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_unqueued_dirty,
699 unsigned long *ret_nr_writeback,
700 unsigned long *ret_nr_immediate,
703 LIST_HEAD(ret_pages);
704 LIST_HEAD(free_pages);
706 unsigned long nr_unqueued_dirty = 0;
707 unsigned long nr_dirty = 0;
708 unsigned long nr_congested = 0;
709 unsigned long nr_reclaimed = 0;
710 unsigned long nr_writeback = 0;
711 unsigned long nr_immediate = 0;
715 mem_cgroup_uncharge_start();
716 while (!list_empty(page_list)) {
717 struct address_space *mapping;
720 enum page_references references = PAGEREF_RECLAIM_CLEAN;
721 bool dirty, writeback;
725 page = lru_to_page(page_list);
726 list_del(&page->lru);
728 if (!trylock_page(page))
731 VM_BUG_ON(PageActive(page));
732 VM_BUG_ON(page_zone(page) != zone);
736 if (unlikely(!page_evictable(page)))
739 if (!sc->may_unmap && page_mapped(page))
742 /* Double the slab pressure for mapped and swapcache pages */
743 if (page_mapped(page) || PageSwapCache(page))
746 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
747 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
750 * The number of dirty pages determines if a zone is marked
751 * reclaim_congested which affects wait_iff_congested. kswapd
752 * will stall and start writing pages if the tail of the LRU
753 * is all dirty unqueued pages.
755 page_check_dirty_writeback(page, &dirty, &writeback);
756 if (dirty || writeback)
759 if (dirty && !writeback)
762 /* Treat this page as congested if underlying BDI is */
763 mapping = page_mapping(page);
764 if (mapping && bdi_write_congested(mapping->backing_dev_info))
768 * If a page at the tail of the LRU is under writeback, there
769 * are three cases to consider.
771 * 1) If reclaim is encountering an excessive number of pages
772 * under writeback and this page is both under writeback and
773 * PageReclaim then it indicates that pages are being queued
774 * for IO but are being recycled through the LRU before the
775 * IO can complete. Waiting on the page itself risks an
776 * indefinite stall if it is impossible to writeback the
777 * page due to IO error or disconnected storage so instead
778 * note that the LRU is being scanned too quickly and the
779 * caller can stall after page list has been processed.
781 * 2) Global reclaim encounters a page, memcg encounters a
782 * page that is not marked for immediate reclaim or
783 * the caller does not have __GFP_IO. In this case mark
784 * the page for immediate reclaim and continue scanning.
786 * __GFP_IO is checked because a loop driver thread might
787 * enter reclaim, and deadlock if it waits on a page for
788 * which it is needed to do the write (loop masks off
789 * __GFP_IO|__GFP_FS for this reason); but more thought
790 * would probably show more reasons.
792 * Don't require __GFP_FS, since we're not going into the
793 * FS, just waiting on its writeback completion. Worryingly,
794 * ext4 gfs2 and xfs allocate pages with
795 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
796 * may_enter_fs here is liable to OOM on them.
798 * 3) memcg encounters a page that is not already marked
799 * PageReclaim. memcg does not have any dirty pages
800 * throttling so we could easily OOM just because too many
801 * pages are in writeback and there is nothing else to
802 * reclaim. Wait for the writeback to complete.
804 if (PageWriteback(page)) {
806 if (current_is_kswapd() &&
808 zone_is_reclaim_writeback(zone)) {
813 } else if (global_reclaim(sc) ||
814 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
816 * This is slightly racy - end_page_writeback()
817 * might have just cleared PageReclaim, then
818 * setting PageReclaim here end up interpreted
819 * as PageReadahead - but that does not matter
820 * enough to care. What we do want is for this
821 * page to have PageReclaim set next time memcg
822 * reclaim reaches the tests above, so it will
823 * then wait_on_page_writeback() to avoid OOM;
824 * and it's also appropriate in global reclaim.
826 SetPageReclaim(page);
833 wait_on_page_writeback(page);
838 references = page_check_references(page, sc);
840 switch (references) {
841 case PAGEREF_ACTIVATE:
842 goto activate_locked;
845 case PAGEREF_RECLAIM:
846 case PAGEREF_RECLAIM_CLEAN:
847 ; /* try to reclaim the page below */
851 * Anonymous process memory has backing store?
852 * Try to allocate it some swap space here.
854 if (PageAnon(page) && !PageSwapCache(page)) {
855 if (!(sc->gfp_mask & __GFP_IO))
857 if (!add_to_swap(page, page_list))
858 goto activate_locked;
861 /* Adding to swap updated mapping */
862 mapping = page_mapping(page);
866 * The page is mapped into the page tables of one or more
867 * processes. Try to unmap it here.
869 if (page_mapped(page) && mapping) {
870 switch (try_to_unmap(page, ttu_flags)) {
872 goto activate_locked;
878 ; /* try to free the page below */
882 if (PageDirty(page)) {
884 * Only kswapd can writeback filesystem pages to
885 * avoid risk of stack overflow but only writeback
886 * if many dirty pages have been encountered.
888 if (page_is_file_cache(page) &&
889 (!current_is_kswapd() ||
890 !zone_is_reclaim_dirty(zone))) {
892 * Immediately reclaim when written back.
893 * Similar in principal to deactivate_page()
894 * except we already have the page isolated
895 * and know it's dirty
897 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
898 SetPageReclaim(page);
903 if (references == PAGEREF_RECLAIM_CLEAN)
907 if (!sc->may_writepage)
910 /* Page is dirty, try to write it out here */
911 switch (pageout(page, mapping, sc)) {
915 goto activate_locked;
917 if (PageWriteback(page))
923 * A synchronous write - probably a ramdisk. Go
924 * ahead and try to reclaim the page.
926 if (!trylock_page(page))
928 if (PageDirty(page) || PageWriteback(page))
930 mapping = page_mapping(page);
932 ; /* try to free the page below */
937 * If the page has buffers, try to free the buffer mappings
938 * associated with this page. If we succeed we try to free
941 * We do this even if the page is PageDirty().
942 * try_to_release_page() does not perform I/O, but it is
943 * possible for a page to have PageDirty set, but it is actually
944 * clean (all its buffers are clean). This happens if the
945 * buffers were written out directly, with submit_bh(). ext3
946 * will do this, as well as the blockdev mapping.
947 * try_to_release_page() will discover that cleanness and will
948 * drop the buffers and mark the page clean - it can be freed.
950 * Rarely, pages can have buffers and no ->mapping. These are
951 * the pages which were not successfully invalidated in
952 * truncate_complete_page(). We try to drop those buffers here
953 * and if that worked, and the page is no longer mapped into
954 * process address space (page_count == 1) it can be freed.
955 * Otherwise, leave the page on the LRU so it is swappable.
957 if (page_has_private(page)) {
958 if (!try_to_release_page(page, sc->gfp_mask))
959 goto activate_locked;
960 if (!mapping && page_count(page) == 1) {
962 if (put_page_testzero(page))
966 * rare race with speculative reference.
967 * the speculative reference will free
968 * this page shortly, so we may
969 * increment nr_reclaimed here (and
970 * leave it off the LRU).
978 if (!mapping || !__remove_mapping(mapping, page))
982 * At this point, we have no other references and there is
983 * no way to pick any more up (removed from LRU, removed
984 * from pagecache). Can use non-atomic bitops now (and
985 * we obviously don't have to worry about waking up a process
986 * waiting on the page lock, because there are no references.
988 __clear_page_locked(page);
993 * Is there need to periodically free_page_list? It would
994 * appear not as the counts should be low
996 list_add(&page->lru, &free_pages);
1000 if (PageSwapCache(page))
1001 try_to_free_swap(page);
1003 putback_lru_page(page);
1007 /* Not a candidate for swapping, so reclaim swap space. */
1008 if (PageSwapCache(page) && vm_swap_full())
1009 try_to_free_swap(page);
1010 VM_BUG_ON(PageActive(page));
1011 SetPageActive(page);
1016 list_add(&page->lru, &ret_pages);
1017 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1021 * Tag a zone as congested if all the dirty pages encountered were
1022 * backed by a congested BDI. In this case, reclaimers should just
1023 * back off and wait for congestion to clear because further reclaim
1024 * will encounter the same problem
1026 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1027 zone_set_flag(zone, ZONE_CONGESTED);
1029 free_hot_cold_page_list(&free_pages, 1);
1031 list_splice(&ret_pages, page_list);
1032 count_vm_events(PGACTIVATE, pgactivate);
1033 mem_cgroup_uncharge_end();
1034 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1035 *ret_nr_writeback += nr_writeback;
1036 *ret_nr_immediate += nr_immediate;
1037 return nr_reclaimed;
1040 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1041 struct list_head *page_list)
1043 struct scan_control sc = {
1044 .gfp_mask = GFP_KERNEL,
1045 .priority = DEF_PRIORITY,
1048 unsigned long ret, dummy1, dummy2, dummy3;
1049 struct page *page, *next;
1050 LIST_HEAD(clean_pages);
1052 list_for_each_entry_safe(page, next, page_list, lru) {
1053 if (page_is_file_cache(page) && !PageDirty(page)) {
1054 ClearPageActive(page);
1055 list_move(&page->lru, &clean_pages);
1059 ret = shrink_page_list(&clean_pages, zone, &sc,
1060 TTU_UNMAP|TTU_IGNORE_ACCESS,
1061 &dummy1, &dummy2, &dummy3, true);
1062 list_splice(&clean_pages, page_list);
1063 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1068 * Attempt to remove the specified page from its LRU. Only take this page
1069 * if it is of the appropriate PageActive status. Pages which are being
1070 * freed elsewhere are also ignored.
1072 * page: page to consider
1073 * mode: one of the LRU isolation modes defined above
1075 * returns 0 on success, -ve errno on failure.
1077 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1081 /* Only take pages on the LRU. */
1085 /* Compaction should not handle unevictable pages but CMA can do so */
1086 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1092 * To minimise LRU disruption, the caller can indicate that it only
1093 * wants to isolate pages it will be able to operate on without
1094 * blocking - clean pages for the most part.
1096 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1097 * is used by reclaim when it is cannot write to backing storage
1099 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1100 * that it is possible to migrate without blocking
1102 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1103 /* All the caller can do on PageWriteback is block */
1104 if (PageWriteback(page))
1107 if (PageDirty(page)) {
1108 struct address_space *mapping;
1110 /* ISOLATE_CLEAN means only clean pages */
1111 if (mode & ISOLATE_CLEAN)
1115 * Only pages without mappings or that have a
1116 * ->migratepage callback are possible to migrate
1119 mapping = page_mapping(page);
1120 if (mapping && !mapping->a_ops->migratepage)
1125 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1128 if (likely(get_page_unless_zero(page))) {
1130 * Be careful not to clear PageLRU until after we're
1131 * sure the page is not being freed elsewhere -- the
1132 * page release code relies on it.
1142 * zone->lru_lock is heavily contended. Some of the functions that
1143 * shrink the lists perform better by taking out a batch of pages
1144 * and working on them outside the LRU lock.
1146 * For pagecache intensive workloads, this function is the hottest
1147 * spot in the kernel (apart from copy_*_user functions).
1149 * Appropriate locks must be held before calling this function.
1151 * @nr_to_scan: The number of pages to look through on the list.
1152 * @lruvec: The LRU vector to pull pages from.
1153 * @dst: The temp list to put pages on to.
1154 * @nr_scanned: The number of pages that were scanned.
1155 * @sc: The scan_control struct for this reclaim session
1156 * @mode: One of the LRU isolation modes
1157 * @lru: LRU list id for isolating
1159 * returns how many pages were moved onto *@dst.
1161 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1162 struct lruvec *lruvec, struct list_head *dst,
1163 unsigned long *nr_scanned, struct scan_control *sc,
1164 isolate_mode_t mode, enum lru_list lru)
1166 struct list_head *src = &lruvec->lists[lru];
1167 unsigned long nr_taken = 0;
1170 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1174 page = lru_to_page(src);
1175 prefetchw_prev_lru_page(page, src, flags);
1177 VM_BUG_ON(!PageLRU(page));
1179 switch (__isolate_lru_page(page, mode)) {
1181 nr_pages = hpage_nr_pages(page);
1182 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1183 list_move(&page->lru, dst);
1184 nr_taken += nr_pages;
1188 /* else it is being freed elsewhere */
1189 list_move(&page->lru, src);
1198 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1199 nr_taken, mode, is_file_lru(lru));
1204 * isolate_lru_page - tries to isolate a page from its LRU list
1205 * @page: page to isolate from its LRU list
1207 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1208 * vmstat statistic corresponding to whatever LRU list the page was on.
1210 * Returns 0 if the page was removed from an LRU list.
1211 * Returns -EBUSY if the page was not on an LRU list.
1213 * The returned page will have PageLRU() cleared. If it was found on
1214 * the active list, it will have PageActive set. If it was found on
1215 * the unevictable list, it will have the PageUnevictable bit set. That flag
1216 * may need to be cleared by the caller before letting the page go.
1218 * The vmstat statistic corresponding to the list on which the page was
1219 * found will be decremented.
1222 * (1) Must be called with an elevated refcount on the page. This is a
1223 * fundamentnal difference from isolate_lru_pages (which is called
1224 * without a stable reference).
1225 * (2) the lru_lock must not be held.
1226 * (3) interrupts must be enabled.
1228 int isolate_lru_page(struct page *page)
1232 VM_BUG_ON(!page_count(page));
1234 if (PageLRU(page)) {
1235 struct zone *zone = page_zone(page);
1236 struct lruvec *lruvec;
1238 spin_lock_irq(&zone->lru_lock);
1239 lruvec = mem_cgroup_page_lruvec(page, zone);
1240 if (PageLRU(page)) {
1241 int lru = page_lru(page);
1244 del_page_from_lru_list(page, lruvec, lru);
1247 spin_unlock_irq(&zone->lru_lock);
1253 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1254 * then get resheduled. When there are massive number of tasks doing page
1255 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1256 * the LRU list will go small and be scanned faster than necessary, leading to
1257 * unnecessary swapping, thrashing and OOM.
1259 static int too_many_isolated(struct zone *zone, int file,
1260 struct scan_control *sc)
1262 unsigned long inactive, isolated;
1264 if (current_is_kswapd())
1267 if (!global_reclaim(sc))
1271 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1272 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1274 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1275 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1279 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1280 * won't get blocked by normal direct-reclaimers, forming a circular
1283 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1286 return isolated > inactive;
1289 static noinline_for_stack void
1290 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1292 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1293 struct zone *zone = lruvec_zone(lruvec);
1294 LIST_HEAD(pages_to_free);
1297 * Put back any unfreeable pages.
1299 while (!list_empty(page_list)) {
1300 struct page *page = lru_to_page(page_list);
1303 VM_BUG_ON(PageLRU(page));
1304 list_del(&page->lru);
1305 if (unlikely(!page_evictable(page))) {
1306 spin_unlock_irq(&zone->lru_lock);
1307 putback_lru_page(page);
1308 spin_lock_irq(&zone->lru_lock);
1312 lruvec = mem_cgroup_page_lruvec(page, zone);
1315 lru = page_lru(page);
1316 add_page_to_lru_list(page, lruvec, lru);
1318 if (is_active_lru(lru)) {
1319 int file = is_file_lru(lru);
1320 int numpages = hpage_nr_pages(page);
1321 reclaim_stat->recent_rotated[file] += numpages;
1323 if (put_page_testzero(page)) {
1324 __ClearPageLRU(page);
1325 __ClearPageActive(page);
1326 del_page_from_lru_list(page, lruvec, lru);
1328 if (unlikely(PageCompound(page))) {
1329 spin_unlock_irq(&zone->lru_lock);
1330 (*get_compound_page_dtor(page))(page);
1331 spin_lock_irq(&zone->lru_lock);
1333 list_add(&page->lru, &pages_to_free);
1338 * To save our caller's stack, now use input list for pages to free.
1340 list_splice(&pages_to_free, page_list);
1344 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1345 * of reclaimed pages
1347 static noinline_for_stack unsigned long
1348 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1349 struct scan_control *sc, enum lru_list lru)
1351 LIST_HEAD(page_list);
1352 unsigned long nr_scanned;
1353 unsigned long nr_reclaimed = 0;
1354 unsigned long nr_taken;
1355 unsigned long nr_unqueued_dirty = 0;
1356 unsigned long nr_writeback = 0;
1357 unsigned long nr_immediate = 0;
1358 isolate_mode_t isolate_mode = 0;
1359 int file = is_file_lru(lru);
1360 struct zone *zone = lruvec_zone(lruvec);
1361 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1363 while (unlikely(too_many_isolated(zone, file, sc))) {
1364 congestion_wait(BLK_RW_ASYNC, HZ/10);
1366 /* We are about to die and free our memory. Return now. */
1367 if (fatal_signal_pending(current))
1368 return SWAP_CLUSTER_MAX;
1374 isolate_mode |= ISOLATE_UNMAPPED;
1375 if (!sc->may_writepage)
1376 isolate_mode |= ISOLATE_CLEAN;
1378 spin_lock_irq(&zone->lru_lock);
1380 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1381 &nr_scanned, sc, isolate_mode, lru);
1383 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1384 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1386 if (global_reclaim(sc)) {
1387 zone->pages_scanned += nr_scanned;
1388 if (current_is_kswapd())
1389 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1391 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1393 spin_unlock_irq(&zone->lru_lock);
1398 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1399 &nr_unqueued_dirty, &nr_writeback, &nr_immediate,
1402 spin_lock_irq(&zone->lru_lock);
1404 reclaim_stat->recent_scanned[file] += nr_taken;
1406 if (global_reclaim(sc)) {
1407 if (current_is_kswapd())
1408 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1411 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1415 putback_inactive_pages(lruvec, &page_list);
1417 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1419 spin_unlock_irq(&zone->lru_lock);
1421 free_hot_cold_page_list(&page_list, 1);
1424 * If reclaim is isolating dirty pages under writeback, it implies
1425 * that the long-lived page allocation rate is exceeding the page
1426 * laundering rate. Either the global limits are not being effective
1427 * at throttling processes due to the page distribution throughout
1428 * zones or there is heavy usage of a slow backing device. The
1429 * only option is to throttle from reclaim context which is not ideal
1430 * as there is no guarantee the dirtying process is throttled in the
1431 * same way balance_dirty_pages() manages.
1433 * This scales the number of dirty pages that must be under writeback
1434 * before throttling depending on priority. It is a simple backoff
1435 * function that has the most effect in the range DEF_PRIORITY to
1436 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1437 * in trouble and reclaim is considered to be in trouble.
1439 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1440 * DEF_PRIORITY-1 50% must be PageWriteback
1441 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1443 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1444 * isolated page is PageWriteback
1446 if (nr_writeback && nr_writeback >=
1447 (nr_taken >> (DEF_PRIORITY - sc->priority))) {
1448 zone_set_flag(zone, ZONE_WRITEBACK);
1449 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1453 * memcg will stall in page writeback so only consider forcibly
1454 * stalling for global reclaim
1456 if (global_reclaim(sc)) {
1458 * If dirty pages are scanned that are not queued for IO, it
1459 * implies that flushers are not keeping up. In this case, flag
1460 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1461 * pages from reclaim context. It will forcibly stall in the
1464 if (nr_unqueued_dirty == nr_taken)
1465 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1468 * In addition, if kswapd scans pages marked marked for
1469 * immediate reclaim and under writeback (nr_immediate), it
1470 * implies that pages are cycling through the LRU faster than
1471 * they are written so also forcibly stall.
1473 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1474 congestion_wait(BLK_RW_ASYNC, HZ/10);
1477 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1479 nr_scanned, nr_reclaimed,
1481 trace_shrink_flags(file));
1482 return nr_reclaimed;
1486 * This moves pages from the active list to the inactive list.
1488 * We move them the other way if the page is referenced by one or more
1489 * processes, from rmap.
1491 * If the pages are mostly unmapped, the processing is fast and it is
1492 * appropriate to hold zone->lru_lock across the whole operation. But if
1493 * the pages are mapped, the processing is slow (page_referenced()) so we
1494 * should drop zone->lru_lock around each page. It's impossible to balance
1495 * this, so instead we remove the pages from the LRU while processing them.
1496 * It is safe to rely on PG_active against the non-LRU pages in here because
1497 * nobody will play with that bit on a non-LRU page.
1499 * The downside is that we have to touch page->_count against each page.
1500 * But we had to alter page->flags anyway.
1503 static void move_active_pages_to_lru(struct lruvec *lruvec,
1504 struct list_head *list,
1505 struct list_head *pages_to_free,
1508 struct zone *zone = lruvec_zone(lruvec);
1509 unsigned long pgmoved = 0;
1513 while (!list_empty(list)) {
1514 page = lru_to_page(list);
1515 lruvec = mem_cgroup_page_lruvec(page, zone);
1517 VM_BUG_ON(PageLRU(page));
1520 nr_pages = hpage_nr_pages(page);
1521 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1522 list_move(&page->lru, &lruvec->lists[lru]);
1523 pgmoved += nr_pages;
1525 if (put_page_testzero(page)) {
1526 __ClearPageLRU(page);
1527 __ClearPageActive(page);
1528 del_page_from_lru_list(page, lruvec, lru);
1530 if (unlikely(PageCompound(page))) {
1531 spin_unlock_irq(&zone->lru_lock);
1532 (*get_compound_page_dtor(page))(page);
1533 spin_lock_irq(&zone->lru_lock);
1535 list_add(&page->lru, pages_to_free);
1538 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1539 if (!is_active_lru(lru))
1540 __count_vm_events(PGDEACTIVATE, pgmoved);
1543 static void shrink_active_list(unsigned long nr_to_scan,
1544 struct lruvec *lruvec,
1545 struct scan_control *sc,
1548 unsigned long nr_taken;
1549 unsigned long nr_scanned;
1550 unsigned long vm_flags;
1551 LIST_HEAD(l_hold); /* The pages which were snipped off */
1552 LIST_HEAD(l_active);
1553 LIST_HEAD(l_inactive);
1555 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1556 unsigned long nr_rotated = 0;
1557 isolate_mode_t isolate_mode = 0;
1558 int file = is_file_lru(lru);
1559 struct zone *zone = lruvec_zone(lruvec);
1564 isolate_mode |= ISOLATE_UNMAPPED;
1565 if (!sc->may_writepage)
1566 isolate_mode |= ISOLATE_CLEAN;
1568 spin_lock_irq(&zone->lru_lock);
1570 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1571 &nr_scanned, sc, isolate_mode, lru);
1572 if (global_reclaim(sc))
1573 zone->pages_scanned += nr_scanned;
1575 reclaim_stat->recent_scanned[file] += nr_taken;
1577 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1578 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1579 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1580 spin_unlock_irq(&zone->lru_lock);
1582 while (!list_empty(&l_hold)) {
1584 page = lru_to_page(&l_hold);
1585 list_del(&page->lru);
1587 if (unlikely(!page_evictable(page))) {
1588 putback_lru_page(page);
1592 if (unlikely(buffer_heads_over_limit)) {
1593 if (page_has_private(page) && trylock_page(page)) {
1594 if (page_has_private(page))
1595 try_to_release_page(page, 0);
1600 if (page_referenced(page, 0, sc->target_mem_cgroup,
1602 nr_rotated += hpage_nr_pages(page);
1604 * Identify referenced, file-backed active pages and
1605 * give them one more trip around the active list. So
1606 * that executable code get better chances to stay in
1607 * memory under moderate memory pressure. Anon pages
1608 * are not likely to be evicted by use-once streaming
1609 * IO, plus JVM can create lots of anon VM_EXEC pages,
1610 * so we ignore them here.
1612 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1613 list_add(&page->lru, &l_active);
1618 ClearPageActive(page); /* we are de-activating */
1619 list_add(&page->lru, &l_inactive);
1623 * Move pages back to the lru list.
1625 spin_lock_irq(&zone->lru_lock);
1627 * Count referenced pages from currently used mappings as rotated,
1628 * even though only some of them are actually re-activated. This
1629 * helps balance scan pressure between file and anonymous pages in
1632 reclaim_stat->recent_rotated[file] += nr_rotated;
1634 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1635 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1636 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1637 spin_unlock_irq(&zone->lru_lock);
1639 free_hot_cold_page_list(&l_hold, 1);
1643 static int inactive_anon_is_low_global(struct zone *zone)
1645 unsigned long active, inactive;
1647 active = zone_page_state(zone, NR_ACTIVE_ANON);
1648 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1650 if (inactive * zone->inactive_ratio < active)
1657 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1658 * @lruvec: LRU vector to check
1660 * Returns true if the zone does not have enough inactive anon pages,
1661 * meaning some active anon pages need to be deactivated.
1663 static int inactive_anon_is_low(struct lruvec *lruvec)
1666 * If we don't have swap space, anonymous page deactivation
1669 if (!total_swap_pages)
1672 if (!mem_cgroup_disabled())
1673 return mem_cgroup_inactive_anon_is_low(lruvec);
1675 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1678 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1685 * inactive_file_is_low - check if file pages need to be deactivated
1686 * @lruvec: LRU vector to check
1688 * When the system is doing streaming IO, memory pressure here
1689 * ensures that active file pages get deactivated, until more
1690 * than half of the file pages are on the inactive list.
1692 * Once we get to that situation, protect the system's working
1693 * set from being evicted by disabling active file page aging.
1695 * This uses a different ratio than the anonymous pages, because
1696 * the page cache uses a use-once replacement algorithm.
1698 static int inactive_file_is_low(struct lruvec *lruvec)
1700 unsigned long inactive;
1701 unsigned long active;
1703 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1704 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1706 return active > inactive;
1709 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1711 if (is_file_lru(lru))
1712 return inactive_file_is_low(lruvec);
1714 return inactive_anon_is_low(lruvec);
1717 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1718 struct lruvec *lruvec, struct scan_control *sc)
1720 if (is_active_lru(lru)) {
1721 if (inactive_list_is_low(lruvec, lru))
1722 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1726 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1729 static int vmscan_swappiness(struct scan_control *sc)
1731 if (global_reclaim(sc))
1732 return vm_swappiness;
1733 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1744 * Determine how aggressively the anon and file LRU lists should be
1745 * scanned. The relative value of each set of LRU lists is determined
1746 * by looking at the fraction of the pages scanned we did rotate back
1747 * onto the active list instead of evict.
1749 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1750 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1752 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1755 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1757 u64 denominator = 0; /* gcc */
1758 struct zone *zone = lruvec_zone(lruvec);
1759 unsigned long anon_prio, file_prio;
1760 enum scan_balance scan_balance;
1761 unsigned long anon, file, free;
1762 bool force_scan = false;
1763 unsigned long ap, fp;
1767 * If the zone or memcg is small, nr[l] can be 0. This
1768 * results in no scanning on this priority and a potential
1769 * priority drop. Global direct reclaim can go to the next
1770 * zone and tends to have no problems. Global kswapd is for
1771 * zone balancing and it needs to scan a minimum amount. When
1772 * reclaiming for a memcg, a priority drop can cause high
1773 * latencies, so it's better to scan a minimum amount there as
1776 if (current_is_kswapd() && zone->all_unreclaimable)
1778 if (!global_reclaim(sc))
1781 /* If we have no swap space, do not bother scanning anon pages. */
1782 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1783 scan_balance = SCAN_FILE;
1788 * Global reclaim will swap to prevent OOM even with no
1789 * swappiness, but memcg users want to use this knob to
1790 * disable swapping for individual groups completely when
1791 * using the memory controller's swap limit feature would be
1794 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1795 scan_balance = SCAN_FILE;
1800 * Do not apply any pressure balancing cleverness when the
1801 * system is close to OOM, scan both anon and file equally
1802 * (unless the swappiness setting disagrees with swapping).
1804 if (!sc->priority && vmscan_swappiness(sc)) {
1805 scan_balance = SCAN_EQUAL;
1809 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1810 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1811 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1812 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1815 * If it's foreseeable that reclaiming the file cache won't be
1816 * enough to get the zone back into a desirable shape, we have
1817 * to swap. Better start now and leave the - probably heavily
1818 * thrashing - remaining file pages alone.
1820 if (global_reclaim(sc)) {
1821 free = zone_page_state(zone, NR_FREE_PAGES);
1822 if (unlikely(file + free <= high_wmark_pages(zone))) {
1823 scan_balance = SCAN_ANON;
1829 * There is enough inactive page cache, do not reclaim
1830 * anything from the anonymous working set right now.
1832 if (!inactive_file_is_low(lruvec)) {
1833 scan_balance = SCAN_FILE;
1837 scan_balance = SCAN_FRACT;
1840 * With swappiness at 100, anonymous and file have the same priority.
1841 * This scanning priority is essentially the inverse of IO cost.
1843 anon_prio = vmscan_swappiness(sc);
1844 file_prio = 200 - anon_prio;
1847 * OK, so we have swap space and a fair amount of page cache
1848 * pages. We use the recently rotated / recently scanned
1849 * ratios to determine how valuable each cache is.
1851 * Because workloads change over time (and to avoid overflow)
1852 * we keep these statistics as a floating average, which ends
1853 * up weighing recent references more than old ones.
1855 * anon in [0], file in [1]
1857 spin_lock_irq(&zone->lru_lock);
1858 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1859 reclaim_stat->recent_scanned[0] /= 2;
1860 reclaim_stat->recent_rotated[0] /= 2;
1863 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1864 reclaim_stat->recent_scanned[1] /= 2;
1865 reclaim_stat->recent_rotated[1] /= 2;
1869 * The amount of pressure on anon vs file pages is inversely
1870 * proportional to the fraction of recently scanned pages on
1871 * each list that were recently referenced and in active use.
1873 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1874 ap /= reclaim_stat->recent_rotated[0] + 1;
1876 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1877 fp /= reclaim_stat->recent_rotated[1] + 1;
1878 spin_unlock_irq(&zone->lru_lock);
1882 denominator = ap + fp + 1;
1884 for_each_evictable_lru(lru) {
1885 int file = is_file_lru(lru);
1889 size = get_lru_size(lruvec, lru);
1890 scan = size >> sc->priority;
1892 if (!scan && force_scan)
1893 scan = min(size, SWAP_CLUSTER_MAX);
1895 switch (scan_balance) {
1897 /* Scan lists relative to size */
1901 * Scan types proportional to swappiness and
1902 * their relative recent reclaim efficiency.
1904 scan = div64_u64(scan * fraction[file], denominator);
1908 /* Scan one type exclusively */
1909 if ((scan_balance == SCAN_FILE) != file)
1913 /* Look ma, no brain */
1921 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1923 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1925 unsigned long nr[NR_LRU_LISTS];
1926 unsigned long targets[NR_LRU_LISTS];
1927 unsigned long nr_to_scan;
1929 unsigned long nr_reclaimed = 0;
1930 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1931 struct blk_plug plug;
1932 bool scan_adjusted = false;
1934 get_scan_count(lruvec, sc, nr);
1936 /* Record the original scan target for proportional adjustments later */
1937 memcpy(targets, nr, sizeof(nr));
1939 blk_start_plug(&plug);
1940 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1941 nr[LRU_INACTIVE_FILE]) {
1942 unsigned long nr_anon, nr_file, percentage;
1943 unsigned long nr_scanned;
1945 for_each_evictable_lru(lru) {
1947 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1948 nr[lru] -= nr_to_scan;
1950 nr_reclaimed += shrink_list(lru, nr_to_scan,
1955 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
1959 * For global direct reclaim, reclaim only the number of pages
1960 * requested. Less care is taken to scan proportionally as it
1961 * is more important to minimise direct reclaim stall latency
1962 * than it is to properly age the LRU lists.
1964 if (global_reclaim(sc) && !current_is_kswapd())
1968 * For kswapd and memcg, reclaim at least the number of pages
1969 * requested. Ensure that the anon and file LRUs shrink
1970 * proportionally what was requested by get_scan_count(). We
1971 * stop reclaiming one LRU and reduce the amount scanning
1972 * proportional to the original scan target.
1974 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
1975 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
1977 if (nr_file > nr_anon) {
1978 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
1979 targets[LRU_ACTIVE_ANON] + 1;
1981 percentage = nr_anon * 100 / scan_target;
1983 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
1984 targets[LRU_ACTIVE_FILE] + 1;
1986 percentage = nr_file * 100 / scan_target;
1989 /* Stop scanning the smaller of the LRU */
1991 nr[lru + LRU_ACTIVE] = 0;
1994 * Recalculate the other LRU scan count based on its original
1995 * scan target and the percentage scanning already complete
1997 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
1998 nr_scanned = targets[lru] - nr[lru];
1999 nr[lru] = targets[lru] * (100 - percentage) / 100;
2000 nr[lru] -= min(nr[lru], nr_scanned);
2003 nr_scanned = targets[lru] - nr[lru];
2004 nr[lru] = targets[lru] * (100 - percentage) / 100;
2005 nr[lru] -= min(nr[lru], nr_scanned);
2007 scan_adjusted = true;
2009 blk_finish_plug(&plug);
2010 sc->nr_reclaimed += nr_reclaimed;
2013 * Even if we did not try to evict anon pages at all, we want to
2014 * rebalance the anon lru active/inactive ratio.
2016 if (inactive_anon_is_low(lruvec))
2017 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2018 sc, LRU_ACTIVE_ANON);
2020 throttle_vm_writeout(sc->gfp_mask);
2023 /* Use reclaim/compaction for costly allocs or under memory pressure */
2024 static bool in_reclaim_compaction(struct scan_control *sc)
2026 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2027 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2028 sc->priority < DEF_PRIORITY - 2))
2035 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2036 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2037 * true if more pages should be reclaimed such that when the page allocator
2038 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2039 * It will give up earlier than that if there is difficulty reclaiming pages.
2041 static inline bool should_continue_reclaim(struct zone *zone,
2042 unsigned long nr_reclaimed,
2043 unsigned long nr_scanned,
2044 struct scan_control *sc)
2046 unsigned long pages_for_compaction;
2047 unsigned long inactive_lru_pages;
2049 /* If not in reclaim/compaction mode, stop */
2050 if (!in_reclaim_compaction(sc))
2053 /* Consider stopping depending on scan and reclaim activity */
2054 if (sc->gfp_mask & __GFP_REPEAT) {
2056 * For __GFP_REPEAT allocations, stop reclaiming if the
2057 * full LRU list has been scanned and we are still failing
2058 * to reclaim pages. This full LRU scan is potentially
2059 * expensive but a __GFP_REPEAT caller really wants to succeed
2061 if (!nr_reclaimed && !nr_scanned)
2065 * For non-__GFP_REPEAT allocations which can presumably
2066 * fail without consequence, stop if we failed to reclaim
2067 * any pages from the last SWAP_CLUSTER_MAX number of
2068 * pages that were scanned. This will return to the
2069 * caller faster at the risk reclaim/compaction and
2070 * the resulting allocation attempt fails
2077 * If we have not reclaimed enough pages for compaction and the
2078 * inactive lists are large enough, continue reclaiming
2080 pages_for_compaction = (2UL << sc->order);
2081 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2082 if (get_nr_swap_pages() > 0)
2083 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2084 if (sc->nr_reclaimed < pages_for_compaction &&
2085 inactive_lru_pages > pages_for_compaction)
2088 /* If compaction would go ahead or the allocation would succeed, stop */
2089 switch (compaction_suitable(zone, sc->order)) {
2090 case COMPACT_PARTIAL:
2091 case COMPACT_CONTINUE:
2098 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2100 unsigned long nr_reclaimed, nr_scanned;
2103 struct mem_cgroup *root = sc->target_mem_cgroup;
2104 struct mem_cgroup_reclaim_cookie reclaim = {
2106 .priority = sc->priority,
2108 struct mem_cgroup *memcg;
2110 nr_reclaimed = sc->nr_reclaimed;
2111 nr_scanned = sc->nr_scanned;
2113 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2115 struct lruvec *lruvec;
2117 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2119 shrink_lruvec(lruvec, sc);
2122 * Direct reclaim and kswapd have to scan all memory
2123 * cgroups to fulfill the overall scan target for the
2126 * Limit reclaim, on the other hand, only cares about
2127 * nr_to_reclaim pages to be reclaimed and it will
2128 * retry with decreasing priority if one round over the
2129 * whole hierarchy is not sufficient.
2131 if (!global_reclaim(sc) &&
2132 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2133 mem_cgroup_iter_break(root, memcg);
2136 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2139 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2140 sc->nr_scanned - nr_scanned,
2141 sc->nr_reclaimed - nr_reclaimed);
2143 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2144 sc->nr_scanned - nr_scanned, sc));
2147 /* Returns true if compaction should go ahead for a high-order request */
2148 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2150 unsigned long balance_gap, watermark;
2153 /* Do not consider compaction for orders reclaim is meant to satisfy */
2154 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2158 * Compaction takes time to run and there are potentially other
2159 * callers using the pages just freed. Continue reclaiming until
2160 * there is a buffer of free pages available to give compaction
2161 * a reasonable chance of completing and allocating the page
2163 balance_gap = min(low_wmark_pages(zone),
2164 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2165 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2166 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2167 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2170 * If compaction is deferred, reclaim up to a point where
2171 * compaction will have a chance of success when re-enabled
2173 if (compaction_deferred(zone, sc->order))
2174 return watermark_ok;
2176 /* If compaction is not ready to start, keep reclaiming */
2177 if (!compaction_suitable(zone, sc->order))
2180 return watermark_ok;
2184 * This is the direct reclaim path, for page-allocating processes. We only
2185 * try to reclaim pages from zones which will satisfy the caller's allocation
2188 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2190 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2192 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2193 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2194 * zone defense algorithm.
2196 * If a zone is deemed to be full of pinned pages then just give it a light
2197 * scan then give up on it.
2199 * This function returns true if a zone is being reclaimed for a costly
2200 * high-order allocation and compaction is ready to begin. This indicates to
2201 * the caller that it should consider retrying the allocation instead of
2204 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2208 unsigned long nr_soft_reclaimed;
2209 unsigned long nr_soft_scanned;
2210 bool aborted_reclaim = false;
2213 * If the number of buffer_heads in the machine exceeds the maximum
2214 * allowed level, force direct reclaim to scan the highmem zone as
2215 * highmem pages could be pinning lowmem pages storing buffer_heads
2217 if (buffer_heads_over_limit)
2218 sc->gfp_mask |= __GFP_HIGHMEM;
2220 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2221 gfp_zone(sc->gfp_mask), sc->nodemask) {
2222 if (!populated_zone(zone))
2225 * Take care memory controller reclaiming has small influence
2228 if (global_reclaim(sc)) {
2229 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2231 if (zone->all_unreclaimable &&
2232 sc->priority != DEF_PRIORITY)
2233 continue; /* Let kswapd poll it */
2234 if (IS_ENABLED(CONFIG_COMPACTION)) {
2236 * If we already have plenty of memory free for
2237 * compaction in this zone, don't free any more.
2238 * Even though compaction is invoked for any
2239 * non-zero order, only frequent costly order
2240 * reclamation is disruptive enough to become a
2241 * noticeable problem, like transparent huge
2244 if (compaction_ready(zone, sc)) {
2245 aborted_reclaim = true;
2250 * This steals pages from memory cgroups over softlimit
2251 * and returns the number of reclaimed pages and
2252 * scanned pages. This works for global memory pressure
2253 * and balancing, not for a memcg's limit.
2255 nr_soft_scanned = 0;
2256 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2257 sc->order, sc->gfp_mask,
2259 sc->nr_reclaimed += nr_soft_reclaimed;
2260 sc->nr_scanned += nr_soft_scanned;
2261 /* need some check for avoid more shrink_zone() */
2264 shrink_zone(zone, sc);
2267 return aborted_reclaim;
2270 static bool zone_reclaimable(struct zone *zone)
2272 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2275 /* All zones in zonelist are unreclaimable? */
2276 static bool all_unreclaimable(struct zonelist *zonelist,
2277 struct scan_control *sc)
2282 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2283 gfp_zone(sc->gfp_mask), sc->nodemask) {
2284 if (!populated_zone(zone))
2286 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2288 if (!zone->all_unreclaimable)
2296 * This is the main entry point to direct page reclaim.
2298 * If a full scan of the inactive list fails to free enough memory then we
2299 * are "out of memory" and something needs to be killed.
2301 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2302 * high - the zone may be full of dirty or under-writeback pages, which this
2303 * caller can't do much about. We kick the writeback threads and take explicit
2304 * naps in the hope that some of these pages can be written. But if the
2305 * allocating task holds filesystem locks which prevent writeout this might not
2306 * work, and the allocation attempt will fail.
2308 * returns: 0, if no pages reclaimed
2309 * else, the number of pages reclaimed
2311 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2312 struct scan_control *sc,
2313 struct shrink_control *shrink)
2315 unsigned long total_scanned = 0;
2316 struct reclaim_state *reclaim_state = current->reclaim_state;
2319 unsigned long writeback_threshold;
2320 bool aborted_reclaim;
2322 delayacct_freepages_start();
2324 if (global_reclaim(sc))
2325 count_vm_event(ALLOCSTALL);
2328 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2331 aborted_reclaim = shrink_zones(zonelist, sc);
2334 * Don't shrink slabs when reclaiming memory from
2335 * over limit cgroups
2337 if (global_reclaim(sc)) {
2338 unsigned long lru_pages = 0;
2339 for_each_zone_zonelist(zone, z, zonelist,
2340 gfp_zone(sc->gfp_mask)) {
2341 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2344 lru_pages += zone_reclaimable_pages(zone);
2347 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2348 if (reclaim_state) {
2349 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2350 reclaim_state->reclaimed_slab = 0;
2353 total_scanned += sc->nr_scanned;
2354 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2358 * If we're getting trouble reclaiming, start doing
2359 * writepage even in laptop mode.
2361 if (sc->priority < DEF_PRIORITY - 2)
2362 sc->may_writepage = 1;
2365 * Try to write back as many pages as we just scanned. This
2366 * tends to cause slow streaming writers to write data to the
2367 * disk smoothly, at the dirtying rate, which is nice. But
2368 * that's undesirable in laptop mode, where we *want* lumpy
2369 * writeout. So in laptop mode, write out the whole world.
2371 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2372 if (total_scanned > writeback_threshold) {
2373 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2374 WB_REASON_TRY_TO_FREE_PAGES);
2375 sc->may_writepage = 1;
2378 /* Take a nap, wait for some writeback to complete */
2379 if (!sc->hibernation_mode && sc->nr_scanned &&
2380 sc->priority < DEF_PRIORITY - 2) {
2381 struct zone *preferred_zone;
2383 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2384 &cpuset_current_mems_allowed,
2386 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2388 } while (--sc->priority >= 0);
2391 delayacct_freepages_end();
2393 if (sc->nr_reclaimed)
2394 return sc->nr_reclaimed;
2397 * As hibernation is going on, kswapd is freezed so that it can't mark
2398 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2401 if (oom_killer_disabled)
2404 /* Aborted reclaim to try compaction? don't OOM, then */
2405 if (aborted_reclaim)
2408 /* top priority shrink_zones still had more to do? don't OOM, then */
2409 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2415 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2418 unsigned long pfmemalloc_reserve = 0;
2419 unsigned long free_pages = 0;
2423 for (i = 0; i <= ZONE_NORMAL; i++) {
2424 zone = &pgdat->node_zones[i];
2425 pfmemalloc_reserve += min_wmark_pages(zone);
2426 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2429 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2431 /* kswapd must be awake if processes are being throttled */
2432 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2433 pgdat->classzone_idx = min(pgdat->classzone_idx,
2434 (enum zone_type)ZONE_NORMAL);
2435 wake_up_interruptible(&pgdat->kswapd_wait);
2442 * Throttle direct reclaimers if backing storage is backed by the network
2443 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2444 * depleted. kswapd will continue to make progress and wake the processes
2445 * when the low watermark is reached.
2447 * Returns true if a fatal signal was delivered during throttling. If this
2448 * happens, the page allocator should not consider triggering the OOM killer.
2450 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2451 nodemask_t *nodemask)
2454 int high_zoneidx = gfp_zone(gfp_mask);
2458 * Kernel threads should not be throttled as they may be indirectly
2459 * responsible for cleaning pages necessary for reclaim to make forward
2460 * progress. kjournald for example may enter direct reclaim while
2461 * committing a transaction where throttling it could forcing other
2462 * processes to block on log_wait_commit().
2464 if (current->flags & PF_KTHREAD)
2468 * If a fatal signal is pending, this process should not throttle.
2469 * It should return quickly so it can exit and free its memory
2471 if (fatal_signal_pending(current))
2474 /* Check if the pfmemalloc reserves are ok */
2475 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2476 pgdat = zone->zone_pgdat;
2477 if (pfmemalloc_watermark_ok(pgdat))
2480 /* Account for the throttling */
2481 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2484 * If the caller cannot enter the filesystem, it's possible that it
2485 * is due to the caller holding an FS lock or performing a journal
2486 * transaction in the case of a filesystem like ext[3|4]. In this case,
2487 * it is not safe to block on pfmemalloc_wait as kswapd could be
2488 * blocked waiting on the same lock. Instead, throttle for up to a
2489 * second before continuing.
2491 if (!(gfp_mask & __GFP_FS)) {
2492 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2493 pfmemalloc_watermark_ok(pgdat), HZ);
2498 /* Throttle until kswapd wakes the process */
2499 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2500 pfmemalloc_watermark_ok(pgdat));
2503 if (fatal_signal_pending(current))
2510 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2511 gfp_t gfp_mask, nodemask_t *nodemask)
2513 unsigned long nr_reclaimed;
2514 struct scan_control sc = {
2515 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2516 .may_writepage = !laptop_mode,
2517 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2521 .priority = DEF_PRIORITY,
2522 .target_mem_cgroup = NULL,
2523 .nodemask = nodemask,
2525 struct shrink_control shrink = {
2526 .gfp_mask = sc.gfp_mask,
2530 * Do not enter reclaim if fatal signal was delivered while throttled.
2531 * 1 is returned so that the page allocator does not OOM kill at this
2534 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2537 trace_mm_vmscan_direct_reclaim_begin(order,
2541 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2543 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2545 return nr_reclaimed;
2550 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2551 gfp_t gfp_mask, bool noswap,
2553 unsigned long *nr_scanned)
2555 struct scan_control sc = {
2557 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2558 .may_writepage = !laptop_mode,
2560 .may_swap = !noswap,
2563 .target_mem_cgroup = memcg,
2565 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2567 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2568 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2570 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2575 * NOTE: Although we can get the priority field, using it
2576 * here is not a good idea, since it limits the pages we can scan.
2577 * if we don't reclaim here, the shrink_zone from balance_pgdat
2578 * will pick up pages from other mem cgroup's as well. We hack
2579 * the priority and make it zero.
2581 shrink_lruvec(lruvec, &sc);
2583 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2585 *nr_scanned = sc.nr_scanned;
2586 return sc.nr_reclaimed;
2589 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2593 struct zonelist *zonelist;
2594 unsigned long nr_reclaimed;
2596 struct scan_control sc = {
2597 .may_writepage = !laptop_mode,
2599 .may_swap = !noswap,
2600 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2602 .priority = DEF_PRIORITY,
2603 .target_mem_cgroup = memcg,
2604 .nodemask = NULL, /* we don't care the placement */
2605 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2606 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2608 struct shrink_control shrink = {
2609 .gfp_mask = sc.gfp_mask,
2613 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2614 * take care of from where we get pages. So the node where we start the
2615 * scan does not need to be the current node.
2617 nid = mem_cgroup_select_victim_node(memcg);
2619 zonelist = NODE_DATA(nid)->node_zonelists;
2621 trace_mm_vmscan_memcg_reclaim_begin(0,
2625 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2627 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2629 return nr_reclaimed;
2633 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2635 struct mem_cgroup *memcg;
2637 if (!total_swap_pages)
2640 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2642 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2644 if (inactive_anon_is_low(lruvec))
2645 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2646 sc, LRU_ACTIVE_ANON);
2648 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2652 static bool zone_balanced(struct zone *zone, int order,
2653 unsigned long balance_gap, int classzone_idx)
2655 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2656 balance_gap, classzone_idx, 0))
2659 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2660 !compaction_suitable(zone, order))
2667 * pgdat_balanced() is used when checking if a node is balanced.
2669 * For order-0, all zones must be balanced!
2671 * For high-order allocations only zones that meet watermarks and are in a
2672 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2673 * total of balanced pages must be at least 25% of the zones allowed by
2674 * classzone_idx for the node to be considered balanced. Forcing all zones to
2675 * be balanced for high orders can cause excessive reclaim when there are
2677 * The choice of 25% is due to
2678 * o a 16M DMA zone that is balanced will not balance a zone on any
2679 * reasonable sized machine
2680 * o On all other machines, the top zone must be at least a reasonable
2681 * percentage of the middle zones. For example, on 32-bit x86, highmem
2682 * would need to be at least 256M for it to be balance a whole node.
2683 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2684 * to balance a node on its own. These seemed like reasonable ratios.
2686 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2688 unsigned long managed_pages = 0;
2689 unsigned long balanced_pages = 0;
2692 /* Check the watermark levels */
2693 for (i = 0; i <= classzone_idx; i++) {
2694 struct zone *zone = pgdat->node_zones + i;
2696 if (!populated_zone(zone))
2699 managed_pages += zone->managed_pages;
2702 * A special case here:
2704 * balance_pgdat() skips over all_unreclaimable after
2705 * DEF_PRIORITY. Effectively, it considers them balanced so
2706 * they must be considered balanced here as well!
2708 if (zone->all_unreclaimable) {
2709 balanced_pages += zone->managed_pages;
2713 if (zone_balanced(zone, order, 0, i))
2714 balanced_pages += zone->managed_pages;
2720 return balanced_pages >= (managed_pages >> 2);
2726 * Prepare kswapd for sleeping. This verifies that there are no processes
2727 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2729 * Returns true if kswapd is ready to sleep
2731 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2734 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2739 * There is a potential race between when kswapd checks its watermarks
2740 * and a process gets throttled. There is also a potential race if
2741 * processes get throttled, kswapd wakes, a large process exits therby
2742 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2743 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2744 * so wake them now if necessary. If necessary, processes will wake
2745 * kswapd and get throttled again
2747 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2748 wake_up(&pgdat->pfmemalloc_wait);
2752 return pgdat_balanced(pgdat, order, classzone_idx);
2756 * kswapd shrinks the zone by the number of pages required to reach
2757 * the high watermark.
2759 * Returns true if kswapd scanned at least the requested number of pages to
2760 * reclaim or if the lack of progress was due to pages under writeback.
2761 * This is used to determine if the scanning priority needs to be raised.
2763 static bool kswapd_shrink_zone(struct zone *zone,
2765 struct scan_control *sc,
2766 unsigned long lru_pages,
2767 unsigned long *nr_attempted)
2769 unsigned long nr_slab;
2770 int testorder = sc->order;
2771 unsigned long balance_gap;
2772 struct reclaim_state *reclaim_state = current->reclaim_state;
2773 struct shrink_control shrink = {
2774 .gfp_mask = sc->gfp_mask,
2776 bool lowmem_pressure;
2778 /* Reclaim above the high watermark. */
2779 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2782 * Kswapd reclaims only single pages with compaction enabled. Trying
2783 * too hard to reclaim until contiguous free pages have become
2784 * available can hurt performance by evicting too much useful data
2785 * from memory. Do not reclaim more than needed for compaction.
2787 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2788 compaction_suitable(zone, sc->order) !=
2793 * We put equal pressure on every zone, unless one zone has way too
2794 * many pages free already. The "too many pages" is defined as the
2795 * high wmark plus a "gap" where the gap is either the low
2796 * watermark or 1% of the zone, whichever is smaller.
2798 balance_gap = min(low_wmark_pages(zone),
2799 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2800 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2803 * If there is no low memory pressure or the zone is balanced then no
2804 * reclaim is necessary
2806 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2807 if (!lowmem_pressure && zone_balanced(zone, testorder,
2808 balance_gap, classzone_idx))
2811 shrink_zone(zone, sc);
2813 reclaim_state->reclaimed_slab = 0;
2814 nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2815 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2817 /* Account for the number of pages attempted to reclaim */
2818 *nr_attempted += sc->nr_to_reclaim;
2820 if (nr_slab == 0 && !zone_reclaimable(zone))
2821 zone->all_unreclaimable = 1;
2823 zone_clear_flag(zone, ZONE_WRITEBACK);
2826 * If a zone reaches its high watermark, consider it to be no longer
2827 * congested. It's possible there are dirty pages backed by congested
2828 * BDIs but as pressure is relieved, speculatively avoid congestion
2831 if (!zone->all_unreclaimable &&
2832 zone_balanced(zone, testorder, 0, classzone_idx)) {
2833 zone_clear_flag(zone, ZONE_CONGESTED);
2834 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2837 return sc->nr_scanned >= sc->nr_to_reclaim;
2841 * For kswapd, balance_pgdat() will work across all this node's zones until
2842 * they are all at high_wmark_pages(zone).
2844 * Returns the final order kswapd was reclaiming at
2846 * There is special handling here for zones which are full of pinned pages.
2847 * This can happen if the pages are all mlocked, or if they are all used by
2848 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2849 * What we do is to detect the case where all pages in the zone have been
2850 * scanned twice and there has been zero successful reclaim. Mark the zone as
2851 * dead and from now on, only perform a short scan. Basically we're polling
2852 * the zone for when the problem goes away.
2854 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2855 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2856 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2857 * lower zones regardless of the number of free pages in the lower zones. This
2858 * interoperates with the page allocator fallback scheme to ensure that aging
2859 * of pages is balanced across the zones.
2861 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2865 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2866 unsigned long nr_soft_reclaimed;
2867 unsigned long nr_soft_scanned;
2868 struct scan_control sc = {
2869 .gfp_mask = GFP_KERNEL,
2870 .priority = DEF_PRIORITY,
2873 .may_writepage = !laptop_mode,
2875 .target_mem_cgroup = NULL,
2877 count_vm_event(PAGEOUTRUN);
2880 unsigned long lru_pages = 0;
2881 unsigned long nr_attempted = 0;
2882 bool raise_priority = true;
2883 bool pgdat_needs_compaction = (order > 0);
2885 sc.nr_reclaimed = 0;
2888 * Scan in the highmem->dma direction for the highest
2889 * zone which needs scanning
2891 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2892 struct zone *zone = pgdat->node_zones + i;
2894 if (!populated_zone(zone))
2897 if (zone->all_unreclaimable &&
2898 sc.priority != DEF_PRIORITY)
2902 * Do some background aging of the anon list, to give
2903 * pages a chance to be referenced before reclaiming.
2905 age_active_anon(zone, &sc);
2908 * If the number of buffer_heads in the machine
2909 * exceeds the maximum allowed level and this node
2910 * has a highmem zone, force kswapd to reclaim from
2911 * it to relieve lowmem pressure.
2913 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2918 if (!zone_balanced(zone, order, 0, 0)) {
2923 * If balanced, clear the dirty and congested
2926 zone_clear_flag(zone, ZONE_CONGESTED);
2927 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2934 for (i = 0; i <= end_zone; i++) {
2935 struct zone *zone = pgdat->node_zones + i;
2937 if (!populated_zone(zone))
2940 lru_pages += zone_reclaimable_pages(zone);
2943 * If any zone is currently balanced then kswapd will
2944 * not call compaction as it is expected that the
2945 * necessary pages are already available.
2947 if (pgdat_needs_compaction &&
2948 zone_watermark_ok(zone, order,
2949 low_wmark_pages(zone),
2951 pgdat_needs_compaction = false;
2955 * If we're getting trouble reclaiming, start doing writepage
2956 * even in laptop mode.
2958 if (sc.priority < DEF_PRIORITY - 2)
2959 sc.may_writepage = 1;
2962 * Now scan the zone in the dma->highmem direction, stopping
2963 * at the last zone which needs scanning.
2965 * We do this because the page allocator works in the opposite
2966 * direction. This prevents the page allocator from allocating
2967 * pages behind kswapd's direction of progress, which would
2968 * cause too much scanning of the lower zones.
2970 for (i = 0; i <= end_zone; i++) {
2971 struct zone *zone = pgdat->node_zones + i;
2973 if (!populated_zone(zone))
2976 if (zone->all_unreclaimable &&
2977 sc.priority != DEF_PRIORITY)
2982 nr_soft_scanned = 0;
2984 * Call soft limit reclaim before calling shrink_zone.
2986 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2989 sc.nr_reclaimed += nr_soft_reclaimed;
2992 * There should be no need to raise the scanning
2993 * priority if enough pages are already being scanned
2994 * that that high watermark would be met at 100%
2997 if (kswapd_shrink_zone(zone, end_zone, &sc,
2998 lru_pages, &nr_attempted))
2999 raise_priority = false;
3003 * If the low watermark is met there is no need for processes
3004 * to be throttled on pfmemalloc_wait as they should not be
3005 * able to safely make forward progress. Wake them
3007 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3008 pfmemalloc_watermark_ok(pgdat))
3009 wake_up(&pgdat->pfmemalloc_wait);
3012 * Fragmentation may mean that the system cannot be rebalanced
3013 * for high-order allocations in all zones. If twice the
3014 * allocation size has been reclaimed and the zones are still
3015 * not balanced then recheck the watermarks at order-0 to
3016 * prevent kswapd reclaiming excessively. Assume that a
3017 * process requested a high-order can direct reclaim/compact.
3019 if (order && sc.nr_reclaimed >= 2UL << order)
3020 order = sc.order = 0;
3022 /* Check if kswapd should be suspending */
3023 if (try_to_freeze() || kthread_should_stop())
3027 * Compact if necessary and kswapd is reclaiming at least the
3028 * high watermark number of pages as requsted
3030 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3031 compact_pgdat(pgdat, order);
3034 * Raise priority if scanning rate is too low or there was no
3035 * progress in reclaiming pages
3037 if (raise_priority || !sc.nr_reclaimed)
3039 } while (sc.priority >= 1 &&
3040 !pgdat_balanced(pgdat, order, *classzone_idx));
3044 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3045 * makes a decision on the order we were last reclaiming at. However,
3046 * if another caller entered the allocator slow path while kswapd
3047 * was awake, order will remain at the higher level
3049 *classzone_idx = end_zone;
3053 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3058 if (freezing(current) || kthread_should_stop())
3061 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3063 /* Try to sleep for a short interval */
3064 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3065 remaining = schedule_timeout(HZ/10);
3066 finish_wait(&pgdat->kswapd_wait, &wait);
3067 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3071 * After a short sleep, check if it was a premature sleep. If not, then
3072 * go fully to sleep until explicitly woken up.
3074 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3075 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3078 * vmstat counters are not perfectly accurate and the estimated
3079 * value for counters such as NR_FREE_PAGES can deviate from the
3080 * true value by nr_online_cpus * threshold. To avoid the zone
3081 * watermarks being breached while under pressure, we reduce the
3082 * per-cpu vmstat threshold while kswapd is awake and restore
3083 * them before going back to sleep.
3085 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3088 * Compaction records what page blocks it recently failed to
3089 * isolate pages from and skips them in the future scanning.
3090 * When kswapd is going to sleep, it is reasonable to assume
3091 * that pages and compaction may succeed so reset the cache.
3093 reset_isolation_suitable(pgdat);
3095 if (!kthread_should_stop())
3098 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3101 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3103 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3105 finish_wait(&pgdat->kswapd_wait, &wait);
3109 * The background pageout daemon, started as a kernel thread
3110 * from the init process.
3112 * This basically trickles out pages so that we have _some_
3113 * free memory available even if there is no other activity
3114 * that frees anything up. This is needed for things like routing
3115 * etc, where we otherwise might have all activity going on in
3116 * asynchronous contexts that cannot page things out.
3118 * If there are applications that are active memory-allocators
3119 * (most normal use), this basically shouldn't matter.
3121 static int kswapd(void *p)
3123 unsigned long order, new_order;
3124 unsigned balanced_order;
3125 int classzone_idx, new_classzone_idx;
3126 int balanced_classzone_idx;
3127 pg_data_t *pgdat = (pg_data_t*)p;
3128 struct task_struct *tsk = current;
3130 struct reclaim_state reclaim_state = {
3131 .reclaimed_slab = 0,
3133 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3135 lockdep_set_current_reclaim_state(GFP_KERNEL);
3137 if (!cpumask_empty(cpumask))
3138 set_cpus_allowed_ptr(tsk, cpumask);
3139 current->reclaim_state = &reclaim_state;
3142 * Tell the memory management that we're a "memory allocator",
3143 * and that if we need more memory we should get access to it
3144 * regardless (see "__alloc_pages()"). "kswapd" should
3145 * never get caught in the normal page freeing logic.
3147 * (Kswapd normally doesn't need memory anyway, but sometimes
3148 * you need a small amount of memory in order to be able to
3149 * page out something else, and this flag essentially protects
3150 * us from recursively trying to free more memory as we're
3151 * trying to free the first piece of memory in the first place).
3153 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3156 order = new_order = 0;
3158 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3159 balanced_classzone_idx = classzone_idx;
3164 * If the last balance_pgdat was unsuccessful it's unlikely a
3165 * new request of a similar or harder type will succeed soon
3166 * so consider going to sleep on the basis we reclaimed at
3168 if (balanced_classzone_idx >= new_classzone_idx &&
3169 balanced_order == new_order) {
3170 new_order = pgdat->kswapd_max_order;
3171 new_classzone_idx = pgdat->classzone_idx;
3172 pgdat->kswapd_max_order = 0;
3173 pgdat->classzone_idx = pgdat->nr_zones - 1;
3176 if (order < new_order || classzone_idx > new_classzone_idx) {
3178 * Don't sleep if someone wants a larger 'order'
3179 * allocation or has tigher zone constraints
3182 classzone_idx = new_classzone_idx;
3184 kswapd_try_to_sleep(pgdat, balanced_order,
3185 balanced_classzone_idx);
3186 order = pgdat->kswapd_max_order;
3187 classzone_idx = pgdat->classzone_idx;
3189 new_classzone_idx = classzone_idx;
3190 pgdat->kswapd_max_order = 0;
3191 pgdat->classzone_idx = pgdat->nr_zones - 1;
3194 ret = try_to_freeze();
3195 if (kthread_should_stop())
3199 * We can speed up thawing tasks if we don't call balance_pgdat
3200 * after returning from the refrigerator
3203 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3204 balanced_classzone_idx = classzone_idx;
3205 balanced_order = balance_pgdat(pgdat, order,
3206 &balanced_classzone_idx);
3210 current->reclaim_state = NULL;
3215 * A zone is low on free memory, so wake its kswapd task to service it.
3217 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3221 if (!populated_zone(zone))
3224 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3226 pgdat = zone->zone_pgdat;
3227 if (pgdat->kswapd_max_order < order) {
3228 pgdat->kswapd_max_order = order;
3229 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3231 if (!waitqueue_active(&pgdat->kswapd_wait))
3233 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3236 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3237 wake_up_interruptible(&pgdat->kswapd_wait);
3241 * The reclaimable count would be mostly accurate.
3242 * The less reclaimable pages may be
3243 * - mlocked pages, which will be moved to unevictable list when encountered
3244 * - mapped pages, which may require several travels to be reclaimed
3245 * - dirty pages, which is not "instantly" reclaimable
3247 unsigned long global_reclaimable_pages(void)
3251 nr = global_page_state(NR_ACTIVE_FILE) +
3252 global_page_state(NR_INACTIVE_FILE);
3254 if (get_nr_swap_pages() > 0)
3255 nr += global_page_state(NR_ACTIVE_ANON) +
3256 global_page_state(NR_INACTIVE_ANON);
3261 unsigned long zone_reclaimable_pages(struct zone *zone)
3265 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3266 zone_page_state(zone, NR_INACTIVE_FILE);
3268 if (get_nr_swap_pages() > 0)
3269 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3270 zone_page_state(zone, NR_INACTIVE_ANON);
3275 #ifdef CONFIG_HIBERNATION
3277 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3280 * Rather than trying to age LRUs the aim is to preserve the overall
3281 * LRU order by reclaiming preferentially
3282 * inactive > active > active referenced > active mapped
3284 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3286 struct reclaim_state reclaim_state;
3287 struct scan_control sc = {
3288 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3292 .nr_to_reclaim = nr_to_reclaim,
3293 .hibernation_mode = 1,
3295 .priority = DEF_PRIORITY,
3297 struct shrink_control shrink = {
3298 .gfp_mask = sc.gfp_mask,
3300 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3301 struct task_struct *p = current;
3302 unsigned long nr_reclaimed;
3304 p->flags |= PF_MEMALLOC;
3305 lockdep_set_current_reclaim_state(sc.gfp_mask);
3306 reclaim_state.reclaimed_slab = 0;
3307 p->reclaim_state = &reclaim_state;
3309 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3311 p->reclaim_state = NULL;
3312 lockdep_clear_current_reclaim_state();
3313 p->flags &= ~PF_MEMALLOC;
3315 return nr_reclaimed;
3317 #endif /* CONFIG_HIBERNATION */
3319 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3320 not required for correctness. So if the last cpu in a node goes
3321 away, we get changed to run anywhere: as the first one comes back,
3322 restore their cpu bindings. */
3323 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3328 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3329 for_each_node_state(nid, N_MEMORY) {
3330 pg_data_t *pgdat = NODE_DATA(nid);
3331 const struct cpumask *mask;
3333 mask = cpumask_of_node(pgdat->node_id);
3335 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3336 /* One of our CPUs online: restore mask */
3337 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3344 * This kswapd start function will be called by init and node-hot-add.
3345 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3347 int kswapd_run(int nid)
3349 pg_data_t *pgdat = NODE_DATA(nid);
3355 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3356 if (IS_ERR(pgdat->kswapd)) {
3357 /* failure at boot is fatal */
3358 BUG_ON(system_state == SYSTEM_BOOTING);
3359 pr_err("Failed to start kswapd on node %d\n", nid);
3360 ret = PTR_ERR(pgdat->kswapd);
3361 pgdat->kswapd = NULL;
3367 * Called by memory hotplug when all memory in a node is offlined. Caller must
3368 * hold lock_memory_hotplug().
3370 void kswapd_stop(int nid)
3372 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3375 kthread_stop(kswapd);
3376 NODE_DATA(nid)->kswapd = NULL;
3380 static int __init kswapd_init(void)
3385 for_each_node_state(nid, N_MEMORY)
3387 hotcpu_notifier(cpu_callback, 0);
3391 module_init(kswapd_init)
3397 * If non-zero call zone_reclaim when the number of free pages falls below
3400 int zone_reclaim_mode __read_mostly;
3402 #define RECLAIM_OFF 0
3403 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3404 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3405 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3408 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3409 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3412 #define ZONE_RECLAIM_PRIORITY 4
3415 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3418 int sysctl_min_unmapped_ratio = 1;
3421 * If the number of slab pages in a zone grows beyond this percentage then
3422 * slab reclaim needs to occur.
3424 int sysctl_min_slab_ratio = 5;
3426 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3428 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3429 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3430 zone_page_state(zone, NR_ACTIVE_FILE);
3433 * It's possible for there to be more file mapped pages than
3434 * accounted for by the pages on the file LRU lists because
3435 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3437 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3440 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3441 static long zone_pagecache_reclaimable(struct zone *zone)
3443 long nr_pagecache_reclaimable;
3447 * If RECLAIM_SWAP is set, then all file pages are considered
3448 * potentially reclaimable. Otherwise, we have to worry about
3449 * pages like swapcache and zone_unmapped_file_pages() provides
3452 if (zone_reclaim_mode & RECLAIM_SWAP)
3453 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3455 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3457 /* If we can't clean pages, remove dirty pages from consideration */
3458 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3459 delta += zone_page_state(zone, NR_FILE_DIRTY);
3461 /* Watch for any possible underflows due to delta */
3462 if (unlikely(delta > nr_pagecache_reclaimable))
3463 delta = nr_pagecache_reclaimable;
3465 return nr_pagecache_reclaimable - delta;
3469 * Try to free up some pages from this zone through reclaim.
3471 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3473 /* Minimum pages needed in order to stay on node */
3474 const unsigned long nr_pages = 1 << order;
3475 struct task_struct *p = current;
3476 struct reclaim_state reclaim_state;
3477 struct scan_control sc = {
3478 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3479 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3481 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3482 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3484 .priority = ZONE_RECLAIM_PRIORITY,
3486 struct shrink_control shrink = {
3487 .gfp_mask = sc.gfp_mask,
3489 unsigned long nr_slab_pages0, nr_slab_pages1;
3493 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3494 * and we also need to be able to write out pages for RECLAIM_WRITE
3497 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3498 lockdep_set_current_reclaim_state(gfp_mask);
3499 reclaim_state.reclaimed_slab = 0;
3500 p->reclaim_state = &reclaim_state;
3502 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3504 * Free memory by calling shrink zone with increasing
3505 * priorities until we have enough memory freed.
3508 shrink_zone(zone, &sc);
3509 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3512 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3513 if (nr_slab_pages0 > zone->min_slab_pages) {
3515 * shrink_slab() does not currently allow us to determine how
3516 * many pages were freed in this zone. So we take the current
3517 * number of slab pages and shake the slab until it is reduced
3518 * by the same nr_pages that we used for reclaiming unmapped
3521 * Note that shrink_slab will free memory on all zones and may
3525 unsigned long lru_pages = zone_reclaimable_pages(zone);
3527 /* No reclaimable slab or very low memory pressure */
3528 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3531 /* Freed enough memory */
3532 nr_slab_pages1 = zone_page_state(zone,
3533 NR_SLAB_RECLAIMABLE);
3534 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3539 * Update nr_reclaimed by the number of slab pages we
3540 * reclaimed from this zone.
3542 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3543 if (nr_slab_pages1 < nr_slab_pages0)
3544 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3547 p->reclaim_state = NULL;
3548 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3549 lockdep_clear_current_reclaim_state();
3550 return sc.nr_reclaimed >= nr_pages;
3553 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3559 * Zone reclaim reclaims unmapped file backed pages and
3560 * slab pages if we are over the defined limits.
3562 * A small portion of unmapped file backed pages is needed for
3563 * file I/O otherwise pages read by file I/O will be immediately
3564 * thrown out if the zone is overallocated. So we do not reclaim
3565 * if less than a specified percentage of the zone is used by
3566 * unmapped file backed pages.
3568 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3569 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3570 return ZONE_RECLAIM_FULL;
3572 if (zone->all_unreclaimable)
3573 return ZONE_RECLAIM_FULL;
3576 * Do not scan if the allocation should not be delayed.
3578 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3579 return ZONE_RECLAIM_NOSCAN;
3582 * Only run zone reclaim on the local zone or on zones that do not
3583 * have associated processors. This will favor the local processor
3584 * over remote processors and spread off node memory allocations
3585 * as wide as possible.
3587 node_id = zone_to_nid(zone);
3588 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3589 return ZONE_RECLAIM_NOSCAN;
3591 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3592 return ZONE_RECLAIM_NOSCAN;
3594 ret = __zone_reclaim(zone, gfp_mask, order);
3595 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3598 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3605 * page_evictable - test whether a page is evictable
3606 * @page: the page to test
3608 * Test whether page is evictable--i.e., should be placed on active/inactive
3609 * lists vs unevictable list.
3611 * Reasons page might not be evictable:
3612 * (1) page's mapping marked unevictable
3613 * (2) page is part of an mlocked VMA
3616 int page_evictable(struct page *page)
3618 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3623 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3624 * @pages: array of pages to check
3625 * @nr_pages: number of pages to check
3627 * Checks pages for evictability and moves them to the appropriate lru list.
3629 * This function is only used for SysV IPC SHM_UNLOCK.
3631 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3633 struct lruvec *lruvec;
3634 struct zone *zone = NULL;
3639 for (i = 0; i < nr_pages; i++) {
3640 struct page *page = pages[i];
3641 struct zone *pagezone;
3644 pagezone = page_zone(page);
3645 if (pagezone != zone) {
3647 spin_unlock_irq(&zone->lru_lock);
3649 spin_lock_irq(&zone->lru_lock);
3651 lruvec = mem_cgroup_page_lruvec(page, zone);
3653 if (!PageLRU(page) || !PageUnevictable(page))
3656 if (page_evictable(page)) {
3657 enum lru_list lru = page_lru_base_type(page);
3659 VM_BUG_ON(PageActive(page));
3660 ClearPageUnevictable(page);
3661 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3662 add_page_to_lru_list(page, lruvec, lru);
3668 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3669 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3670 spin_unlock_irq(&zone->lru_lock);
3673 #endif /* CONFIG_SHMEM */
3675 static void warn_scan_unevictable_pages(void)
3677 printk_once(KERN_WARNING
3678 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3679 "disabled for lack of a legitimate use case. If you have "
3680 "one, please send an email to linux-mm@kvack.org.\n",
3685 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3686 * all nodes' unevictable lists for evictable pages
3688 unsigned long scan_unevictable_pages;
3690 int scan_unevictable_handler(struct ctl_table *table, int write,
3691 void __user *buffer,
3692 size_t *length, loff_t *ppos)
3694 warn_scan_unevictable_pages();
3695 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3696 scan_unevictable_pages = 0;
3702 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3703 * a specified node's per zone unevictable lists for evictable pages.
3706 static ssize_t read_scan_unevictable_node(struct device *dev,
3707 struct device_attribute *attr,
3710 warn_scan_unevictable_pages();
3711 return sprintf(buf, "0\n"); /* always zero; should fit... */
3714 static ssize_t write_scan_unevictable_node(struct device *dev,
3715 struct device_attribute *attr,
3716 const char *buf, size_t count)
3718 warn_scan_unevictable_pages();
3723 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3724 read_scan_unevictable_node,
3725 write_scan_unevictable_node);
3727 int scan_unevictable_register_node(struct node *node)
3729 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3732 void scan_unevictable_unregister_node(struct node *node)
3734 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);