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 global_reclaim(struct scan_control *sc)
149 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
151 if (!mem_cgroup_disabled())
152 return mem_cgroup_get_lru_size(lruvec, lru);
154 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
158 * Add a shrinker callback to be called from the vm
160 void register_shrinker(struct shrinker *shrinker)
162 atomic_long_set(&shrinker->nr_in_batch, 0);
163 down_write(&shrinker_rwsem);
164 list_add_tail(&shrinker->list, &shrinker_list);
165 up_write(&shrinker_rwsem);
167 EXPORT_SYMBOL(register_shrinker);
172 void unregister_shrinker(struct shrinker *shrinker)
174 down_write(&shrinker_rwsem);
175 list_del(&shrinker->list);
176 up_write(&shrinker_rwsem);
178 EXPORT_SYMBOL(unregister_shrinker);
180 static inline int do_shrinker_shrink(struct shrinker *shrinker,
181 struct shrink_control *sc,
182 unsigned long nr_to_scan)
184 sc->nr_to_scan = nr_to_scan;
185 return (*shrinker->shrink)(shrinker, sc);
188 #define SHRINK_BATCH 128
190 * Call the shrink functions to age shrinkable caches
192 * Here we assume it costs one seek to replace a lru page and that it also
193 * takes a seek to recreate a cache object. With this in mind we age equal
194 * percentages of the lru and ageable caches. This should balance the seeks
195 * generated by these structures.
197 * If the vm encountered mapped pages on the LRU it increase the pressure on
198 * slab to avoid swapping.
200 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
202 * `lru_pages' represents the number of on-LRU pages in all the zones which
203 * are eligible for the caller's allocation attempt. It is used for balancing
204 * slab reclaim versus page reclaim.
206 * Returns the number of slab objects which we shrunk.
208 unsigned long shrink_slab(struct shrink_control *shrink,
209 unsigned long nr_pages_scanned,
210 unsigned long lru_pages)
212 struct shrinker *shrinker;
213 unsigned long ret = 0;
215 if (nr_pages_scanned == 0)
216 nr_pages_scanned = SWAP_CLUSTER_MAX;
218 if (!down_read_trylock(&shrinker_rwsem)) {
219 /* Assume we'll be able to shrink next time */
224 list_for_each_entry(shrinker, &shrinker_list, list) {
225 unsigned long long delta;
231 long batch_size = shrinker->batch ? shrinker->batch
234 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
243 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
246 delta = (4 * nr_pages_scanned) / shrinker->seeks;
248 do_div(delta, lru_pages + 1);
250 if (total_scan < 0) {
251 printk(KERN_ERR "shrink_slab: %pF negative objects to "
253 shrinker->shrink, total_scan);
254 total_scan = max_pass;
258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
263 * max_pass. This is bad for sustaining a working set in
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
269 if (delta < max_pass / 4)
270 total_scan = min(total_scan, max_pass / 2);
273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
277 if (total_scan > max_pass * 2)
278 total_scan = max_pass * 2;
280 trace_mm_shrink_slab_start(shrinker, shrink, nr,
281 nr_pages_scanned, lru_pages,
282 max_pass, delta, total_scan);
284 while (total_scan >= batch_size) {
287 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
288 shrink_ret = do_shrinker_shrink(shrinker, shrink,
290 if (shrink_ret == -1)
292 if (shrink_ret < nr_before)
293 ret += nr_before - shrink_ret;
294 count_vm_events(SLABS_SCANNED, batch_size);
295 total_scan -= batch_size;
301 * move the unused scan count back into the shrinker in a
302 * manner that handles concurrent updates. If we exhausted the
303 * scan, there is no need to do an update.
306 new_nr = atomic_long_add_return(total_scan,
307 &shrinker->nr_in_batch);
309 new_nr = atomic_long_read(&shrinker->nr_in_batch);
311 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
313 up_read(&shrinker_rwsem);
319 static inline int is_page_cache_freeable(struct page *page)
322 * A freeable page cache page is referenced only by the caller
323 * that isolated the page, the page cache radix tree and
324 * optional buffer heads at page->private.
326 return page_count(page) - page_has_private(page) == 2;
329 static int may_write_to_queue(struct backing_dev_info *bdi,
330 struct scan_control *sc)
332 if (current->flags & PF_SWAPWRITE)
334 if (!bdi_write_congested(bdi))
336 if (bdi == current->backing_dev_info)
342 * We detected a synchronous write error writing a page out. Probably
343 * -ENOSPC. We need to propagate that into the address_space for a subsequent
344 * fsync(), msync() or close().
346 * The tricky part is that after writepage we cannot touch the mapping: nothing
347 * prevents it from being freed up. But we have a ref on the page and once
348 * that page is locked, the mapping is pinned.
350 * We're allowed to run sleeping lock_page() here because we know the caller has
353 static void handle_write_error(struct address_space *mapping,
354 struct page *page, int error)
357 if (page_mapping(page) == mapping)
358 mapping_set_error(mapping, error);
362 /* possible outcome of pageout() */
364 /* failed to write page out, page is locked */
366 /* move page to the active list, page is locked */
368 /* page has been sent to the disk successfully, page is unlocked */
370 /* page is clean and locked */
375 * pageout is called by shrink_page_list() for each dirty page.
376 * Calls ->writepage().
378 static pageout_t pageout(struct page *page, struct address_space *mapping,
379 struct scan_control *sc)
382 * If the page is dirty, only perform writeback if that write
383 * will be non-blocking. To prevent this allocation from being
384 * stalled by pagecache activity. But note that there may be
385 * stalls if we need to run get_block(). We could test
386 * PagePrivate for that.
388 * If this process is currently in __generic_file_aio_write() against
389 * this page's queue, we can perform writeback even if that
392 * If the page is swapcache, write it back even if that would
393 * block, for some throttling. This happens by accident, because
394 * swap_backing_dev_info is bust: it doesn't reflect the
395 * congestion state of the swapdevs. Easy to fix, if needed.
397 if (!is_page_cache_freeable(page))
401 * Some data journaling orphaned pages can have
402 * page->mapping == NULL while being dirty with clean buffers.
404 if (page_has_private(page)) {
405 if (try_to_free_buffers(page)) {
406 ClearPageDirty(page);
407 printk("%s: orphaned page\n", __func__);
413 if (mapping->a_ops->writepage == NULL)
414 return PAGE_ACTIVATE;
415 if (!may_write_to_queue(mapping->backing_dev_info, sc))
418 if (clear_page_dirty_for_io(page)) {
420 struct writeback_control wbc = {
421 .sync_mode = WB_SYNC_NONE,
422 .nr_to_write = SWAP_CLUSTER_MAX,
424 .range_end = LLONG_MAX,
428 SetPageReclaim(page);
429 res = mapping->a_ops->writepage(page, &wbc);
431 handle_write_error(mapping, page, res);
432 if (res == AOP_WRITEPAGE_ACTIVATE) {
433 ClearPageReclaim(page);
434 return PAGE_ACTIVATE;
437 if (!PageWriteback(page)) {
438 /* synchronous write or broken a_ops? */
439 ClearPageReclaim(page);
441 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
442 inc_zone_page_state(page, NR_VMSCAN_WRITE);
450 * Same as remove_mapping, but if the page is removed from the mapping, it
451 * gets returned with a refcount of 0.
453 static int __remove_mapping(struct address_space *mapping, struct page *page)
455 BUG_ON(!PageLocked(page));
456 BUG_ON(mapping != page_mapping(page));
458 spin_lock_irq(&mapping->tree_lock);
460 * The non racy check for a busy page.
462 * Must be careful with the order of the tests. When someone has
463 * a ref to the page, it may be possible that they dirty it then
464 * drop the reference. So if PageDirty is tested before page_count
465 * here, then the following race may occur:
467 * get_user_pages(&page);
468 * [user mapping goes away]
470 * !PageDirty(page) [good]
471 * SetPageDirty(page);
473 * !page_count(page) [good, discard it]
475 * [oops, our write_to data is lost]
477 * Reversing the order of the tests ensures such a situation cannot
478 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
479 * load is not satisfied before that of page->_count.
481 * Note that if SetPageDirty is always performed via set_page_dirty,
482 * and thus under tree_lock, then this ordering is not required.
484 if (!page_freeze_refs(page, 2))
486 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
487 if (unlikely(PageDirty(page))) {
488 page_unfreeze_refs(page, 2);
492 if (PageSwapCache(page)) {
493 swp_entry_t swap = { .val = page_private(page) };
494 __delete_from_swap_cache(page);
495 spin_unlock_irq(&mapping->tree_lock);
496 swapcache_free(swap, page);
498 void (*freepage)(struct page *);
500 freepage = mapping->a_ops->freepage;
502 __delete_from_page_cache(page);
503 spin_unlock_irq(&mapping->tree_lock);
504 mem_cgroup_uncharge_cache_page(page);
506 if (freepage != NULL)
513 spin_unlock_irq(&mapping->tree_lock);
518 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
519 * someone else has a ref on the page, abort and return 0. If it was
520 * successfully detached, return 1. Assumes the caller has a single ref on
523 int remove_mapping(struct address_space *mapping, struct page *page)
525 if (__remove_mapping(mapping, page)) {
527 * Unfreezing the refcount with 1 rather than 2 effectively
528 * drops the pagecache ref for us without requiring another
531 page_unfreeze_refs(page, 1);
538 * putback_lru_page - put previously isolated page onto appropriate LRU list
539 * @page: page to be put back to appropriate lru list
541 * Add previously isolated @page to appropriate LRU list.
542 * Page may still be unevictable for other reasons.
544 * lru_lock must not be held, interrupts must be enabled.
546 void putback_lru_page(struct page *page)
549 int active = !!TestClearPageActive(page);
550 int was_unevictable = PageUnevictable(page);
552 VM_BUG_ON(PageLRU(page));
555 ClearPageUnevictable(page);
557 if (page_evictable(page)) {
559 * For evictable pages, we can use the cache.
560 * In event of a race, worst case is we end up with an
561 * unevictable page on [in]active list.
562 * We know how to handle that.
564 lru = active + page_lru_base_type(page);
565 lru_cache_add_lru(page, lru);
568 * Put unevictable pages directly on zone's unevictable
571 lru = LRU_UNEVICTABLE;
572 add_page_to_unevictable_list(page);
574 * When racing with an mlock or AS_UNEVICTABLE clearing
575 * (page is unlocked) make sure that if the other thread
576 * does not observe our setting of PG_lru and fails
577 * isolation/check_move_unevictable_pages,
578 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
579 * the page back to the evictable list.
581 * The other side is TestClearPageMlocked() or shmem_lock().
587 * page's status can change while we move it among lru. If an evictable
588 * page is on unevictable list, it never be freed. To avoid that,
589 * check after we added it to the list, again.
591 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
592 if (!isolate_lru_page(page)) {
596 /* This means someone else dropped this page from LRU
597 * So, it will be freed or putback to LRU again. There is
598 * nothing to do here.
602 if (was_unevictable && lru != LRU_UNEVICTABLE)
603 count_vm_event(UNEVICTABLE_PGRESCUED);
604 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
605 count_vm_event(UNEVICTABLE_PGCULLED);
607 put_page(page); /* drop ref from isolate */
610 enum page_references {
612 PAGEREF_RECLAIM_CLEAN,
617 static enum page_references page_check_references(struct page *page,
618 struct scan_control *sc)
620 int referenced_ptes, referenced_page;
621 unsigned long vm_flags;
623 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
625 referenced_page = TestClearPageReferenced(page);
628 * Mlock lost the isolation race with us. Let try_to_unmap()
629 * move the page to the unevictable list.
631 if (vm_flags & VM_LOCKED)
632 return PAGEREF_RECLAIM;
634 if (referenced_ptes) {
635 if (PageSwapBacked(page))
636 return PAGEREF_ACTIVATE;
638 * All mapped pages start out with page table
639 * references from the instantiating fault, so we need
640 * to look twice if a mapped file page is used more
643 * Mark it and spare it for another trip around the
644 * inactive list. Another page table reference will
645 * lead to its activation.
647 * Note: the mark is set for activated pages as well
648 * so that recently deactivated but used pages are
651 SetPageReferenced(page);
653 if (referenced_page || referenced_ptes > 1)
654 return PAGEREF_ACTIVATE;
657 * Activate file-backed executable pages after first usage.
659 if (vm_flags & VM_EXEC)
660 return PAGEREF_ACTIVATE;
665 /* Reclaim if clean, defer dirty pages to writeback */
666 if (referenced_page && !PageSwapBacked(page))
667 return PAGEREF_RECLAIM_CLEAN;
669 return PAGEREF_RECLAIM;
672 /* Check if a page is dirty or under writeback */
673 static void page_check_dirty_writeback(struct page *page,
674 bool *dirty, bool *writeback)
677 * Anonymous pages are not handled by flushers and must be written
678 * from reclaim context. Do not stall reclaim based on them
680 if (!page_is_file_cache(page)) {
686 /* By default assume that the page flags are accurate */
687 *dirty = PageDirty(page);
688 *writeback = PageWriteback(page);
692 * shrink_page_list() returns the number of reclaimed pages
694 static unsigned long shrink_page_list(struct list_head *page_list,
696 struct scan_control *sc,
697 enum ttu_flags ttu_flags,
698 unsigned long *ret_nr_dirty,
699 unsigned long *ret_nr_unqueued_dirty,
700 unsigned long *ret_nr_congested,
701 unsigned long *ret_nr_writeback,
702 unsigned long *ret_nr_immediate,
705 LIST_HEAD(ret_pages);
706 LIST_HEAD(free_pages);
708 unsigned long nr_unqueued_dirty = 0;
709 unsigned long nr_dirty = 0;
710 unsigned long nr_congested = 0;
711 unsigned long nr_reclaimed = 0;
712 unsigned long nr_writeback = 0;
713 unsigned long nr_immediate = 0;
717 mem_cgroup_uncharge_start();
718 while (!list_empty(page_list)) {
719 struct address_space *mapping;
722 enum page_references references = PAGEREF_RECLAIM_CLEAN;
723 bool dirty, writeback;
727 page = lru_to_page(page_list);
728 list_del(&page->lru);
730 if (!trylock_page(page))
733 VM_BUG_ON(PageActive(page));
734 VM_BUG_ON(page_zone(page) != zone);
738 if (unlikely(!page_evictable(page)))
741 if (!sc->may_unmap && page_mapped(page))
744 /* Double the slab pressure for mapped and swapcache pages */
745 if (page_mapped(page) || PageSwapCache(page))
748 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
749 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
752 * The number of dirty pages determines if a zone is marked
753 * reclaim_congested which affects wait_iff_congested. kswapd
754 * will stall and start writing pages if the tail of the LRU
755 * is all dirty unqueued pages.
757 page_check_dirty_writeback(page, &dirty, &writeback);
758 if (dirty || writeback)
761 if (dirty && !writeback)
765 * Treat this page as congested if the underlying BDI is or if
766 * pages are cycling through the LRU so quickly that the
767 * pages marked for immediate reclaim are making it to the
768 * end of the LRU a second time.
770 mapping = page_mapping(page);
771 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
772 (writeback && PageReclaim(page)))
776 * If a page at the tail of the LRU is under writeback, there
777 * are three cases to consider.
779 * 1) If reclaim is encountering an excessive number of pages
780 * under writeback and this page is both under writeback and
781 * PageReclaim then it indicates that pages are being queued
782 * for IO but are being recycled through the LRU before the
783 * IO can complete. Waiting on the page itself risks an
784 * indefinite stall if it is impossible to writeback the
785 * page due to IO error or disconnected storage so instead
786 * note that the LRU is being scanned too quickly and the
787 * caller can stall after page list has been processed.
789 * 2) Global reclaim encounters a page, memcg encounters a
790 * page that is not marked for immediate reclaim or
791 * the caller does not have __GFP_IO. In this case mark
792 * the page for immediate reclaim and continue scanning.
794 * __GFP_IO is checked because a loop driver thread might
795 * enter reclaim, and deadlock if it waits on a page for
796 * which it is needed to do the write (loop masks off
797 * __GFP_IO|__GFP_FS for this reason); but more thought
798 * would probably show more reasons.
800 * Don't require __GFP_FS, since we're not going into the
801 * FS, just waiting on its writeback completion. Worryingly,
802 * ext4 gfs2 and xfs allocate pages with
803 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
804 * may_enter_fs here is liable to OOM on them.
806 * 3) memcg encounters a page that is not already marked
807 * PageReclaim. memcg does not have any dirty pages
808 * throttling so we could easily OOM just because too many
809 * pages are in writeback and there is nothing else to
810 * reclaim. Wait for the writeback to complete.
812 if (PageWriteback(page)) {
814 if (current_is_kswapd() &&
816 zone_is_reclaim_writeback(zone)) {
821 } else if (global_reclaim(sc) ||
822 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
824 * This is slightly racy - end_page_writeback()
825 * might have just cleared PageReclaim, then
826 * setting PageReclaim here end up interpreted
827 * as PageReadahead - but that does not matter
828 * enough to care. What we do want is for this
829 * page to have PageReclaim set next time memcg
830 * reclaim reaches the tests above, so it will
831 * then wait_on_page_writeback() to avoid OOM;
832 * and it's also appropriate in global reclaim.
834 SetPageReclaim(page);
841 wait_on_page_writeback(page);
846 references = page_check_references(page, sc);
848 switch (references) {
849 case PAGEREF_ACTIVATE:
850 goto activate_locked;
853 case PAGEREF_RECLAIM:
854 case PAGEREF_RECLAIM_CLEAN:
855 ; /* try to reclaim the page below */
859 * Anonymous process memory has backing store?
860 * Try to allocate it some swap space here.
862 if (PageAnon(page) && !PageSwapCache(page)) {
863 if (!(sc->gfp_mask & __GFP_IO))
865 if (!add_to_swap(page, page_list))
866 goto activate_locked;
869 /* Adding to swap updated mapping */
870 mapping = page_mapping(page);
874 * The page is mapped into the page tables of one or more
875 * processes. Try to unmap it here.
877 if (page_mapped(page) && mapping) {
878 switch (try_to_unmap(page, ttu_flags)) {
880 goto activate_locked;
886 ; /* try to free the page below */
890 if (PageDirty(page)) {
892 * Only kswapd can writeback filesystem pages to
893 * avoid risk of stack overflow but only writeback
894 * if many dirty pages have been encountered.
896 if (page_is_file_cache(page) &&
897 (!current_is_kswapd() ||
898 !zone_is_reclaim_dirty(zone))) {
900 * Immediately reclaim when written back.
901 * Similar in principal to deactivate_page()
902 * except we already have the page isolated
903 * and know it's dirty
905 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
906 SetPageReclaim(page);
911 if (references == PAGEREF_RECLAIM_CLEAN)
915 if (!sc->may_writepage)
918 /* Page is dirty, try to write it out here */
919 switch (pageout(page, mapping, sc)) {
923 goto activate_locked;
925 if (PageWriteback(page))
931 * A synchronous write - probably a ramdisk. Go
932 * ahead and try to reclaim the page.
934 if (!trylock_page(page))
936 if (PageDirty(page) || PageWriteback(page))
938 mapping = page_mapping(page);
940 ; /* try to free the page below */
945 * If the page has buffers, try to free the buffer mappings
946 * associated with this page. If we succeed we try to free
949 * We do this even if the page is PageDirty().
950 * try_to_release_page() does not perform I/O, but it is
951 * possible for a page to have PageDirty set, but it is actually
952 * clean (all its buffers are clean). This happens if the
953 * buffers were written out directly, with submit_bh(). ext3
954 * will do this, as well as the blockdev mapping.
955 * try_to_release_page() will discover that cleanness and will
956 * drop the buffers and mark the page clean - it can be freed.
958 * Rarely, pages can have buffers and no ->mapping. These are
959 * the pages which were not successfully invalidated in
960 * truncate_complete_page(). We try to drop those buffers here
961 * and if that worked, and the page is no longer mapped into
962 * process address space (page_count == 1) it can be freed.
963 * Otherwise, leave the page on the LRU so it is swappable.
965 if (page_has_private(page)) {
966 if (!try_to_release_page(page, sc->gfp_mask))
967 goto activate_locked;
968 if (!mapping && page_count(page) == 1) {
970 if (put_page_testzero(page))
974 * rare race with speculative reference.
975 * the speculative reference will free
976 * this page shortly, so we may
977 * increment nr_reclaimed here (and
978 * leave it off the LRU).
986 if (!mapping || !__remove_mapping(mapping, page))
990 * At this point, we have no other references and there is
991 * no way to pick any more up (removed from LRU, removed
992 * from pagecache). Can use non-atomic bitops now (and
993 * we obviously don't have to worry about waking up a process
994 * waiting on the page lock, because there are no references.
996 __clear_page_locked(page);
1001 * Is there need to periodically free_page_list? It would
1002 * appear not as the counts should be low
1004 list_add(&page->lru, &free_pages);
1008 if (PageSwapCache(page))
1009 try_to_free_swap(page);
1011 putback_lru_page(page);
1015 /* Not a candidate for swapping, so reclaim swap space. */
1016 if (PageSwapCache(page) && vm_swap_full())
1017 try_to_free_swap(page);
1018 VM_BUG_ON(PageActive(page));
1019 SetPageActive(page);
1024 list_add(&page->lru, &ret_pages);
1025 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1028 free_hot_cold_page_list(&free_pages, 1);
1030 list_splice(&ret_pages, page_list);
1031 count_vm_events(PGACTIVATE, pgactivate);
1032 mem_cgroup_uncharge_end();
1033 *ret_nr_dirty += nr_dirty;
1034 *ret_nr_congested += nr_congested;
1035 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1036 *ret_nr_writeback += nr_writeback;
1037 *ret_nr_immediate += nr_immediate;
1038 return nr_reclaimed;
1041 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1042 struct list_head *page_list)
1044 struct scan_control sc = {
1045 .gfp_mask = GFP_KERNEL,
1046 .priority = DEF_PRIORITY,
1049 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1050 struct page *page, *next;
1051 LIST_HEAD(clean_pages);
1053 list_for_each_entry_safe(page, next, page_list, lru) {
1054 if (page_is_file_cache(page) && !PageDirty(page)) {
1055 ClearPageActive(page);
1056 list_move(&page->lru, &clean_pages);
1060 ret = shrink_page_list(&clean_pages, zone, &sc,
1061 TTU_UNMAP|TTU_IGNORE_ACCESS,
1062 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1063 list_splice(&clean_pages, page_list);
1064 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1069 * Attempt to remove the specified page from its LRU. Only take this page
1070 * if it is of the appropriate PageActive status. Pages which are being
1071 * freed elsewhere are also ignored.
1073 * page: page to consider
1074 * mode: one of the LRU isolation modes defined above
1076 * returns 0 on success, -ve errno on failure.
1078 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1082 /* Only take pages on the LRU. */
1086 /* Compaction should not handle unevictable pages but CMA can do so */
1087 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1093 * To minimise LRU disruption, the caller can indicate that it only
1094 * wants to isolate pages it will be able to operate on without
1095 * blocking - clean pages for the most part.
1097 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1098 * is used by reclaim when it is cannot write to backing storage
1100 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1101 * that it is possible to migrate without blocking
1103 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1104 /* All the caller can do on PageWriteback is block */
1105 if (PageWriteback(page))
1108 if (PageDirty(page)) {
1109 struct address_space *mapping;
1111 /* ISOLATE_CLEAN means only clean pages */
1112 if (mode & ISOLATE_CLEAN)
1116 * Only pages without mappings or that have a
1117 * ->migratepage callback are possible to migrate
1120 mapping = page_mapping(page);
1121 if (mapping && !mapping->a_ops->migratepage)
1126 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1129 if (likely(get_page_unless_zero(page))) {
1131 * Be careful not to clear PageLRU until after we're
1132 * sure the page is not being freed elsewhere -- the
1133 * page release code relies on it.
1143 * zone->lru_lock is heavily contended. Some of the functions that
1144 * shrink the lists perform better by taking out a batch of pages
1145 * and working on them outside the LRU lock.
1147 * For pagecache intensive workloads, this function is the hottest
1148 * spot in the kernel (apart from copy_*_user functions).
1150 * Appropriate locks must be held before calling this function.
1152 * @nr_to_scan: The number of pages to look through on the list.
1153 * @lruvec: The LRU vector to pull pages from.
1154 * @dst: The temp list to put pages on to.
1155 * @nr_scanned: The number of pages that were scanned.
1156 * @sc: The scan_control struct for this reclaim session
1157 * @mode: One of the LRU isolation modes
1158 * @lru: LRU list id for isolating
1160 * returns how many pages were moved onto *@dst.
1162 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1163 struct lruvec *lruvec, struct list_head *dst,
1164 unsigned long *nr_scanned, struct scan_control *sc,
1165 isolate_mode_t mode, enum lru_list lru)
1167 struct list_head *src = &lruvec->lists[lru];
1168 unsigned long nr_taken = 0;
1171 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1175 page = lru_to_page(src);
1176 prefetchw_prev_lru_page(page, src, flags);
1178 VM_BUG_ON(!PageLRU(page));
1180 switch (__isolate_lru_page(page, mode)) {
1182 nr_pages = hpage_nr_pages(page);
1183 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1184 list_move(&page->lru, dst);
1185 nr_taken += nr_pages;
1189 /* else it is being freed elsewhere */
1190 list_move(&page->lru, src);
1199 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1200 nr_taken, mode, is_file_lru(lru));
1205 * isolate_lru_page - tries to isolate a page from its LRU list
1206 * @page: page to isolate from its LRU list
1208 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1209 * vmstat statistic corresponding to whatever LRU list the page was on.
1211 * Returns 0 if the page was removed from an LRU list.
1212 * Returns -EBUSY if the page was not on an LRU list.
1214 * The returned page will have PageLRU() cleared. If it was found on
1215 * the active list, it will have PageActive set. If it was found on
1216 * the unevictable list, it will have the PageUnevictable bit set. That flag
1217 * may need to be cleared by the caller before letting the page go.
1219 * The vmstat statistic corresponding to the list on which the page was
1220 * found will be decremented.
1223 * (1) Must be called with an elevated refcount on the page. This is a
1224 * fundamentnal difference from isolate_lru_pages (which is called
1225 * without a stable reference).
1226 * (2) the lru_lock must not be held.
1227 * (3) interrupts must be enabled.
1229 int isolate_lru_page(struct page *page)
1233 VM_BUG_ON(!page_count(page));
1235 if (PageLRU(page)) {
1236 struct zone *zone = page_zone(page);
1237 struct lruvec *lruvec;
1239 spin_lock_irq(&zone->lru_lock);
1240 lruvec = mem_cgroup_page_lruvec(page, zone);
1241 if (PageLRU(page)) {
1242 int lru = page_lru(page);
1245 del_page_from_lru_list(page, lruvec, lru);
1248 spin_unlock_irq(&zone->lru_lock);
1254 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1255 * then get resheduled. When there are massive number of tasks doing page
1256 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1257 * the LRU list will go small and be scanned faster than necessary, leading to
1258 * unnecessary swapping, thrashing and OOM.
1260 static int too_many_isolated(struct zone *zone, int file,
1261 struct scan_control *sc)
1263 unsigned long inactive, isolated;
1265 if (current_is_kswapd())
1268 if (!global_reclaim(sc))
1272 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1273 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1275 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1276 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1280 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1281 * won't get blocked by normal direct-reclaimers, forming a circular
1284 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1287 return isolated > inactive;
1290 static noinline_for_stack void
1291 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1293 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1294 struct zone *zone = lruvec_zone(lruvec);
1295 LIST_HEAD(pages_to_free);
1298 * Put back any unfreeable pages.
1300 while (!list_empty(page_list)) {
1301 struct page *page = lru_to_page(page_list);
1304 VM_BUG_ON(PageLRU(page));
1305 list_del(&page->lru);
1306 if (unlikely(!page_evictable(page))) {
1307 spin_unlock_irq(&zone->lru_lock);
1308 putback_lru_page(page);
1309 spin_lock_irq(&zone->lru_lock);
1313 lruvec = mem_cgroup_page_lruvec(page, zone);
1316 lru = page_lru(page);
1317 add_page_to_lru_list(page, lruvec, lru);
1319 if (is_active_lru(lru)) {
1320 int file = is_file_lru(lru);
1321 int numpages = hpage_nr_pages(page);
1322 reclaim_stat->recent_rotated[file] += numpages;
1324 if (put_page_testzero(page)) {
1325 __ClearPageLRU(page);
1326 __ClearPageActive(page);
1327 del_page_from_lru_list(page, lruvec, lru);
1329 if (unlikely(PageCompound(page))) {
1330 spin_unlock_irq(&zone->lru_lock);
1331 (*get_compound_page_dtor(page))(page);
1332 spin_lock_irq(&zone->lru_lock);
1334 list_add(&page->lru, &pages_to_free);
1339 * To save our caller's stack, now use input list for pages to free.
1341 list_splice(&pages_to_free, page_list);
1345 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1346 * of reclaimed pages
1348 static noinline_for_stack unsigned long
1349 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1350 struct scan_control *sc, enum lru_list lru)
1352 LIST_HEAD(page_list);
1353 unsigned long nr_scanned;
1354 unsigned long nr_reclaimed = 0;
1355 unsigned long nr_taken;
1356 unsigned long nr_dirty = 0;
1357 unsigned long nr_congested = 0;
1358 unsigned long nr_unqueued_dirty = 0;
1359 unsigned long nr_writeback = 0;
1360 unsigned long nr_immediate = 0;
1361 isolate_mode_t isolate_mode = 0;
1362 int file = is_file_lru(lru);
1363 struct zone *zone = lruvec_zone(lruvec);
1364 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1366 while (unlikely(too_many_isolated(zone, file, sc))) {
1367 congestion_wait(BLK_RW_ASYNC, HZ/10);
1369 /* We are about to die and free our memory. Return now. */
1370 if (fatal_signal_pending(current))
1371 return SWAP_CLUSTER_MAX;
1377 isolate_mode |= ISOLATE_UNMAPPED;
1378 if (!sc->may_writepage)
1379 isolate_mode |= ISOLATE_CLEAN;
1381 spin_lock_irq(&zone->lru_lock);
1383 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1384 &nr_scanned, sc, isolate_mode, lru);
1386 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1387 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1389 if (global_reclaim(sc)) {
1390 zone->pages_scanned += nr_scanned;
1391 if (current_is_kswapd())
1392 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1394 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1396 spin_unlock_irq(&zone->lru_lock);
1401 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1402 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1403 &nr_writeback, &nr_immediate,
1406 spin_lock_irq(&zone->lru_lock);
1408 reclaim_stat->recent_scanned[file] += nr_taken;
1410 if (global_reclaim(sc)) {
1411 if (current_is_kswapd())
1412 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1415 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1419 putback_inactive_pages(lruvec, &page_list);
1421 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1423 spin_unlock_irq(&zone->lru_lock);
1425 free_hot_cold_page_list(&page_list, 1);
1428 * If reclaim is isolating dirty pages under writeback, it implies
1429 * that the long-lived page allocation rate is exceeding the page
1430 * laundering rate. Either the global limits are not being effective
1431 * at throttling processes due to the page distribution throughout
1432 * zones or there is heavy usage of a slow backing device. The
1433 * only option is to throttle from reclaim context which is not ideal
1434 * as there is no guarantee the dirtying process is throttled in the
1435 * same way balance_dirty_pages() manages.
1437 * This scales the number of dirty pages that must be under writeback
1438 * before a zone gets flagged ZONE_WRITEBACK. It is a simple backoff
1439 * function that has the most effect in the range DEF_PRIORITY to
1440 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1441 * in trouble and reclaim is considered to be in trouble.
1443 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1444 * DEF_PRIORITY-1 50% must be PageWriteback
1445 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1447 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1448 * isolated page is PageWriteback
1450 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1451 * of pages under pages flagged for immediate reclaim and stall if any
1452 * are encountered in the nr_immediate check below.
1454 if (nr_writeback && nr_writeback >=
1455 (nr_taken >> (DEF_PRIORITY - sc->priority)))
1456 zone_set_flag(zone, ZONE_WRITEBACK);
1459 * memcg will stall in page writeback so only consider forcibly
1460 * stalling for global reclaim
1462 if (global_reclaim(sc)) {
1464 * Tag a zone as congested if all the dirty pages scanned were
1465 * backed by a congested BDI and wait_iff_congested will stall.
1467 if (nr_dirty && nr_dirty == nr_congested)
1468 zone_set_flag(zone, ZONE_CONGESTED);
1471 * If dirty pages are scanned that are not queued for IO, it
1472 * implies that flushers are not keeping up. In this case, flag
1473 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1474 * pages from reclaim context. It will forcibly stall in the
1477 if (nr_unqueued_dirty == nr_taken)
1478 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1481 * In addition, if kswapd scans pages marked marked for
1482 * immediate reclaim and under writeback (nr_immediate), it
1483 * implies that pages are cycling through the LRU faster than
1484 * they are written so also forcibly stall.
1486 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1487 congestion_wait(BLK_RW_ASYNC, HZ/10);
1491 * Stall direct reclaim for IO completions if underlying BDIs or zone
1492 * is congested. Allow kswapd to continue until it starts encountering
1493 * unqueued dirty pages or cycling through the LRU too quickly.
1495 if (!sc->hibernation_mode && !current_is_kswapd())
1496 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1498 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1500 nr_scanned, nr_reclaimed,
1502 trace_shrink_flags(file));
1503 return nr_reclaimed;
1507 * This moves pages from the active list to the inactive list.
1509 * We move them the other way if the page is referenced by one or more
1510 * processes, from rmap.
1512 * If the pages are mostly unmapped, the processing is fast and it is
1513 * appropriate to hold zone->lru_lock across the whole operation. But if
1514 * the pages are mapped, the processing is slow (page_referenced()) so we
1515 * should drop zone->lru_lock around each page. It's impossible to balance
1516 * this, so instead we remove the pages from the LRU while processing them.
1517 * It is safe to rely on PG_active against the non-LRU pages in here because
1518 * nobody will play with that bit on a non-LRU page.
1520 * The downside is that we have to touch page->_count against each page.
1521 * But we had to alter page->flags anyway.
1524 static void move_active_pages_to_lru(struct lruvec *lruvec,
1525 struct list_head *list,
1526 struct list_head *pages_to_free,
1529 struct zone *zone = lruvec_zone(lruvec);
1530 unsigned long pgmoved = 0;
1534 while (!list_empty(list)) {
1535 page = lru_to_page(list);
1536 lruvec = mem_cgroup_page_lruvec(page, zone);
1538 VM_BUG_ON(PageLRU(page));
1541 nr_pages = hpage_nr_pages(page);
1542 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1543 list_move(&page->lru, &lruvec->lists[lru]);
1544 pgmoved += nr_pages;
1546 if (put_page_testzero(page)) {
1547 __ClearPageLRU(page);
1548 __ClearPageActive(page);
1549 del_page_from_lru_list(page, lruvec, lru);
1551 if (unlikely(PageCompound(page))) {
1552 spin_unlock_irq(&zone->lru_lock);
1553 (*get_compound_page_dtor(page))(page);
1554 spin_lock_irq(&zone->lru_lock);
1556 list_add(&page->lru, pages_to_free);
1559 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1560 if (!is_active_lru(lru))
1561 __count_vm_events(PGDEACTIVATE, pgmoved);
1564 static void shrink_active_list(unsigned long nr_to_scan,
1565 struct lruvec *lruvec,
1566 struct scan_control *sc,
1569 unsigned long nr_taken;
1570 unsigned long nr_scanned;
1571 unsigned long vm_flags;
1572 LIST_HEAD(l_hold); /* The pages which were snipped off */
1573 LIST_HEAD(l_active);
1574 LIST_HEAD(l_inactive);
1576 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1577 unsigned long nr_rotated = 0;
1578 isolate_mode_t isolate_mode = 0;
1579 int file = is_file_lru(lru);
1580 struct zone *zone = lruvec_zone(lruvec);
1585 isolate_mode |= ISOLATE_UNMAPPED;
1586 if (!sc->may_writepage)
1587 isolate_mode |= ISOLATE_CLEAN;
1589 spin_lock_irq(&zone->lru_lock);
1591 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1592 &nr_scanned, sc, isolate_mode, lru);
1593 if (global_reclaim(sc))
1594 zone->pages_scanned += nr_scanned;
1596 reclaim_stat->recent_scanned[file] += nr_taken;
1598 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1599 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1600 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1601 spin_unlock_irq(&zone->lru_lock);
1603 while (!list_empty(&l_hold)) {
1605 page = lru_to_page(&l_hold);
1606 list_del(&page->lru);
1608 if (unlikely(!page_evictable(page))) {
1609 putback_lru_page(page);
1613 if (unlikely(buffer_heads_over_limit)) {
1614 if (page_has_private(page) && trylock_page(page)) {
1615 if (page_has_private(page))
1616 try_to_release_page(page, 0);
1621 if (page_referenced(page, 0, sc->target_mem_cgroup,
1623 nr_rotated += hpage_nr_pages(page);
1625 * Identify referenced, file-backed active pages and
1626 * give them one more trip around the active list. So
1627 * that executable code get better chances to stay in
1628 * memory under moderate memory pressure. Anon pages
1629 * are not likely to be evicted by use-once streaming
1630 * IO, plus JVM can create lots of anon VM_EXEC pages,
1631 * so we ignore them here.
1633 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1634 list_add(&page->lru, &l_active);
1639 ClearPageActive(page); /* we are de-activating */
1640 list_add(&page->lru, &l_inactive);
1644 * Move pages back to the lru list.
1646 spin_lock_irq(&zone->lru_lock);
1648 * Count referenced pages from currently used mappings as rotated,
1649 * even though only some of them are actually re-activated. This
1650 * helps balance scan pressure between file and anonymous pages in
1653 reclaim_stat->recent_rotated[file] += nr_rotated;
1655 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1656 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1657 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1658 spin_unlock_irq(&zone->lru_lock);
1660 free_hot_cold_page_list(&l_hold, 1);
1664 static int inactive_anon_is_low_global(struct zone *zone)
1666 unsigned long active, inactive;
1668 active = zone_page_state(zone, NR_ACTIVE_ANON);
1669 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1671 if (inactive * zone->inactive_ratio < active)
1678 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1679 * @lruvec: LRU vector to check
1681 * Returns true if the zone does not have enough inactive anon pages,
1682 * meaning some active anon pages need to be deactivated.
1684 static int inactive_anon_is_low(struct lruvec *lruvec)
1687 * If we don't have swap space, anonymous page deactivation
1690 if (!total_swap_pages)
1693 if (!mem_cgroup_disabled())
1694 return mem_cgroup_inactive_anon_is_low(lruvec);
1696 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1699 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1706 * inactive_file_is_low - check if file pages need to be deactivated
1707 * @lruvec: LRU vector to check
1709 * When the system is doing streaming IO, memory pressure here
1710 * ensures that active file pages get deactivated, until more
1711 * than half of the file pages are on the inactive list.
1713 * Once we get to that situation, protect the system's working
1714 * set from being evicted by disabling active file page aging.
1716 * This uses a different ratio than the anonymous pages, because
1717 * the page cache uses a use-once replacement algorithm.
1719 static int inactive_file_is_low(struct lruvec *lruvec)
1721 unsigned long inactive;
1722 unsigned long active;
1724 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1725 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1727 return active > inactive;
1730 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1732 if (is_file_lru(lru))
1733 return inactive_file_is_low(lruvec);
1735 return inactive_anon_is_low(lruvec);
1738 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1739 struct lruvec *lruvec, struct scan_control *sc)
1741 if (is_active_lru(lru)) {
1742 if (inactive_list_is_low(lruvec, lru))
1743 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1747 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1750 static int vmscan_swappiness(struct scan_control *sc)
1752 if (global_reclaim(sc))
1753 return vm_swappiness;
1754 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1765 * Determine how aggressively the anon and file LRU lists should be
1766 * scanned. The relative value of each set of LRU lists is determined
1767 * by looking at the fraction of the pages scanned we did rotate back
1768 * onto the active list instead of evict.
1770 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1771 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1773 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1776 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1778 u64 denominator = 0; /* gcc */
1779 struct zone *zone = lruvec_zone(lruvec);
1780 unsigned long anon_prio, file_prio;
1781 enum scan_balance scan_balance;
1782 unsigned long anon, file, free;
1783 bool force_scan = false;
1784 unsigned long ap, fp;
1788 * If the zone or memcg is small, nr[l] can be 0. This
1789 * results in no scanning on this priority and a potential
1790 * priority drop. Global direct reclaim can go to the next
1791 * zone and tends to have no problems. Global kswapd is for
1792 * zone balancing and it needs to scan a minimum amount. When
1793 * reclaiming for a memcg, a priority drop can cause high
1794 * latencies, so it's better to scan a minimum amount there as
1797 if (current_is_kswapd() && zone->all_unreclaimable)
1799 if (!global_reclaim(sc))
1802 /* If we have no swap space, do not bother scanning anon pages. */
1803 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1804 scan_balance = SCAN_FILE;
1809 * Global reclaim will swap to prevent OOM even with no
1810 * swappiness, but memcg users want to use this knob to
1811 * disable swapping for individual groups completely when
1812 * using the memory controller's swap limit feature would be
1815 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1816 scan_balance = SCAN_FILE;
1821 * Do not apply any pressure balancing cleverness when the
1822 * system is close to OOM, scan both anon and file equally
1823 * (unless the swappiness setting disagrees with swapping).
1825 if (!sc->priority && vmscan_swappiness(sc)) {
1826 scan_balance = SCAN_EQUAL;
1830 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1831 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1832 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1833 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1836 * If it's foreseeable that reclaiming the file cache won't be
1837 * enough to get the zone back into a desirable shape, we have
1838 * to swap. Better start now and leave the - probably heavily
1839 * thrashing - remaining file pages alone.
1841 if (global_reclaim(sc)) {
1842 free = zone_page_state(zone, NR_FREE_PAGES);
1843 if (unlikely(file + free <= high_wmark_pages(zone))) {
1844 scan_balance = SCAN_ANON;
1850 * There is enough inactive page cache, do not reclaim
1851 * anything from the anonymous working set right now.
1853 if (!inactive_file_is_low(lruvec)) {
1854 scan_balance = SCAN_FILE;
1858 scan_balance = SCAN_FRACT;
1861 * With swappiness at 100, anonymous and file have the same priority.
1862 * This scanning priority is essentially the inverse of IO cost.
1864 anon_prio = vmscan_swappiness(sc);
1865 file_prio = 200 - anon_prio;
1868 * OK, so we have swap space and a fair amount of page cache
1869 * pages. We use the recently rotated / recently scanned
1870 * ratios to determine how valuable each cache is.
1872 * Because workloads change over time (and to avoid overflow)
1873 * we keep these statistics as a floating average, which ends
1874 * up weighing recent references more than old ones.
1876 * anon in [0], file in [1]
1878 spin_lock_irq(&zone->lru_lock);
1879 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1880 reclaim_stat->recent_scanned[0] /= 2;
1881 reclaim_stat->recent_rotated[0] /= 2;
1884 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1885 reclaim_stat->recent_scanned[1] /= 2;
1886 reclaim_stat->recent_rotated[1] /= 2;
1890 * The amount of pressure on anon vs file pages is inversely
1891 * proportional to the fraction of recently scanned pages on
1892 * each list that were recently referenced and in active use.
1894 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1895 ap /= reclaim_stat->recent_rotated[0] + 1;
1897 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1898 fp /= reclaim_stat->recent_rotated[1] + 1;
1899 spin_unlock_irq(&zone->lru_lock);
1903 denominator = ap + fp + 1;
1905 for_each_evictable_lru(lru) {
1906 int file = is_file_lru(lru);
1910 size = get_lru_size(lruvec, lru);
1911 scan = size >> sc->priority;
1913 if (!scan && force_scan)
1914 scan = min(size, SWAP_CLUSTER_MAX);
1916 switch (scan_balance) {
1918 /* Scan lists relative to size */
1922 * Scan types proportional to swappiness and
1923 * their relative recent reclaim efficiency.
1925 scan = div64_u64(scan * fraction[file], denominator);
1929 /* Scan one type exclusively */
1930 if ((scan_balance == SCAN_FILE) != file)
1934 /* Look ma, no brain */
1942 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1944 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1946 unsigned long nr[NR_LRU_LISTS];
1947 unsigned long targets[NR_LRU_LISTS];
1948 unsigned long nr_to_scan;
1950 unsigned long nr_reclaimed = 0;
1951 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1952 struct blk_plug plug;
1953 bool scan_adjusted = false;
1955 get_scan_count(lruvec, sc, nr);
1957 /* Record the original scan target for proportional adjustments later */
1958 memcpy(targets, nr, sizeof(nr));
1960 blk_start_plug(&plug);
1961 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1962 nr[LRU_INACTIVE_FILE]) {
1963 unsigned long nr_anon, nr_file, percentage;
1964 unsigned long nr_scanned;
1966 for_each_evictable_lru(lru) {
1968 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1969 nr[lru] -= nr_to_scan;
1971 nr_reclaimed += shrink_list(lru, nr_to_scan,
1976 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
1980 * For global direct reclaim, reclaim only the number of pages
1981 * requested. Less care is taken to scan proportionally as it
1982 * is more important to minimise direct reclaim stall latency
1983 * than it is to properly age the LRU lists.
1985 if (global_reclaim(sc) && !current_is_kswapd())
1989 * For kswapd and memcg, reclaim at least the number of pages
1990 * requested. Ensure that the anon and file LRUs shrink
1991 * proportionally what was requested by get_scan_count(). We
1992 * stop reclaiming one LRU and reduce the amount scanning
1993 * proportional to the original scan target.
1995 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
1996 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
1998 if (nr_file > nr_anon) {
1999 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2000 targets[LRU_ACTIVE_ANON] + 1;
2002 percentage = nr_anon * 100 / scan_target;
2004 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2005 targets[LRU_ACTIVE_FILE] + 1;
2007 percentage = nr_file * 100 / scan_target;
2010 /* Stop scanning the smaller of the LRU */
2012 nr[lru + LRU_ACTIVE] = 0;
2015 * Recalculate the other LRU scan count based on its original
2016 * scan target and the percentage scanning already complete
2018 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2019 nr_scanned = targets[lru] - nr[lru];
2020 nr[lru] = targets[lru] * (100 - percentage) / 100;
2021 nr[lru] -= min(nr[lru], nr_scanned);
2024 nr_scanned = targets[lru] - nr[lru];
2025 nr[lru] = targets[lru] * (100 - percentage) / 100;
2026 nr[lru] -= min(nr[lru], nr_scanned);
2028 scan_adjusted = true;
2030 blk_finish_plug(&plug);
2031 sc->nr_reclaimed += nr_reclaimed;
2034 * Even if we did not try to evict anon pages at all, we want to
2035 * rebalance the anon lru active/inactive ratio.
2037 if (inactive_anon_is_low(lruvec))
2038 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2039 sc, LRU_ACTIVE_ANON);
2041 throttle_vm_writeout(sc->gfp_mask);
2044 /* Use reclaim/compaction for costly allocs or under memory pressure */
2045 static bool in_reclaim_compaction(struct scan_control *sc)
2047 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2048 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2049 sc->priority < DEF_PRIORITY - 2))
2056 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2057 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2058 * true if more pages should be reclaimed such that when the page allocator
2059 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2060 * It will give up earlier than that if there is difficulty reclaiming pages.
2062 static inline bool should_continue_reclaim(struct zone *zone,
2063 unsigned long nr_reclaimed,
2064 unsigned long nr_scanned,
2065 struct scan_control *sc)
2067 unsigned long pages_for_compaction;
2068 unsigned long inactive_lru_pages;
2070 /* If not in reclaim/compaction mode, stop */
2071 if (!in_reclaim_compaction(sc))
2074 /* Consider stopping depending on scan and reclaim activity */
2075 if (sc->gfp_mask & __GFP_REPEAT) {
2077 * For __GFP_REPEAT allocations, stop reclaiming if the
2078 * full LRU list has been scanned and we are still failing
2079 * to reclaim pages. This full LRU scan is potentially
2080 * expensive but a __GFP_REPEAT caller really wants to succeed
2082 if (!nr_reclaimed && !nr_scanned)
2086 * For non-__GFP_REPEAT allocations which can presumably
2087 * fail without consequence, stop if we failed to reclaim
2088 * any pages from the last SWAP_CLUSTER_MAX number of
2089 * pages that were scanned. This will return to the
2090 * caller faster at the risk reclaim/compaction and
2091 * the resulting allocation attempt fails
2098 * If we have not reclaimed enough pages for compaction and the
2099 * inactive lists are large enough, continue reclaiming
2101 pages_for_compaction = (2UL << sc->order);
2102 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2103 if (get_nr_swap_pages() > 0)
2104 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2105 if (sc->nr_reclaimed < pages_for_compaction &&
2106 inactive_lru_pages > pages_for_compaction)
2109 /* If compaction would go ahead or the allocation would succeed, stop */
2110 switch (compaction_suitable(zone, sc->order)) {
2111 case COMPACT_PARTIAL:
2112 case COMPACT_CONTINUE:
2119 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2121 unsigned long nr_reclaimed, nr_scanned;
2124 struct mem_cgroup *root = sc->target_mem_cgroup;
2125 struct mem_cgroup_reclaim_cookie reclaim = {
2127 .priority = sc->priority,
2129 struct mem_cgroup *memcg;
2131 nr_reclaimed = sc->nr_reclaimed;
2132 nr_scanned = sc->nr_scanned;
2134 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2136 struct lruvec *lruvec;
2138 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2140 shrink_lruvec(lruvec, sc);
2143 * Direct reclaim and kswapd have to scan all memory
2144 * cgroups to fulfill the overall scan target for the
2147 * Limit reclaim, on the other hand, only cares about
2148 * nr_to_reclaim pages to be reclaimed and it will
2149 * retry with decreasing priority if one round over the
2150 * whole hierarchy is not sufficient.
2152 if (!global_reclaim(sc) &&
2153 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2154 mem_cgroup_iter_break(root, memcg);
2157 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2160 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2161 sc->nr_scanned - nr_scanned,
2162 sc->nr_reclaimed - nr_reclaimed);
2164 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2165 sc->nr_scanned - nr_scanned, sc));
2168 /* Returns true if compaction should go ahead for a high-order request */
2169 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2171 unsigned long balance_gap, watermark;
2174 /* Do not consider compaction for orders reclaim is meant to satisfy */
2175 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2179 * Compaction takes time to run and there are potentially other
2180 * callers using the pages just freed. Continue reclaiming until
2181 * there is a buffer of free pages available to give compaction
2182 * a reasonable chance of completing and allocating the page
2184 balance_gap = min(low_wmark_pages(zone),
2185 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2186 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2187 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2188 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2191 * If compaction is deferred, reclaim up to a point where
2192 * compaction will have a chance of success when re-enabled
2194 if (compaction_deferred(zone, sc->order))
2195 return watermark_ok;
2197 /* If compaction is not ready to start, keep reclaiming */
2198 if (!compaction_suitable(zone, sc->order))
2201 return watermark_ok;
2205 * This is the direct reclaim path, for page-allocating processes. We only
2206 * try to reclaim pages from zones which will satisfy the caller's allocation
2209 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2211 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2213 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2214 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2215 * zone defense algorithm.
2217 * If a zone is deemed to be full of pinned pages then just give it a light
2218 * scan then give up on it.
2220 * This function returns true if a zone is being reclaimed for a costly
2221 * high-order allocation and compaction is ready to begin. This indicates to
2222 * the caller that it should consider retrying the allocation instead of
2225 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2229 unsigned long nr_soft_reclaimed;
2230 unsigned long nr_soft_scanned;
2231 bool aborted_reclaim = false;
2234 * If the number of buffer_heads in the machine exceeds the maximum
2235 * allowed level, force direct reclaim to scan the highmem zone as
2236 * highmem pages could be pinning lowmem pages storing buffer_heads
2238 if (buffer_heads_over_limit)
2239 sc->gfp_mask |= __GFP_HIGHMEM;
2241 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2242 gfp_zone(sc->gfp_mask), sc->nodemask) {
2243 if (!populated_zone(zone))
2246 * Take care memory controller reclaiming has small influence
2249 if (global_reclaim(sc)) {
2250 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2252 if (zone->all_unreclaimable &&
2253 sc->priority != DEF_PRIORITY)
2254 continue; /* Let kswapd poll it */
2255 if (IS_ENABLED(CONFIG_COMPACTION)) {
2257 * If we already have plenty of memory free for
2258 * compaction in this zone, don't free any more.
2259 * Even though compaction is invoked for any
2260 * non-zero order, only frequent costly order
2261 * reclamation is disruptive enough to become a
2262 * noticeable problem, like transparent huge
2265 if (compaction_ready(zone, sc)) {
2266 aborted_reclaim = true;
2271 * This steals pages from memory cgroups over softlimit
2272 * and returns the number of reclaimed pages and
2273 * scanned pages. This works for global memory pressure
2274 * and balancing, not for a memcg's limit.
2276 nr_soft_scanned = 0;
2277 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2278 sc->order, sc->gfp_mask,
2280 sc->nr_reclaimed += nr_soft_reclaimed;
2281 sc->nr_scanned += nr_soft_scanned;
2282 /* need some check for avoid more shrink_zone() */
2285 shrink_zone(zone, sc);
2288 return aborted_reclaim;
2291 static bool zone_reclaimable(struct zone *zone)
2293 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2296 /* All zones in zonelist are unreclaimable? */
2297 static bool all_unreclaimable(struct zonelist *zonelist,
2298 struct scan_control *sc)
2303 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2304 gfp_zone(sc->gfp_mask), sc->nodemask) {
2305 if (!populated_zone(zone))
2307 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2309 if (!zone->all_unreclaimable)
2317 * This is the main entry point to direct page reclaim.
2319 * If a full scan of the inactive list fails to free enough memory then we
2320 * are "out of memory" and something needs to be killed.
2322 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2323 * high - the zone may be full of dirty or under-writeback pages, which this
2324 * caller can't do much about. We kick the writeback threads and take explicit
2325 * naps in the hope that some of these pages can be written. But if the
2326 * allocating task holds filesystem locks which prevent writeout this might not
2327 * work, and the allocation attempt will fail.
2329 * returns: 0, if no pages reclaimed
2330 * else, the number of pages reclaimed
2332 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2333 struct scan_control *sc,
2334 struct shrink_control *shrink)
2336 unsigned long total_scanned = 0;
2337 struct reclaim_state *reclaim_state = current->reclaim_state;
2340 unsigned long writeback_threshold;
2341 bool aborted_reclaim;
2343 delayacct_freepages_start();
2345 if (global_reclaim(sc))
2346 count_vm_event(ALLOCSTALL);
2349 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2352 aborted_reclaim = shrink_zones(zonelist, sc);
2355 * Don't shrink slabs when reclaiming memory from
2356 * over limit cgroups
2358 if (global_reclaim(sc)) {
2359 unsigned long lru_pages = 0;
2360 for_each_zone_zonelist(zone, z, zonelist,
2361 gfp_zone(sc->gfp_mask)) {
2362 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2365 lru_pages += zone_reclaimable_pages(zone);
2368 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2369 if (reclaim_state) {
2370 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2371 reclaim_state->reclaimed_slab = 0;
2374 total_scanned += sc->nr_scanned;
2375 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2379 * If we're getting trouble reclaiming, start doing
2380 * writepage even in laptop mode.
2382 if (sc->priority < DEF_PRIORITY - 2)
2383 sc->may_writepage = 1;
2386 * Try to write back as many pages as we just scanned. This
2387 * tends to cause slow streaming writers to write data to the
2388 * disk smoothly, at the dirtying rate, which is nice. But
2389 * that's undesirable in laptop mode, where we *want* lumpy
2390 * writeout. So in laptop mode, write out the whole world.
2392 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2393 if (total_scanned > writeback_threshold) {
2394 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2395 WB_REASON_TRY_TO_FREE_PAGES);
2396 sc->may_writepage = 1;
2398 } while (--sc->priority >= 0);
2401 delayacct_freepages_end();
2403 if (sc->nr_reclaimed)
2404 return sc->nr_reclaimed;
2407 * As hibernation is going on, kswapd is freezed so that it can't mark
2408 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2411 if (oom_killer_disabled)
2414 /* Aborted reclaim to try compaction? don't OOM, then */
2415 if (aborted_reclaim)
2418 /* top priority shrink_zones still had more to do? don't OOM, then */
2419 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2425 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2428 unsigned long pfmemalloc_reserve = 0;
2429 unsigned long free_pages = 0;
2433 for (i = 0; i <= ZONE_NORMAL; i++) {
2434 zone = &pgdat->node_zones[i];
2435 pfmemalloc_reserve += min_wmark_pages(zone);
2436 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2439 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2441 /* kswapd must be awake if processes are being throttled */
2442 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2443 pgdat->classzone_idx = min(pgdat->classzone_idx,
2444 (enum zone_type)ZONE_NORMAL);
2445 wake_up_interruptible(&pgdat->kswapd_wait);
2452 * Throttle direct reclaimers if backing storage is backed by the network
2453 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2454 * depleted. kswapd will continue to make progress and wake the processes
2455 * when the low watermark is reached.
2457 * Returns true if a fatal signal was delivered during throttling. If this
2458 * happens, the page allocator should not consider triggering the OOM killer.
2460 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2461 nodemask_t *nodemask)
2464 int high_zoneidx = gfp_zone(gfp_mask);
2468 * Kernel threads should not be throttled as they may be indirectly
2469 * responsible for cleaning pages necessary for reclaim to make forward
2470 * progress. kjournald for example may enter direct reclaim while
2471 * committing a transaction where throttling it could forcing other
2472 * processes to block on log_wait_commit().
2474 if (current->flags & PF_KTHREAD)
2478 * If a fatal signal is pending, this process should not throttle.
2479 * It should return quickly so it can exit and free its memory
2481 if (fatal_signal_pending(current))
2484 /* Check if the pfmemalloc reserves are ok */
2485 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2486 pgdat = zone->zone_pgdat;
2487 if (pfmemalloc_watermark_ok(pgdat))
2490 /* Account for the throttling */
2491 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2494 * If the caller cannot enter the filesystem, it's possible that it
2495 * is due to the caller holding an FS lock or performing a journal
2496 * transaction in the case of a filesystem like ext[3|4]. In this case,
2497 * it is not safe to block on pfmemalloc_wait as kswapd could be
2498 * blocked waiting on the same lock. Instead, throttle for up to a
2499 * second before continuing.
2501 if (!(gfp_mask & __GFP_FS)) {
2502 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2503 pfmemalloc_watermark_ok(pgdat), HZ);
2508 /* Throttle until kswapd wakes the process */
2509 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2510 pfmemalloc_watermark_ok(pgdat));
2513 if (fatal_signal_pending(current))
2520 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2521 gfp_t gfp_mask, nodemask_t *nodemask)
2523 unsigned long nr_reclaimed;
2524 struct scan_control sc = {
2525 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2526 .may_writepage = !laptop_mode,
2527 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2531 .priority = DEF_PRIORITY,
2532 .target_mem_cgroup = NULL,
2533 .nodemask = nodemask,
2535 struct shrink_control shrink = {
2536 .gfp_mask = sc.gfp_mask,
2540 * Do not enter reclaim if fatal signal was delivered while throttled.
2541 * 1 is returned so that the page allocator does not OOM kill at this
2544 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2547 trace_mm_vmscan_direct_reclaim_begin(order,
2551 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2553 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2555 return nr_reclaimed;
2560 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2561 gfp_t gfp_mask, bool noswap,
2563 unsigned long *nr_scanned)
2565 struct scan_control sc = {
2567 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2568 .may_writepage = !laptop_mode,
2570 .may_swap = !noswap,
2573 .target_mem_cgroup = memcg,
2575 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2577 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2578 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2580 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2585 * NOTE: Although we can get the priority field, using it
2586 * here is not a good idea, since it limits the pages we can scan.
2587 * if we don't reclaim here, the shrink_zone from balance_pgdat
2588 * will pick up pages from other mem cgroup's as well. We hack
2589 * the priority and make it zero.
2591 shrink_lruvec(lruvec, &sc);
2593 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2595 *nr_scanned = sc.nr_scanned;
2596 return sc.nr_reclaimed;
2599 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2603 struct zonelist *zonelist;
2604 unsigned long nr_reclaimed;
2606 struct scan_control sc = {
2607 .may_writepage = !laptop_mode,
2609 .may_swap = !noswap,
2610 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2612 .priority = DEF_PRIORITY,
2613 .target_mem_cgroup = memcg,
2614 .nodemask = NULL, /* we don't care the placement */
2615 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2616 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2618 struct shrink_control shrink = {
2619 .gfp_mask = sc.gfp_mask,
2623 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2624 * take care of from where we get pages. So the node where we start the
2625 * scan does not need to be the current node.
2627 nid = mem_cgroup_select_victim_node(memcg);
2629 zonelist = NODE_DATA(nid)->node_zonelists;
2631 trace_mm_vmscan_memcg_reclaim_begin(0,
2635 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2637 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2639 return nr_reclaimed;
2643 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2645 struct mem_cgroup *memcg;
2647 if (!total_swap_pages)
2650 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2652 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2654 if (inactive_anon_is_low(lruvec))
2655 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2656 sc, LRU_ACTIVE_ANON);
2658 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2662 static bool zone_balanced(struct zone *zone, int order,
2663 unsigned long balance_gap, int classzone_idx)
2665 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2666 balance_gap, classzone_idx, 0))
2669 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2670 !compaction_suitable(zone, order))
2677 * pgdat_balanced() is used when checking if a node is balanced.
2679 * For order-0, all zones must be balanced!
2681 * For high-order allocations only zones that meet watermarks and are in a
2682 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2683 * total of balanced pages must be at least 25% of the zones allowed by
2684 * classzone_idx for the node to be considered balanced. Forcing all zones to
2685 * be balanced for high orders can cause excessive reclaim when there are
2687 * The choice of 25% is due to
2688 * o a 16M DMA zone that is balanced will not balance a zone on any
2689 * reasonable sized machine
2690 * o On all other machines, the top zone must be at least a reasonable
2691 * percentage of the middle zones. For example, on 32-bit x86, highmem
2692 * would need to be at least 256M for it to be balance a whole node.
2693 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2694 * to balance a node on its own. These seemed like reasonable ratios.
2696 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2698 unsigned long managed_pages = 0;
2699 unsigned long balanced_pages = 0;
2702 /* Check the watermark levels */
2703 for (i = 0; i <= classzone_idx; i++) {
2704 struct zone *zone = pgdat->node_zones + i;
2706 if (!populated_zone(zone))
2709 managed_pages += zone->managed_pages;
2712 * A special case here:
2714 * balance_pgdat() skips over all_unreclaimable after
2715 * DEF_PRIORITY. Effectively, it considers them balanced so
2716 * they must be considered balanced here as well!
2718 if (zone->all_unreclaimable) {
2719 balanced_pages += zone->managed_pages;
2723 if (zone_balanced(zone, order, 0, i))
2724 balanced_pages += zone->managed_pages;
2730 return balanced_pages >= (managed_pages >> 2);
2736 * Prepare kswapd for sleeping. This verifies that there are no processes
2737 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2739 * Returns true if kswapd is ready to sleep
2741 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2744 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2749 * There is a potential race between when kswapd checks its watermarks
2750 * and a process gets throttled. There is also a potential race if
2751 * processes get throttled, kswapd wakes, a large process exits therby
2752 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2753 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2754 * so wake them now if necessary. If necessary, processes will wake
2755 * kswapd and get throttled again
2757 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2758 wake_up(&pgdat->pfmemalloc_wait);
2762 return pgdat_balanced(pgdat, order, classzone_idx);
2766 * kswapd shrinks the zone by the number of pages required to reach
2767 * the high watermark.
2769 * Returns true if kswapd scanned at least the requested number of pages to
2770 * reclaim or if the lack of progress was due to pages under writeback.
2771 * This is used to determine if the scanning priority needs to be raised.
2773 static bool kswapd_shrink_zone(struct zone *zone,
2775 struct scan_control *sc,
2776 unsigned long lru_pages,
2777 unsigned long *nr_attempted)
2779 unsigned long nr_slab;
2780 int testorder = sc->order;
2781 unsigned long balance_gap;
2782 struct reclaim_state *reclaim_state = current->reclaim_state;
2783 struct shrink_control shrink = {
2784 .gfp_mask = sc->gfp_mask,
2786 bool lowmem_pressure;
2788 /* Reclaim above the high watermark. */
2789 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2792 * Kswapd reclaims only single pages with compaction enabled. Trying
2793 * too hard to reclaim until contiguous free pages have become
2794 * available can hurt performance by evicting too much useful data
2795 * from memory. Do not reclaim more than needed for compaction.
2797 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2798 compaction_suitable(zone, sc->order) !=
2803 * We put equal pressure on every zone, unless one zone has way too
2804 * many pages free already. The "too many pages" is defined as the
2805 * high wmark plus a "gap" where the gap is either the low
2806 * watermark or 1% of the zone, whichever is smaller.
2808 balance_gap = min(low_wmark_pages(zone),
2809 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2810 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2813 * If there is no low memory pressure or the zone is balanced then no
2814 * reclaim is necessary
2816 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2817 if (!lowmem_pressure && zone_balanced(zone, testorder,
2818 balance_gap, classzone_idx))
2821 shrink_zone(zone, sc);
2823 reclaim_state->reclaimed_slab = 0;
2824 nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2825 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2827 /* Account for the number of pages attempted to reclaim */
2828 *nr_attempted += sc->nr_to_reclaim;
2830 if (nr_slab == 0 && !zone_reclaimable(zone))
2831 zone->all_unreclaimable = 1;
2833 zone_clear_flag(zone, ZONE_WRITEBACK);
2836 * If a zone reaches its high watermark, consider it to be no longer
2837 * congested. It's possible there are dirty pages backed by congested
2838 * BDIs but as pressure is relieved, speculatively avoid congestion
2841 if (!zone->all_unreclaimable &&
2842 zone_balanced(zone, testorder, 0, classzone_idx)) {
2843 zone_clear_flag(zone, ZONE_CONGESTED);
2844 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2847 return sc->nr_scanned >= sc->nr_to_reclaim;
2851 * For kswapd, balance_pgdat() will work across all this node's zones until
2852 * they are all at high_wmark_pages(zone).
2854 * Returns the final order kswapd was reclaiming at
2856 * There is special handling here for zones which are full of pinned pages.
2857 * This can happen if the pages are all mlocked, or if they are all used by
2858 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2859 * What we do is to detect the case where all pages in the zone have been
2860 * scanned twice and there has been zero successful reclaim. Mark the zone as
2861 * dead and from now on, only perform a short scan. Basically we're polling
2862 * the zone for when the problem goes away.
2864 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2865 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2866 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2867 * lower zones regardless of the number of free pages in the lower zones. This
2868 * interoperates with the page allocator fallback scheme to ensure that aging
2869 * of pages is balanced across the zones.
2871 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2875 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2876 unsigned long nr_soft_reclaimed;
2877 unsigned long nr_soft_scanned;
2878 struct scan_control sc = {
2879 .gfp_mask = GFP_KERNEL,
2880 .priority = DEF_PRIORITY,
2883 .may_writepage = !laptop_mode,
2885 .target_mem_cgroup = NULL,
2887 count_vm_event(PAGEOUTRUN);
2890 unsigned long lru_pages = 0;
2891 unsigned long nr_attempted = 0;
2892 bool raise_priority = true;
2893 bool pgdat_needs_compaction = (order > 0);
2895 sc.nr_reclaimed = 0;
2898 * Scan in the highmem->dma direction for the highest
2899 * zone which needs scanning
2901 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2902 struct zone *zone = pgdat->node_zones + i;
2904 if (!populated_zone(zone))
2907 if (zone->all_unreclaimable &&
2908 sc.priority != DEF_PRIORITY)
2912 * Do some background aging of the anon list, to give
2913 * pages a chance to be referenced before reclaiming.
2915 age_active_anon(zone, &sc);
2918 * If the number of buffer_heads in the machine
2919 * exceeds the maximum allowed level and this node
2920 * has a highmem zone, force kswapd to reclaim from
2921 * it to relieve lowmem pressure.
2923 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2928 if (!zone_balanced(zone, order, 0, 0)) {
2933 * If balanced, clear the dirty and congested
2936 zone_clear_flag(zone, ZONE_CONGESTED);
2937 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2944 for (i = 0; i <= end_zone; i++) {
2945 struct zone *zone = pgdat->node_zones + i;
2947 if (!populated_zone(zone))
2950 lru_pages += zone_reclaimable_pages(zone);
2953 * If any zone is currently balanced then kswapd will
2954 * not call compaction as it is expected that the
2955 * necessary pages are already available.
2957 if (pgdat_needs_compaction &&
2958 zone_watermark_ok(zone, order,
2959 low_wmark_pages(zone),
2961 pgdat_needs_compaction = false;
2965 * If we're getting trouble reclaiming, start doing writepage
2966 * even in laptop mode.
2968 if (sc.priority < DEF_PRIORITY - 2)
2969 sc.may_writepage = 1;
2972 * Now scan the zone in the dma->highmem direction, stopping
2973 * at the last zone which needs scanning.
2975 * We do this because the page allocator works in the opposite
2976 * direction. This prevents the page allocator from allocating
2977 * pages behind kswapd's direction of progress, which would
2978 * cause too much scanning of the lower zones.
2980 for (i = 0; i <= end_zone; i++) {
2981 struct zone *zone = pgdat->node_zones + i;
2983 if (!populated_zone(zone))
2986 if (zone->all_unreclaimable &&
2987 sc.priority != DEF_PRIORITY)
2992 nr_soft_scanned = 0;
2994 * Call soft limit reclaim before calling shrink_zone.
2996 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2999 sc.nr_reclaimed += nr_soft_reclaimed;
3002 * There should be no need to raise the scanning
3003 * priority if enough pages are already being scanned
3004 * that that high watermark would be met at 100%
3007 if (kswapd_shrink_zone(zone, end_zone, &sc,
3008 lru_pages, &nr_attempted))
3009 raise_priority = false;
3013 * If the low watermark is met there is no need for processes
3014 * to be throttled on pfmemalloc_wait as they should not be
3015 * able to safely make forward progress. Wake them
3017 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3018 pfmemalloc_watermark_ok(pgdat))
3019 wake_up(&pgdat->pfmemalloc_wait);
3022 * Fragmentation may mean that the system cannot be rebalanced
3023 * for high-order allocations in all zones. If twice the
3024 * allocation size has been reclaimed and the zones are still
3025 * not balanced then recheck the watermarks at order-0 to
3026 * prevent kswapd reclaiming excessively. Assume that a
3027 * process requested a high-order can direct reclaim/compact.
3029 if (order && sc.nr_reclaimed >= 2UL << order)
3030 order = sc.order = 0;
3032 /* Check if kswapd should be suspending */
3033 if (try_to_freeze() || kthread_should_stop())
3037 * Compact if necessary and kswapd is reclaiming at least the
3038 * high watermark number of pages as requsted
3040 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3041 compact_pgdat(pgdat, order);
3044 * Raise priority if scanning rate is too low or there was no
3045 * progress in reclaiming pages
3047 if (raise_priority || !sc.nr_reclaimed)
3049 } while (sc.priority >= 1 &&
3050 !pgdat_balanced(pgdat, order, *classzone_idx));
3054 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3055 * makes a decision on the order we were last reclaiming at. However,
3056 * if another caller entered the allocator slow path while kswapd
3057 * was awake, order will remain at the higher level
3059 *classzone_idx = end_zone;
3063 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3068 if (freezing(current) || kthread_should_stop())
3071 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3073 /* Try to sleep for a short interval */
3074 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3075 remaining = schedule_timeout(HZ/10);
3076 finish_wait(&pgdat->kswapd_wait, &wait);
3077 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3081 * After a short sleep, check if it was a premature sleep. If not, then
3082 * go fully to sleep until explicitly woken up.
3084 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3085 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3088 * vmstat counters are not perfectly accurate and the estimated
3089 * value for counters such as NR_FREE_PAGES can deviate from the
3090 * true value by nr_online_cpus * threshold. To avoid the zone
3091 * watermarks being breached while under pressure, we reduce the
3092 * per-cpu vmstat threshold while kswapd is awake and restore
3093 * them before going back to sleep.
3095 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3098 * Compaction records what page blocks it recently failed to
3099 * isolate pages from and skips them in the future scanning.
3100 * When kswapd is going to sleep, it is reasonable to assume
3101 * that pages and compaction may succeed so reset the cache.
3103 reset_isolation_suitable(pgdat);
3105 if (!kthread_should_stop())
3108 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3111 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3113 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3115 finish_wait(&pgdat->kswapd_wait, &wait);
3119 * The background pageout daemon, started as a kernel thread
3120 * from the init process.
3122 * This basically trickles out pages so that we have _some_
3123 * free memory available even if there is no other activity
3124 * that frees anything up. This is needed for things like routing
3125 * etc, where we otherwise might have all activity going on in
3126 * asynchronous contexts that cannot page things out.
3128 * If there are applications that are active memory-allocators
3129 * (most normal use), this basically shouldn't matter.
3131 static int kswapd(void *p)
3133 unsigned long order, new_order;
3134 unsigned balanced_order;
3135 int classzone_idx, new_classzone_idx;
3136 int balanced_classzone_idx;
3137 pg_data_t *pgdat = (pg_data_t*)p;
3138 struct task_struct *tsk = current;
3140 struct reclaim_state reclaim_state = {
3141 .reclaimed_slab = 0,
3143 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3145 lockdep_set_current_reclaim_state(GFP_KERNEL);
3147 if (!cpumask_empty(cpumask))
3148 set_cpus_allowed_ptr(tsk, cpumask);
3149 current->reclaim_state = &reclaim_state;
3152 * Tell the memory management that we're a "memory allocator",
3153 * and that if we need more memory we should get access to it
3154 * regardless (see "__alloc_pages()"). "kswapd" should
3155 * never get caught in the normal page freeing logic.
3157 * (Kswapd normally doesn't need memory anyway, but sometimes
3158 * you need a small amount of memory in order to be able to
3159 * page out something else, and this flag essentially protects
3160 * us from recursively trying to free more memory as we're
3161 * trying to free the first piece of memory in the first place).
3163 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3166 order = new_order = 0;
3168 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3169 balanced_classzone_idx = classzone_idx;
3174 * If the last balance_pgdat was unsuccessful it's unlikely a
3175 * new request of a similar or harder type will succeed soon
3176 * so consider going to sleep on the basis we reclaimed at
3178 if (balanced_classzone_idx >= new_classzone_idx &&
3179 balanced_order == new_order) {
3180 new_order = pgdat->kswapd_max_order;
3181 new_classzone_idx = pgdat->classzone_idx;
3182 pgdat->kswapd_max_order = 0;
3183 pgdat->classzone_idx = pgdat->nr_zones - 1;
3186 if (order < new_order || classzone_idx > new_classzone_idx) {
3188 * Don't sleep if someone wants a larger 'order'
3189 * allocation or has tigher zone constraints
3192 classzone_idx = new_classzone_idx;
3194 kswapd_try_to_sleep(pgdat, balanced_order,
3195 balanced_classzone_idx);
3196 order = pgdat->kswapd_max_order;
3197 classzone_idx = pgdat->classzone_idx;
3199 new_classzone_idx = classzone_idx;
3200 pgdat->kswapd_max_order = 0;
3201 pgdat->classzone_idx = pgdat->nr_zones - 1;
3204 ret = try_to_freeze();
3205 if (kthread_should_stop())
3209 * We can speed up thawing tasks if we don't call balance_pgdat
3210 * after returning from the refrigerator
3213 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3214 balanced_classzone_idx = classzone_idx;
3215 balanced_order = balance_pgdat(pgdat, order,
3216 &balanced_classzone_idx);
3220 current->reclaim_state = NULL;
3225 * A zone is low on free memory, so wake its kswapd task to service it.
3227 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3231 if (!populated_zone(zone))
3234 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3236 pgdat = zone->zone_pgdat;
3237 if (pgdat->kswapd_max_order < order) {
3238 pgdat->kswapd_max_order = order;
3239 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3241 if (!waitqueue_active(&pgdat->kswapd_wait))
3243 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3246 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3247 wake_up_interruptible(&pgdat->kswapd_wait);
3251 * The reclaimable count would be mostly accurate.
3252 * The less reclaimable pages may be
3253 * - mlocked pages, which will be moved to unevictable list when encountered
3254 * - mapped pages, which may require several travels to be reclaimed
3255 * - dirty pages, which is not "instantly" reclaimable
3257 unsigned long global_reclaimable_pages(void)
3261 nr = global_page_state(NR_ACTIVE_FILE) +
3262 global_page_state(NR_INACTIVE_FILE);
3264 if (get_nr_swap_pages() > 0)
3265 nr += global_page_state(NR_ACTIVE_ANON) +
3266 global_page_state(NR_INACTIVE_ANON);
3271 unsigned long zone_reclaimable_pages(struct zone *zone)
3275 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3276 zone_page_state(zone, NR_INACTIVE_FILE);
3278 if (get_nr_swap_pages() > 0)
3279 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3280 zone_page_state(zone, NR_INACTIVE_ANON);
3285 #ifdef CONFIG_HIBERNATION
3287 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3290 * Rather than trying to age LRUs the aim is to preserve the overall
3291 * LRU order by reclaiming preferentially
3292 * inactive > active > active referenced > active mapped
3294 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3296 struct reclaim_state reclaim_state;
3297 struct scan_control sc = {
3298 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3302 .nr_to_reclaim = nr_to_reclaim,
3303 .hibernation_mode = 1,
3305 .priority = DEF_PRIORITY,
3307 struct shrink_control shrink = {
3308 .gfp_mask = sc.gfp_mask,
3310 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3311 struct task_struct *p = current;
3312 unsigned long nr_reclaimed;
3314 p->flags |= PF_MEMALLOC;
3315 lockdep_set_current_reclaim_state(sc.gfp_mask);
3316 reclaim_state.reclaimed_slab = 0;
3317 p->reclaim_state = &reclaim_state;
3319 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3321 p->reclaim_state = NULL;
3322 lockdep_clear_current_reclaim_state();
3323 p->flags &= ~PF_MEMALLOC;
3325 return nr_reclaimed;
3327 #endif /* CONFIG_HIBERNATION */
3329 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3330 not required for correctness. So if the last cpu in a node goes
3331 away, we get changed to run anywhere: as the first one comes back,
3332 restore their cpu bindings. */
3333 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3338 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3339 for_each_node_state(nid, N_MEMORY) {
3340 pg_data_t *pgdat = NODE_DATA(nid);
3341 const struct cpumask *mask;
3343 mask = cpumask_of_node(pgdat->node_id);
3345 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3346 /* One of our CPUs online: restore mask */
3347 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3354 * This kswapd start function will be called by init and node-hot-add.
3355 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3357 int kswapd_run(int nid)
3359 pg_data_t *pgdat = NODE_DATA(nid);
3365 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3366 if (IS_ERR(pgdat->kswapd)) {
3367 /* failure at boot is fatal */
3368 BUG_ON(system_state == SYSTEM_BOOTING);
3369 pr_err("Failed to start kswapd on node %d\n", nid);
3370 ret = PTR_ERR(pgdat->kswapd);
3371 pgdat->kswapd = NULL;
3377 * Called by memory hotplug when all memory in a node is offlined. Caller must
3378 * hold lock_memory_hotplug().
3380 void kswapd_stop(int nid)
3382 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3385 kthread_stop(kswapd);
3386 NODE_DATA(nid)->kswapd = NULL;
3390 static int __init kswapd_init(void)
3395 for_each_node_state(nid, N_MEMORY)
3397 hotcpu_notifier(cpu_callback, 0);
3401 module_init(kswapd_init)
3407 * If non-zero call zone_reclaim when the number of free pages falls below
3410 int zone_reclaim_mode __read_mostly;
3412 #define RECLAIM_OFF 0
3413 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3414 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3415 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3418 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3419 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3422 #define ZONE_RECLAIM_PRIORITY 4
3425 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3428 int sysctl_min_unmapped_ratio = 1;
3431 * If the number of slab pages in a zone grows beyond this percentage then
3432 * slab reclaim needs to occur.
3434 int sysctl_min_slab_ratio = 5;
3436 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3438 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3439 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3440 zone_page_state(zone, NR_ACTIVE_FILE);
3443 * It's possible for there to be more file mapped pages than
3444 * accounted for by the pages on the file LRU lists because
3445 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3447 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3450 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3451 static long zone_pagecache_reclaimable(struct zone *zone)
3453 long nr_pagecache_reclaimable;
3457 * If RECLAIM_SWAP is set, then all file pages are considered
3458 * potentially reclaimable. Otherwise, we have to worry about
3459 * pages like swapcache and zone_unmapped_file_pages() provides
3462 if (zone_reclaim_mode & RECLAIM_SWAP)
3463 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3465 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3467 /* If we can't clean pages, remove dirty pages from consideration */
3468 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3469 delta += zone_page_state(zone, NR_FILE_DIRTY);
3471 /* Watch for any possible underflows due to delta */
3472 if (unlikely(delta > nr_pagecache_reclaimable))
3473 delta = nr_pagecache_reclaimable;
3475 return nr_pagecache_reclaimable - delta;
3479 * Try to free up some pages from this zone through reclaim.
3481 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3483 /* Minimum pages needed in order to stay on node */
3484 const unsigned long nr_pages = 1 << order;
3485 struct task_struct *p = current;
3486 struct reclaim_state reclaim_state;
3487 struct scan_control sc = {
3488 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3489 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3491 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3492 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3494 .priority = ZONE_RECLAIM_PRIORITY,
3496 struct shrink_control shrink = {
3497 .gfp_mask = sc.gfp_mask,
3499 unsigned long nr_slab_pages0, nr_slab_pages1;
3503 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3504 * and we also need to be able to write out pages for RECLAIM_WRITE
3507 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3508 lockdep_set_current_reclaim_state(gfp_mask);
3509 reclaim_state.reclaimed_slab = 0;
3510 p->reclaim_state = &reclaim_state;
3512 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3514 * Free memory by calling shrink zone with increasing
3515 * priorities until we have enough memory freed.
3518 shrink_zone(zone, &sc);
3519 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3522 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3523 if (nr_slab_pages0 > zone->min_slab_pages) {
3525 * shrink_slab() does not currently allow us to determine how
3526 * many pages were freed in this zone. So we take the current
3527 * number of slab pages and shake the slab until it is reduced
3528 * by the same nr_pages that we used for reclaiming unmapped
3531 * Note that shrink_slab will free memory on all zones and may
3535 unsigned long lru_pages = zone_reclaimable_pages(zone);
3537 /* No reclaimable slab or very low memory pressure */
3538 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3541 /* Freed enough memory */
3542 nr_slab_pages1 = zone_page_state(zone,
3543 NR_SLAB_RECLAIMABLE);
3544 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3549 * Update nr_reclaimed by the number of slab pages we
3550 * reclaimed from this zone.
3552 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3553 if (nr_slab_pages1 < nr_slab_pages0)
3554 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3557 p->reclaim_state = NULL;
3558 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3559 lockdep_clear_current_reclaim_state();
3560 return sc.nr_reclaimed >= nr_pages;
3563 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3569 * Zone reclaim reclaims unmapped file backed pages and
3570 * slab pages if we are over the defined limits.
3572 * A small portion of unmapped file backed pages is needed for
3573 * file I/O otherwise pages read by file I/O will be immediately
3574 * thrown out if the zone is overallocated. So we do not reclaim
3575 * if less than a specified percentage of the zone is used by
3576 * unmapped file backed pages.
3578 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3579 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3580 return ZONE_RECLAIM_FULL;
3582 if (zone->all_unreclaimable)
3583 return ZONE_RECLAIM_FULL;
3586 * Do not scan if the allocation should not be delayed.
3588 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3589 return ZONE_RECLAIM_NOSCAN;
3592 * Only run zone reclaim on the local zone or on zones that do not
3593 * have associated processors. This will favor the local processor
3594 * over remote processors and spread off node memory allocations
3595 * as wide as possible.
3597 node_id = zone_to_nid(zone);
3598 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3599 return ZONE_RECLAIM_NOSCAN;
3601 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3602 return ZONE_RECLAIM_NOSCAN;
3604 ret = __zone_reclaim(zone, gfp_mask, order);
3605 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3608 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3615 * page_evictable - test whether a page is evictable
3616 * @page: the page to test
3618 * Test whether page is evictable--i.e., should be placed on active/inactive
3619 * lists vs unevictable list.
3621 * Reasons page might not be evictable:
3622 * (1) page's mapping marked unevictable
3623 * (2) page is part of an mlocked VMA
3626 int page_evictable(struct page *page)
3628 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3633 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3634 * @pages: array of pages to check
3635 * @nr_pages: number of pages to check
3637 * Checks pages for evictability and moves them to the appropriate lru list.
3639 * This function is only used for SysV IPC SHM_UNLOCK.
3641 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3643 struct lruvec *lruvec;
3644 struct zone *zone = NULL;
3649 for (i = 0; i < nr_pages; i++) {
3650 struct page *page = pages[i];
3651 struct zone *pagezone;
3654 pagezone = page_zone(page);
3655 if (pagezone != zone) {
3657 spin_unlock_irq(&zone->lru_lock);
3659 spin_lock_irq(&zone->lru_lock);
3661 lruvec = mem_cgroup_page_lruvec(page, zone);
3663 if (!PageLRU(page) || !PageUnevictable(page))
3666 if (page_evictable(page)) {
3667 enum lru_list lru = page_lru_base_type(page);
3669 VM_BUG_ON(PageActive(page));
3670 ClearPageUnevictable(page);
3671 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3672 add_page_to_lru_list(page, lruvec, lru);
3678 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3679 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3680 spin_unlock_irq(&zone->lru_lock);
3683 #endif /* CONFIG_SHMEM */
3685 static void warn_scan_unevictable_pages(void)
3687 printk_once(KERN_WARNING
3688 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3689 "disabled for lack of a legitimate use case. If you have "
3690 "one, please send an email to linux-mm@kvack.org.\n",
3695 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3696 * all nodes' unevictable lists for evictable pages
3698 unsigned long scan_unevictable_pages;
3700 int scan_unevictable_handler(struct ctl_table *table, int write,
3701 void __user *buffer,
3702 size_t *length, loff_t *ppos)
3704 warn_scan_unevictable_pages();
3705 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3706 scan_unevictable_pages = 0;
3712 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3713 * a specified node's per zone unevictable lists for evictable pages.
3716 static ssize_t read_scan_unevictable_node(struct device *dev,
3717 struct device_attribute *attr,
3720 warn_scan_unevictable_pages();
3721 return sprintf(buf, "0\n"); /* always zero; should fit... */
3724 static ssize_t write_scan_unevictable_node(struct device *dev,
3725 struct device_attribute *attr,
3726 const char *buf, size_t count)
3728 warn_scan_unevictable_pages();
3733 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3734 read_scan_unevictable_node,
3735 write_scan_unevictable_node);
3737 int scan_unevictable_register_node(struct node *node)
3739 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3742 void scan_unevictable_unregister_node(struct node *node)
3744 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);