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 int register_shrinker(struct shrinker *shrinker)
162 size_t size = sizeof(*shrinker->nr_deferred);
165 * If we only have one possible node in the system anyway, save
166 * ourselves the trouble and disable NUMA aware behavior. This way we
167 * will save memory and some small loop time later.
169 if (nr_node_ids == 1)
170 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
172 if (shrinker->flags & SHRINKER_NUMA_AWARE)
175 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
176 if (!shrinker->nr_deferred)
179 down_write(&shrinker_rwsem);
180 list_add_tail(&shrinker->list, &shrinker_list);
181 up_write(&shrinker_rwsem);
184 EXPORT_SYMBOL(register_shrinker);
189 void unregister_shrinker(struct shrinker *shrinker)
191 down_write(&shrinker_rwsem);
192 list_del(&shrinker->list);
193 up_write(&shrinker_rwsem);
195 EXPORT_SYMBOL(unregister_shrinker);
197 #define SHRINK_BATCH 128
200 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
201 unsigned long nr_pages_scanned, unsigned long lru_pages)
203 unsigned long freed = 0;
204 unsigned long long delta;
209 int nid = shrinkctl->nid;
210 long batch_size = shrinker->batch ? shrinker->batch
213 max_pass = shrinker->count_objects(shrinker, shrinkctl);
218 * copy the current shrinker scan count into a local variable
219 * and zero it so that other concurrent shrinker invocations
220 * don't also do this scanning work.
222 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
225 delta = (4 * nr_pages_scanned) / shrinker->seeks;
227 do_div(delta, lru_pages + 1);
229 if (total_scan < 0) {
231 "shrink_slab: %pF negative objects to delete nr=%ld\n",
232 shrinker->scan_objects, total_scan);
233 total_scan = max_pass;
237 * We need to avoid excessive windup on filesystem shrinkers
238 * due to large numbers of GFP_NOFS allocations causing the
239 * shrinkers to return -1 all the time. This results in a large
240 * nr being built up so when a shrink that can do some work
241 * comes along it empties the entire cache due to nr >>>
242 * max_pass. This is bad for sustaining a working set in
245 * Hence only allow the shrinker to scan the entire cache when
246 * a large delta change is calculated directly.
248 if (delta < max_pass / 4)
249 total_scan = min(total_scan, max_pass / 2);
252 * Avoid risking looping forever due to too large nr value:
253 * never try to free more than twice the estimate number of
256 if (total_scan > max_pass * 2)
257 total_scan = max_pass * 2;
259 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
260 nr_pages_scanned, lru_pages,
261 max_pass, delta, total_scan);
263 while (total_scan >= batch_size) {
266 shrinkctl->nr_to_scan = batch_size;
267 ret = shrinker->scan_objects(shrinker, shrinkctl);
268 if (ret == SHRINK_STOP)
272 count_vm_events(SLABS_SCANNED, batch_size);
273 total_scan -= batch_size;
279 * move the unused scan count back into the shrinker in a
280 * manner that handles concurrent updates. If we exhausted the
281 * scan, there is no need to do an update.
284 new_nr = atomic_long_add_return(total_scan,
285 &shrinker->nr_deferred[nid]);
287 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
289 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
294 * Call the shrink functions to age shrinkable caches
296 * Here we assume it costs one seek to replace a lru page and that it also
297 * takes a seek to recreate a cache object. With this in mind we age equal
298 * percentages of the lru and ageable caches. This should balance the seeks
299 * generated by these structures.
301 * If the vm encountered mapped pages on the LRU it increase the pressure on
302 * slab to avoid swapping.
304 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
306 * `lru_pages' represents the number of on-LRU pages in all the zones which
307 * are eligible for the caller's allocation attempt. It is used for balancing
308 * slab reclaim versus page reclaim.
310 * Returns the number of slab objects which we shrunk.
312 unsigned long shrink_slab(struct shrink_control *shrinkctl,
313 unsigned long nr_pages_scanned,
314 unsigned long lru_pages)
316 struct shrinker *shrinker;
317 unsigned long freed = 0;
319 if (nr_pages_scanned == 0)
320 nr_pages_scanned = SWAP_CLUSTER_MAX;
322 if (!down_read_trylock(&shrinker_rwsem)) {
324 * If we would return 0, our callers would understand that we
325 * have nothing else to shrink and give up trying. By returning
326 * 1 we keep it going and assume we'll be able to shrink next
333 list_for_each_entry(shrinker, &shrinker_list, list) {
334 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
335 if (!node_online(shrinkctl->nid))
338 if (!(shrinker->flags & SHRINKER_NUMA_AWARE) &&
339 (shrinkctl->nid != 0))
342 freed += shrink_slab_node(shrinkctl, shrinker,
343 nr_pages_scanned, lru_pages);
347 up_read(&shrinker_rwsem);
353 static inline int is_page_cache_freeable(struct page *page)
356 * A freeable page cache page is referenced only by the caller
357 * that isolated the page, the page cache radix tree and
358 * optional buffer heads at page->private.
360 return page_count(page) - page_has_private(page) == 2;
363 static int may_write_to_queue(struct backing_dev_info *bdi,
364 struct scan_control *sc)
366 if (current->flags & PF_SWAPWRITE)
368 if (!bdi_write_congested(bdi))
370 if (bdi == current->backing_dev_info)
376 * We detected a synchronous write error writing a page out. Probably
377 * -ENOSPC. We need to propagate that into the address_space for a subsequent
378 * fsync(), msync() or close().
380 * The tricky part is that after writepage we cannot touch the mapping: nothing
381 * prevents it from being freed up. But we have a ref on the page and once
382 * that page is locked, the mapping is pinned.
384 * We're allowed to run sleeping lock_page() here because we know the caller has
387 static void handle_write_error(struct address_space *mapping,
388 struct page *page, int error)
391 if (page_mapping(page) == mapping)
392 mapping_set_error(mapping, error);
396 /* possible outcome of pageout() */
398 /* failed to write page out, page is locked */
400 /* move page to the active list, page is locked */
402 /* page has been sent to the disk successfully, page is unlocked */
404 /* page is clean and locked */
409 * pageout is called by shrink_page_list() for each dirty page.
410 * Calls ->writepage().
412 static pageout_t pageout(struct page *page, struct address_space *mapping,
413 struct scan_control *sc)
416 * If the page is dirty, only perform writeback if that write
417 * will be non-blocking. To prevent this allocation from being
418 * stalled by pagecache activity. But note that there may be
419 * stalls if we need to run get_block(). We could test
420 * PagePrivate for that.
422 * If this process is currently in __generic_file_aio_write() against
423 * this page's queue, we can perform writeback even if that
426 * If the page is swapcache, write it back even if that would
427 * block, for some throttling. This happens by accident, because
428 * swap_backing_dev_info is bust: it doesn't reflect the
429 * congestion state of the swapdevs. Easy to fix, if needed.
431 if (!is_page_cache_freeable(page))
435 * Some data journaling orphaned pages can have
436 * page->mapping == NULL while being dirty with clean buffers.
438 if (page_has_private(page)) {
439 if (try_to_free_buffers(page)) {
440 ClearPageDirty(page);
441 printk("%s: orphaned page\n", __func__);
447 if (mapping->a_ops->writepage == NULL)
448 return PAGE_ACTIVATE;
449 if (!may_write_to_queue(mapping->backing_dev_info, sc))
452 if (clear_page_dirty_for_io(page)) {
454 struct writeback_control wbc = {
455 .sync_mode = WB_SYNC_NONE,
456 .nr_to_write = SWAP_CLUSTER_MAX,
458 .range_end = LLONG_MAX,
462 SetPageReclaim(page);
463 res = mapping->a_ops->writepage(page, &wbc);
465 handle_write_error(mapping, page, res);
466 if (res == AOP_WRITEPAGE_ACTIVATE) {
467 ClearPageReclaim(page);
468 return PAGE_ACTIVATE;
471 if (!PageWriteback(page)) {
472 /* synchronous write or broken a_ops? */
473 ClearPageReclaim(page);
475 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
476 inc_zone_page_state(page, NR_VMSCAN_WRITE);
484 * Same as remove_mapping, but if the page is removed from the mapping, it
485 * gets returned with a refcount of 0.
487 static int __remove_mapping(struct address_space *mapping, struct page *page)
489 BUG_ON(!PageLocked(page));
490 BUG_ON(mapping != page_mapping(page));
492 spin_lock_irq(&mapping->tree_lock);
494 * The non racy check for a busy page.
496 * Must be careful with the order of the tests. When someone has
497 * a ref to the page, it may be possible that they dirty it then
498 * drop the reference. So if PageDirty is tested before page_count
499 * here, then the following race may occur:
501 * get_user_pages(&page);
502 * [user mapping goes away]
504 * !PageDirty(page) [good]
505 * SetPageDirty(page);
507 * !page_count(page) [good, discard it]
509 * [oops, our write_to data is lost]
511 * Reversing the order of the tests ensures such a situation cannot
512 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
513 * load is not satisfied before that of page->_count.
515 * Note that if SetPageDirty is always performed via set_page_dirty,
516 * and thus under tree_lock, then this ordering is not required.
518 if (!page_freeze_refs(page, 2))
520 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
521 if (unlikely(PageDirty(page))) {
522 page_unfreeze_refs(page, 2);
526 if (PageSwapCache(page)) {
527 swp_entry_t swap = { .val = page_private(page) };
528 __delete_from_swap_cache(page);
529 spin_unlock_irq(&mapping->tree_lock);
530 swapcache_free(swap, page);
532 void (*freepage)(struct page *);
534 freepage = mapping->a_ops->freepage;
536 __delete_from_page_cache(page);
537 spin_unlock_irq(&mapping->tree_lock);
538 mem_cgroup_uncharge_cache_page(page);
540 if (freepage != NULL)
547 spin_unlock_irq(&mapping->tree_lock);
552 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
553 * someone else has a ref on the page, abort and return 0. If it was
554 * successfully detached, return 1. Assumes the caller has a single ref on
557 int remove_mapping(struct address_space *mapping, struct page *page)
559 if (__remove_mapping(mapping, page)) {
561 * Unfreezing the refcount with 1 rather than 2 effectively
562 * drops the pagecache ref for us without requiring another
565 page_unfreeze_refs(page, 1);
572 * putback_lru_page - put previously isolated page onto appropriate LRU list
573 * @page: page to be put back to appropriate lru list
575 * Add previously isolated @page to appropriate LRU list.
576 * Page may still be unevictable for other reasons.
578 * lru_lock must not be held, interrupts must be enabled.
580 void putback_lru_page(struct page *page)
583 int was_unevictable = PageUnevictable(page);
585 VM_BUG_ON(PageLRU(page));
588 ClearPageUnevictable(page);
590 if (page_evictable(page)) {
592 * For evictable pages, we can use the cache.
593 * In event of a race, worst case is we end up with an
594 * unevictable page on [in]active list.
595 * We know how to handle that.
597 lru = page_lru_base_type(page);
601 * Put unevictable pages directly on zone's unevictable
604 lru = LRU_UNEVICTABLE;
605 add_page_to_unevictable_list(page);
607 * When racing with an mlock or AS_UNEVICTABLE clearing
608 * (page is unlocked) make sure that if the other thread
609 * does not observe our setting of PG_lru and fails
610 * isolation/check_move_unevictable_pages,
611 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
612 * the page back to the evictable list.
614 * The other side is TestClearPageMlocked() or shmem_lock().
620 * page's status can change while we move it among lru. If an evictable
621 * page is on unevictable list, it never be freed. To avoid that,
622 * check after we added it to the list, again.
624 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
625 if (!isolate_lru_page(page)) {
629 /* This means someone else dropped this page from LRU
630 * So, it will be freed or putback to LRU again. There is
631 * nothing to do here.
635 if (was_unevictable && lru != LRU_UNEVICTABLE)
636 count_vm_event(UNEVICTABLE_PGRESCUED);
637 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
638 count_vm_event(UNEVICTABLE_PGCULLED);
640 put_page(page); /* drop ref from isolate */
643 enum page_references {
645 PAGEREF_RECLAIM_CLEAN,
650 static enum page_references page_check_references(struct page *page,
651 struct scan_control *sc)
653 int referenced_ptes, referenced_page;
654 unsigned long vm_flags;
656 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
658 referenced_page = TestClearPageReferenced(page);
661 * Mlock lost the isolation race with us. Let try_to_unmap()
662 * move the page to the unevictable list.
664 if (vm_flags & VM_LOCKED)
665 return PAGEREF_RECLAIM;
667 if (referenced_ptes) {
668 if (PageSwapBacked(page))
669 return PAGEREF_ACTIVATE;
671 * All mapped pages start out with page table
672 * references from the instantiating fault, so we need
673 * to look twice if a mapped file page is used more
676 * Mark it and spare it for another trip around the
677 * inactive list. Another page table reference will
678 * lead to its activation.
680 * Note: the mark is set for activated pages as well
681 * so that recently deactivated but used pages are
684 SetPageReferenced(page);
686 if (referenced_page || referenced_ptes > 1)
687 return PAGEREF_ACTIVATE;
690 * Activate file-backed executable pages after first usage.
692 if (vm_flags & VM_EXEC)
693 return PAGEREF_ACTIVATE;
698 /* Reclaim if clean, defer dirty pages to writeback */
699 if (referenced_page && !PageSwapBacked(page))
700 return PAGEREF_RECLAIM_CLEAN;
702 return PAGEREF_RECLAIM;
705 /* Check if a page is dirty or under writeback */
706 static void page_check_dirty_writeback(struct page *page,
707 bool *dirty, bool *writeback)
709 struct address_space *mapping;
712 * Anonymous pages are not handled by flushers and must be written
713 * from reclaim context. Do not stall reclaim based on them
715 if (!page_is_file_cache(page)) {
721 /* By default assume that the page flags are accurate */
722 *dirty = PageDirty(page);
723 *writeback = PageWriteback(page);
725 /* Verify dirty/writeback state if the filesystem supports it */
726 if (!page_has_private(page))
729 mapping = page_mapping(page);
730 if (mapping && mapping->a_ops->is_dirty_writeback)
731 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
735 * shrink_page_list() returns the number of reclaimed pages
737 static unsigned long shrink_page_list(struct list_head *page_list,
739 struct scan_control *sc,
740 enum ttu_flags ttu_flags,
741 unsigned long *ret_nr_dirty,
742 unsigned long *ret_nr_unqueued_dirty,
743 unsigned long *ret_nr_congested,
744 unsigned long *ret_nr_writeback,
745 unsigned long *ret_nr_immediate,
748 LIST_HEAD(ret_pages);
749 LIST_HEAD(free_pages);
751 unsigned long nr_unqueued_dirty = 0;
752 unsigned long nr_dirty = 0;
753 unsigned long nr_congested = 0;
754 unsigned long nr_reclaimed = 0;
755 unsigned long nr_writeback = 0;
756 unsigned long nr_immediate = 0;
760 mem_cgroup_uncharge_start();
761 while (!list_empty(page_list)) {
762 struct address_space *mapping;
765 enum page_references references = PAGEREF_RECLAIM_CLEAN;
766 bool dirty, writeback;
770 page = lru_to_page(page_list);
771 list_del(&page->lru);
773 if (!trylock_page(page))
776 VM_BUG_ON(PageActive(page));
777 VM_BUG_ON(page_zone(page) != zone);
781 if (unlikely(!page_evictable(page)))
784 if (!sc->may_unmap && page_mapped(page))
787 /* Double the slab pressure for mapped and swapcache pages */
788 if (page_mapped(page) || PageSwapCache(page))
791 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
792 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
795 * The number of dirty pages determines if a zone is marked
796 * reclaim_congested which affects wait_iff_congested. kswapd
797 * will stall and start writing pages if the tail of the LRU
798 * is all dirty unqueued pages.
800 page_check_dirty_writeback(page, &dirty, &writeback);
801 if (dirty || writeback)
804 if (dirty && !writeback)
808 * Treat this page as congested if the underlying BDI is or if
809 * pages are cycling through the LRU so quickly that the
810 * pages marked for immediate reclaim are making it to the
811 * end of the LRU a second time.
813 mapping = page_mapping(page);
814 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
815 (writeback && PageReclaim(page)))
819 * If a page at the tail of the LRU is under writeback, there
820 * are three cases to consider.
822 * 1) If reclaim is encountering an excessive number of pages
823 * under writeback and this page is both under writeback and
824 * PageReclaim then it indicates that pages are being queued
825 * for IO but are being recycled through the LRU before the
826 * IO can complete. Waiting on the page itself risks an
827 * indefinite stall if it is impossible to writeback the
828 * page due to IO error or disconnected storage so instead
829 * note that the LRU is being scanned too quickly and the
830 * caller can stall after page list has been processed.
832 * 2) Global reclaim encounters a page, memcg encounters a
833 * page that is not marked for immediate reclaim or
834 * the caller does not have __GFP_IO. In this case mark
835 * the page for immediate reclaim and continue scanning.
837 * __GFP_IO is checked because a loop driver thread might
838 * enter reclaim, and deadlock if it waits on a page for
839 * which it is needed to do the write (loop masks off
840 * __GFP_IO|__GFP_FS for this reason); but more thought
841 * would probably show more reasons.
843 * Don't require __GFP_FS, since we're not going into the
844 * FS, just waiting on its writeback completion. Worryingly,
845 * ext4 gfs2 and xfs allocate pages with
846 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
847 * may_enter_fs here is liable to OOM on them.
849 * 3) memcg encounters a page that is not already marked
850 * PageReclaim. memcg does not have any dirty pages
851 * throttling so we could easily OOM just because too many
852 * pages are in writeback and there is nothing else to
853 * reclaim. Wait for the writeback to complete.
855 if (PageWriteback(page)) {
857 if (current_is_kswapd() &&
859 zone_is_reclaim_writeback(zone)) {
864 } else if (global_reclaim(sc) ||
865 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
867 * This is slightly racy - end_page_writeback()
868 * might have just cleared PageReclaim, then
869 * setting PageReclaim here end up interpreted
870 * as PageReadahead - but that does not matter
871 * enough to care. What we do want is for this
872 * page to have PageReclaim set next time memcg
873 * reclaim reaches the tests above, so it will
874 * then wait_on_page_writeback() to avoid OOM;
875 * and it's also appropriate in global reclaim.
877 SetPageReclaim(page);
884 wait_on_page_writeback(page);
889 references = page_check_references(page, sc);
891 switch (references) {
892 case PAGEREF_ACTIVATE:
893 goto activate_locked;
896 case PAGEREF_RECLAIM:
897 case PAGEREF_RECLAIM_CLEAN:
898 ; /* try to reclaim the page below */
902 * Anonymous process memory has backing store?
903 * Try to allocate it some swap space here.
905 if (PageAnon(page) && !PageSwapCache(page)) {
906 if (!(sc->gfp_mask & __GFP_IO))
908 if (!add_to_swap(page, page_list))
909 goto activate_locked;
912 /* Adding to swap updated mapping */
913 mapping = page_mapping(page);
917 * The page is mapped into the page tables of one or more
918 * processes. Try to unmap it here.
920 if (page_mapped(page) && mapping) {
921 switch (try_to_unmap(page, ttu_flags)) {
923 goto activate_locked;
929 ; /* try to free the page below */
933 if (PageDirty(page)) {
935 * Only kswapd can writeback filesystem pages to
936 * avoid risk of stack overflow but only writeback
937 * if many dirty pages have been encountered.
939 if (page_is_file_cache(page) &&
940 (!current_is_kswapd() ||
941 !zone_is_reclaim_dirty(zone))) {
943 * Immediately reclaim when written back.
944 * Similar in principal to deactivate_page()
945 * except we already have the page isolated
946 * and know it's dirty
948 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
949 SetPageReclaim(page);
954 if (references == PAGEREF_RECLAIM_CLEAN)
958 if (!sc->may_writepage)
961 /* Page is dirty, try to write it out here */
962 switch (pageout(page, mapping, sc)) {
966 goto activate_locked;
968 if (PageWriteback(page))
974 * A synchronous write - probably a ramdisk. Go
975 * ahead and try to reclaim the page.
977 if (!trylock_page(page))
979 if (PageDirty(page) || PageWriteback(page))
981 mapping = page_mapping(page);
983 ; /* try to free the page below */
988 * If the page has buffers, try to free the buffer mappings
989 * associated with this page. If we succeed we try to free
992 * We do this even if the page is PageDirty().
993 * try_to_release_page() does not perform I/O, but it is
994 * possible for a page to have PageDirty set, but it is actually
995 * clean (all its buffers are clean). This happens if the
996 * buffers were written out directly, with submit_bh(). ext3
997 * will do this, as well as the blockdev mapping.
998 * try_to_release_page() will discover that cleanness and will
999 * drop the buffers and mark the page clean - it can be freed.
1001 * Rarely, pages can have buffers and no ->mapping. These are
1002 * the pages which were not successfully invalidated in
1003 * truncate_complete_page(). We try to drop those buffers here
1004 * and if that worked, and the page is no longer mapped into
1005 * process address space (page_count == 1) it can be freed.
1006 * Otherwise, leave the page on the LRU so it is swappable.
1008 if (page_has_private(page)) {
1009 if (!try_to_release_page(page, sc->gfp_mask))
1010 goto activate_locked;
1011 if (!mapping && page_count(page) == 1) {
1013 if (put_page_testzero(page))
1017 * rare race with speculative reference.
1018 * the speculative reference will free
1019 * this page shortly, so we may
1020 * increment nr_reclaimed here (and
1021 * leave it off the LRU).
1029 if (!mapping || !__remove_mapping(mapping, page))
1033 * At this point, we have no other references and there is
1034 * no way to pick any more up (removed from LRU, removed
1035 * from pagecache). Can use non-atomic bitops now (and
1036 * we obviously don't have to worry about waking up a process
1037 * waiting on the page lock, because there are no references.
1039 __clear_page_locked(page);
1044 * Is there need to periodically free_page_list? It would
1045 * appear not as the counts should be low
1047 list_add(&page->lru, &free_pages);
1051 if (PageSwapCache(page))
1052 try_to_free_swap(page);
1054 putback_lru_page(page);
1058 /* Not a candidate for swapping, so reclaim swap space. */
1059 if (PageSwapCache(page) && vm_swap_full())
1060 try_to_free_swap(page);
1061 VM_BUG_ON(PageActive(page));
1062 SetPageActive(page);
1067 list_add(&page->lru, &ret_pages);
1068 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1071 free_hot_cold_page_list(&free_pages, 1);
1073 list_splice(&ret_pages, page_list);
1074 count_vm_events(PGACTIVATE, pgactivate);
1075 mem_cgroup_uncharge_end();
1076 *ret_nr_dirty += nr_dirty;
1077 *ret_nr_congested += nr_congested;
1078 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1079 *ret_nr_writeback += nr_writeback;
1080 *ret_nr_immediate += nr_immediate;
1081 return nr_reclaimed;
1084 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1085 struct list_head *page_list)
1087 struct scan_control sc = {
1088 .gfp_mask = GFP_KERNEL,
1089 .priority = DEF_PRIORITY,
1092 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1093 struct page *page, *next;
1094 LIST_HEAD(clean_pages);
1096 list_for_each_entry_safe(page, next, page_list, lru) {
1097 if (page_is_file_cache(page) && !PageDirty(page)) {
1098 ClearPageActive(page);
1099 list_move(&page->lru, &clean_pages);
1103 ret = shrink_page_list(&clean_pages, zone, &sc,
1104 TTU_UNMAP|TTU_IGNORE_ACCESS,
1105 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1106 list_splice(&clean_pages, page_list);
1107 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1112 * Attempt to remove the specified page from its LRU. Only take this page
1113 * if it is of the appropriate PageActive status. Pages which are being
1114 * freed elsewhere are also ignored.
1116 * page: page to consider
1117 * mode: one of the LRU isolation modes defined above
1119 * returns 0 on success, -ve errno on failure.
1121 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1125 /* Only take pages on the LRU. */
1129 /* Compaction should not handle unevictable pages but CMA can do so */
1130 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1136 * To minimise LRU disruption, the caller can indicate that it only
1137 * wants to isolate pages it will be able to operate on without
1138 * blocking - clean pages for the most part.
1140 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1141 * is used by reclaim when it is cannot write to backing storage
1143 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1144 * that it is possible to migrate without blocking
1146 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1147 /* All the caller can do on PageWriteback is block */
1148 if (PageWriteback(page))
1151 if (PageDirty(page)) {
1152 struct address_space *mapping;
1154 /* ISOLATE_CLEAN means only clean pages */
1155 if (mode & ISOLATE_CLEAN)
1159 * Only pages without mappings or that have a
1160 * ->migratepage callback are possible to migrate
1163 mapping = page_mapping(page);
1164 if (mapping && !mapping->a_ops->migratepage)
1169 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1172 if (likely(get_page_unless_zero(page))) {
1174 * Be careful not to clear PageLRU until after we're
1175 * sure the page is not being freed elsewhere -- the
1176 * page release code relies on it.
1186 * zone->lru_lock is heavily contended. Some of the functions that
1187 * shrink the lists perform better by taking out a batch of pages
1188 * and working on them outside the LRU lock.
1190 * For pagecache intensive workloads, this function is the hottest
1191 * spot in the kernel (apart from copy_*_user functions).
1193 * Appropriate locks must be held before calling this function.
1195 * @nr_to_scan: The number of pages to look through on the list.
1196 * @lruvec: The LRU vector to pull pages from.
1197 * @dst: The temp list to put pages on to.
1198 * @nr_scanned: The number of pages that were scanned.
1199 * @sc: The scan_control struct for this reclaim session
1200 * @mode: One of the LRU isolation modes
1201 * @lru: LRU list id for isolating
1203 * returns how many pages were moved onto *@dst.
1205 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1206 struct lruvec *lruvec, struct list_head *dst,
1207 unsigned long *nr_scanned, struct scan_control *sc,
1208 isolate_mode_t mode, enum lru_list lru)
1210 struct list_head *src = &lruvec->lists[lru];
1211 unsigned long nr_taken = 0;
1214 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1218 page = lru_to_page(src);
1219 prefetchw_prev_lru_page(page, src, flags);
1221 VM_BUG_ON(!PageLRU(page));
1223 switch (__isolate_lru_page(page, mode)) {
1225 nr_pages = hpage_nr_pages(page);
1226 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1227 list_move(&page->lru, dst);
1228 nr_taken += nr_pages;
1232 /* else it is being freed elsewhere */
1233 list_move(&page->lru, src);
1242 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1243 nr_taken, mode, is_file_lru(lru));
1248 * isolate_lru_page - tries to isolate a page from its LRU list
1249 * @page: page to isolate from its LRU list
1251 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1252 * vmstat statistic corresponding to whatever LRU list the page was on.
1254 * Returns 0 if the page was removed from an LRU list.
1255 * Returns -EBUSY if the page was not on an LRU list.
1257 * The returned page will have PageLRU() cleared. If it was found on
1258 * the active list, it will have PageActive set. If it was found on
1259 * the unevictable list, it will have the PageUnevictable bit set. That flag
1260 * may need to be cleared by the caller before letting the page go.
1262 * The vmstat statistic corresponding to the list on which the page was
1263 * found will be decremented.
1266 * (1) Must be called with an elevated refcount on the page. This is a
1267 * fundamentnal difference from isolate_lru_pages (which is called
1268 * without a stable reference).
1269 * (2) the lru_lock must not be held.
1270 * (3) interrupts must be enabled.
1272 int isolate_lru_page(struct page *page)
1276 VM_BUG_ON(!page_count(page));
1278 if (PageLRU(page)) {
1279 struct zone *zone = page_zone(page);
1280 struct lruvec *lruvec;
1282 spin_lock_irq(&zone->lru_lock);
1283 lruvec = mem_cgroup_page_lruvec(page, zone);
1284 if (PageLRU(page)) {
1285 int lru = page_lru(page);
1288 del_page_from_lru_list(page, lruvec, lru);
1291 spin_unlock_irq(&zone->lru_lock);
1297 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1298 * then get resheduled. When there are massive number of tasks doing page
1299 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1300 * the LRU list will go small and be scanned faster than necessary, leading to
1301 * unnecessary swapping, thrashing and OOM.
1303 static int too_many_isolated(struct zone *zone, int file,
1304 struct scan_control *sc)
1306 unsigned long inactive, isolated;
1308 if (current_is_kswapd())
1311 if (!global_reclaim(sc))
1315 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1316 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1318 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1319 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1323 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1324 * won't get blocked by normal direct-reclaimers, forming a circular
1327 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1330 return isolated > inactive;
1333 static noinline_for_stack void
1334 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1336 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1337 struct zone *zone = lruvec_zone(lruvec);
1338 LIST_HEAD(pages_to_free);
1341 * Put back any unfreeable pages.
1343 while (!list_empty(page_list)) {
1344 struct page *page = lru_to_page(page_list);
1347 VM_BUG_ON(PageLRU(page));
1348 list_del(&page->lru);
1349 if (unlikely(!page_evictable(page))) {
1350 spin_unlock_irq(&zone->lru_lock);
1351 putback_lru_page(page);
1352 spin_lock_irq(&zone->lru_lock);
1356 lruvec = mem_cgroup_page_lruvec(page, zone);
1359 lru = page_lru(page);
1360 add_page_to_lru_list(page, lruvec, lru);
1362 if (is_active_lru(lru)) {
1363 int file = is_file_lru(lru);
1364 int numpages = hpage_nr_pages(page);
1365 reclaim_stat->recent_rotated[file] += numpages;
1367 if (put_page_testzero(page)) {
1368 __ClearPageLRU(page);
1369 __ClearPageActive(page);
1370 del_page_from_lru_list(page, lruvec, lru);
1372 if (unlikely(PageCompound(page))) {
1373 spin_unlock_irq(&zone->lru_lock);
1374 (*get_compound_page_dtor(page))(page);
1375 spin_lock_irq(&zone->lru_lock);
1377 list_add(&page->lru, &pages_to_free);
1382 * To save our caller's stack, now use input list for pages to free.
1384 list_splice(&pages_to_free, page_list);
1388 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1389 * of reclaimed pages
1391 static noinline_for_stack unsigned long
1392 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1393 struct scan_control *sc, enum lru_list lru)
1395 LIST_HEAD(page_list);
1396 unsigned long nr_scanned;
1397 unsigned long nr_reclaimed = 0;
1398 unsigned long nr_taken;
1399 unsigned long nr_dirty = 0;
1400 unsigned long nr_congested = 0;
1401 unsigned long nr_unqueued_dirty = 0;
1402 unsigned long nr_writeback = 0;
1403 unsigned long nr_immediate = 0;
1404 isolate_mode_t isolate_mode = 0;
1405 int file = is_file_lru(lru);
1406 struct zone *zone = lruvec_zone(lruvec);
1407 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1409 while (unlikely(too_many_isolated(zone, file, sc))) {
1410 congestion_wait(BLK_RW_ASYNC, HZ/10);
1412 /* We are about to die and free our memory. Return now. */
1413 if (fatal_signal_pending(current))
1414 return SWAP_CLUSTER_MAX;
1420 isolate_mode |= ISOLATE_UNMAPPED;
1421 if (!sc->may_writepage)
1422 isolate_mode |= ISOLATE_CLEAN;
1424 spin_lock_irq(&zone->lru_lock);
1426 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1427 &nr_scanned, sc, isolate_mode, lru);
1429 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1430 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1432 if (global_reclaim(sc)) {
1433 zone->pages_scanned += nr_scanned;
1434 if (current_is_kswapd())
1435 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1437 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1439 spin_unlock_irq(&zone->lru_lock);
1444 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1445 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1446 &nr_writeback, &nr_immediate,
1449 spin_lock_irq(&zone->lru_lock);
1451 reclaim_stat->recent_scanned[file] += nr_taken;
1453 if (global_reclaim(sc)) {
1454 if (current_is_kswapd())
1455 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1458 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1462 putback_inactive_pages(lruvec, &page_list);
1464 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1466 spin_unlock_irq(&zone->lru_lock);
1468 free_hot_cold_page_list(&page_list, 1);
1471 * If reclaim is isolating dirty pages under writeback, it implies
1472 * that the long-lived page allocation rate is exceeding the page
1473 * laundering rate. Either the global limits are not being effective
1474 * at throttling processes due to the page distribution throughout
1475 * zones or there is heavy usage of a slow backing device. The
1476 * only option is to throttle from reclaim context which is not ideal
1477 * as there is no guarantee the dirtying process is throttled in the
1478 * same way balance_dirty_pages() manages.
1480 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1481 * of pages under pages flagged for immediate reclaim and stall if any
1482 * are encountered in the nr_immediate check below.
1484 if (nr_writeback && nr_writeback == nr_taken)
1485 zone_set_flag(zone, ZONE_WRITEBACK);
1488 * memcg will stall in page writeback so only consider forcibly
1489 * stalling for global reclaim
1491 if (global_reclaim(sc)) {
1493 * Tag a zone as congested if all the dirty pages scanned were
1494 * backed by a congested BDI and wait_iff_congested will stall.
1496 if (nr_dirty && nr_dirty == nr_congested)
1497 zone_set_flag(zone, ZONE_CONGESTED);
1500 * If dirty pages are scanned that are not queued for IO, it
1501 * implies that flushers are not keeping up. In this case, flag
1502 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1503 * pages from reclaim context. It will forcibly stall in the
1506 if (nr_unqueued_dirty == nr_taken)
1507 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1510 * In addition, if kswapd scans pages marked marked for
1511 * immediate reclaim and under writeback (nr_immediate), it
1512 * implies that pages are cycling through the LRU faster than
1513 * they are written so also forcibly stall.
1515 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1516 congestion_wait(BLK_RW_ASYNC, HZ/10);
1520 * Stall direct reclaim for IO completions if underlying BDIs or zone
1521 * is congested. Allow kswapd to continue until it starts encountering
1522 * unqueued dirty pages or cycling through the LRU too quickly.
1524 if (!sc->hibernation_mode && !current_is_kswapd())
1525 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1527 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1529 nr_scanned, nr_reclaimed,
1531 trace_shrink_flags(file));
1532 return nr_reclaimed;
1536 * This moves pages from the active list to the inactive list.
1538 * We move them the other way if the page is referenced by one or more
1539 * processes, from rmap.
1541 * If the pages are mostly unmapped, the processing is fast and it is
1542 * appropriate to hold zone->lru_lock across the whole operation. But if
1543 * the pages are mapped, the processing is slow (page_referenced()) so we
1544 * should drop zone->lru_lock around each page. It's impossible to balance
1545 * this, so instead we remove the pages from the LRU while processing them.
1546 * It is safe to rely on PG_active against the non-LRU pages in here because
1547 * nobody will play with that bit on a non-LRU page.
1549 * The downside is that we have to touch page->_count against each page.
1550 * But we had to alter page->flags anyway.
1553 static void move_active_pages_to_lru(struct lruvec *lruvec,
1554 struct list_head *list,
1555 struct list_head *pages_to_free,
1558 struct zone *zone = lruvec_zone(lruvec);
1559 unsigned long pgmoved = 0;
1563 while (!list_empty(list)) {
1564 page = lru_to_page(list);
1565 lruvec = mem_cgroup_page_lruvec(page, zone);
1567 VM_BUG_ON(PageLRU(page));
1570 nr_pages = hpage_nr_pages(page);
1571 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1572 list_move(&page->lru, &lruvec->lists[lru]);
1573 pgmoved += nr_pages;
1575 if (put_page_testzero(page)) {
1576 __ClearPageLRU(page);
1577 __ClearPageActive(page);
1578 del_page_from_lru_list(page, lruvec, lru);
1580 if (unlikely(PageCompound(page))) {
1581 spin_unlock_irq(&zone->lru_lock);
1582 (*get_compound_page_dtor(page))(page);
1583 spin_lock_irq(&zone->lru_lock);
1585 list_add(&page->lru, pages_to_free);
1588 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1589 if (!is_active_lru(lru))
1590 __count_vm_events(PGDEACTIVATE, pgmoved);
1593 static void shrink_active_list(unsigned long nr_to_scan,
1594 struct lruvec *lruvec,
1595 struct scan_control *sc,
1598 unsigned long nr_taken;
1599 unsigned long nr_scanned;
1600 unsigned long vm_flags;
1601 LIST_HEAD(l_hold); /* The pages which were snipped off */
1602 LIST_HEAD(l_active);
1603 LIST_HEAD(l_inactive);
1605 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1606 unsigned long nr_rotated = 0;
1607 isolate_mode_t isolate_mode = 0;
1608 int file = is_file_lru(lru);
1609 struct zone *zone = lruvec_zone(lruvec);
1614 isolate_mode |= ISOLATE_UNMAPPED;
1615 if (!sc->may_writepage)
1616 isolate_mode |= ISOLATE_CLEAN;
1618 spin_lock_irq(&zone->lru_lock);
1620 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1621 &nr_scanned, sc, isolate_mode, lru);
1622 if (global_reclaim(sc))
1623 zone->pages_scanned += nr_scanned;
1625 reclaim_stat->recent_scanned[file] += nr_taken;
1627 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1628 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1629 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1630 spin_unlock_irq(&zone->lru_lock);
1632 while (!list_empty(&l_hold)) {
1634 page = lru_to_page(&l_hold);
1635 list_del(&page->lru);
1637 if (unlikely(!page_evictable(page))) {
1638 putback_lru_page(page);
1642 if (unlikely(buffer_heads_over_limit)) {
1643 if (page_has_private(page) && trylock_page(page)) {
1644 if (page_has_private(page))
1645 try_to_release_page(page, 0);
1650 if (page_referenced(page, 0, sc->target_mem_cgroup,
1652 nr_rotated += hpage_nr_pages(page);
1654 * Identify referenced, file-backed active pages and
1655 * give them one more trip around the active list. So
1656 * that executable code get better chances to stay in
1657 * memory under moderate memory pressure. Anon pages
1658 * are not likely to be evicted by use-once streaming
1659 * IO, plus JVM can create lots of anon VM_EXEC pages,
1660 * so we ignore them here.
1662 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1663 list_add(&page->lru, &l_active);
1668 ClearPageActive(page); /* we are de-activating */
1669 list_add(&page->lru, &l_inactive);
1673 * Move pages back to the lru list.
1675 spin_lock_irq(&zone->lru_lock);
1677 * Count referenced pages from currently used mappings as rotated,
1678 * even though only some of them are actually re-activated. This
1679 * helps balance scan pressure between file and anonymous pages in
1682 reclaim_stat->recent_rotated[file] += nr_rotated;
1684 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1685 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1686 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1687 spin_unlock_irq(&zone->lru_lock);
1689 free_hot_cold_page_list(&l_hold, 1);
1693 static int inactive_anon_is_low_global(struct zone *zone)
1695 unsigned long active, inactive;
1697 active = zone_page_state(zone, NR_ACTIVE_ANON);
1698 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1700 if (inactive * zone->inactive_ratio < active)
1707 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1708 * @lruvec: LRU vector to check
1710 * Returns true if the zone does not have enough inactive anon pages,
1711 * meaning some active anon pages need to be deactivated.
1713 static int inactive_anon_is_low(struct lruvec *lruvec)
1716 * If we don't have swap space, anonymous page deactivation
1719 if (!total_swap_pages)
1722 if (!mem_cgroup_disabled())
1723 return mem_cgroup_inactive_anon_is_low(lruvec);
1725 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1728 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1735 * inactive_file_is_low - check if file pages need to be deactivated
1736 * @lruvec: LRU vector to check
1738 * When the system is doing streaming IO, memory pressure here
1739 * ensures that active file pages get deactivated, until more
1740 * than half of the file pages are on the inactive list.
1742 * Once we get to that situation, protect the system's working
1743 * set from being evicted by disabling active file page aging.
1745 * This uses a different ratio than the anonymous pages, because
1746 * the page cache uses a use-once replacement algorithm.
1748 static int inactive_file_is_low(struct lruvec *lruvec)
1750 unsigned long inactive;
1751 unsigned long active;
1753 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1754 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1756 return active > inactive;
1759 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1761 if (is_file_lru(lru))
1762 return inactive_file_is_low(lruvec);
1764 return inactive_anon_is_low(lruvec);
1767 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1768 struct lruvec *lruvec, struct scan_control *sc)
1770 if (is_active_lru(lru)) {
1771 if (inactive_list_is_low(lruvec, lru))
1772 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1776 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1779 static int vmscan_swappiness(struct scan_control *sc)
1781 if (global_reclaim(sc))
1782 return vm_swappiness;
1783 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1794 * Determine how aggressively the anon and file LRU lists should be
1795 * scanned. The relative value of each set of LRU lists is determined
1796 * by looking at the fraction of the pages scanned we did rotate back
1797 * onto the active list instead of evict.
1799 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1800 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1802 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1805 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1807 u64 denominator = 0; /* gcc */
1808 struct zone *zone = lruvec_zone(lruvec);
1809 unsigned long anon_prio, file_prio;
1810 enum scan_balance scan_balance;
1811 unsigned long anon, file, free;
1812 bool force_scan = false;
1813 unsigned long ap, fp;
1817 * If the zone or memcg is small, nr[l] can be 0. This
1818 * results in no scanning on this priority and a potential
1819 * priority drop. Global direct reclaim can go to the next
1820 * zone and tends to have no problems. Global kswapd is for
1821 * zone balancing and it needs to scan a minimum amount. When
1822 * reclaiming for a memcg, a priority drop can cause high
1823 * latencies, so it's better to scan a minimum amount there as
1826 if (current_is_kswapd() && zone->all_unreclaimable)
1828 if (!global_reclaim(sc))
1831 /* If we have no swap space, do not bother scanning anon pages. */
1832 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1833 scan_balance = SCAN_FILE;
1838 * Global reclaim will swap to prevent OOM even with no
1839 * swappiness, but memcg users want to use this knob to
1840 * disable swapping for individual groups completely when
1841 * using the memory controller's swap limit feature would be
1844 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1845 scan_balance = SCAN_FILE;
1850 * Do not apply any pressure balancing cleverness when the
1851 * system is close to OOM, scan both anon and file equally
1852 * (unless the swappiness setting disagrees with swapping).
1854 if (!sc->priority && vmscan_swappiness(sc)) {
1855 scan_balance = SCAN_EQUAL;
1859 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1860 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1861 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1862 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1865 * If it's foreseeable that reclaiming the file cache won't be
1866 * enough to get the zone back into a desirable shape, we have
1867 * to swap. Better start now and leave the - probably heavily
1868 * thrashing - remaining file pages alone.
1870 if (global_reclaim(sc)) {
1871 free = zone_page_state(zone, NR_FREE_PAGES);
1872 if (unlikely(file + free <= high_wmark_pages(zone))) {
1873 scan_balance = SCAN_ANON;
1879 * There is enough inactive page cache, do not reclaim
1880 * anything from the anonymous working set right now.
1882 if (!inactive_file_is_low(lruvec)) {
1883 scan_balance = SCAN_FILE;
1887 scan_balance = SCAN_FRACT;
1890 * With swappiness at 100, anonymous and file have the same priority.
1891 * This scanning priority is essentially the inverse of IO cost.
1893 anon_prio = vmscan_swappiness(sc);
1894 file_prio = 200 - anon_prio;
1897 * OK, so we have swap space and a fair amount of page cache
1898 * pages. We use the recently rotated / recently scanned
1899 * ratios to determine how valuable each cache is.
1901 * Because workloads change over time (and to avoid overflow)
1902 * we keep these statistics as a floating average, which ends
1903 * up weighing recent references more than old ones.
1905 * anon in [0], file in [1]
1907 spin_lock_irq(&zone->lru_lock);
1908 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1909 reclaim_stat->recent_scanned[0] /= 2;
1910 reclaim_stat->recent_rotated[0] /= 2;
1913 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1914 reclaim_stat->recent_scanned[1] /= 2;
1915 reclaim_stat->recent_rotated[1] /= 2;
1919 * The amount of pressure on anon vs file pages is inversely
1920 * proportional to the fraction of recently scanned pages on
1921 * each list that were recently referenced and in active use.
1923 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1924 ap /= reclaim_stat->recent_rotated[0] + 1;
1926 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1927 fp /= reclaim_stat->recent_rotated[1] + 1;
1928 spin_unlock_irq(&zone->lru_lock);
1932 denominator = ap + fp + 1;
1934 for_each_evictable_lru(lru) {
1935 int file = is_file_lru(lru);
1939 size = get_lru_size(lruvec, lru);
1940 scan = size >> sc->priority;
1942 if (!scan && force_scan)
1943 scan = min(size, SWAP_CLUSTER_MAX);
1945 switch (scan_balance) {
1947 /* Scan lists relative to size */
1951 * Scan types proportional to swappiness and
1952 * their relative recent reclaim efficiency.
1954 scan = div64_u64(scan * fraction[file], denominator);
1958 /* Scan one type exclusively */
1959 if ((scan_balance == SCAN_FILE) != file)
1963 /* Look ma, no brain */
1971 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1973 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1975 unsigned long nr[NR_LRU_LISTS];
1976 unsigned long targets[NR_LRU_LISTS];
1977 unsigned long nr_to_scan;
1979 unsigned long nr_reclaimed = 0;
1980 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1981 struct blk_plug plug;
1982 bool scan_adjusted = false;
1984 get_scan_count(lruvec, sc, nr);
1986 /* Record the original scan target for proportional adjustments later */
1987 memcpy(targets, nr, sizeof(nr));
1989 blk_start_plug(&plug);
1990 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1991 nr[LRU_INACTIVE_FILE]) {
1992 unsigned long nr_anon, nr_file, percentage;
1993 unsigned long nr_scanned;
1995 for_each_evictable_lru(lru) {
1997 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1998 nr[lru] -= nr_to_scan;
2000 nr_reclaimed += shrink_list(lru, nr_to_scan,
2005 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2009 * For global direct reclaim, reclaim only the number of pages
2010 * requested. Less care is taken to scan proportionally as it
2011 * is more important to minimise direct reclaim stall latency
2012 * than it is to properly age the LRU lists.
2014 if (global_reclaim(sc) && !current_is_kswapd())
2018 * For kswapd and memcg, reclaim at least the number of pages
2019 * requested. Ensure that the anon and file LRUs shrink
2020 * proportionally what was requested by get_scan_count(). We
2021 * stop reclaiming one LRU and reduce the amount scanning
2022 * proportional to the original scan target.
2024 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2025 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2027 if (nr_file > nr_anon) {
2028 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2029 targets[LRU_ACTIVE_ANON] + 1;
2031 percentage = nr_anon * 100 / scan_target;
2033 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2034 targets[LRU_ACTIVE_FILE] + 1;
2036 percentage = nr_file * 100 / scan_target;
2039 /* Stop scanning the smaller of the LRU */
2041 nr[lru + LRU_ACTIVE] = 0;
2044 * Recalculate the other LRU scan count based on its original
2045 * scan target and the percentage scanning already complete
2047 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2048 nr_scanned = targets[lru] - nr[lru];
2049 nr[lru] = targets[lru] * (100 - percentage) / 100;
2050 nr[lru] -= min(nr[lru], nr_scanned);
2053 nr_scanned = targets[lru] - nr[lru];
2054 nr[lru] = targets[lru] * (100 - percentage) / 100;
2055 nr[lru] -= min(nr[lru], nr_scanned);
2057 scan_adjusted = true;
2059 blk_finish_plug(&plug);
2060 sc->nr_reclaimed += nr_reclaimed;
2063 * Even if we did not try to evict anon pages at all, we want to
2064 * rebalance the anon lru active/inactive ratio.
2066 if (inactive_anon_is_low(lruvec))
2067 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2068 sc, LRU_ACTIVE_ANON);
2070 throttle_vm_writeout(sc->gfp_mask);
2073 /* Use reclaim/compaction for costly allocs or under memory pressure */
2074 static bool in_reclaim_compaction(struct scan_control *sc)
2076 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2077 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2078 sc->priority < DEF_PRIORITY - 2))
2085 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2086 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2087 * true if more pages should be reclaimed such that when the page allocator
2088 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2089 * It will give up earlier than that if there is difficulty reclaiming pages.
2091 static inline bool should_continue_reclaim(struct zone *zone,
2092 unsigned long nr_reclaimed,
2093 unsigned long nr_scanned,
2094 struct scan_control *sc)
2096 unsigned long pages_for_compaction;
2097 unsigned long inactive_lru_pages;
2099 /* If not in reclaim/compaction mode, stop */
2100 if (!in_reclaim_compaction(sc))
2103 /* Consider stopping depending on scan and reclaim activity */
2104 if (sc->gfp_mask & __GFP_REPEAT) {
2106 * For __GFP_REPEAT allocations, stop reclaiming if the
2107 * full LRU list has been scanned and we are still failing
2108 * to reclaim pages. This full LRU scan is potentially
2109 * expensive but a __GFP_REPEAT caller really wants to succeed
2111 if (!nr_reclaimed && !nr_scanned)
2115 * For non-__GFP_REPEAT allocations which can presumably
2116 * fail without consequence, stop if we failed to reclaim
2117 * any pages from the last SWAP_CLUSTER_MAX number of
2118 * pages that were scanned. This will return to the
2119 * caller faster at the risk reclaim/compaction and
2120 * the resulting allocation attempt fails
2127 * If we have not reclaimed enough pages for compaction and the
2128 * inactive lists are large enough, continue reclaiming
2130 pages_for_compaction = (2UL << sc->order);
2131 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2132 if (get_nr_swap_pages() > 0)
2133 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2134 if (sc->nr_reclaimed < pages_for_compaction &&
2135 inactive_lru_pages > pages_for_compaction)
2138 /* If compaction would go ahead or the allocation would succeed, stop */
2139 switch (compaction_suitable(zone, sc->order)) {
2140 case COMPACT_PARTIAL:
2141 case COMPACT_CONTINUE:
2148 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2150 unsigned long nr_reclaimed, nr_scanned;
2153 struct mem_cgroup *root = sc->target_mem_cgroup;
2154 struct mem_cgroup_reclaim_cookie reclaim = {
2156 .priority = sc->priority,
2158 struct mem_cgroup *memcg;
2160 nr_reclaimed = sc->nr_reclaimed;
2161 nr_scanned = sc->nr_scanned;
2163 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2165 struct lruvec *lruvec;
2167 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2169 shrink_lruvec(lruvec, sc);
2172 * Direct reclaim and kswapd have to scan all memory
2173 * cgroups to fulfill the overall scan target for the
2176 * Limit reclaim, on the other hand, only cares about
2177 * nr_to_reclaim pages to be reclaimed and it will
2178 * retry with decreasing priority if one round over the
2179 * whole hierarchy is not sufficient.
2181 if (!global_reclaim(sc) &&
2182 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2183 mem_cgroup_iter_break(root, memcg);
2186 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2189 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2190 sc->nr_scanned - nr_scanned,
2191 sc->nr_reclaimed - nr_reclaimed);
2193 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2194 sc->nr_scanned - nr_scanned, sc));
2197 /* Returns true if compaction should go ahead for a high-order request */
2198 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2200 unsigned long balance_gap, watermark;
2203 /* Do not consider compaction for orders reclaim is meant to satisfy */
2204 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2208 * Compaction takes time to run and there are potentially other
2209 * callers using the pages just freed. Continue reclaiming until
2210 * there is a buffer of free pages available to give compaction
2211 * a reasonable chance of completing and allocating the page
2213 balance_gap = min(low_wmark_pages(zone),
2214 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2215 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2216 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2217 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2220 * If compaction is deferred, reclaim up to a point where
2221 * compaction will have a chance of success when re-enabled
2223 if (compaction_deferred(zone, sc->order))
2224 return watermark_ok;
2226 /* If compaction is not ready to start, keep reclaiming */
2227 if (!compaction_suitable(zone, sc->order))
2230 return watermark_ok;
2234 * This is the direct reclaim path, for page-allocating processes. We only
2235 * try to reclaim pages from zones which will satisfy the caller's allocation
2238 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2240 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2242 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2243 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2244 * zone defense algorithm.
2246 * If a zone is deemed to be full of pinned pages then just give it a light
2247 * scan then give up on it.
2249 * This function returns true if a zone is being reclaimed for a costly
2250 * high-order allocation and compaction is ready to begin. This indicates to
2251 * the caller that it should consider retrying the allocation instead of
2254 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2258 unsigned long nr_soft_reclaimed;
2259 unsigned long nr_soft_scanned;
2260 bool aborted_reclaim = false;
2263 * If the number of buffer_heads in the machine exceeds the maximum
2264 * allowed level, force direct reclaim to scan the highmem zone as
2265 * highmem pages could be pinning lowmem pages storing buffer_heads
2267 if (buffer_heads_over_limit)
2268 sc->gfp_mask |= __GFP_HIGHMEM;
2270 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2271 gfp_zone(sc->gfp_mask), sc->nodemask) {
2272 if (!populated_zone(zone))
2275 * Take care memory controller reclaiming has small influence
2278 if (global_reclaim(sc)) {
2279 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2281 if (zone->all_unreclaimable &&
2282 sc->priority != DEF_PRIORITY)
2283 continue; /* Let kswapd poll it */
2284 if (IS_ENABLED(CONFIG_COMPACTION)) {
2286 * If we already have plenty of memory free for
2287 * compaction in this zone, don't free any more.
2288 * Even though compaction is invoked for any
2289 * non-zero order, only frequent costly order
2290 * reclamation is disruptive enough to become a
2291 * noticeable problem, like transparent huge
2294 if (compaction_ready(zone, sc)) {
2295 aborted_reclaim = true;
2300 * This steals pages from memory cgroups over softlimit
2301 * and returns the number of reclaimed pages and
2302 * scanned pages. This works for global memory pressure
2303 * and balancing, not for a memcg's limit.
2305 nr_soft_scanned = 0;
2306 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2307 sc->order, sc->gfp_mask,
2309 sc->nr_reclaimed += nr_soft_reclaimed;
2310 sc->nr_scanned += nr_soft_scanned;
2311 /* need some check for avoid more shrink_zone() */
2314 shrink_zone(zone, sc);
2317 return aborted_reclaim;
2320 static bool zone_reclaimable(struct zone *zone)
2322 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2325 /* All zones in zonelist are unreclaimable? */
2326 static bool all_unreclaimable(struct zonelist *zonelist,
2327 struct scan_control *sc)
2332 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2333 gfp_zone(sc->gfp_mask), sc->nodemask) {
2334 if (!populated_zone(zone))
2336 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2338 if (!zone->all_unreclaimable)
2346 * This is the main entry point to direct page reclaim.
2348 * If a full scan of the inactive list fails to free enough memory then we
2349 * are "out of memory" and something needs to be killed.
2351 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2352 * high - the zone may be full of dirty or under-writeback pages, which this
2353 * caller can't do much about. We kick the writeback threads and take explicit
2354 * naps in the hope that some of these pages can be written. But if the
2355 * allocating task holds filesystem locks which prevent writeout this might not
2356 * work, and the allocation attempt will fail.
2358 * returns: 0, if no pages reclaimed
2359 * else, the number of pages reclaimed
2361 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2362 struct scan_control *sc,
2363 struct shrink_control *shrink)
2365 unsigned long total_scanned = 0;
2366 struct reclaim_state *reclaim_state = current->reclaim_state;
2369 unsigned long writeback_threshold;
2370 bool aborted_reclaim;
2372 delayacct_freepages_start();
2374 if (global_reclaim(sc))
2375 count_vm_event(ALLOCSTALL);
2378 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2381 aborted_reclaim = shrink_zones(zonelist, sc);
2384 * Don't shrink slabs when reclaiming memory from over limit
2385 * cgroups but do shrink slab at least once when aborting
2386 * reclaim for compaction to avoid unevenly scanning file/anon
2387 * LRU pages over slab pages.
2389 if (global_reclaim(sc)) {
2390 unsigned long lru_pages = 0;
2392 nodes_clear(shrink->nodes_to_scan);
2393 for_each_zone_zonelist(zone, z, zonelist,
2394 gfp_zone(sc->gfp_mask)) {
2395 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2398 lru_pages += zone_reclaimable_pages(zone);
2399 node_set(zone_to_nid(zone),
2400 shrink->nodes_to_scan);
2403 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2404 if (reclaim_state) {
2405 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2406 reclaim_state->reclaimed_slab = 0;
2409 total_scanned += sc->nr_scanned;
2410 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2414 * If we're getting trouble reclaiming, start doing
2415 * writepage even in laptop mode.
2417 if (sc->priority < DEF_PRIORITY - 2)
2418 sc->may_writepage = 1;
2421 * Try to write back as many pages as we just scanned. This
2422 * tends to cause slow streaming writers to write data to the
2423 * disk smoothly, at the dirtying rate, which is nice. But
2424 * that's undesirable in laptop mode, where we *want* lumpy
2425 * writeout. So in laptop mode, write out the whole world.
2427 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2428 if (total_scanned > writeback_threshold) {
2429 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2430 WB_REASON_TRY_TO_FREE_PAGES);
2431 sc->may_writepage = 1;
2433 } while (--sc->priority >= 0 && !aborted_reclaim);
2436 delayacct_freepages_end();
2438 if (sc->nr_reclaimed)
2439 return sc->nr_reclaimed;
2442 * As hibernation is going on, kswapd is freezed so that it can't mark
2443 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2446 if (oom_killer_disabled)
2449 /* Aborted reclaim to try compaction? don't OOM, then */
2450 if (aborted_reclaim)
2453 /* top priority shrink_zones still had more to do? don't OOM, then */
2454 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2460 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2463 unsigned long pfmemalloc_reserve = 0;
2464 unsigned long free_pages = 0;
2468 for (i = 0; i <= ZONE_NORMAL; i++) {
2469 zone = &pgdat->node_zones[i];
2470 pfmemalloc_reserve += min_wmark_pages(zone);
2471 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2474 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2476 /* kswapd must be awake if processes are being throttled */
2477 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2478 pgdat->classzone_idx = min(pgdat->classzone_idx,
2479 (enum zone_type)ZONE_NORMAL);
2480 wake_up_interruptible(&pgdat->kswapd_wait);
2487 * Throttle direct reclaimers if backing storage is backed by the network
2488 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2489 * depleted. kswapd will continue to make progress and wake the processes
2490 * when the low watermark is reached.
2492 * Returns true if a fatal signal was delivered during throttling. If this
2493 * happens, the page allocator should not consider triggering the OOM killer.
2495 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2496 nodemask_t *nodemask)
2499 int high_zoneidx = gfp_zone(gfp_mask);
2503 * Kernel threads should not be throttled as they may be indirectly
2504 * responsible for cleaning pages necessary for reclaim to make forward
2505 * progress. kjournald for example may enter direct reclaim while
2506 * committing a transaction where throttling it could forcing other
2507 * processes to block on log_wait_commit().
2509 if (current->flags & PF_KTHREAD)
2513 * If a fatal signal is pending, this process should not throttle.
2514 * It should return quickly so it can exit and free its memory
2516 if (fatal_signal_pending(current))
2519 /* Check if the pfmemalloc reserves are ok */
2520 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2521 pgdat = zone->zone_pgdat;
2522 if (pfmemalloc_watermark_ok(pgdat))
2525 /* Account for the throttling */
2526 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2529 * If the caller cannot enter the filesystem, it's possible that it
2530 * is due to the caller holding an FS lock or performing a journal
2531 * transaction in the case of a filesystem like ext[3|4]. In this case,
2532 * it is not safe to block on pfmemalloc_wait as kswapd could be
2533 * blocked waiting on the same lock. Instead, throttle for up to a
2534 * second before continuing.
2536 if (!(gfp_mask & __GFP_FS)) {
2537 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2538 pfmemalloc_watermark_ok(pgdat), HZ);
2543 /* Throttle until kswapd wakes the process */
2544 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2545 pfmemalloc_watermark_ok(pgdat));
2548 if (fatal_signal_pending(current))
2555 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2556 gfp_t gfp_mask, nodemask_t *nodemask)
2558 unsigned long nr_reclaimed;
2559 struct scan_control sc = {
2560 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2561 .may_writepage = !laptop_mode,
2562 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2566 .priority = DEF_PRIORITY,
2567 .target_mem_cgroup = NULL,
2568 .nodemask = nodemask,
2570 struct shrink_control shrink = {
2571 .gfp_mask = sc.gfp_mask,
2575 * Do not enter reclaim if fatal signal was delivered while throttled.
2576 * 1 is returned so that the page allocator does not OOM kill at this
2579 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2582 trace_mm_vmscan_direct_reclaim_begin(order,
2586 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2588 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2590 return nr_reclaimed;
2595 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2596 gfp_t gfp_mask, bool noswap,
2598 unsigned long *nr_scanned)
2600 struct scan_control sc = {
2602 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2603 .may_writepage = !laptop_mode,
2605 .may_swap = !noswap,
2608 .target_mem_cgroup = memcg,
2610 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2612 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2613 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2615 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2620 * NOTE: Although we can get the priority field, using it
2621 * here is not a good idea, since it limits the pages we can scan.
2622 * if we don't reclaim here, the shrink_zone from balance_pgdat
2623 * will pick up pages from other mem cgroup's as well. We hack
2624 * the priority and make it zero.
2626 shrink_lruvec(lruvec, &sc);
2628 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2630 *nr_scanned = sc.nr_scanned;
2631 return sc.nr_reclaimed;
2634 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2638 struct zonelist *zonelist;
2639 unsigned long nr_reclaimed;
2641 struct scan_control sc = {
2642 .may_writepage = !laptop_mode,
2644 .may_swap = !noswap,
2645 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2647 .priority = DEF_PRIORITY,
2648 .target_mem_cgroup = memcg,
2649 .nodemask = NULL, /* we don't care the placement */
2650 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2651 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2653 struct shrink_control shrink = {
2654 .gfp_mask = sc.gfp_mask,
2658 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2659 * take care of from where we get pages. So the node where we start the
2660 * scan does not need to be the current node.
2662 nid = mem_cgroup_select_victim_node(memcg);
2664 zonelist = NODE_DATA(nid)->node_zonelists;
2666 trace_mm_vmscan_memcg_reclaim_begin(0,
2670 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2672 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2674 return nr_reclaimed;
2678 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2680 struct mem_cgroup *memcg;
2682 if (!total_swap_pages)
2685 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2687 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2689 if (inactive_anon_is_low(lruvec))
2690 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2691 sc, LRU_ACTIVE_ANON);
2693 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2697 static bool zone_balanced(struct zone *zone, int order,
2698 unsigned long balance_gap, int classzone_idx)
2700 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2701 balance_gap, classzone_idx, 0))
2704 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2705 !compaction_suitable(zone, order))
2712 * pgdat_balanced() is used when checking if a node is balanced.
2714 * For order-0, all zones must be balanced!
2716 * For high-order allocations only zones that meet watermarks and are in a
2717 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2718 * total of balanced pages must be at least 25% of the zones allowed by
2719 * classzone_idx for the node to be considered balanced. Forcing all zones to
2720 * be balanced for high orders can cause excessive reclaim when there are
2722 * The choice of 25% is due to
2723 * o a 16M DMA zone that is balanced will not balance a zone on any
2724 * reasonable sized machine
2725 * o On all other machines, the top zone must be at least a reasonable
2726 * percentage of the middle zones. For example, on 32-bit x86, highmem
2727 * would need to be at least 256M for it to be balance a whole node.
2728 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2729 * to balance a node on its own. These seemed like reasonable ratios.
2731 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2733 unsigned long managed_pages = 0;
2734 unsigned long balanced_pages = 0;
2737 /* Check the watermark levels */
2738 for (i = 0; i <= classzone_idx; i++) {
2739 struct zone *zone = pgdat->node_zones + i;
2741 if (!populated_zone(zone))
2744 managed_pages += zone->managed_pages;
2747 * A special case here:
2749 * balance_pgdat() skips over all_unreclaimable after
2750 * DEF_PRIORITY. Effectively, it considers them balanced so
2751 * they must be considered balanced here as well!
2753 if (zone->all_unreclaimable) {
2754 balanced_pages += zone->managed_pages;
2758 if (zone_balanced(zone, order, 0, i))
2759 balanced_pages += zone->managed_pages;
2765 return balanced_pages >= (managed_pages >> 2);
2771 * Prepare kswapd for sleeping. This verifies that there are no processes
2772 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2774 * Returns true if kswapd is ready to sleep
2776 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2779 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2784 * There is a potential race between when kswapd checks its watermarks
2785 * and a process gets throttled. There is also a potential race if
2786 * processes get throttled, kswapd wakes, a large process exits therby
2787 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2788 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2789 * so wake them now if necessary. If necessary, processes will wake
2790 * kswapd and get throttled again
2792 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2793 wake_up(&pgdat->pfmemalloc_wait);
2797 return pgdat_balanced(pgdat, order, classzone_idx);
2801 * kswapd shrinks the zone by the number of pages required to reach
2802 * the high watermark.
2804 * Returns true if kswapd scanned at least the requested number of pages to
2805 * reclaim or if the lack of progress was due to pages under writeback.
2806 * This is used to determine if the scanning priority needs to be raised.
2808 static bool kswapd_shrink_zone(struct zone *zone,
2810 struct scan_control *sc,
2811 unsigned long lru_pages,
2812 unsigned long *nr_attempted)
2814 unsigned long nr_slab;
2815 int testorder = sc->order;
2816 unsigned long balance_gap;
2817 struct reclaim_state *reclaim_state = current->reclaim_state;
2818 struct shrink_control shrink = {
2819 .gfp_mask = sc->gfp_mask,
2821 bool lowmem_pressure;
2823 /* Reclaim above the high watermark. */
2824 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2827 * Kswapd reclaims only single pages with compaction enabled. Trying
2828 * too hard to reclaim until contiguous free pages have become
2829 * available can hurt performance by evicting too much useful data
2830 * from memory. Do not reclaim more than needed for compaction.
2832 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2833 compaction_suitable(zone, sc->order) !=
2838 * We put equal pressure on every zone, unless one zone has way too
2839 * many pages free already. The "too many pages" is defined as the
2840 * high wmark plus a "gap" where the gap is either the low
2841 * watermark or 1% of the zone, whichever is smaller.
2843 balance_gap = min(low_wmark_pages(zone),
2844 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2845 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2848 * If there is no low memory pressure or the zone is balanced then no
2849 * reclaim is necessary
2851 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2852 if (!lowmem_pressure && zone_balanced(zone, testorder,
2853 balance_gap, classzone_idx))
2856 shrink_zone(zone, sc);
2857 nodes_clear(shrink.nodes_to_scan);
2858 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2860 reclaim_state->reclaimed_slab = 0;
2861 nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2862 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2864 /* Account for the number of pages attempted to reclaim */
2865 *nr_attempted += sc->nr_to_reclaim;
2867 if (nr_slab == 0 && !zone_reclaimable(zone))
2868 zone->all_unreclaimable = 1;
2870 zone_clear_flag(zone, ZONE_WRITEBACK);
2873 * If a zone reaches its high watermark, consider it to be no longer
2874 * congested. It's possible there are dirty pages backed by congested
2875 * BDIs but as pressure is relieved, speculatively avoid congestion
2878 if (!zone->all_unreclaimable &&
2879 zone_balanced(zone, testorder, 0, classzone_idx)) {
2880 zone_clear_flag(zone, ZONE_CONGESTED);
2881 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2884 return sc->nr_scanned >= sc->nr_to_reclaim;
2888 * For kswapd, balance_pgdat() will work across all this node's zones until
2889 * they are all at high_wmark_pages(zone).
2891 * Returns the final order kswapd was reclaiming at
2893 * There is special handling here for zones which are full of pinned pages.
2894 * This can happen if the pages are all mlocked, or if they are all used by
2895 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2896 * What we do is to detect the case where all pages in the zone have been
2897 * scanned twice and there has been zero successful reclaim. Mark the zone as
2898 * dead and from now on, only perform a short scan. Basically we're polling
2899 * the zone for when the problem goes away.
2901 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2902 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2903 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2904 * lower zones regardless of the number of free pages in the lower zones. This
2905 * interoperates with the page allocator fallback scheme to ensure that aging
2906 * of pages is balanced across the zones.
2908 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2912 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2913 unsigned long nr_soft_reclaimed;
2914 unsigned long nr_soft_scanned;
2915 struct scan_control sc = {
2916 .gfp_mask = GFP_KERNEL,
2917 .priority = DEF_PRIORITY,
2920 .may_writepage = !laptop_mode,
2922 .target_mem_cgroup = NULL,
2924 count_vm_event(PAGEOUTRUN);
2927 unsigned long lru_pages = 0;
2928 unsigned long nr_attempted = 0;
2929 bool raise_priority = true;
2930 bool pgdat_needs_compaction = (order > 0);
2932 sc.nr_reclaimed = 0;
2935 * Scan in the highmem->dma direction for the highest
2936 * zone which needs scanning
2938 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2939 struct zone *zone = pgdat->node_zones + i;
2941 if (!populated_zone(zone))
2944 if (zone->all_unreclaimable &&
2945 sc.priority != DEF_PRIORITY)
2949 * Do some background aging of the anon list, to give
2950 * pages a chance to be referenced before reclaiming.
2952 age_active_anon(zone, &sc);
2955 * If the number of buffer_heads in the machine
2956 * exceeds the maximum allowed level and this node
2957 * has a highmem zone, force kswapd to reclaim from
2958 * it to relieve lowmem pressure.
2960 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2965 if (!zone_balanced(zone, order, 0, 0)) {
2970 * If balanced, clear the dirty and congested
2973 zone_clear_flag(zone, ZONE_CONGESTED);
2974 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2981 for (i = 0; i <= end_zone; i++) {
2982 struct zone *zone = pgdat->node_zones + i;
2984 if (!populated_zone(zone))
2987 lru_pages += zone_reclaimable_pages(zone);
2990 * If any zone is currently balanced then kswapd will
2991 * not call compaction as it is expected that the
2992 * necessary pages are already available.
2994 if (pgdat_needs_compaction &&
2995 zone_watermark_ok(zone, order,
2996 low_wmark_pages(zone),
2998 pgdat_needs_compaction = false;
3002 * If we're getting trouble reclaiming, start doing writepage
3003 * even in laptop mode.
3005 if (sc.priority < DEF_PRIORITY - 2)
3006 sc.may_writepage = 1;
3009 * Now scan the zone in the dma->highmem direction, stopping
3010 * at the last zone which needs scanning.
3012 * We do this because the page allocator works in the opposite
3013 * direction. This prevents the page allocator from allocating
3014 * pages behind kswapd's direction of progress, which would
3015 * cause too much scanning of the lower zones.
3017 for (i = 0; i <= end_zone; i++) {
3018 struct zone *zone = pgdat->node_zones + i;
3020 if (!populated_zone(zone))
3023 if (zone->all_unreclaimable &&
3024 sc.priority != DEF_PRIORITY)
3029 nr_soft_scanned = 0;
3031 * Call soft limit reclaim before calling shrink_zone.
3033 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3036 sc.nr_reclaimed += nr_soft_reclaimed;
3039 * There should be no need to raise the scanning
3040 * priority if enough pages are already being scanned
3041 * that that high watermark would be met at 100%
3044 if (kswapd_shrink_zone(zone, end_zone, &sc,
3045 lru_pages, &nr_attempted))
3046 raise_priority = false;
3050 * If the low watermark is met there is no need for processes
3051 * to be throttled on pfmemalloc_wait as they should not be
3052 * able to safely make forward progress. Wake them
3054 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3055 pfmemalloc_watermark_ok(pgdat))
3056 wake_up(&pgdat->pfmemalloc_wait);
3059 * Fragmentation may mean that the system cannot be rebalanced
3060 * for high-order allocations in all zones. If twice the
3061 * allocation size has been reclaimed and the zones are still
3062 * not balanced then recheck the watermarks at order-0 to
3063 * prevent kswapd reclaiming excessively. Assume that a
3064 * process requested a high-order can direct reclaim/compact.
3066 if (order && sc.nr_reclaimed >= 2UL << order)
3067 order = sc.order = 0;
3069 /* Check if kswapd should be suspending */
3070 if (try_to_freeze() || kthread_should_stop())
3074 * Compact if necessary and kswapd is reclaiming at least the
3075 * high watermark number of pages as requsted
3077 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3078 compact_pgdat(pgdat, order);
3081 * Raise priority if scanning rate is too low or there was no
3082 * progress in reclaiming pages
3084 if (raise_priority || !sc.nr_reclaimed)
3086 } while (sc.priority >= 1 &&
3087 !pgdat_balanced(pgdat, order, *classzone_idx));
3091 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3092 * makes a decision on the order we were last reclaiming at. However,
3093 * if another caller entered the allocator slow path while kswapd
3094 * was awake, order will remain at the higher level
3096 *classzone_idx = end_zone;
3100 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3105 if (freezing(current) || kthread_should_stop())
3108 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3110 /* Try to sleep for a short interval */
3111 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3112 remaining = schedule_timeout(HZ/10);
3113 finish_wait(&pgdat->kswapd_wait, &wait);
3114 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3118 * After a short sleep, check if it was a premature sleep. If not, then
3119 * go fully to sleep until explicitly woken up.
3121 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3122 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3125 * vmstat counters are not perfectly accurate and the estimated
3126 * value for counters such as NR_FREE_PAGES can deviate from the
3127 * true value by nr_online_cpus * threshold. To avoid the zone
3128 * watermarks being breached while under pressure, we reduce the
3129 * per-cpu vmstat threshold while kswapd is awake and restore
3130 * them before going back to sleep.
3132 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3135 * Compaction records what page blocks it recently failed to
3136 * isolate pages from and skips them in the future scanning.
3137 * When kswapd is going to sleep, it is reasonable to assume
3138 * that pages and compaction may succeed so reset the cache.
3140 reset_isolation_suitable(pgdat);
3142 if (!kthread_should_stop())
3145 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3148 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3150 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3152 finish_wait(&pgdat->kswapd_wait, &wait);
3156 * The background pageout daemon, started as a kernel thread
3157 * from the init process.
3159 * This basically trickles out pages so that we have _some_
3160 * free memory available even if there is no other activity
3161 * that frees anything up. This is needed for things like routing
3162 * etc, where we otherwise might have all activity going on in
3163 * asynchronous contexts that cannot page things out.
3165 * If there are applications that are active memory-allocators
3166 * (most normal use), this basically shouldn't matter.
3168 static int kswapd(void *p)
3170 unsigned long order, new_order;
3171 unsigned balanced_order;
3172 int classzone_idx, new_classzone_idx;
3173 int balanced_classzone_idx;
3174 pg_data_t *pgdat = (pg_data_t*)p;
3175 struct task_struct *tsk = current;
3177 struct reclaim_state reclaim_state = {
3178 .reclaimed_slab = 0,
3180 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3182 lockdep_set_current_reclaim_state(GFP_KERNEL);
3184 if (!cpumask_empty(cpumask))
3185 set_cpus_allowed_ptr(tsk, cpumask);
3186 current->reclaim_state = &reclaim_state;
3189 * Tell the memory management that we're a "memory allocator",
3190 * and that if we need more memory we should get access to it
3191 * regardless (see "__alloc_pages()"). "kswapd" should
3192 * never get caught in the normal page freeing logic.
3194 * (Kswapd normally doesn't need memory anyway, but sometimes
3195 * you need a small amount of memory in order to be able to
3196 * page out something else, and this flag essentially protects
3197 * us from recursively trying to free more memory as we're
3198 * trying to free the first piece of memory in the first place).
3200 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3203 order = new_order = 0;
3205 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3206 balanced_classzone_idx = classzone_idx;
3211 * If the last balance_pgdat was unsuccessful it's unlikely a
3212 * new request of a similar or harder type will succeed soon
3213 * so consider going to sleep on the basis we reclaimed at
3215 if (balanced_classzone_idx >= new_classzone_idx &&
3216 balanced_order == new_order) {
3217 new_order = pgdat->kswapd_max_order;
3218 new_classzone_idx = pgdat->classzone_idx;
3219 pgdat->kswapd_max_order = 0;
3220 pgdat->classzone_idx = pgdat->nr_zones - 1;
3223 if (order < new_order || classzone_idx > new_classzone_idx) {
3225 * Don't sleep if someone wants a larger 'order'
3226 * allocation or has tigher zone constraints
3229 classzone_idx = new_classzone_idx;
3231 kswapd_try_to_sleep(pgdat, balanced_order,
3232 balanced_classzone_idx);
3233 order = pgdat->kswapd_max_order;
3234 classzone_idx = pgdat->classzone_idx;
3236 new_classzone_idx = classzone_idx;
3237 pgdat->kswapd_max_order = 0;
3238 pgdat->classzone_idx = pgdat->nr_zones - 1;
3241 ret = try_to_freeze();
3242 if (kthread_should_stop())
3246 * We can speed up thawing tasks if we don't call balance_pgdat
3247 * after returning from the refrigerator
3250 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3251 balanced_classzone_idx = classzone_idx;
3252 balanced_order = balance_pgdat(pgdat, order,
3253 &balanced_classzone_idx);
3257 current->reclaim_state = NULL;
3262 * A zone is low on free memory, so wake its kswapd task to service it.
3264 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3268 if (!populated_zone(zone))
3271 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3273 pgdat = zone->zone_pgdat;
3274 if (pgdat->kswapd_max_order < order) {
3275 pgdat->kswapd_max_order = order;
3276 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3278 if (!waitqueue_active(&pgdat->kswapd_wait))
3280 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3283 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3284 wake_up_interruptible(&pgdat->kswapd_wait);
3288 * The reclaimable count would be mostly accurate.
3289 * The less reclaimable pages may be
3290 * - mlocked pages, which will be moved to unevictable list when encountered
3291 * - mapped pages, which may require several travels to be reclaimed
3292 * - dirty pages, which is not "instantly" reclaimable
3294 unsigned long global_reclaimable_pages(void)
3298 nr = global_page_state(NR_ACTIVE_FILE) +
3299 global_page_state(NR_INACTIVE_FILE);
3301 if (get_nr_swap_pages() > 0)
3302 nr += global_page_state(NR_ACTIVE_ANON) +
3303 global_page_state(NR_INACTIVE_ANON);
3308 unsigned long zone_reclaimable_pages(struct zone *zone)
3312 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3313 zone_page_state(zone, NR_INACTIVE_FILE);
3315 if (get_nr_swap_pages() > 0)
3316 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3317 zone_page_state(zone, NR_INACTIVE_ANON);
3322 #ifdef CONFIG_HIBERNATION
3324 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3327 * Rather than trying to age LRUs the aim is to preserve the overall
3328 * LRU order by reclaiming preferentially
3329 * inactive > active > active referenced > active mapped
3331 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3333 struct reclaim_state reclaim_state;
3334 struct scan_control sc = {
3335 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3339 .nr_to_reclaim = nr_to_reclaim,
3340 .hibernation_mode = 1,
3342 .priority = DEF_PRIORITY,
3344 struct shrink_control shrink = {
3345 .gfp_mask = sc.gfp_mask,
3347 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3348 struct task_struct *p = current;
3349 unsigned long nr_reclaimed;
3351 p->flags |= PF_MEMALLOC;
3352 lockdep_set_current_reclaim_state(sc.gfp_mask);
3353 reclaim_state.reclaimed_slab = 0;
3354 p->reclaim_state = &reclaim_state;
3356 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3358 p->reclaim_state = NULL;
3359 lockdep_clear_current_reclaim_state();
3360 p->flags &= ~PF_MEMALLOC;
3362 return nr_reclaimed;
3364 #endif /* CONFIG_HIBERNATION */
3366 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3367 not required for correctness. So if the last cpu in a node goes
3368 away, we get changed to run anywhere: as the first one comes back,
3369 restore their cpu bindings. */
3370 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3375 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3376 for_each_node_state(nid, N_MEMORY) {
3377 pg_data_t *pgdat = NODE_DATA(nid);
3378 const struct cpumask *mask;
3380 mask = cpumask_of_node(pgdat->node_id);
3382 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3383 /* One of our CPUs online: restore mask */
3384 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3391 * This kswapd start function will be called by init and node-hot-add.
3392 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3394 int kswapd_run(int nid)
3396 pg_data_t *pgdat = NODE_DATA(nid);
3402 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3403 if (IS_ERR(pgdat->kswapd)) {
3404 /* failure at boot is fatal */
3405 BUG_ON(system_state == SYSTEM_BOOTING);
3406 pr_err("Failed to start kswapd on node %d\n", nid);
3407 ret = PTR_ERR(pgdat->kswapd);
3408 pgdat->kswapd = NULL;
3414 * Called by memory hotplug when all memory in a node is offlined. Caller must
3415 * hold lock_memory_hotplug().
3417 void kswapd_stop(int nid)
3419 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3422 kthread_stop(kswapd);
3423 NODE_DATA(nid)->kswapd = NULL;
3427 static int __init kswapd_init(void)
3432 for_each_node_state(nid, N_MEMORY)
3434 hotcpu_notifier(cpu_callback, 0);
3438 module_init(kswapd_init)
3444 * If non-zero call zone_reclaim when the number of free pages falls below
3447 int zone_reclaim_mode __read_mostly;
3449 #define RECLAIM_OFF 0
3450 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3451 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3452 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3455 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3456 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3459 #define ZONE_RECLAIM_PRIORITY 4
3462 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3465 int sysctl_min_unmapped_ratio = 1;
3468 * If the number of slab pages in a zone grows beyond this percentage then
3469 * slab reclaim needs to occur.
3471 int sysctl_min_slab_ratio = 5;
3473 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3475 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3476 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3477 zone_page_state(zone, NR_ACTIVE_FILE);
3480 * It's possible for there to be more file mapped pages than
3481 * accounted for by the pages on the file LRU lists because
3482 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3484 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3487 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3488 static long zone_pagecache_reclaimable(struct zone *zone)
3490 long nr_pagecache_reclaimable;
3494 * If RECLAIM_SWAP is set, then all file pages are considered
3495 * potentially reclaimable. Otherwise, we have to worry about
3496 * pages like swapcache and zone_unmapped_file_pages() provides
3499 if (zone_reclaim_mode & RECLAIM_SWAP)
3500 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3502 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3504 /* If we can't clean pages, remove dirty pages from consideration */
3505 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3506 delta += zone_page_state(zone, NR_FILE_DIRTY);
3508 /* Watch for any possible underflows due to delta */
3509 if (unlikely(delta > nr_pagecache_reclaimable))
3510 delta = nr_pagecache_reclaimable;
3512 return nr_pagecache_reclaimable - delta;
3516 * Try to free up some pages from this zone through reclaim.
3518 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3520 /* Minimum pages needed in order to stay on node */
3521 const unsigned long nr_pages = 1 << order;
3522 struct task_struct *p = current;
3523 struct reclaim_state reclaim_state;
3524 struct scan_control sc = {
3525 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3526 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3528 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3529 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3531 .priority = ZONE_RECLAIM_PRIORITY,
3533 struct shrink_control shrink = {
3534 .gfp_mask = sc.gfp_mask,
3536 unsigned long nr_slab_pages0, nr_slab_pages1;
3540 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3541 * and we also need to be able to write out pages for RECLAIM_WRITE
3544 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3545 lockdep_set_current_reclaim_state(gfp_mask);
3546 reclaim_state.reclaimed_slab = 0;
3547 p->reclaim_state = &reclaim_state;
3549 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3551 * Free memory by calling shrink zone with increasing
3552 * priorities until we have enough memory freed.
3555 shrink_zone(zone, &sc);
3556 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3559 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3560 if (nr_slab_pages0 > zone->min_slab_pages) {
3562 * shrink_slab() does not currently allow us to determine how
3563 * many pages were freed in this zone. So we take the current
3564 * number of slab pages and shake the slab until it is reduced
3565 * by the same nr_pages that we used for reclaiming unmapped
3568 nodes_clear(shrink.nodes_to_scan);
3569 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3571 unsigned long lru_pages = zone_reclaimable_pages(zone);
3573 /* No reclaimable slab or very low memory pressure */
3574 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3577 /* Freed enough memory */
3578 nr_slab_pages1 = zone_page_state(zone,
3579 NR_SLAB_RECLAIMABLE);
3580 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3585 * Update nr_reclaimed by the number of slab pages we
3586 * reclaimed from this zone.
3588 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3589 if (nr_slab_pages1 < nr_slab_pages0)
3590 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3593 p->reclaim_state = NULL;
3594 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3595 lockdep_clear_current_reclaim_state();
3596 return sc.nr_reclaimed >= nr_pages;
3599 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3605 * Zone reclaim reclaims unmapped file backed pages and
3606 * slab pages if we are over the defined limits.
3608 * A small portion of unmapped file backed pages is needed for
3609 * file I/O otherwise pages read by file I/O will be immediately
3610 * thrown out if the zone is overallocated. So we do not reclaim
3611 * if less than a specified percentage of the zone is used by
3612 * unmapped file backed pages.
3614 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3615 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3616 return ZONE_RECLAIM_FULL;
3618 if (zone->all_unreclaimable)
3619 return ZONE_RECLAIM_FULL;
3622 * Do not scan if the allocation should not be delayed.
3624 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3625 return ZONE_RECLAIM_NOSCAN;
3628 * Only run zone reclaim on the local zone or on zones that do not
3629 * have associated processors. This will favor the local processor
3630 * over remote processors and spread off node memory allocations
3631 * as wide as possible.
3633 node_id = zone_to_nid(zone);
3634 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3635 return ZONE_RECLAIM_NOSCAN;
3637 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3638 return ZONE_RECLAIM_NOSCAN;
3640 ret = __zone_reclaim(zone, gfp_mask, order);
3641 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3644 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3651 * page_evictable - test whether a page is evictable
3652 * @page: the page to test
3654 * Test whether page is evictable--i.e., should be placed on active/inactive
3655 * lists vs unevictable list.
3657 * Reasons page might not be evictable:
3658 * (1) page's mapping marked unevictable
3659 * (2) page is part of an mlocked VMA
3662 int page_evictable(struct page *page)
3664 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3669 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3670 * @pages: array of pages to check
3671 * @nr_pages: number of pages to check
3673 * Checks pages for evictability and moves them to the appropriate lru list.
3675 * This function is only used for SysV IPC SHM_UNLOCK.
3677 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3679 struct lruvec *lruvec;
3680 struct zone *zone = NULL;
3685 for (i = 0; i < nr_pages; i++) {
3686 struct page *page = pages[i];
3687 struct zone *pagezone;
3690 pagezone = page_zone(page);
3691 if (pagezone != zone) {
3693 spin_unlock_irq(&zone->lru_lock);
3695 spin_lock_irq(&zone->lru_lock);
3697 lruvec = mem_cgroup_page_lruvec(page, zone);
3699 if (!PageLRU(page) || !PageUnevictable(page))
3702 if (page_evictable(page)) {
3703 enum lru_list lru = page_lru_base_type(page);
3705 VM_BUG_ON(PageActive(page));
3706 ClearPageUnevictable(page);
3707 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3708 add_page_to_lru_list(page, lruvec, lru);
3714 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3715 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3716 spin_unlock_irq(&zone->lru_lock);
3719 #endif /* CONFIG_SHMEM */
3721 static void warn_scan_unevictable_pages(void)
3723 printk_once(KERN_WARNING
3724 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3725 "disabled for lack of a legitimate use case. If you have "
3726 "one, please send an email to linux-mm@kvack.org.\n",
3731 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3732 * all nodes' unevictable lists for evictable pages
3734 unsigned long scan_unevictable_pages;
3736 int scan_unevictable_handler(struct ctl_table *table, int write,
3737 void __user *buffer,
3738 size_t *length, loff_t *ppos)
3740 warn_scan_unevictable_pages();
3741 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3742 scan_unevictable_pages = 0;
3748 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3749 * a specified node's per zone unevictable lists for evictable pages.
3752 static ssize_t read_scan_unevictable_node(struct device *dev,
3753 struct device_attribute *attr,
3756 warn_scan_unevictable_pages();
3757 return sprintf(buf, "0\n"); /* always zero; should fit... */
3760 static ssize_t write_scan_unevictable_node(struct device *dev,
3761 struct device_attribute *attr,
3762 const char *buf, size_t count)
3764 warn_scan_unevictable_pages();
3769 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3770 read_scan_unevictable_node,
3771 write_scan_unevictable_node);
3773 int scan_unevictable_register_node(struct node *node)
3775 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3778 void scan_unevictable_unregister_node(struct node *node)
3780 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);