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/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 * reclaim_mode determines how the inactive list is shrunk
58 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59 * RECLAIM_MODE_ASYNC: Do not block
60 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
61 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
62 * order-0 pages and then compact the zone
64 typedef unsigned __bitwise__ reclaim_mode_t;
65 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
66 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
67 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
68 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
71 /* Incremented by the number of inactive pages that were scanned */
72 unsigned long nr_scanned;
74 /* Number of pages freed so far during a call to shrink_zones() */
75 unsigned long nr_reclaimed;
77 /* How many pages shrink_list() should reclaim */
78 unsigned long nr_to_reclaim;
80 unsigned long hibernation_mode;
82 /* This context's GFP mask */
87 /* Can mapped pages be reclaimed? */
90 /* Can pages be swapped as part of reclaim? */
96 * Intend to reclaim enough continuous memory rather than reclaim
97 * enough amount of memory. i.e, mode for high order allocation.
99 reclaim_mode_t reclaim_mode;
102 * The memory cgroup that hit its limit and as a result is the
103 * primary target of this reclaim invocation.
105 struct mem_cgroup *target_mem_cgroup;
108 * Nodemask of nodes allowed by the caller. If NULL, all nodes
111 nodemask_t *nodemask;
114 struct mem_cgroup_zone {
115 struct mem_cgroup *mem_cgroup;
119 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
121 #ifdef ARCH_HAS_PREFETCH
122 #define prefetch_prev_lru_page(_page, _base, _field) \
124 if ((_page)->lru.prev != _base) { \
127 prev = lru_to_page(&(_page->lru)); \
128 prefetch(&prev->_field); \
132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
135 #ifdef ARCH_HAS_PREFETCHW
136 #define prefetchw_prev_lru_page(_page, _base, _field) \
138 if ((_page)->lru.prev != _base) { \
141 prev = lru_to_page(&(_page->lru)); \
142 prefetchw(&prev->_field); \
146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
150 * From 0 .. 100. Higher means more swappy.
152 int vm_swappiness = 60;
153 long vm_total_pages; /* The total number of pages which the VM controls */
155 static LIST_HEAD(shrinker_list);
156 static DECLARE_RWSEM(shrinker_rwsem);
158 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
159 static bool global_reclaim(struct scan_control *sc)
161 return !sc->target_mem_cgroup;
164 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
166 return !mz->mem_cgroup;
169 static bool global_reclaim(struct scan_control *sc)
174 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
180 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
182 if (!scanning_global_lru(mz))
183 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
185 return &mz->zone->reclaim_stat;
188 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
191 if (!scanning_global_lru(mz))
192 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
193 zone_to_nid(mz->zone),
197 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
202 * Add a shrinker callback to be called from the vm
204 void register_shrinker(struct shrinker *shrinker)
206 atomic_long_set(&shrinker->nr_in_batch, 0);
207 down_write(&shrinker_rwsem);
208 list_add_tail(&shrinker->list, &shrinker_list);
209 up_write(&shrinker_rwsem);
211 EXPORT_SYMBOL(register_shrinker);
216 void unregister_shrinker(struct shrinker *shrinker)
218 down_write(&shrinker_rwsem);
219 list_del(&shrinker->list);
220 up_write(&shrinker_rwsem);
222 EXPORT_SYMBOL(unregister_shrinker);
224 static inline int do_shrinker_shrink(struct shrinker *shrinker,
225 struct shrink_control *sc,
226 unsigned long nr_to_scan)
228 sc->nr_to_scan = nr_to_scan;
229 return (*shrinker->shrink)(shrinker, sc);
232 #define SHRINK_BATCH 128
234 * Call the shrink functions to age shrinkable caches
236 * Here we assume it costs one seek to replace a lru page and that it also
237 * takes a seek to recreate a cache object. With this in mind we age equal
238 * percentages of the lru and ageable caches. This should balance the seeks
239 * generated by these structures.
241 * If the vm encountered mapped pages on the LRU it increase the pressure on
242 * slab to avoid swapping.
244 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
246 * `lru_pages' represents the number of on-LRU pages in all the zones which
247 * are eligible for the caller's allocation attempt. It is used for balancing
248 * slab reclaim versus page reclaim.
250 * Returns the number of slab objects which we shrunk.
252 unsigned long shrink_slab(struct shrink_control *shrink,
253 unsigned long nr_pages_scanned,
254 unsigned long lru_pages)
256 struct shrinker *shrinker;
257 unsigned long ret = 0;
259 if (nr_pages_scanned == 0)
260 nr_pages_scanned = SWAP_CLUSTER_MAX;
262 if (!down_read_trylock(&shrinker_rwsem)) {
263 /* Assume we'll be able to shrink next time */
268 list_for_each_entry(shrinker, &shrinker_list, list) {
269 unsigned long long delta;
275 long batch_size = shrinker->batch ? shrinker->batch
278 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
283 * copy the current shrinker scan count into a local variable
284 * and zero it so that other concurrent shrinker invocations
285 * don't also do this scanning work.
287 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
290 delta = (4 * nr_pages_scanned) / shrinker->seeks;
292 do_div(delta, lru_pages + 1);
294 if (total_scan < 0) {
295 printk(KERN_ERR "shrink_slab: %pF negative objects to "
297 shrinker->shrink, total_scan);
298 total_scan = max_pass;
302 * We need to avoid excessive windup on filesystem shrinkers
303 * due to large numbers of GFP_NOFS allocations causing the
304 * shrinkers to return -1 all the time. This results in a large
305 * nr being built up so when a shrink that can do some work
306 * comes along it empties the entire cache due to nr >>>
307 * max_pass. This is bad for sustaining a working set in
310 * Hence only allow the shrinker to scan the entire cache when
311 * a large delta change is calculated directly.
313 if (delta < max_pass / 4)
314 total_scan = min(total_scan, max_pass / 2);
317 * Avoid risking looping forever due to too large nr value:
318 * never try to free more than twice the estimate number of
321 if (total_scan > max_pass * 2)
322 total_scan = max_pass * 2;
324 trace_mm_shrink_slab_start(shrinker, shrink, nr,
325 nr_pages_scanned, lru_pages,
326 max_pass, delta, total_scan);
328 while (total_scan >= batch_size) {
331 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
332 shrink_ret = do_shrinker_shrink(shrinker, shrink,
334 if (shrink_ret == -1)
336 if (shrink_ret < nr_before)
337 ret += nr_before - shrink_ret;
338 count_vm_events(SLABS_SCANNED, batch_size);
339 total_scan -= batch_size;
345 * move the unused scan count back into the shrinker in a
346 * manner that handles concurrent updates. If we exhausted the
347 * scan, there is no need to do an update.
350 new_nr = atomic_long_add_return(total_scan,
351 &shrinker->nr_in_batch);
353 new_nr = atomic_long_read(&shrinker->nr_in_batch);
355 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
357 up_read(&shrinker_rwsem);
363 static void set_reclaim_mode(int priority, struct scan_control *sc,
366 /* Sync reclaim used only for compaction */
367 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
370 * Restrict reclaim/compaction to costly allocations or when
371 * under memory pressure
373 if (COMPACTION_BUILD && sc->order &&
374 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
375 priority < DEF_PRIORITY - 2))
376 sc->reclaim_mode = RECLAIM_MODE_COMPACTION | syncmode;
378 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
381 static void reset_reclaim_mode(struct scan_control *sc)
383 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
386 static inline int is_page_cache_freeable(struct page *page)
389 * A freeable page cache page is referenced only by the caller
390 * that isolated the page, the page cache radix tree and
391 * optional buffer heads at page->private.
393 return page_count(page) - page_has_private(page) == 2;
396 static int may_write_to_queue(struct backing_dev_info *bdi,
397 struct scan_control *sc)
399 if (current->flags & PF_SWAPWRITE)
401 if (!bdi_write_congested(bdi))
403 if (bdi == current->backing_dev_info)
409 * We detected a synchronous write error writing a page out. Probably
410 * -ENOSPC. We need to propagate that into the address_space for a subsequent
411 * fsync(), msync() or close().
413 * The tricky part is that after writepage we cannot touch the mapping: nothing
414 * prevents it from being freed up. But we have a ref on the page and once
415 * that page is locked, the mapping is pinned.
417 * We're allowed to run sleeping lock_page() here because we know the caller has
420 static void handle_write_error(struct address_space *mapping,
421 struct page *page, int error)
424 if (page_mapping(page) == mapping)
425 mapping_set_error(mapping, error);
429 /* possible outcome of pageout() */
431 /* failed to write page out, page is locked */
433 /* move page to the active list, page is locked */
435 /* page has been sent to the disk successfully, page is unlocked */
437 /* page is clean and locked */
442 * pageout is called by shrink_page_list() for each dirty page.
443 * Calls ->writepage().
445 static pageout_t pageout(struct page *page, struct address_space *mapping,
446 struct scan_control *sc)
449 * If the page is dirty, only perform writeback if that write
450 * will be non-blocking. To prevent this allocation from being
451 * stalled by pagecache activity. But note that there may be
452 * stalls if we need to run get_block(). We could test
453 * PagePrivate for that.
455 * If this process is currently in __generic_file_aio_write() against
456 * this page's queue, we can perform writeback even if that
459 * If the page is swapcache, write it back even if that would
460 * block, for some throttling. This happens by accident, because
461 * swap_backing_dev_info is bust: it doesn't reflect the
462 * congestion state of the swapdevs. Easy to fix, if needed.
464 if (!is_page_cache_freeable(page))
468 * Some data journaling orphaned pages can have
469 * page->mapping == NULL while being dirty with clean buffers.
471 if (page_has_private(page)) {
472 if (try_to_free_buffers(page)) {
473 ClearPageDirty(page);
474 printk("%s: orphaned page\n", __func__);
480 if (mapping->a_ops->writepage == NULL)
481 return PAGE_ACTIVATE;
482 if (!may_write_to_queue(mapping->backing_dev_info, sc))
485 if (clear_page_dirty_for_io(page)) {
487 struct writeback_control wbc = {
488 .sync_mode = WB_SYNC_NONE,
489 .nr_to_write = SWAP_CLUSTER_MAX,
491 .range_end = LLONG_MAX,
495 SetPageReclaim(page);
496 res = mapping->a_ops->writepage(page, &wbc);
498 handle_write_error(mapping, page, res);
499 if (res == AOP_WRITEPAGE_ACTIVATE) {
500 ClearPageReclaim(page);
501 return PAGE_ACTIVATE;
504 if (!PageWriteback(page)) {
505 /* synchronous write or broken a_ops? */
506 ClearPageReclaim(page);
508 trace_mm_vmscan_writepage(page,
509 trace_reclaim_flags(page, sc->reclaim_mode));
510 inc_zone_page_state(page, NR_VMSCAN_WRITE);
518 * Same as remove_mapping, but if the page is removed from the mapping, it
519 * gets returned with a refcount of 0.
521 static int __remove_mapping(struct address_space *mapping, struct page *page)
523 BUG_ON(!PageLocked(page));
524 BUG_ON(mapping != page_mapping(page));
526 spin_lock_irq(&mapping->tree_lock);
528 * The non racy check for a busy page.
530 * Must be careful with the order of the tests. When someone has
531 * a ref to the page, it may be possible that they dirty it then
532 * drop the reference. So if PageDirty is tested before page_count
533 * here, then the following race may occur:
535 * get_user_pages(&page);
536 * [user mapping goes away]
538 * !PageDirty(page) [good]
539 * SetPageDirty(page);
541 * !page_count(page) [good, discard it]
543 * [oops, our write_to data is lost]
545 * Reversing the order of the tests ensures such a situation cannot
546 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
547 * load is not satisfied before that of page->_count.
549 * Note that if SetPageDirty is always performed via set_page_dirty,
550 * and thus under tree_lock, then this ordering is not required.
552 if (!page_freeze_refs(page, 2))
554 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
555 if (unlikely(PageDirty(page))) {
556 page_unfreeze_refs(page, 2);
560 if (PageSwapCache(page)) {
561 swp_entry_t swap = { .val = page_private(page) };
562 __delete_from_swap_cache(page);
563 spin_unlock_irq(&mapping->tree_lock);
564 swapcache_free(swap, page);
566 void (*freepage)(struct page *);
568 freepage = mapping->a_ops->freepage;
570 __delete_from_page_cache(page);
571 spin_unlock_irq(&mapping->tree_lock);
572 mem_cgroup_uncharge_cache_page(page);
574 if (freepage != NULL)
581 spin_unlock_irq(&mapping->tree_lock);
586 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
587 * someone else has a ref on the page, abort and return 0. If it was
588 * successfully detached, return 1. Assumes the caller has a single ref on
591 int remove_mapping(struct address_space *mapping, struct page *page)
593 if (__remove_mapping(mapping, page)) {
595 * Unfreezing the refcount with 1 rather than 2 effectively
596 * drops the pagecache ref for us without requiring another
599 page_unfreeze_refs(page, 1);
606 * putback_lru_page - put previously isolated page onto appropriate LRU list
607 * @page: page to be put back to appropriate lru list
609 * Add previously isolated @page to appropriate LRU list.
610 * Page may still be unevictable for other reasons.
612 * lru_lock must not be held, interrupts must be enabled.
614 void putback_lru_page(struct page *page)
617 int active = !!TestClearPageActive(page);
618 int was_unevictable = PageUnevictable(page);
620 VM_BUG_ON(PageLRU(page));
623 ClearPageUnevictable(page);
625 if (page_evictable(page, NULL)) {
627 * For evictable pages, we can use the cache.
628 * In event of a race, worst case is we end up with an
629 * unevictable page on [in]active list.
630 * We know how to handle that.
632 lru = active + page_lru_base_type(page);
633 lru_cache_add_lru(page, lru);
636 * Put unevictable pages directly on zone's unevictable
639 lru = LRU_UNEVICTABLE;
640 add_page_to_unevictable_list(page);
642 * When racing with an mlock or AS_UNEVICTABLE clearing
643 * (page is unlocked) make sure that if the other thread
644 * does not observe our setting of PG_lru and fails
645 * isolation/check_move_unevictable_pages,
646 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
647 * the page back to the evictable list.
649 * The other side is TestClearPageMlocked() or shmem_lock().
655 * page's status can change while we move it among lru. If an evictable
656 * page is on unevictable list, it never be freed. To avoid that,
657 * check after we added it to the list, again.
659 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
660 if (!isolate_lru_page(page)) {
664 /* This means someone else dropped this page from LRU
665 * So, it will be freed or putback to LRU again. There is
666 * nothing to do here.
670 if (was_unevictable && lru != LRU_UNEVICTABLE)
671 count_vm_event(UNEVICTABLE_PGRESCUED);
672 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
673 count_vm_event(UNEVICTABLE_PGCULLED);
675 put_page(page); /* drop ref from isolate */
678 enum page_references {
680 PAGEREF_RECLAIM_CLEAN,
685 static enum page_references page_check_references(struct page *page,
686 struct mem_cgroup_zone *mz,
687 struct scan_control *sc)
689 int referenced_ptes, referenced_page;
690 unsigned long vm_flags;
692 referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
693 referenced_page = TestClearPageReferenced(page);
696 * Mlock lost the isolation race with us. Let try_to_unmap()
697 * move the page to the unevictable list.
699 if (vm_flags & VM_LOCKED)
700 return PAGEREF_RECLAIM;
702 if (referenced_ptes) {
704 return PAGEREF_ACTIVATE;
706 * All mapped pages start out with page table
707 * references from the instantiating fault, so we need
708 * to look twice if a mapped file page is used more
711 * Mark it and spare it for another trip around the
712 * inactive list. Another page table reference will
713 * lead to its activation.
715 * Note: the mark is set for activated pages as well
716 * so that recently deactivated but used pages are
719 SetPageReferenced(page);
721 if (referenced_page || referenced_ptes > 1)
722 return PAGEREF_ACTIVATE;
725 * Activate file-backed executable pages after first usage.
727 if (vm_flags & VM_EXEC)
728 return PAGEREF_ACTIVATE;
733 /* Reclaim if clean, defer dirty pages to writeback */
734 if (referenced_page && !PageSwapBacked(page))
735 return PAGEREF_RECLAIM_CLEAN;
737 return PAGEREF_RECLAIM;
741 * shrink_page_list() returns the number of reclaimed pages
743 static unsigned long shrink_page_list(struct list_head *page_list,
744 struct mem_cgroup_zone *mz,
745 struct scan_control *sc,
747 unsigned long *ret_nr_dirty,
748 unsigned long *ret_nr_writeback)
750 LIST_HEAD(ret_pages);
751 LIST_HEAD(free_pages);
753 unsigned long nr_dirty = 0;
754 unsigned long nr_congested = 0;
755 unsigned long nr_reclaimed = 0;
756 unsigned long nr_writeback = 0;
760 while (!list_empty(page_list)) {
761 enum page_references references;
762 struct address_space *mapping;
768 page = lru_to_page(page_list);
769 list_del(&page->lru);
771 if (!trylock_page(page))
774 VM_BUG_ON(PageActive(page));
775 VM_BUG_ON(page_zone(page) != mz->zone);
779 if (unlikely(!page_evictable(page, NULL)))
782 if (!sc->may_unmap && page_mapped(page))
785 /* Double the slab pressure for mapped and swapcache pages */
786 if (page_mapped(page) || PageSwapCache(page))
789 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
790 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
792 if (PageWriteback(page)) {
795 * Synchronous reclaim cannot queue pages for
796 * writeback due to the possibility of stack overflow
797 * but if it encounters a page under writeback, wait
798 * for the IO to complete.
800 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
802 wait_on_page_writeback(page);
805 goto keep_reclaim_mode;
809 references = page_check_references(page, mz, sc);
810 switch (references) {
811 case PAGEREF_ACTIVATE:
812 goto activate_locked;
815 case PAGEREF_RECLAIM:
816 case PAGEREF_RECLAIM_CLEAN:
817 ; /* try to reclaim the page below */
821 * Anonymous process memory has backing store?
822 * Try to allocate it some swap space here.
824 if (PageAnon(page) && !PageSwapCache(page)) {
825 if (!(sc->gfp_mask & __GFP_IO))
827 if (!add_to_swap(page))
828 goto activate_locked;
832 mapping = page_mapping(page);
835 * The page is mapped into the page tables of one or more
836 * processes. Try to unmap it here.
838 if (page_mapped(page) && mapping) {
839 switch (try_to_unmap(page, TTU_UNMAP)) {
841 goto activate_locked;
847 ; /* try to free the page below */
851 if (PageDirty(page)) {
855 * Only kswapd can writeback filesystem pages to
856 * avoid risk of stack overflow but do not writeback
857 * unless under significant pressure.
859 if (page_is_file_cache(page) &&
860 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
862 * Immediately reclaim when written back.
863 * Similar in principal to deactivate_page()
864 * except we already have the page isolated
865 * and know it's dirty
867 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
868 SetPageReclaim(page);
873 if (references == PAGEREF_RECLAIM_CLEAN)
877 if (!sc->may_writepage)
880 /* Page is dirty, try to write it out here */
881 switch (pageout(page, mapping, sc)) {
886 goto activate_locked;
888 if (PageWriteback(page))
889 goto keep_reclaim_mode;
894 * A synchronous write - probably a ramdisk. Go
895 * ahead and try to reclaim the page.
897 if (!trylock_page(page))
899 if (PageDirty(page) || PageWriteback(page))
901 mapping = page_mapping(page);
903 ; /* try to free the page below */
908 * If the page has buffers, try to free the buffer mappings
909 * associated with this page. If we succeed we try to free
912 * We do this even if the page is PageDirty().
913 * try_to_release_page() does not perform I/O, but it is
914 * possible for a page to have PageDirty set, but it is actually
915 * clean (all its buffers are clean). This happens if the
916 * buffers were written out directly, with submit_bh(). ext3
917 * will do this, as well as the blockdev mapping.
918 * try_to_release_page() will discover that cleanness and will
919 * drop the buffers and mark the page clean - it can be freed.
921 * Rarely, pages can have buffers and no ->mapping. These are
922 * the pages which were not successfully invalidated in
923 * truncate_complete_page(). We try to drop those buffers here
924 * and if that worked, and the page is no longer mapped into
925 * process address space (page_count == 1) it can be freed.
926 * Otherwise, leave the page on the LRU so it is swappable.
928 if (page_has_private(page)) {
929 if (!try_to_release_page(page, sc->gfp_mask))
930 goto activate_locked;
931 if (!mapping && page_count(page) == 1) {
933 if (put_page_testzero(page))
937 * rare race with speculative reference.
938 * the speculative reference will free
939 * this page shortly, so we may
940 * increment nr_reclaimed here (and
941 * leave it off the LRU).
949 if (!mapping || !__remove_mapping(mapping, page))
953 * At this point, we have no other references and there is
954 * no way to pick any more up (removed from LRU, removed
955 * from pagecache). Can use non-atomic bitops now (and
956 * we obviously don't have to worry about waking up a process
957 * waiting on the page lock, because there are no references.
959 __clear_page_locked(page);
964 * Is there need to periodically free_page_list? It would
965 * appear not as the counts should be low
967 list_add(&page->lru, &free_pages);
971 if (PageSwapCache(page))
972 try_to_free_swap(page);
974 putback_lru_page(page);
975 reset_reclaim_mode(sc);
979 /* Not a candidate for swapping, so reclaim swap space. */
980 if (PageSwapCache(page) && vm_swap_full())
981 try_to_free_swap(page);
982 VM_BUG_ON(PageActive(page));
988 reset_reclaim_mode(sc);
990 list_add(&page->lru, &ret_pages);
991 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
995 * Tag a zone as congested if all the dirty pages encountered were
996 * backed by a congested BDI. In this case, reclaimers should just
997 * back off and wait for congestion to clear because further reclaim
998 * will encounter the same problem
1000 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1001 zone_set_flag(mz->zone, ZONE_CONGESTED);
1003 free_hot_cold_page_list(&free_pages, 1);
1005 list_splice(&ret_pages, page_list);
1006 count_vm_events(PGACTIVATE, pgactivate);
1007 *ret_nr_dirty += nr_dirty;
1008 *ret_nr_writeback += nr_writeback;
1009 return nr_reclaimed;
1013 * Attempt to remove the specified page from its LRU. Only take this page
1014 * if it is of the appropriate PageActive status. Pages which are being
1015 * freed elsewhere are also ignored.
1017 * page: page to consider
1018 * mode: one of the LRU isolation modes defined above
1020 * returns 0 on success, -ve errno on failure.
1022 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1027 /* Only take pages on the LRU. */
1031 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1032 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1035 * When checking the active state, we need to be sure we are
1036 * dealing with comparible boolean values. Take the logical not
1039 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1042 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1045 /* Do not give back unevictable pages for compaction */
1046 if (PageUnevictable(page))
1052 * To minimise LRU disruption, the caller can indicate that it only
1053 * wants to isolate pages it will be able to operate on without
1054 * blocking - clean pages for the most part.
1056 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1057 * is used by reclaim when it is cannot write to backing storage
1059 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1060 * that it is possible to migrate without blocking
1062 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1063 /* All the caller can do on PageWriteback is block */
1064 if (PageWriteback(page))
1067 if (PageDirty(page)) {
1068 struct address_space *mapping;
1070 /* ISOLATE_CLEAN means only clean pages */
1071 if (mode & ISOLATE_CLEAN)
1075 * Only pages without mappings or that have a
1076 * ->migratepage callback are possible to migrate
1079 mapping = page_mapping(page);
1080 if (mapping && !mapping->a_ops->migratepage)
1085 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1088 if (likely(get_page_unless_zero(page))) {
1090 * Be careful not to clear PageLRU until after we're
1091 * sure the page is not being freed elsewhere -- the
1092 * page release code relies on it.
1102 * zone->lru_lock is heavily contended. Some of the functions that
1103 * shrink the lists perform better by taking out a batch of pages
1104 * and working on them outside the LRU lock.
1106 * For pagecache intensive workloads, this function is the hottest
1107 * spot in the kernel (apart from copy_*_user functions).
1109 * Appropriate locks must be held before calling this function.
1111 * @nr_to_scan: The number of pages to look through on the list.
1112 * @mz: The mem_cgroup_zone to pull pages from.
1113 * @dst: The temp list to put pages on to.
1114 * @nr_scanned: The number of pages that were scanned.
1115 * @sc: The scan_control struct for this reclaim session
1116 * @mode: One of the LRU isolation modes
1117 * @active: True [1] if isolating active pages
1118 * @file: True [1] if isolating file [!anon] pages
1120 * returns how many pages were moved onto *@dst.
1122 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1123 struct mem_cgroup_zone *mz, struct list_head *dst,
1124 unsigned long *nr_scanned, struct scan_control *sc,
1125 isolate_mode_t mode, int active, int file)
1127 struct lruvec *lruvec;
1128 struct list_head *src;
1129 unsigned long nr_taken = 0;
1133 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1138 src = &lruvec->lists[lru];
1140 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1143 page = lru_to_page(src);
1144 prefetchw_prev_lru_page(page, src, flags);
1146 VM_BUG_ON(!PageLRU(page));
1148 switch (__isolate_lru_page(page, mode, file)) {
1150 mem_cgroup_lru_del(page);
1151 list_move(&page->lru, dst);
1152 nr_taken += hpage_nr_pages(page);
1156 /* else it is being freed elsewhere */
1157 list_move(&page->lru, src);
1167 trace_mm_vmscan_lru_isolate(sc->order,
1175 * isolate_lru_page - tries to isolate a page from its LRU list
1176 * @page: page to isolate from its LRU list
1178 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1179 * vmstat statistic corresponding to whatever LRU list the page was on.
1181 * Returns 0 if the page was removed from an LRU list.
1182 * Returns -EBUSY if the page was not on an LRU list.
1184 * The returned page will have PageLRU() cleared. If it was found on
1185 * the active list, it will have PageActive set. If it was found on
1186 * the unevictable list, it will have the PageUnevictable bit set. That flag
1187 * may need to be cleared by the caller before letting the page go.
1189 * The vmstat statistic corresponding to the list on which the page was
1190 * found will be decremented.
1193 * (1) Must be called with an elevated refcount on the page. This is a
1194 * fundamentnal difference from isolate_lru_pages (which is called
1195 * without a stable reference).
1196 * (2) the lru_lock must not be held.
1197 * (3) interrupts must be enabled.
1199 int isolate_lru_page(struct page *page)
1203 VM_BUG_ON(!page_count(page));
1205 if (PageLRU(page)) {
1206 struct zone *zone = page_zone(page);
1208 spin_lock_irq(&zone->lru_lock);
1209 if (PageLRU(page)) {
1210 int lru = page_lru(page);
1215 del_page_from_lru_list(zone, page, lru);
1217 spin_unlock_irq(&zone->lru_lock);
1223 * Are there way too many processes in the direct reclaim path already?
1225 static int too_many_isolated(struct zone *zone, int file,
1226 struct scan_control *sc)
1228 unsigned long inactive, isolated;
1230 if (current_is_kswapd())
1233 if (!global_reclaim(sc))
1237 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1238 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1240 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1241 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1244 return isolated > inactive;
1247 static noinline_for_stack void
1248 putback_inactive_pages(struct mem_cgroup_zone *mz,
1249 struct list_head *page_list)
1251 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1252 struct zone *zone = mz->zone;
1253 LIST_HEAD(pages_to_free);
1256 * Put back any unfreeable pages.
1258 while (!list_empty(page_list)) {
1259 struct page *page = lru_to_page(page_list);
1262 VM_BUG_ON(PageLRU(page));
1263 list_del(&page->lru);
1264 if (unlikely(!page_evictable(page, NULL))) {
1265 spin_unlock_irq(&zone->lru_lock);
1266 putback_lru_page(page);
1267 spin_lock_irq(&zone->lru_lock);
1271 lru = page_lru(page);
1272 add_page_to_lru_list(zone, page, lru);
1273 if (is_active_lru(lru)) {
1274 int file = is_file_lru(lru);
1275 int numpages = hpage_nr_pages(page);
1276 reclaim_stat->recent_rotated[file] += numpages;
1278 if (put_page_testzero(page)) {
1279 __ClearPageLRU(page);
1280 __ClearPageActive(page);
1281 del_page_from_lru_list(zone, page, lru);
1283 if (unlikely(PageCompound(page))) {
1284 spin_unlock_irq(&zone->lru_lock);
1285 (*get_compound_page_dtor(page))(page);
1286 spin_lock_irq(&zone->lru_lock);
1288 list_add(&page->lru, &pages_to_free);
1293 * To save our caller's stack, now use input list for pages to free.
1295 list_splice(&pages_to_free, page_list);
1298 static noinline_for_stack void
1299 update_isolated_counts(struct mem_cgroup_zone *mz,
1300 struct list_head *page_list,
1301 unsigned long *nr_anon,
1302 unsigned long *nr_file)
1304 struct zone *zone = mz->zone;
1305 unsigned int count[NR_LRU_LISTS] = { 0, };
1306 unsigned long nr_active = 0;
1311 * Count pages and clear active flags
1313 list_for_each_entry(page, page_list, lru) {
1314 int numpages = hpage_nr_pages(page);
1315 lru = page_lru_base_type(page);
1316 if (PageActive(page)) {
1318 ClearPageActive(page);
1319 nr_active += numpages;
1321 count[lru] += numpages;
1325 __count_vm_events(PGDEACTIVATE, nr_active);
1327 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1328 -count[LRU_ACTIVE_FILE]);
1329 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1330 -count[LRU_INACTIVE_FILE]);
1331 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1332 -count[LRU_ACTIVE_ANON]);
1333 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1334 -count[LRU_INACTIVE_ANON]);
1336 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1337 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1339 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1340 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1345 * Returns true if a direct reclaim should wait on pages under writeback.
1347 * If we are direct reclaiming for contiguous pages and we do not reclaim
1348 * everything in the list, try again and wait for writeback IO to complete.
1349 * This will stall high-order allocations noticeably. Only do that when really
1350 * need to free the pages under high memory pressure.
1352 static inline bool should_reclaim_stall(unsigned long nr_taken,
1353 unsigned long nr_freed,
1355 struct scan_control *sc)
1359 /* kswapd should not stall on sync IO */
1360 if (current_is_kswapd())
1363 /* Only stall for memory compaction */
1364 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1367 /* If we have reclaimed everything on the isolated list, no stall */
1368 if (nr_freed == nr_taken)
1372 * For high-order allocations, there are two stall thresholds.
1373 * High-cost allocations stall immediately where as lower
1374 * order allocations such as stacks require the scanning
1375 * priority to be much higher before stalling.
1377 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1378 stall_priority = DEF_PRIORITY;
1380 stall_priority = DEF_PRIORITY / 3;
1382 return priority <= stall_priority;
1386 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1387 * of reclaimed pages
1389 static noinline_for_stack unsigned long
1390 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1391 struct scan_control *sc, int priority, int file)
1393 LIST_HEAD(page_list);
1394 unsigned long nr_scanned;
1395 unsigned long nr_reclaimed = 0;
1396 unsigned long nr_taken;
1397 unsigned long nr_anon;
1398 unsigned long nr_file;
1399 unsigned long nr_dirty = 0;
1400 unsigned long nr_writeback = 0;
1401 isolate_mode_t isolate_mode = ISOLATE_INACTIVE;
1402 struct zone *zone = mz->zone;
1403 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1405 while (unlikely(too_many_isolated(zone, file, sc))) {
1406 congestion_wait(BLK_RW_ASYNC, HZ/10);
1408 /* We are about to die and free our memory. Return now. */
1409 if (fatal_signal_pending(current))
1410 return SWAP_CLUSTER_MAX;
1413 set_reclaim_mode(priority, sc, false);
1418 isolate_mode |= ISOLATE_UNMAPPED;
1419 if (!sc->may_writepage)
1420 isolate_mode |= ISOLATE_CLEAN;
1422 spin_lock_irq(&zone->lru_lock);
1424 nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list, &nr_scanned,
1425 sc, isolate_mode, 0, file);
1426 if (global_reclaim(sc)) {
1427 zone->pages_scanned += nr_scanned;
1428 if (current_is_kswapd())
1429 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1432 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1435 spin_unlock_irq(&zone->lru_lock);
1440 update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1442 nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1443 &nr_dirty, &nr_writeback);
1445 /* Check if we should syncronously wait for writeback */
1446 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1447 set_reclaim_mode(priority, sc, true);
1448 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1449 priority, &nr_dirty, &nr_writeback);
1452 spin_lock_irq(&zone->lru_lock);
1454 reclaim_stat->recent_scanned[0] += nr_anon;
1455 reclaim_stat->recent_scanned[1] += nr_file;
1457 if (global_reclaim(sc)) {
1458 if (current_is_kswapd())
1459 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1462 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1466 putback_inactive_pages(mz, &page_list);
1468 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1469 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1471 spin_unlock_irq(&zone->lru_lock);
1473 free_hot_cold_page_list(&page_list, 1);
1476 * If reclaim is isolating dirty pages under writeback, it implies
1477 * that the long-lived page allocation rate is exceeding the page
1478 * laundering rate. Either the global limits are not being effective
1479 * at throttling processes due to the page distribution throughout
1480 * zones or there is heavy usage of a slow backing device. The
1481 * only option is to throttle from reclaim context which is not ideal
1482 * as there is no guarantee the dirtying process is throttled in the
1483 * same way balance_dirty_pages() manages.
1485 * This scales the number of dirty pages that must be under writeback
1486 * before throttling depending on priority. It is a simple backoff
1487 * function that has the most effect in the range DEF_PRIORITY to
1488 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1489 * in trouble and reclaim is considered to be in trouble.
1491 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1492 * DEF_PRIORITY-1 50% must be PageWriteback
1493 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1495 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1496 * isolated page is PageWriteback
1498 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1499 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1501 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1503 nr_scanned, nr_reclaimed,
1505 trace_shrink_flags(file, sc->reclaim_mode));
1506 return nr_reclaimed;
1510 * This moves pages from the active list to the inactive list.
1512 * We move them the other way if the page is referenced by one or more
1513 * processes, from rmap.
1515 * If the pages are mostly unmapped, the processing is fast and it is
1516 * appropriate to hold zone->lru_lock across the whole operation. But if
1517 * the pages are mapped, the processing is slow (page_referenced()) so we
1518 * should drop zone->lru_lock around each page. It's impossible to balance
1519 * this, so instead we remove the pages from the LRU while processing them.
1520 * It is safe to rely on PG_active against the non-LRU pages in here because
1521 * nobody will play with that bit on a non-LRU page.
1523 * The downside is that we have to touch page->_count against each page.
1524 * But we had to alter page->flags anyway.
1527 static void move_active_pages_to_lru(struct zone *zone,
1528 struct list_head *list,
1529 struct list_head *pages_to_free,
1532 unsigned long pgmoved = 0;
1535 while (!list_empty(list)) {
1536 struct lruvec *lruvec;
1538 page = lru_to_page(list);
1540 VM_BUG_ON(PageLRU(page));
1543 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1544 list_move(&page->lru, &lruvec->lists[lru]);
1545 pgmoved += hpage_nr_pages(page);
1547 if (put_page_testzero(page)) {
1548 __ClearPageLRU(page);
1549 __ClearPageActive(page);
1550 del_page_from_lru_list(zone, page, lru);
1552 if (unlikely(PageCompound(page))) {
1553 spin_unlock_irq(&zone->lru_lock);
1554 (*get_compound_page_dtor(page))(page);
1555 spin_lock_irq(&zone->lru_lock);
1557 list_add(&page->lru, pages_to_free);
1560 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1561 if (!is_active_lru(lru))
1562 __count_vm_events(PGDEACTIVATE, pgmoved);
1565 static void shrink_active_list(unsigned long nr_to_scan,
1566 struct mem_cgroup_zone *mz,
1567 struct scan_control *sc,
1568 int priority, int file)
1570 unsigned long nr_taken;
1571 unsigned long nr_scanned;
1572 unsigned long vm_flags;
1573 LIST_HEAD(l_hold); /* The pages which were snipped off */
1574 LIST_HEAD(l_active);
1575 LIST_HEAD(l_inactive);
1577 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1578 unsigned long nr_rotated = 0;
1579 isolate_mode_t isolate_mode = ISOLATE_ACTIVE;
1580 struct zone *zone = mz->zone;
1584 reset_reclaim_mode(sc);
1587 isolate_mode |= ISOLATE_UNMAPPED;
1588 if (!sc->may_writepage)
1589 isolate_mode |= ISOLATE_CLEAN;
1591 spin_lock_irq(&zone->lru_lock);
1593 nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold, &nr_scanned, sc,
1594 isolate_mode, 1, file);
1595 if (global_reclaim(sc))
1596 zone->pages_scanned += nr_scanned;
1598 reclaim_stat->recent_scanned[file] += nr_taken;
1600 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1602 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1604 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1605 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1606 spin_unlock_irq(&zone->lru_lock);
1608 while (!list_empty(&l_hold)) {
1610 page = lru_to_page(&l_hold);
1611 list_del(&page->lru);
1613 if (unlikely(!page_evictable(page, NULL))) {
1614 putback_lru_page(page);
1618 if (unlikely(buffer_heads_over_limit)) {
1619 if (page_has_private(page) && trylock_page(page)) {
1620 if (page_has_private(page))
1621 try_to_release_page(page, 0);
1626 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1627 nr_rotated += hpage_nr_pages(page);
1629 * Identify referenced, file-backed active pages and
1630 * give them one more trip around the active list. So
1631 * that executable code get better chances to stay in
1632 * memory under moderate memory pressure. Anon pages
1633 * are not likely to be evicted by use-once streaming
1634 * IO, plus JVM can create lots of anon VM_EXEC pages,
1635 * so we ignore them here.
1637 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1638 list_add(&page->lru, &l_active);
1643 ClearPageActive(page); /* we are de-activating */
1644 list_add(&page->lru, &l_inactive);
1648 * Move pages back to the lru list.
1650 spin_lock_irq(&zone->lru_lock);
1652 * Count referenced pages from currently used mappings as rotated,
1653 * even though only some of them are actually re-activated. This
1654 * helps balance scan pressure between file and anonymous pages in
1657 reclaim_stat->recent_rotated[file] += nr_rotated;
1659 move_active_pages_to_lru(zone, &l_active, &l_hold,
1660 LRU_ACTIVE + file * LRU_FILE);
1661 move_active_pages_to_lru(zone, &l_inactive, &l_hold,
1662 LRU_BASE + file * LRU_FILE);
1663 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1664 spin_unlock_irq(&zone->lru_lock);
1666 free_hot_cold_page_list(&l_hold, 1);
1670 static int inactive_anon_is_low_global(struct zone *zone)
1672 unsigned long active, inactive;
1674 active = zone_page_state(zone, NR_ACTIVE_ANON);
1675 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1677 if (inactive * zone->inactive_ratio < active)
1684 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1685 * @zone: zone to check
1686 * @sc: scan control of this context
1688 * Returns true if the zone does not have enough inactive anon pages,
1689 * meaning some active anon pages need to be deactivated.
1691 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1694 * If we don't have swap space, anonymous page deactivation
1697 if (!total_swap_pages)
1700 if (!scanning_global_lru(mz))
1701 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1704 return inactive_anon_is_low_global(mz->zone);
1707 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1713 static int inactive_file_is_low_global(struct zone *zone)
1715 unsigned long active, inactive;
1717 active = zone_page_state(zone, NR_ACTIVE_FILE);
1718 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1720 return (active > inactive);
1724 * inactive_file_is_low - check if file pages need to be deactivated
1725 * @mz: memory cgroup and zone to check
1727 * When the system is doing streaming IO, memory pressure here
1728 * ensures that active file pages get deactivated, until more
1729 * than half of the file pages are on the inactive list.
1731 * Once we get to that situation, protect the system's working
1732 * set from being evicted by disabling active file page aging.
1734 * This uses a different ratio than the anonymous pages, because
1735 * the page cache uses a use-once replacement algorithm.
1737 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1739 if (!scanning_global_lru(mz))
1740 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1743 return inactive_file_is_low_global(mz->zone);
1746 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1749 return inactive_file_is_low(mz);
1751 return inactive_anon_is_low(mz);
1754 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1755 struct mem_cgroup_zone *mz,
1756 struct scan_control *sc, int priority)
1758 int file = is_file_lru(lru);
1760 if (is_active_lru(lru)) {
1761 if (inactive_list_is_low(mz, file))
1762 shrink_active_list(nr_to_scan, mz, sc, priority, file);
1766 return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1769 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1770 struct scan_control *sc)
1772 if (global_reclaim(sc))
1773 return vm_swappiness;
1774 return mem_cgroup_swappiness(mz->mem_cgroup);
1778 * Determine how aggressively the anon and file LRU lists should be
1779 * scanned. The relative value of each set of LRU lists is determined
1780 * by looking at the fraction of the pages scanned we did rotate back
1781 * onto the active list instead of evict.
1783 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1785 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1786 unsigned long *nr, int priority)
1788 unsigned long anon, file, free;
1789 unsigned long anon_prio, file_prio;
1790 unsigned long ap, fp;
1791 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1792 u64 fraction[2], denominator;
1795 bool force_scan = false;
1798 * If the zone or memcg is small, nr[l] can be 0. This
1799 * results in no scanning on this priority and a potential
1800 * priority drop. Global direct reclaim can go to the next
1801 * zone and tends to have no problems. Global kswapd is for
1802 * zone balancing and it needs to scan a minimum amount. When
1803 * reclaiming for a memcg, a priority drop can cause high
1804 * latencies, so it's better to scan a minimum amount there as
1807 if (current_is_kswapd() && mz->zone->all_unreclaimable)
1809 if (!global_reclaim(sc))
1812 /* If we have no swap space, do not bother scanning anon pages. */
1813 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1821 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1822 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1823 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1824 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1826 if (global_reclaim(sc)) {
1827 free = zone_page_state(mz->zone, NR_FREE_PAGES);
1828 /* If we have very few page cache pages,
1829 force-scan anon pages. */
1830 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1839 * With swappiness at 100, anonymous and file have the same priority.
1840 * This scanning priority is essentially the inverse of IO cost.
1842 anon_prio = vmscan_swappiness(mz, sc);
1843 file_prio = 200 - vmscan_swappiness(mz, sc);
1846 * OK, so we have swap space and a fair amount of page cache
1847 * pages. We use the recently rotated / recently scanned
1848 * ratios to determine how valuable each cache is.
1850 * Because workloads change over time (and to avoid overflow)
1851 * we keep these statistics as a floating average, which ends
1852 * up weighing recent references more than old ones.
1854 * anon in [0], file in [1]
1856 spin_lock_irq(&mz->zone->lru_lock);
1857 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1858 reclaim_stat->recent_scanned[0] /= 2;
1859 reclaim_stat->recent_rotated[0] /= 2;
1862 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1863 reclaim_stat->recent_scanned[1] /= 2;
1864 reclaim_stat->recent_rotated[1] /= 2;
1868 * The amount of pressure on anon vs file pages is inversely
1869 * proportional to the fraction of recently scanned pages on
1870 * each list that were recently referenced and in active use.
1872 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1873 ap /= reclaim_stat->recent_rotated[0] + 1;
1875 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1876 fp /= reclaim_stat->recent_rotated[1] + 1;
1877 spin_unlock_irq(&mz->zone->lru_lock);
1881 denominator = ap + fp + 1;
1883 for_each_evictable_lru(lru) {
1884 int file = is_file_lru(lru);
1887 scan = zone_nr_lru_pages(mz, lru);
1888 if (priority || noswap) {
1890 if (!scan && force_scan)
1891 scan = SWAP_CLUSTER_MAX;
1892 scan = div64_u64(scan * fraction[file], denominator);
1899 * Reclaim/compaction depends on a number of pages being freed. To avoid
1900 * disruption to the system, a small number of order-0 pages continue to be
1901 * rotated and reclaimed in the normal fashion. However, by the time we get
1902 * back to the allocator and call try_to_compact_zone(), we ensure that
1903 * there are enough free pages for it to be likely successful
1905 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
1906 unsigned long nr_reclaimed,
1907 unsigned long nr_scanned,
1908 struct scan_control *sc)
1910 unsigned long pages_for_compaction;
1911 unsigned long inactive_lru_pages;
1913 /* If not in reclaim/compaction mode, stop */
1914 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1917 /* Consider stopping depending on scan and reclaim activity */
1918 if (sc->gfp_mask & __GFP_REPEAT) {
1920 * For __GFP_REPEAT allocations, stop reclaiming if the
1921 * full LRU list has been scanned and we are still failing
1922 * to reclaim pages. This full LRU scan is potentially
1923 * expensive but a __GFP_REPEAT caller really wants to succeed
1925 if (!nr_reclaimed && !nr_scanned)
1929 * For non-__GFP_REPEAT allocations which can presumably
1930 * fail without consequence, stop if we failed to reclaim
1931 * any pages from the last SWAP_CLUSTER_MAX number of
1932 * pages that were scanned. This will return to the
1933 * caller faster at the risk reclaim/compaction and
1934 * the resulting allocation attempt fails
1941 * If we have not reclaimed enough pages for compaction and the
1942 * inactive lists are large enough, continue reclaiming
1944 pages_for_compaction = (2UL << sc->order);
1945 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1946 if (nr_swap_pages > 0)
1947 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1948 if (sc->nr_reclaimed < pages_for_compaction &&
1949 inactive_lru_pages > pages_for_compaction)
1952 /* If compaction would go ahead or the allocation would succeed, stop */
1953 switch (compaction_suitable(mz->zone, sc->order)) {
1954 case COMPACT_PARTIAL:
1955 case COMPACT_CONTINUE:
1963 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1965 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
1966 struct scan_control *sc)
1968 unsigned long nr[NR_LRU_LISTS];
1969 unsigned long nr_to_scan;
1971 unsigned long nr_reclaimed, nr_scanned;
1972 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1973 struct blk_plug plug;
1977 nr_scanned = sc->nr_scanned;
1978 get_scan_count(mz, sc, nr, priority);
1980 blk_start_plug(&plug);
1981 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1982 nr[LRU_INACTIVE_FILE]) {
1983 for_each_evictable_lru(lru) {
1985 nr_to_scan = min_t(unsigned long,
1986 nr[lru], SWAP_CLUSTER_MAX);
1987 nr[lru] -= nr_to_scan;
1989 nr_reclaimed += shrink_list(lru, nr_to_scan,
1994 * On large memory systems, scan >> priority can become
1995 * really large. This is fine for the starting priority;
1996 * we want to put equal scanning pressure on each zone.
1997 * However, if the VM has a harder time of freeing pages,
1998 * with multiple processes reclaiming pages, the total
1999 * freeing target can get unreasonably large.
2001 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2004 blk_finish_plug(&plug);
2005 sc->nr_reclaimed += nr_reclaimed;
2008 * Even if we did not try to evict anon pages at all, we want to
2009 * rebalance the anon lru active/inactive ratio.
2011 if (inactive_anon_is_low(mz))
2012 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
2014 /* reclaim/compaction might need reclaim to continue */
2015 if (should_continue_reclaim(mz, nr_reclaimed,
2016 sc->nr_scanned - nr_scanned, sc))
2019 throttle_vm_writeout(sc->gfp_mask);
2022 static void shrink_zone(int priority, struct zone *zone,
2023 struct scan_control *sc)
2025 struct mem_cgroup *root = sc->target_mem_cgroup;
2026 struct mem_cgroup_reclaim_cookie reclaim = {
2028 .priority = priority,
2030 struct mem_cgroup *memcg;
2032 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2034 struct mem_cgroup_zone mz = {
2035 .mem_cgroup = memcg,
2039 shrink_mem_cgroup_zone(priority, &mz, sc);
2041 * Limit reclaim has historically picked one memcg and
2042 * scanned it with decreasing priority levels until
2043 * nr_to_reclaim had been reclaimed. This priority
2044 * cycle is thus over after a single memcg.
2046 * Direct reclaim and kswapd, on the other hand, have
2047 * to scan all memory cgroups to fulfill the overall
2048 * scan target for the zone.
2050 if (!global_reclaim(sc)) {
2051 mem_cgroup_iter_break(root, memcg);
2054 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2058 /* Returns true if compaction should go ahead for a high-order request */
2059 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2061 unsigned long balance_gap, watermark;
2064 /* Do not consider compaction for orders reclaim is meant to satisfy */
2065 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2069 * Compaction takes time to run and there are potentially other
2070 * callers using the pages just freed. Continue reclaiming until
2071 * there is a buffer of free pages available to give compaction
2072 * a reasonable chance of completing and allocating the page
2074 balance_gap = min(low_wmark_pages(zone),
2075 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2076 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2077 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2078 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2081 * If compaction is deferred, reclaim up to a point where
2082 * compaction will have a chance of success when re-enabled
2084 if (compaction_deferred(zone, sc->order))
2085 return watermark_ok;
2087 /* If compaction is not ready to start, keep reclaiming */
2088 if (!compaction_suitable(zone, sc->order))
2091 return watermark_ok;
2095 * This is the direct reclaim path, for page-allocating processes. We only
2096 * try to reclaim pages from zones which will satisfy the caller's allocation
2099 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2101 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2103 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2104 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2105 * zone defense algorithm.
2107 * If a zone is deemed to be full of pinned pages then just give it a light
2108 * scan then give up on it.
2110 * This function returns true if a zone is being reclaimed for a costly
2111 * high-order allocation and compaction is ready to begin. This indicates to
2112 * the caller that it should consider retrying the allocation instead of
2115 static bool shrink_zones(int priority, struct zonelist *zonelist,
2116 struct scan_control *sc)
2120 unsigned long nr_soft_reclaimed;
2121 unsigned long nr_soft_scanned;
2122 bool aborted_reclaim = false;
2125 * If the number of buffer_heads in the machine exceeds the maximum
2126 * allowed level, force direct reclaim to scan the highmem zone as
2127 * highmem pages could be pinning lowmem pages storing buffer_heads
2129 if (buffer_heads_over_limit)
2130 sc->gfp_mask |= __GFP_HIGHMEM;
2132 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2133 gfp_zone(sc->gfp_mask), sc->nodemask) {
2134 if (!populated_zone(zone))
2137 * Take care memory controller reclaiming has small influence
2140 if (global_reclaim(sc)) {
2141 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2143 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2144 continue; /* Let kswapd poll it */
2145 if (COMPACTION_BUILD) {
2147 * If we already have plenty of memory free for
2148 * compaction in this zone, don't free any more.
2149 * Even though compaction is invoked for any
2150 * non-zero order, only frequent costly order
2151 * reclamation is disruptive enough to become a
2152 * noticeable problem, like transparent huge
2155 if (compaction_ready(zone, sc)) {
2156 aborted_reclaim = true;
2161 * This steals pages from memory cgroups over softlimit
2162 * and returns the number of reclaimed pages and
2163 * scanned pages. This works for global memory pressure
2164 * and balancing, not for a memcg's limit.
2166 nr_soft_scanned = 0;
2167 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2168 sc->order, sc->gfp_mask,
2170 sc->nr_reclaimed += nr_soft_reclaimed;
2171 sc->nr_scanned += nr_soft_scanned;
2172 /* need some check for avoid more shrink_zone() */
2175 shrink_zone(priority, zone, sc);
2178 return aborted_reclaim;
2181 static bool zone_reclaimable(struct zone *zone)
2183 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2186 /* All zones in zonelist are unreclaimable? */
2187 static bool all_unreclaimable(struct zonelist *zonelist,
2188 struct scan_control *sc)
2193 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2194 gfp_zone(sc->gfp_mask), sc->nodemask) {
2195 if (!populated_zone(zone))
2197 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2199 if (!zone->all_unreclaimable)
2207 * This is the main entry point to direct page reclaim.
2209 * If a full scan of the inactive list fails to free enough memory then we
2210 * are "out of memory" and something needs to be killed.
2212 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2213 * high - the zone may be full of dirty or under-writeback pages, which this
2214 * caller can't do much about. We kick the writeback threads and take explicit
2215 * naps in the hope that some of these pages can be written. But if the
2216 * allocating task holds filesystem locks which prevent writeout this might not
2217 * work, and the allocation attempt will fail.
2219 * returns: 0, if no pages reclaimed
2220 * else, the number of pages reclaimed
2222 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2223 struct scan_control *sc,
2224 struct shrink_control *shrink)
2227 unsigned long total_scanned = 0;
2228 struct reclaim_state *reclaim_state = current->reclaim_state;
2231 unsigned long writeback_threshold;
2232 bool aborted_reclaim;
2234 delayacct_freepages_start();
2236 if (global_reclaim(sc))
2237 count_vm_event(ALLOCSTALL);
2239 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2241 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2244 * Don't shrink slabs when reclaiming memory from
2245 * over limit cgroups
2247 if (global_reclaim(sc)) {
2248 unsigned long lru_pages = 0;
2249 for_each_zone_zonelist(zone, z, zonelist,
2250 gfp_zone(sc->gfp_mask)) {
2251 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2254 lru_pages += zone_reclaimable_pages(zone);
2257 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2258 if (reclaim_state) {
2259 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2260 reclaim_state->reclaimed_slab = 0;
2263 total_scanned += sc->nr_scanned;
2264 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2268 * Try to write back as many pages as we just scanned. This
2269 * tends to cause slow streaming writers to write data to the
2270 * disk smoothly, at the dirtying rate, which is nice. But
2271 * that's undesirable in laptop mode, where we *want* lumpy
2272 * writeout. So in laptop mode, write out the whole world.
2274 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2275 if (total_scanned > writeback_threshold) {
2276 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2277 WB_REASON_TRY_TO_FREE_PAGES);
2278 sc->may_writepage = 1;
2281 /* Take a nap, wait for some writeback to complete */
2282 if (!sc->hibernation_mode && sc->nr_scanned &&
2283 priority < DEF_PRIORITY - 2) {
2284 struct zone *preferred_zone;
2286 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2287 &cpuset_current_mems_allowed,
2289 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2294 delayacct_freepages_end();
2296 if (sc->nr_reclaimed)
2297 return sc->nr_reclaimed;
2300 * As hibernation is going on, kswapd is freezed so that it can't mark
2301 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2304 if (oom_killer_disabled)
2307 /* Aborted reclaim to try compaction? don't OOM, then */
2308 if (aborted_reclaim)
2311 /* top priority shrink_zones still had more to do? don't OOM, then */
2312 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2318 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2319 gfp_t gfp_mask, nodemask_t *nodemask)
2321 unsigned long nr_reclaimed;
2322 struct scan_control sc = {
2323 .gfp_mask = gfp_mask,
2324 .may_writepage = !laptop_mode,
2325 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2329 .target_mem_cgroup = NULL,
2330 .nodemask = nodemask,
2332 struct shrink_control shrink = {
2333 .gfp_mask = sc.gfp_mask,
2336 trace_mm_vmscan_direct_reclaim_begin(order,
2340 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2342 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2344 return nr_reclaimed;
2347 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2349 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2350 gfp_t gfp_mask, bool noswap,
2352 unsigned long *nr_scanned)
2354 struct scan_control sc = {
2356 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2357 .may_writepage = !laptop_mode,
2359 .may_swap = !noswap,
2361 .target_mem_cgroup = memcg,
2363 struct mem_cgroup_zone mz = {
2364 .mem_cgroup = memcg,
2368 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2369 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2371 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2376 * NOTE: Although we can get the priority field, using it
2377 * here is not a good idea, since it limits the pages we can scan.
2378 * if we don't reclaim here, the shrink_zone from balance_pgdat
2379 * will pick up pages from other mem cgroup's as well. We hack
2380 * the priority and make it zero.
2382 shrink_mem_cgroup_zone(0, &mz, &sc);
2384 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2386 *nr_scanned = sc.nr_scanned;
2387 return sc.nr_reclaimed;
2390 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2394 struct zonelist *zonelist;
2395 unsigned long nr_reclaimed;
2397 struct scan_control sc = {
2398 .may_writepage = !laptop_mode,
2400 .may_swap = !noswap,
2401 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2403 .target_mem_cgroup = memcg,
2404 .nodemask = NULL, /* we don't care the placement */
2405 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2406 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2408 struct shrink_control shrink = {
2409 .gfp_mask = sc.gfp_mask,
2413 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2414 * take care of from where we get pages. So the node where we start the
2415 * scan does not need to be the current node.
2417 nid = mem_cgroup_select_victim_node(memcg);
2419 zonelist = NODE_DATA(nid)->node_zonelists;
2421 trace_mm_vmscan_memcg_reclaim_begin(0,
2425 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2427 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2429 return nr_reclaimed;
2433 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2436 struct mem_cgroup *memcg;
2438 if (!total_swap_pages)
2441 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2443 struct mem_cgroup_zone mz = {
2444 .mem_cgroup = memcg,
2448 if (inactive_anon_is_low(&mz))
2449 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2452 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2457 * pgdat_balanced is used when checking if a node is balanced for high-order
2458 * allocations. Only zones that meet watermarks and are in a zone allowed
2459 * by the callers classzone_idx are added to balanced_pages. The total of
2460 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2461 * for the node to be considered balanced. Forcing all zones to be balanced
2462 * for high orders can cause excessive reclaim when there are imbalanced zones.
2463 * The choice of 25% is due to
2464 * o a 16M DMA zone that is balanced will not balance a zone on any
2465 * reasonable sized machine
2466 * o On all other machines, the top zone must be at least a reasonable
2467 * percentage of the middle zones. For example, on 32-bit x86, highmem
2468 * would need to be at least 256M for it to be balance a whole node.
2469 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2470 * to balance a node on its own. These seemed like reasonable ratios.
2472 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2475 unsigned long present_pages = 0;
2478 for (i = 0; i <= classzone_idx; i++)
2479 present_pages += pgdat->node_zones[i].present_pages;
2481 /* A special case here: if zone has no page, we think it's balanced */
2482 return balanced_pages >= (present_pages >> 2);
2485 /* is kswapd sleeping prematurely? */
2486 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2490 unsigned long balanced = 0;
2491 bool all_zones_ok = true;
2493 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2497 /* Check the watermark levels */
2498 for (i = 0; i <= classzone_idx; i++) {
2499 struct zone *zone = pgdat->node_zones + i;
2501 if (!populated_zone(zone))
2505 * balance_pgdat() skips over all_unreclaimable after
2506 * DEF_PRIORITY. Effectively, it considers them balanced so
2507 * they must be considered balanced here as well if kswapd
2510 if (zone->all_unreclaimable) {
2511 balanced += zone->present_pages;
2515 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2517 all_zones_ok = false;
2519 balanced += zone->present_pages;
2523 * For high-order requests, the balanced zones must contain at least
2524 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2528 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2530 return !all_zones_ok;
2534 * For kswapd, balance_pgdat() will work across all this node's zones until
2535 * they are all at high_wmark_pages(zone).
2537 * Returns the final order kswapd was reclaiming at
2539 * There is special handling here for zones which are full of pinned pages.
2540 * This can happen if the pages are all mlocked, or if they are all used by
2541 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2542 * What we do is to detect the case where all pages in the zone have been
2543 * scanned twice and there has been zero successful reclaim. Mark the zone as
2544 * dead and from now on, only perform a short scan. Basically we're polling
2545 * the zone for when the problem goes away.
2547 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2548 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2549 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2550 * lower zones regardless of the number of free pages in the lower zones. This
2551 * interoperates with the page allocator fallback scheme to ensure that aging
2552 * of pages is balanced across the zones.
2554 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2558 unsigned long balanced;
2561 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2562 unsigned long total_scanned;
2563 struct reclaim_state *reclaim_state = current->reclaim_state;
2564 unsigned long nr_soft_reclaimed;
2565 unsigned long nr_soft_scanned;
2566 struct scan_control sc = {
2567 .gfp_mask = GFP_KERNEL,
2571 * kswapd doesn't want to be bailed out while reclaim. because
2572 * we want to put equal scanning pressure on each zone.
2574 .nr_to_reclaim = ULONG_MAX,
2576 .target_mem_cgroup = NULL,
2578 struct shrink_control shrink = {
2579 .gfp_mask = sc.gfp_mask,
2583 sc.nr_reclaimed = 0;
2584 sc.may_writepage = !laptop_mode;
2585 count_vm_event(PAGEOUTRUN);
2587 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2588 unsigned long lru_pages = 0;
2589 int has_under_min_watermark_zone = 0;
2595 * Scan in the highmem->dma direction for the highest
2596 * zone which needs scanning
2598 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2599 struct zone *zone = pgdat->node_zones + i;
2601 if (!populated_zone(zone))
2604 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2608 * Do some background aging of the anon list, to give
2609 * pages a chance to be referenced before reclaiming.
2611 age_active_anon(zone, &sc, priority);
2614 * If the number of buffer_heads in the machine
2615 * exceeds the maximum allowed level and this node
2616 * has a highmem zone, force kswapd to reclaim from
2617 * it to relieve lowmem pressure.
2619 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2624 if (!zone_watermark_ok_safe(zone, order,
2625 high_wmark_pages(zone), 0, 0)) {
2629 /* If balanced, clear the congested flag */
2630 zone_clear_flag(zone, ZONE_CONGESTED);
2636 for (i = 0; i <= end_zone; i++) {
2637 struct zone *zone = pgdat->node_zones + i;
2639 lru_pages += zone_reclaimable_pages(zone);
2643 * Now scan the zone in the dma->highmem direction, stopping
2644 * at the last zone which needs scanning.
2646 * We do this because the page allocator works in the opposite
2647 * direction. This prevents the page allocator from allocating
2648 * pages behind kswapd's direction of progress, which would
2649 * cause too much scanning of the lower zones.
2651 for (i = 0; i <= end_zone; i++) {
2652 struct zone *zone = pgdat->node_zones + i;
2653 int nr_slab, testorder;
2654 unsigned long balance_gap;
2656 if (!populated_zone(zone))
2659 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2664 nr_soft_scanned = 0;
2666 * Call soft limit reclaim before calling shrink_zone.
2668 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2671 sc.nr_reclaimed += nr_soft_reclaimed;
2672 total_scanned += nr_soft_scanned;
2675 * We put equal pressure on every zone, unless
2676 * one zone has way too many pages free
2677 * already. The "too many pages" is defined
2678 * as the high wmark plus a "gap" where the
2679 * gap is either the low watermark or 1%
2680 * of the zone, whichever is smaller.
2682 balance_gap = min(low_wmark_pages(zone),
2683 (zone->present_pages +
2684 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2685 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2687 * Kswapd reclaims only single pages with compaction
2688 * enabled. Trying too hard to reclaim until contiguous
2689 * free pages have become available can hurt performance
2690 * by evicting too much useful data from memory.
2691 * Do not reclaim more than needed for compaction.
2694 if (COMPACTION_BUILD && order &&
2695 compaction_suitable(zone, order) !=
2699 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2700 !zone_watermark_ok_safe(zone, testorder,
2701 high_wmark_pages(zone) + balance_gap,
2703 shrink_zone(priority, zone, &sc);
2705 reclaim_state->reclaimed_slab = 0;
2706 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2707 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2708 total_scanned += sc.nr_scanned;
2710 if (nr_slab == 0 && !zone_reclaimable(zone))
2711 zone->all_unreclaimable = 1;
2715 * If we've done a decent amount of scanning and
2716 * the reclaim ratio is low, start doing writepage
2717 * even in laptop mode
2719 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2720 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2721 sc.may_writepage = 1;
2723 if (zone->all_unreclaimable) {
2724 if (end_zone && end_zone == i)
2729 if (!zone_watermark_ok_safe(zone, testorder,
2730 high_wmark_pages(zone), end_zone, 0)) {
2733 * We are still under min water mark. This
2734 * means that we have a GFP_ATOMIC allocation
2735 * failure risk. Hurry up!
2737 if (!zone_watermark_ok_safe(zone, order,
2738 min_wmark_pages(zone), end_zone, 0))
2739 has_under_min_watermark_zone = 1;
2742 * If a zone reaches its high watermark,
2743 * consider it to be no longer congested. It's
2744 * possible there are dirty pages backed by
2745 * congested BDIs but as pressure is relieved,
2746 * spectulatively avoid congestion waits
2748 zone_clear_flag(zone, ZONE_CONGESTED);
2749 if (i <= *classzone_idx)
2750 balanced += zone->present_pages;
2754 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2755 break; /* kswapd: all done */
2757 * OK, kswapd is getting into trouble. Take a nap, then take
2758 * another pass across the zones.
2760 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2761 if (has_under_min_watermark_zone)
2762 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2764 congestion_wait(BLK_RW_ASYNC, HZ/10);
2768 * We do this so kswapd doesn't build up large priorities for
2769 * example when it is freeing in parallel with allocators. It
2770 * matches the direct reclaim path behaviour in terms of impact
2771 * on zone->*_priority.
2773 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2779 * order-0: All zones must meet high watermark for a balanced node
2780 * high-order: Balanced zones must make up at least 25% of the node
2781 * for the node to be balanced
2783 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2789 * Fragmentation may mean that the system cannot be
2790 * rebalanced for high-order allocations in all zones.
2791 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2792 * it means the zones have been fully scanned and are still
2793 * not balanced. For high-order allocations, there is
2794 * little point trying all over again as kswapd may
2797 * Instead, recheck all watermarks at order-0 as they
2798 * are the most important. If watermarks are ok, kswapd will go
2799 * back to sleep. High-order users can still perform direct
2800 * reclaim if they wish.
2802 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2803 order = sc.order = 0;
2809 * If kswapd was reclaiming at a higher order, it has the option of
2810 * sleeping without all zones being balanced. Before it does, it must
2811 * ensure that the watermarks for order-0 on *all* zones are met and
2812 * that the congestion flags are cleared. The congestion flag must
2813 * be cleared as kswapd is the only mechanism that clears the flag
2814 * and it is potentially going to sleep here.
2817 int zones_need_compaction = 1;
2819 for (i = 0; i <= end_zone; i++) {
2820 struct zone *zone = pgdat->node_zones + i;
2822 if (!populated_zone(zone))
2825 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2828 /* Would compaction fail due to lack of free memory? */
2829 if (COMPACTION_BUILD &&
2830 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2833 /* Confirm the zone is balanced for order-0 */
2834 if (!zone_watermark_ok(zone, 0,
2835 high_wmark_pages(zone), 0, 0)) {
2836 order = sc.order = 0;
2840 /* Check if the memory needs to be defragmented. */
2841 if (zone_watermark_ok(zone, order,
2842 low_wmark_pages(zone), *classzone_idx, 0))
2843 zones_need_compaction = 0;
2845 /* If balanced, clear the congested flag */
2846 zone_clear_flag(zone, ZONE_CONGESTED);
2849 if (zones_need_compaction)
2850 compact_pgdat(pgdat, order);
2854 * Return the order we were reclaiming at so sleeping_prematurely()
2855 * makes a decision on the order we were last reclaiming at. However,
2856 * if another caller entered the allocator slow path while kswapd
2857 * was awake, order will remain at the higher level
2859 *classzone_idx = end_zone;
2863 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2868 if (freezing(current) || kthread_should_stop())
2871 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2873 /* Try to sleep for a short interval */
2874 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2875 remaining = schedule_timeout(HZ/10);
2876 finish_wait(&pgdat->kswapd_wait, &wait);
2877 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2881 * After a short sleep, check if it was a premature sleep. If not, then
2882 * go fully to sleep until explicitly woken up.
2884 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2885 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2888 * vmstat counters are not perfectly accurate and the estimated
2889 * value for counters such as NR_FREE_PAGES can deviate from the
2890 * true value by nr_online_cpus * threshold. To avoid the zone
2891 * watermarks being breached while under pressure, we reduce the
2892 * per-cpu vmstat threshold while kswapd is awake and restore
2893 * them before going back to sleep.
2895 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2897 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2900 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2902 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2904 finish_wait(&pgdat->kswapd_wait, &wait);
2908 * The background pageout daemon, started as a kernel thread
2909 * from the init process.
2911 * This basically trickles out pages so that we have _some_
2912 * free memory available even if there is no other activity
2913 * that frees anything up. This is needed for things like routing
2914 * etc, where we otherwise might have all activity going on in
2915 * asynchronous contexts that cannot page things out.
2917 * If there are applications that are active memory-allocators
2918 * (most normal use), this basically shouldn't matter.
2920 static int kswapd(void *p)
2922 unsigned long order, new_order;
2923 unsigned balanced_order;
2924 int classzone_idx, new_classzone_idx;
2925 int balanced_classzone_idx;
2926 pg_data_t *pgdat = (pg_data_t*)p;
2927 struct task_struct *tsk = current;
2929 struct reclaim_state reclaim_state = {
2930 .reclaimed_slab = 0,
2932 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2934 lockdep_set_current_reclaim_state(GFP_KERNEL);
2936 if (!cpumask_empty(cpumask))
2937 set_cpus_allowed_ptr(tsk, cpumask);
2938 current->reclaim_state = &reclaim_state;
2941 * Tell the memory management that we're a "memory allocator",
2942 * and that if we need more memory we should get access to it
2943 * regardless (see "__alloc_pages()"). "kswapd" should
2944 * never get caught in the normal page freeing logic.
2946 * (Kswapd normally doesn't need memory anyway, but sometimes
2947 * you need a small amount of memory in order to be able to
2948 * page out something else, and this flag essentially protects
2949 * us from recursively trying to free more memory as we're
2950 * trying to free the first piece of memory in the first place).
2952 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2955 order = new_order = 0;
2957 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2958 balanced_classzone_idx = classzone_idx;
2963 * If the last balance_pgdat was unsuccessful it's unlikely a
2964 * new request of a similar or harder type will succeed soon
2965 * so consider going to sleep on the basis we reclaimed at
2967 if (balanced_classzone_idx >= new_classzone_idx &&
2968 balanced_order == new_order) {
2969 new_order = pgdat->kswapd_max_order;
2970 new_classzone_idx = pgdat->classzone_idx;
2971 pgdat->kswapd_max_order = 0;
2972 pgdat->classzone_idx = pgdat->nr_zones - 1;
2975 if (order < new_order || classzone_idx > new_classzone_idx) {
2977 * Don't sleep if someone wants a larger 'order'
2978 * allocation or has tigher zone constraints
2981 classzone_idx = new_classzone_idx;
2983 kswapd_try_to_sleep(pgdat, balanced_order,
2984 balanced_classzone_idx);
2985 order = pgdat->kswapd_max_order;
2986 classzone_idx = pgdat->classzone_idx;
2988 new_classzone_idx = classzone_idx;
2989 pgdat->kswapd_max_order = 0;
2990 pgdat->classzone_idx = pgdat->nr_zones - 1;
2993 ret = try_to_freeze();
2994 if (kthread_should_stop())
2998 * We can speed up thawing tasks if we don't call balance_pgdat
2999 * after returning from the refrigerator
3002 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3003 balanced_classzone_idx = classzone_idx;
3004 balanced_order = balance_pgdat(pgdat, order,
3005 &balanced_classzone_idx);
3012 * A zone is low on free memory, so wake its kswapd task to service it.
3014 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3018 if (!populated_zone(zone))
3021 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3023 pgdat = zone->zone_pgdat;
3024 if (pgdat->kswapd_max_order < order) {
3025 pgdat->kswapd_max_order = order;
3026 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3028 if (!waitqueue_active(&pgdat->kswapd_wait))
3030 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3033 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3034 wake_up_interruptible(&pgdat->kswapd_wait);
3038 * The reclaimable count would be mostly accurate.
3039 * The less reclaimable pages may be
3040 * - mlocked pages, which will be moved to unevictable list when encountered
3041 * - mapped pages, which may require several travels to be reclaimed
3042 * - dirty pages, which is not "instantly" reclaimable
3044 unsigned long global_reclaimable_pages(void)
3048 nr = global_page_state(NR_ACTIVE_FILE) +
3049 global_page_state(NR_INACTIVE_FILE);
3051 if (nr_swap_pages > 0)
3052 nr += global_page_state(NR_ACTIVE_ANON) +
3053 global_page_state(NR_INACTIVE_ANON);
3058 unsigned long zone_reclaimable_pages(struct zone *zone)
3062 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3063 zone_page_state(zone, NR_INACTIVE_FILE);
3065 if (nr_swap_pages > 0)
3066 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3067 zone_page_state(zone, NR_INACTIVE_ANON);
3072 #ifdef CONFIG_HIBERNATION
3074 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3077 * Rather than trying to age LRUs the aim is to preserve the overall
3078 * LRU order by reclaiming preferentially
3079 * inactive > active > active referenced > active mapped
3081 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3083 struct reclaim_state reclaim_state;
3084 struct scan_control sc = {
3085 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3089 .nr_to_reclaim = nr_to_reclaim,
3090 .hibernation_mode = 1,
3093 struct shrink_control shrink = {
3094 .gfp_mask = sc.gfp_mask,
3096 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3097 struct task_struct *p = current;
3098 unsigned long nr_reclaimed;
3100 p->flags |= PF_MEMALLOC;
3101 lockdep_set_current_reclaim_state(sc.gfp_mask);
3102 reclaim_state.reclaimed_slab = 0;
3103 p->reclaim_state = &reclaim_state;
3105 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3107 p->reclaim_state = NULL;
3108 lockdep_clear_current_reclaim_state();
3109 p->flags &= ~PF_MEMALLOC;
3111 return nr_reclaimed;
3113 #endif /* CONFIG_HIBERNATION */
3115 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3116 not required for correctness. So if the last cpu in a node goes
3117 away, we get changed to run anywhere: as the first one comes back,
3118 restore their cpu bindings. */
3119 static int __devinit cpu_callback(struct notifier_block *nfb,
3120 unsigned long action, void *hcpu)
3124 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3125 for_each_node_state(nid, N_HIGH_MEMORY) {
3126 pg_data_t *pgdat = NODE_DATA(nid);
3127 const struct cpumask *mask;
3129 mask = cpumask_of_node(pgdat->node_id);
3131 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3132 /* One of our CPUs online: restore mask */
3133 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3140 * This kswapd start function will be called by init and node-hot-add.
3141 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3143 int kswapd_run(int nid)
3145 pg_data_t *pgdat = NODE_DATA(nid);
3151 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3152 if (IS_ERR(pgdat->kswapd)) {
3153 /* failure at boot is fatal */
3154 BUG_ON(system_state == SYSTEM_BOOTING);
3155 printk("Failed to start kswapd on node %d\n",nid);
3162 * Called by memory hotplug when all memory in a node is offlined.
3164 void kswapd_stop(int nid)
3166 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3169 kthread_stop(kswapd);
3172 static int __init kswapd_init(void)
3177 for_each_node_state(nid, N_HIGH_MEMORY)
3179 hotcpu_notifier(cpu_callback, 0);
3183 module_init(kswapd_init)
3189 * If non-zero call zone_reclaim when the number of free pages falls below
3192 int zone_reclaim_mode __read_mostly;
3194 #define RECLAIM_OFF 0
3195 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3196 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3197 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3200 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3201 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3204 #define ZONE_RECLAIM_PRIORITY 4
3207 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3210 int sysctl_min_unmapped_ratio = 1;
3213 * If the number of slab pages in a zone grows beyond this percentage then
3214 * slab reclaim needs to occur.
3216 int sysctl_min_slab_ratio = 5;
3218 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3220 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3221 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3222 zone_page_state(zone, NR_ACTIVE_FILE);
3225 * It's possible for there to be more file mapped pages than
3226 * accounted for by the pages on the file LRU lists because
3227 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3229 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3232 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3233 static long zone_pagecache_reclaimable(struct zone *zone)
3235 long nr_pagecache_reclaimable;
3239 * If RECLAIM_SWAP is set, then all file pages are considered
3240 * potentially reclaimable. Otherwise, we have to worry about
3241 * pages like swapcache and zone_unmapped_file_pages() provides
3244 if (zone_reclaim_mode & RECLAIM_SWAP)
3245 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3247 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3249 /* If we can't clean pages, remove dirty pages from consideration */
3250 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3251 delta += zone_page_state(zone, NR_FILE_DIRTY);
3253 /* Watch for any possible underflows due to delta */
3254 if (unlikely(delta > nr_pagecache_reclaimable))
3255 delta = nr_pagecache_reclaimable;
3257 return nr_pagecache_reclaimable - delta;
3261 * Try to free up some pages from this zone through reclaim.
3263 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3265 /* Minimum pages needed in order to stay on node */
3266 const unsigned long nr_pages = 1 << order;
3267 struct task_struct *p = current;
3268 struct reclaim_state reclaim_state;
3270 struct scan_control sc = {
3271 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3272 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3274 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3276 .gfp_mask = gfp_mask,
3279 struct shrink_control shrink = {
3280 .gfp_mask = sc.gfp_mask,
3282 unsigned long nr_slab_pages0, nr_slab_pages1;
3286 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3287 * and we also need to be able to write out pages for RECLAIM_WRITE
3290 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3291 lockdep_set_current_reclaim_state(gfp_mask);
3292 reclaim_state.reclaimed_slab = 0;
3293 p->reclaim_state = &reclaim_state;
3295 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3297 * Free memory by calling shrink zone with increasing
3298 * priorities until we have enough memory freed.
3300 priority = ZONE_RECLAIM_PRIORITY;
3302 shrink_zone(priority, zone, &sc);
3304 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3307 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3308 if (nr_slab_pages0 > zone->min_slab_pages) {
3310 * shrink_slab() does not currently allow us to determine how
3311 * many pages were freed in this zone. So we take the current
3312 * number of slab pages and shake the slab until it is reduced
3313 * by the same nr_pages that we used for reclaiming unmapped
3316 * Note that shrink_slab will free memory on all zones and may
3320 unsigned long lru_pages = zone_reclaimable_pages(zone);
3322 /* No reclaimable slab or very low memory pressure */
3323 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3326 /* Freed enough memory */
3327 nr_slab_pages1 = zone_page_state(zone,
3328 NR_SLAB_RECLAIMABLE);
3329 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3334 * Update nr_reclaimed by the number of slab pages we
3335 * reclaimed from this zone.
3337 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3338 if (nr_slab_pages1 < nr_slab_pages0)
3339 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3342 p->reclaim_state = NULL;
3343 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3344 lockdep_clear_current_reclaim_state();
3345 return sc.nr_reclaimed >= nr_pages;
3348 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3354 * Zone reclaim reclaims unmapped file backed pages and
3355 * slab pages if we are over the defined limits.
3357 * A small portion of unmapped file backed pages is needed for
3358 * file I/O otherwise pages read by file I/O will be immediately
3359 * thrown out if the zone is overallocated. So we do not reclaim
3360 * if less than a specified percentage of the zone is used by
3361 * unmapped file backed pages.
3363 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3364 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3365 return ZONE_RECLAIM_FULL;
3367 if (zone->all_unreclaimable)
3368 return ZONE_RECLAIM_FULL;
3371 * Do not scan if the allocation should not be delayed.
3373 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3374 return ZONE_RECLAIM_NOSCAN;
3377 * Only run zone reclaim on the local zone or on zones that do not
3378 * have associated processors. This will favor the local processor
3379 * over remote processors and spread off node memory allocations
3380 * as wide as possible.
3382 node_id = zone_to_nid(zone);
3383 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3384 return ZONE_RECLAIM_NOSCAN;
3386 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3387 return ZONE_RECLAIM_NOSCAN;
3389 ret = __zone_reclaim(zone, gfp_mask, order);
3390 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3393 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3400 * page_evictable - test whether a page is evictable
3401 * @page: the page to test
3402 * @vma: the VMA in which the page is or will be mapped, may be NULL
3404 * Test whether page is evictable--i.e., should be placed on active/inactive
3405 * lists vs unevictable list. The vma argument is !NULL when called from the
3406 * fault path to determine how to instantate a new page.
3408 * Reasons page might not be evictable:
3409 * (1) page's mapping marked unevictable
3410 * (2) page is part of an mlocked VMA
3413 int page_evictable(struct page *page, struct vm_area_struct *vma)
3416 if (mapping_unevictable(page_mapping(page)))
3419 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3427 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3428 * @pages: array of pages to check
3429 * @nr_pages: number of pages to check
3431 * Checks pages for evictability and moves them to the appropriate lru list.
3433 * This function is only used for SysV IPC SHM_UNLOCK.
3435 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3437 struct lruvec *lruvec;
3438 struct zone *zone = NULL;
3443 for (i = 0; i < nr_pages; i++) {
3444 struct page *page = pages[i];
3445 struct zone *pagezone;
3448 pagezone = page_zone(page);
3449 if (pagezone != zone) {
3451 spin_unlock_irq(&zone->lru_lock);
3453 spin_lock_irq(&zone->lru_lock);
3456 if (!PageLRU(page) || !PageUnevictable(page))
3459 if (page_evictable(page, NULL)) {
3460 enum lru_list lru = page_lru_base_type(page);
3462 VM_BUG_ON(PageActive(page));
3463 ClearPageUnevictable(page);
3464 __dec_zone_state(zone, NR_UNEVICTABLE);
3465 lruvec = mem_cgroup_lru_move_lists(zone, page,
3466 LRU_UNEVICTABLE, lru);
3467 list_move(&page->lru, &lruvec->lists[lru]);
3468 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3474 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3475 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3476 spin_unlock_irq(&zone->lru_lock);
3479 #endif /* CONFIG_SHMEM */
3481 static void warn_scan_unevictable_pages(void)
3483 printk_once(KERN_WARNING
3484 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3485 "disabled for lack of a legitimate use case. If you have "
3486 "one, please send an email to linux-mm@kvack.org.\n",
3491 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3492 * all nodes' unevictable lists for evictable pages
3494 unsigned long scan_unevictable_pages;
3496 int scan_unevictable_handler(struct ctl_table *table, int write,
3497 void __user *buffer,
3498 size_t *length, loff_t *ppos)
3500 warn_scan_unevictable_pages();
3501 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3502 scan_unevictable_pages = 0;
3508 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3509 * a specified node's per zone unevictable lists for evictable pages.
3512 static ssize_t read_scan_unevictable_node(struct device *dev,
3513 struct device_attribute *attr,
3516 warn_scan_unevictable_pages();
3517 return sprintf(buf, "0\n"); /* always zero; should fit... */
3520 static ssize_t write_scan_unevictable_node(struct device *dev,
3521 struct device_attribute *attr,
3522 const char *buf, size_t count)
3524 warn_scan_unevictable_pages();
3529 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3530 read_scan_unevictable_node,
3531 write_scan_unevictable_node);
3533 int scan_unevictable_register_node(struct node *node)
3535 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3538 void scan_unevictable_unregister_node(struct node *node)
3540 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);