to
3170 * page out something else, and this flag essentially protects
3171 * us from recursively trying to free more memory as we're
3172 * trying to free the first piece of memory in the first place).
3174 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3177 order = new_order = 0;
3179 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3180 balanced_classzone_idx = classzone_idx;
3185 * If the last balance_pgdat was unsuccessful it's unlikely a
3186 * new request of a similar or harder type will succeed soon
3187 * so consider going to sleep on the basis we reclaimed at
3189 if (balanced_classzone_idx >= new_classzone_idx &&
3190 balanced_order == new_order) {
3191 new_order = pgdat->kswapd_max_order;
3192 new_classzone_idx = pgdat->classzone_idx;
3193 pgdat->kswapd_max_order = 0;
3194 pgdat->classzone_idx = pgdat->nr_zones - 1;
3197 if (order < new_order || classzone_idx > new_classzone_idx) {
3199 * Don't sleep if someone wants a larger 'order'
3200 * allocation or has tigher zone constraints
3203 classzone_idx = new_classzone_idx;
3205 kswapd_try_to_sleep(pgdat, balanced_order,
3206 balanced_classzone_idx);
3207 order = pgdat->kswapd_max_order;
3208 classzone_idx = pgdat->classzone_idx;
3210 new_classzone_idx = classzone_idx;
3211 pgdat->kswapd_max_order = 0;
3212 pgdat->classzone_idx = pgdat->nr_zones - 1;
3215 ret = try_to_freeze();
3216 if (kthread_should_stop())
3220 * We can speed up thawing tasks if we don't call balance_pgdat
3221 * after returning from the refrigerator
3224 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3225 balanced_classzone_idx = classzone_idx;
3226 balanced_order = balance_pgdat(pgdat, order,
3227 &balanced_classzone_idx);
3231 current->reclaim_state = NULL;
3236 * A zone is low on free memory, so wake its kswapd task to service it.
3238 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3242 if (!populated_zone(zone))
3245 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3247 pgdat = zone->zone_pgdat;
3248 if (pgdat->kswapd_max_order < order) {
3249 pgdat->kswapd_max_order = order;
3250 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3252 if (!waitqueue_active(&pgdat->kswapd_wait))
3254 if (zone_balanced(zone, order, 0, 0))
3257 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3258 wake_up_interruptible(&pgdat->kswapd_wait);
3262 * The reclaimable count would be mostly accurate.
3263 * The less reclaimable pages may be
3264 * - mlocked pages, which will be moved to unevictable list when encountered
3265 * - mapped pages, which may require several travels to be reclaimed
3266 * - dirty pages, which is not "instantly" reclaimable
3268 unsigned long global_reclaimable_pages(void)
3272 nr = global_page_state(NR_ACTIVE_FILE) +
3273 global_page_state(NR_INACTIVE_FILE);
3275 if (get_nr_swap_pages() > 0)
3276 nr += global_page_state(NR_ACTIVE_ANON) +
3277 global_page_state(NR_INACTIVE_ANON);
3282 #ifdef CONFIG_HIBERNATION
3284 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3287 * Rather than trying to age LRUs the aim is to preserve the overall
3288 * LRU order by reclaiming preferentially
3289 * inactive > active > active referenced > active mapped
3291 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3293 struct reclaim_state reclaim_state;
3294 struct scan_control sc = {
3295 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3299 .nr_to_reclaim = nr_to_reclaim,
3300 .hibernation_mode = 1,
3302 .priority = DEF_PRIORITY,
3304 struct shrink_control shrink = {
3305 .gfp_mask = sc.gfp_mask,
3307 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3308 struct task_struct *p = current;
3309 unsigned long nr_reclaimed;
3311 p->flags |= PF_MEMALLOC;
3312 lockdep_set_current_reclaim_state(sc.gfp_mask);
3313 reclaim_state.reclaimed_slab = 0;
3314 p->reclaim_state = &reclaim_state;
3316 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3318 p->reclaim_state = NULL;
3319 lockdep_clear_current_reclaim_state();
3320 p->flags &= ~PF_MEMALLOC;
3322 return nr_reclaimed;
3324 #endif /* CONFIG_HIBERNATION */
3326 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3327 not required for correctness. So if the last cpu in a node goes
3328 away, we get changed to run anywhere: as the first one comes back,
3329 restore their cpu bindings. */
3330 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3335 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3336 for_each_node_state(nid, N_MEMORY) {
3337 pg_data_t *pgdat = NODE_DATA(nid);
3338 const struct cpumask *mask;
3340 mask = cpumask_of_node(pgdat->node_id);
3342 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3343 /* One of our CPUs online: restore mask */
3344 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3351 * This kswapd start function will be called by init and node-hot-add.
3352 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3354 int kswapd_run(int nid)
3356 pg_data_t *pgdat = NODE_DATA(nid);
3362 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3363 if (IS_ERR(pgdat->kswapd)) {
3364 /* failure at boot is fatal */
3365 BUG_ON(system_state == SYSTEM_BOOTING);
3366 pr_err("Failed to start kswapd on node %d\n", nid);
3367 ret = PTR_ERR(pgdat->kswapd);
3368 pgdat->kswapd = NULL;
3374 * Called by memory hotplug when all memory in a node is offlined. Caller must
3375 * hold lock_memory_hotplug().
3377 void kswapd_stop(int nid)
3379 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3382 kthread_stop(kswapd);
3383 NODE_DATA(nid)->kswapd = NULL;
3387 static int __init kswapd_init(void)
3392 for_each_node_state(nid, N_MEMORY)
3394 hotcpu_notifier(cpu_callback, 0);
3398 module_init(kswapd_init)
3404 * If non-zero call zone_reclaim when the number of free pages falls below
3407 int zone_reclaim_mode __read_mostly;
3409 #define RECLAIM_OFF 0
3410 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3411 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3412 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3415 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3416 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3419 #define ZONE_RECLAIM_PRIORITY 4
3422 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3425 int sysctl_min_unmapped_ratio = 1;
3428 * If the number of slab pages in a zone grows beyond this percentage then
3429 * slab reclaim needs to occur.
3431 int sysctl_min_slab_ratio = 5;
3433 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3435 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3436 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3437 zone_page_state(zone, NR_ACTIVE_FILE);
3440 * It's possible for there to be more file mapped pages than
3441 * accounted for by the pages on the file LRU lists because
3442 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3444 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3447 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3448 static long zone_pagecache_reclaimable(struct zone *zone)
3450 long nr_pagecache_reclaimable;
3454 * If RECLAIM_SWAP is set, then all file pages are considered
3455 * potentially reclaimable. Otherwise, we have to worry about
3456 * pages like swapcache and zone_unmapped_file_pages() provides
3459 if (zone_reclaim_mode & RECLAIM_SWAP)
3460 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3462 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3464 /* If we can't clean pages, remove dirty pages from consideration */
3465 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3466 delta += zone_page_state(zone, NR_FILE_DIRTY);
3468 /* Watch for any possible underflows due to delta */
3469 if (unlikely(delta > nr_pagecache_reclaimable))
3470 delta = nr_pagecache_reclaimable;
3472 return nr_pagecache_reclaimable - delta;
3476 * Try to free up some pages from this zone through reclaim.
3478 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3480 /* Minimum pages needed in order to stay on node */
3481 const unsigned long nr_pages = 1 << order;
3482 struct task_struct *p = current;
3483 struct reclaim_state reclaim_state;
3484 struct scan_control sc = {
3485 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3486 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3488 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3489 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3491 .priority = ZONE_RECLAIM_PRIORITY,
3493 struct shrink_control shrink = {
3494 .gfp_mask = sc.gfp_mask,
3496 unsigned long nr_slab_pages0, nr_slab_pages1;
3500 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3501 * and we also need to be able to write out pages for RECLAIM_WRITE
3504 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3505 lockdep_set_current_reclaim_state(gfp_mask);
3506 reclaim_state.reclaimed_slab = 0;
3507 p->reclaim_state = &reclaim_state;
3509 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3511 * Free memory by calling shrink zone with increasing
3512 * priorities until we have enough memory freed.
3515 shrink_zone(zone, &sc);
3516 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3519 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3520 if (nr_slab_pages0 > zone->min_slab_pages) {
3522 * shrink_slab() does not currently allow us to determine how
3523 * many pages were freed in this zone. So we take the current
3524 * number of slab pages and shake the slab until it is reduced
3525 * by the same nr_pages that we used for reclaiming unmapped
3528 * Note that shrink_slab will free memory on all zones and may
3532 unsigned long lru_pages = zone_reclaimable_pages(zone);
3534 /* No reclaimable slab or very low memory pressure */
3535 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3538 /* Freed enough memory */
3539 nr_slab_pages1 = zone_page_state(zone,
3540 NR_SLAB_RECLAIMABLE);
3541 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3546 * Update nr_reclaimed by the number of slab pages we
3547 * reclaimed from this zone.
3549 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3550 if (nr_slab_pages1 < nr_slab_pages0)
3551 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3554 p->reclaim_state = NULL;
3555 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3556 lockdep_clear_current_reclaim_state();
3557 return sc.nr_reclaimed >= nr_pages;
3560 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3566 * Zone reclaim reclaims unmapped file backed pages and
3567 * slab pages if we are over the defined limits.
3569 * A small portion of unmapped file backed pages is needed for
3570 * file I/O otherwise pages read by file I/O will be immediately
3571 * thrown out if the zone is overallocated. So we do not reclaim
3572 * if less than a specified percentage of the zone is used by
3573 * unmapped file backed pages.
3575 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3576 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3577 return ZONE_RECLAIM_FULL;
3579 if (!zone_reclaimable(zone))
3580 return ZONE_RECLAIM_FULL;
3583 * Do not scan if the allocation should not be delayed.
3585 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3586 return ZONE_RECLAIM_NOSCAN;
3589 * Only run zone reclaim on the local zone or on zones that do not
3590 * have associated processors. This will favor the local processor
3591 * over remote processors and spread off node memory allocations
3592 * as wide as possible.
3594 node_id = zone_to_nid(zone);
3595 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3596 return ZONE_RECLAIM_NOSCAN;
3598 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3599 return ZONE_RECLAIM_NOSCAN;
3601 ret = __zone_reclaim(zone, gfp_mask, order);
3602 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3605 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3612 * page_evictable - test whether a page is evictable
3613 * @page: the page to test
3615 * Test whether page is evictable--i.e., should be placed on active/inactive
3616 * lists vs unevictable list.
3618 * Reasons page might not be evictable:
3619 * (1) page's mapping marked unevictable
3620 * (2) page is part of an mlocked VMA
3623 int page_evictable(struct page *page)
3625 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3630 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3631 * @pages: array of pages to check
3632 * @nr_pages: number of pages to check
3634 * Checks pages for evictability and moves them to the appropriate lru list.
3636 * This function is only used for SysV IPC SHM_UNLOCK.
3638 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3640 struct lruvec *lruvec;
3641 struct zone *zone = NULL;
3646 for (i = 0; i < nr_pages; i++) {
3647 struct page *page = pages[i];
3648 struct zone *pagezone;
3651 pagezone = page_zone(page);
3652 if (pagezone != zone) {
3654 spin_unlock_irq(&zone->lru_lock);
3656 spin_lock_irq(&zone->lru_lock);
3658 lruvec = mem_cgroup_page_lruvec(page, zone);
3660 if (!PageLRU(page) || !PageUnevictable(page))
3663 if (page_evictable(page)) {
3664 enum lru_list lru = page_lru_base_type(page);
3666 VM_BUG_ON(PageActive(page));
3667 ClearPageUnevictable(page);
3668 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3669 add_page_to_lru_list(page, lruvec, lru);
3675 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3676 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3677 spin_unlock_irq(&zone->lru_lock);
3680 #endif /* CONFIG_SHMEM */
3682 static void warn_scan_unevictable_pages(void)
3684 printk_once(KERN_WARNING
3685 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3686 "disabled for lack of a legitimate use case. If you have "
3687 "one, please send an email to linux-mm@kvack.org.\n",
3692 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3693 * all nodes' unevictable lists for evictable pages
3695 unsigned long scan_unevictable_pages;
3697 int scan_unevictable_handler(struct ctl_table *table, int write,
3698 void __user *buffer,
3699 size_t *length, loff_t *ppos)
3701 warn_scan_unevictable_pages();
3702 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3703 scan_unevictable_pages = 0;
3709 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3710 * a specified node's per zone unevictable lists for evictable pages.
3713 static ssize_t read_scan_unevictable_node(struct device *dev,
3714 struct device_attribute *attr,
3717 warn_scan_unevictable_pages();
3718 return sprintf(buf, "0\n"); /* always zero; should fit... */
3721 static ssize_t write_scan_unevictable_node(struct device *dev,
3722 struct device_attribute *attr,
3723 const char *buf, size_t count)
3725 warn_scan_unevictable_pages();
3730 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3731 read_scan_unevictable_node,
3732 write_scan_unevictable_node);
3734 int scan_unevictable_register_node(struct node *node)
3736 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3739 void scan_unevictable_unregister_node(struct node *node)
3741 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);