4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29 #include <linux/mm_inline.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46 #include <linux/debugfs.h>
48 #include <asm/tlbflush.h>
49 #include <asm/div64.h>
51 #include <linux/swapops.h>
52 #include <linux/balloon_compaction.h>
53 #include <linux/suspend.h>
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/vmscan.h>
59 #ifdef CONFIG_RUNTIME_COMPCACHE
66 #endif /* CONFIG_RUNTIME_COMPCACHE */
69 /* Incremented by the number of inactive pages that were scanned */
70 unsigned long nr_scanned;
72 /* Number of pages freed so far during a call to shrink_zones() */
73 unsigned long nr_reclaimed;
75 /* How many pages shrink_list() should reclaim */
76 unsigned long nr_to_reclaim;
78 unsigned long hibernation_mode;
80 /* This context's GFP mask */
85 /* Can mapped pages be reclaimed? */
88 /* Can pages be swapped as part of reclaim? */
93 /* Scan (total_size >> priority) pages at once */
97 * The memory cgroup that hit its limit and as a result is the
98 * primary target of this reclaim invocation.
100 struct mem_cgroup *target_mem_cgroup;
103 * Nodemask of nodes allowed by the caller. If NULL, all nodes
106 nodemask_t *nodemask;
108 #ifdef CONFIG_RUNTIME_COMPCACHE
109 struct rtcc_control *rc;
110 #endif /* CONFIG_RUNTIME_COMPCACHE */
113 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
115 #ifdef ARCH_HAS_PREFETCH
116 #define prefetch_prev_lru_page(_page, _base, _field) \
118 if ((_page)->lru.prev != _base) { \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetch(&prev->_field); \
126 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
129 #ifdef ARCH_HAS_PREFETCHW
130 #define prefetchw_prev_lru_page(_page, _base, _field) \
132 if ((_page)->lru.prev != _base) { \
135 prev = lru_to_page(&(_page->lru)); \
136 prefetchw(&prev->_field); \
140 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 * From 0 .. 100. Higher means more swappy.
146 int vm_swappiness = 60;
147 unsigned long vm_total_pages; /* The total number of pages which the VM controls */
149 #ifdef CONFIG_RUNTIME_COMPCACHE
150 extern int get_rtcc_status(void);
151 long nr_kswapd_swapped = 0;
153 static bool rtcc_reclaim(struct scan_control *sc)
155 return (sc->rc != NULL);
157 #endif /* CONFIG_RUNTIME_COMPCACHE */
159 static LIST_HEAD(shrinker_list);
160 static DECLARE_RWSEM(shrinker_rwsem);
163 static bool global_reclaim(struct scan_control *sc)
165 return !sc->target_mem_cgroup;
168 static bool global_reclaim(struct scan_control *sc)
174 bool zone_reclaimable(struct zone *zone)
176 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
179 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
181 if (!mem_cgroup_disabled())
182 return mem_cgroup_get_lru_size(lruvec, lru);
184 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
187 struct dentry *debug_file;
189 static int debug_shrinker_show(struct seq_file *s, void *unused)
191 struct shrinker *shrinker;
192 struct shrink_control sc;
197 down_read(&shrinker_rwsem);
198 list_for_each_entry(shrinker, &shrinker_list, list) {
201 num_objs = shrinker->shrink(shrinker, &sc);
202 seq_printf(s, "%pf %d\n", shrinker->shrink, num_objs);
204 up_read(&shrinker_rwsem);
208 static int debug_shrinker_open(struct inode *inode, struct file *file)
210 return single_open(file, debug_shrinker_show, inode->i_private);
213 static const struct file_operations debug_shrinker_fops = {
214 .open = debug_shrinker_open,
217 .release = single_release,
221 * Add a shrinker callback to be called from the vm
223 void register_shrinker(struct shrinker *shrinker)
225 atomic_long_set(&shrinker->nr_in_batch, 0);
226 down_write(&shrinker_rwsem);
227 list_add_tail(&shrinker->list, &shrinker_list);
228 up_write(&shrinker_rwsem);
230 EXPORT_SYMBOL(register_shrinker);
232 static int __init add_shrinker_debug(void)
234 debugfs_create_file("shrinker", 0644, NULL, NULL,
235 &debug_shrinker_fops);
239 late_initcall(add_shrinker_debug);
244 void unregister_shrinker(struct shrinker *shrinker)
246 down_write(&shrinker_rwsem);
247 list_del(&shrinker->list);
248 up_write(&shrinker_rwsem);
250 EXPORT_SYMBOL(unregister_shrinker);
252 static inline int do_shrinker_shrink(struct shrinker *shrinker,
253 struct shrink_control *sc,
254 unsigned long nr_to_scan)
256 sc->nr_to_scan = nr_to_scan;
257 return (*shrinker->shrink)(shrinker, sc);
260 #define SHRINK_BATCH 128
262 * Call the shrink functions to age shrinkable caches
264 * Here we assume it costs one seek to replace a lru page and that it also
265 * takes a seek to recreate a cache object. With this in mind we age equal
266 * percentages of the lru and ageable caches. This should balance the seeks
267 * generated by these structures.
269 * If the vm encountered mapped pages on the LRU it increase the pressure on
270 * slab to avoid swapping.
272 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
274 * `lru_pages' represents the number of on-LRU pages in all the zones which
275 * are eligible for the caller's allocation attempt. It is used for balancing
276 * slab reclaim versus page reclaim.
278 * Returns the number of slab objects which we shrunk.
280 unsigned long shrink_slab(struct shrink_control *shrink,
281 unsigned long nr_pages_scanned,
282 unsigned long lru_pages)
284 struct shrinker *shrinker;
285 unsigned long ret = 0;
287 if (nr_pages_scanned == 0)
288 nr_pages_scanned = SWAP_CLUSTER_MAX;
290 if (!down_read_trylock(&shrinker_rwsem)) {
291 /* Assume we'll be able to shrink next time */
296 list_for_each_entry(shrinker, &shrinker_list, list) {
297 unsigned long long delta;
303 long batch_size = shrinker->batch ? shrinker->batch
306 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
311 * copy the current shrinker scan count into a local variable
312 * and zero it so that other concurrent shrinker invocations
313 * don't also do this scanning work.
315 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
318 delta = (4 * nr_pages_scanned) / shrinker->seeks;
320 do_div(delta, lru_pages + 1);
322 if (total_scan < 0) {
323 printk(KERN_ERR "shrink_slab: %pF negative objects to "
325 shrinker->shrink, total_scan);
326 total_scan = max_pass;
330 * We need to avoid excessive windup on filesystem shrinkers
331 * due to large numbers of GFP_NOFS allocations causing the
332 * shrinkers to return -1 all the time. This results in a large
333 * nr being built up so when a shrink that can do some work
334 * comes along it empties the entire cache due to nr >>>
335 * max_pass. This is bad for sustaining a working set in
338 * Hence only allow the shrinker to scan the entire cache when
339 * a large delta change is calculated directly.
341 if (delta < max_pass / 4)
342 total_scan = min(total_scan, max_pass / 2);
345 * Avoid risking looping forever due to too large nr value:
346 * never try to free more than twice the estimate number of
349 if (total_scan > max_pass * 2)
350 total_scan = max_pass * 2;
352 trace_mm_shrink_slab_start(shrinker, shrink, nr,
353 nr_pages_scanned, lru_pages,
354 max_pass, delta, total_scan);
356 while (total_scan >= batch_size) {
359 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
360 shrink_ret = do_shrinker_shrink(shrinker, shrink,
362 if (shrink_ret == -1)
364 if (shrink_ret < nr_before)
365 ret += nr_before - shrink_ret;
366 count_vm_events(SLABS_SCANNED, batch_size);
367 total_scan -= batch_size;
373 * move the unused scan count back into the shrinker in a
374 * manner that handles concurrent updates. If we exhausted the
375 * scan, there is no need to do an update.
378 new_nr = atomic_long_add_return(total_scan,
379 &shrinker->nr_in_batch);
381 new_nr = atomic_long_read(&shrinker->nr_in_batch);
383 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
385 up_read(&shrinker_rwsem);
391 static inline int is_page_cache_freeable(struct page *page)
394 * A freeable page cache page is referenced only by the caller
395 * that isolated the page, the page cache radix tree and
396 * optional buffer heads at page->private.
398 return page_count(page) - page_has_private(page) == 2;
401 static int may_write_to_queue(struct backing_dev_info *bdi,
402 struct scan_control *sc)
404 if (current->flags & PF_SWAPWRITE)
406 if (!bdi_write_congested(bdi))
408 if (bdi == current->backing_dev_info)
414 * We detected a synchronous write error writing a page out. Probably
415 * -ENOSPC. We need to propagate that into the address_space for a subsequent
416 * fsync(), msync() or close().
418 * The tricky part is that after writepage we cannot touch the mapping: nothing
419 * prevents it from being freed up. But we have a ref on the page and once
420 * that page is locked, the mapping is pinned.
422 * We're allowed to run sleeping lock_page() here because we know the caller has
425 static void handle_write_error(struct address_space *mapping,
426 struct page *page, int error)
429 if (page_mapping(page) == mapping)
430 mapping_set_error(mapping, error);
434 /* possible outcome of pageout() */
436 /* failed to write page out, page is locked */
438 /* move page to the active list, page is locked */
440 /* page has been sent to the disk successfully, page is unlocked */
442 /* page is clean and locked */
447 * pageout is called by shrink_page_list() for each dirty page.
448 * Calls ->writepage().
450 static pageout_t pageout(struct page *page, struct address_space *mapping,
451 struct scan_control *sc)
454 * If the page is dirty, only perform writeback if that write
455 * will be non-blocking. To prevent this allocation from being
456 * stalled by pagecache activity. But note that there may be
457 * stalls if we need to run get_block(). We could test
458 * PagePrivate for that.
460 * If this process is currently in __generic_file_aio_write() against
461 * this page's queue, we can perform writeback even if that
464 * If the page is swapcache, write it back even if that would
465 * block, for some throttling. This happens by accident, because
466 * swap_backing_dev_info is bust: it doesn't reflect the
467 * congestion state of the swapdevs. Easy to fix, if needed.
469 if (!is_page_cache_freeable(page))
473 * Some data journaling orphaned pages can have
474 * page->mapping == NULL while being dirty with clean buffers.
476 if (page_has_private(page)) {
477 if (try_to_free_buffers(page)) {
478 ClearPageDirty(page);
479 printk("%s: orphaned page\n", __func__);
485 if (mapping->a_ops->writepage == NULL)
486 return PAGE_ACTIVATE;
487 if (!may_write_to_queue(mapping->backing_dev_info, sc))
490 #ifdef CONFIG_CMA_RMQUEUE
491 if(IS_ENABLED(CONFIG_CMA) &&
492 !zone_watermark_ok_safe(page_zone(page), 0, SWAP_CLUSTER_MAX, 0, 0))
497 if (clear_page_dirty_for_io(page)) {
499 struct writeback_control wbc = {
500 .sync_mode = WB_SYNC_NONE,
501 .nr_to_write = SWAP_CLUSTER_MAX,
503 .range_end = LLONG_MAX,
507 SetPageReclaim(page);
508 res = mapping->a_ops->writepage(page, &wbc);
510 handle_write_error(mapping, page, res);
511 if (res == AOP_WRITEPAGE_ACTIVATE) {
512 ClearPageReclaim(page);
513 return PAGE_ACTIVATE;
516 if (!PageWriteback(page)) {
517 /* synchronous write or broken a_ops? */
518 ClearPageReclaim(page);
520 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
521 inc_zone_page_state(page, NR_VMSCAN_WRITE);
529 * Same as remove_mapping, but if the page is removed from the mapping, it
530 * gets returned with a refcount of 0.
532 static int __remove_mapping(struct address_space *mapping, struct page *page)
534 BUG_ON(!PageLocked(page));
535 BUG_ON(mapping != page_mapping(page));
537 spin_lock_irq(&mapping->tree_lock);
539 * The non racy check for a busy page.
541 * Must be careful with the order of the tests. When someone has
542 * a ref to the page, it may be possible that they dirty it then
543 * drop the reference. So if PageDirty is tested before page_count
544 * here, then the following race may occur:
546 * get_user_pages(&page);
547 * [user mapping goes away]
549 * !PageDirty(page) [good]
550 * SetPageDirty(page);
552 * !page_count(page) [good, discard it]
554 * [oops, our write_to data is lost]
556 * Reversing the order of the tests ensures such a situation cannot
557 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
558 * load is not satisfied before that of page->_count.
560 * Note that if SetPageDirty is always performed via set_page_dirty,
561 * and thus under tree_lock, then this ordering is not required.
563 if (!page_freeze_refs(page, 2))
565 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
566 if (unlikely(PageDirty(page))) {
567 page_unfreeze_refs(page, 2);
571 if (PageSwapCache(page)) {
572 swp_entry_t swap = { .val = page_private(page) };
573 __delete_from_swap_cache(page);
574 spin_unlock_irq(&mapping->tree_lock);
575 swapcache_free(swap, page);
577 void (*freepage)(struct page *);
579 freepage = mapping->a_ops->freepage;
581 __delete_from_page_cache(page);
582 spin_unlock_irq(&mapping->tree_lock);
583 mem_cgroup_uncharge_cache_page(page);
585 if (freepage != NULL)
592 spin_unlock_irq(&mapping->tree_lock);
597 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
598 * someone else has a ref on the page, abort and return 0. If it was
599 * successfully detached, return 1. Assumes the caller has a single ref on
602 int remove_mapping(struct address_space *mapping, struct page *page)
604 if (__remove_mapping(mapping, page)) {
606 * Unfreezing the refcount with 1 rather than 2 effectively
607 * drops the pagecache ref for us without requiring another
610 page_unfreeze_refs(page, 1);
617 * putback_lru_page - put previously isolated page onto appropriate LRU list
618 * @page: page to be put back to appropriate lru list
620 * Add previously isolated @page to appropriate LRU list.
621 * Page may still be unevictable for other reasons.
623 * lru_lock must not be held, interrupts must be enabled.
625 void putback_lru_page(struct page *page)
628 int active = !!TestClearPageActive(page);
629 int was_unevictable = PageUnevictable(page);
631 VM_BUG_ON(PageLRU(page));
634 ClearPageUnevictable(page);
636 if (page_evictable(page)) {
638 * For evictable pages, we can use the cache.
639 * In event of a race, worst case is we end up with an
640 * unevictable page on [in]active list.
641 * We know how to handle that.
643 lru = active + page_lru_base_type(page);
644 lru_cache_add_lru(page, lru);
647 * Put unevictable pages directly on zone's unevictable
650 lru = LRU_UNEVICTABLE;
651 add_page_to_unevictable_list(page);
653 * When racing with an mlock or AS_UNEVICTABLE clearing
654 * (page is unlocked) make sure that if the other thread
655 * does not observe our setting of PG_lru and fails
656 * isolation/check_move_unevictable_pages,
657 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
658 * the page back to the evictable list.
660 * The other side is TestClearPageMlocked() or shmem_lock().
666 * page's status can change while we move it among lru. If an evictable
667 * page is on unevictable list, it never be freed. To avoid that,
668 * check after we added it to the list, again.
670 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
671 if (!isolate_lru_page(page)) {
675 /* This means someone else dropped this page from LRU
676 * So, it will be freed or putback to LRU again. There is
677 * nothing to do here.
681 if (was_unevictable && lru != LRU_UNEVICTABLE)
682 count_vm_event(UNEVICTABLE_PGRESCUED);
683 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
684 count_vm_event(UNEVICTABLE_PGCULLED);
686 put_page(page); /* drop ref from isolate */
689 enum page_references {
691 PAGEREF_RECLAIM_CLEAN,
696 static enum page_references page_check_references(struct page *page,
697 struct scan_control *sc)
699 int referenced_ptes, referenced_page;
700 unsigned long vm_flags;
702 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
704 referenced_page = TestClearPageReferenced(page);
707 * Mlock lost the isolation race with us. Let try_to_unmap()
708 * move the page to the unevictable list.
710 if (vm_flags & VM_LOCKED)
711 return PAGEREF_RECLAIM;
713 if (referenced_ptes) {
714 if (PageSwapBacked(page))
715 return PAGEREF_ACTIVATE;
717 * All mapped pages start out with page table
718 * references from the instantiating fault, so we need
719 * to look twice if a mapped file page is used more
722 * Mark it and spare it for another trip around the
723 * inactive list. Another page table reference will
724 * lead to its activation.
726 * Note: the mark is set for activated pages as well
727 * so that recently deactivated but used pages are
730 SetPageReferenced(page);
732 if (referenced_page || referenced_ptes > 1)
733 return PAGEREF_ACTIVATE;
736 * Activate file-backed executable pages after first usage.
738 if (vm_flags & VM_EXEC)
739 return PAGEREF_ACTIVATE;
744 /* Reclaim if clean, defer dirty pages to writeback */
745 if (referenced_page && !PageSwapBacked(page))
746 return PAGEREF_RECLAIM_CLEAN;
748 return PAGEREF_RECLAIM;
752 * shrink_page_list() returns the number of reclaimed pages
754 static unsigned long shrink_page_list(struct list_head *page_list,
756 struct scan_control *sc,
757 enum ttu_flags ttu_flags,
758 unsigned long *ret_nr_dirty,
759 unsigned long *ret_nr_writeback,
762 LIST_HEAD(ret_pages);
763 LIST_HEAD(free_pages);
765 unsigned long nr_dirty = 0;
766 unsigned long nr_congested = 0;
767 unsigned long nr_reclaimed = 0;
768 unsigned long nr_writeback = 0;
772 mem_cgroup_uncharge_start();
773 while (!list_empty(page_list)) {
774 struct address_space *mapping;
777 enum page_references references = PAGEREF_RECLAIM_CLEAN;
781 page = lru_to_page(page_list);
782 list_del(&page->lru);
784 if (!trylock_page(page))
787 VM_BUG_ON(PageActive(page));
788 VM_BUG_ON(page_zone(page) != zone);
792 if (unlikely(!page_evictable(page)))
795 if (!sc->may_unmap && page_mapped(page))
798 /* Double the slab pressure for mapped and swapcache pages */
799 if (page_mapped(page) || PageSwapCache(page))
802 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
803 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
805 if (PageWriteback(page)) {
807 * memcg doesn't have any dirty pages throttling so we
808 * could easily OOM just because too many pages are in
809 * writeback and there is nothing else to reclaim.
811 * Check __GFP_IO, certainly because a loop driver
812 * thread might enter reclaim, and deadlock if it waits
813 * on a page for which it is needed to do the write
814 * (loop masks off __GFP_IO|__GFP_FS for this reason);
815 * but more thought would probably show more reasons.
817 * Don't require __GFP_FS, since we're not going into
818 * the FS, just waiting on its writeback completion.
819 * Worryingly, ext4 gfs2 and xfs allocate pages with
820 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
821 * testing may_enter_fs here is liable to OOM on them.
823 if (global_reclaim(sc) ||
824 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
826 * This is slightly racy - end_page_writeback()
827 * might have just cleared PageReclaim, then
828 * setting PageReclaim here end up interpreted
829 * as PageReadahead - but that does not matter
830 * enough to care. What we do want is for this
831 * page to have PageReclaim set next time memcg
832 * reclaim reaches the tests above, so it will
833 * then wait_on_page_writeback() to avoid OOM;
834 * and it's also appropriate in global reclaim.
836 SetPageReclaim(page);
840 wait_on_page_writeback(page);
844 references = page_check_references(page, sc);
846 switch (references) {
847 case PAGEREF_ACTIVATE:
848 goto activate_locked;
851 case PAGEREF_RECLAIM:
852 case PAGEREF_RECLAIM_CLEAN:
853 ; /* try to reclaim the page below */
857 * Anonymous process memory has backing store?
858 * Try to allocate it some swap space here.
860 if (PageAnon(page) && !PageSwapCache(page)) {
861 if (!(sc->gfp_mask & __GFP_IO))
863 if (!add_to_swap(page, page_list))
864 goto activate_locked;
868 mapping = page_mapping(page);
871 * The page is mapped into the page tables of one or more
872 * processes. Try to unmap it here.
874 if (page_mapped(page) && mapping) {
875 switch (try_to_unmap(page, ttu_flags)) {
877 goto activate_locked;
883 ; /* try to free the page below */
887 if (PageDirty(page)) {
891 * Only kswapd can writeback filesystem pages to
892 * avoid risk of stack overflow but do not writeback
893 * unless under significant pressure.
895 if (page_is_file_cache(page) &&
896 (!current_is_kswapd() ||
897 sc->priority >= DEF_PRIORITY - 2)) {
899 * Immediately reclaim when written back.
900 * Similar in principal to deactivate_page()
901 * except we already have the page isolated
902 * and know it's dirty
904 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
905 SetPageReclaim(page);
910 if (references == PAGEREF_RECLAIM_CLEAN)
914 if (!sc->may_writepage)
917 /* Page is dirty, try to write it out here */
918 switch (pageout(page, mapping, sc)) {
923 goto activate_locked;
925 if (PageWriteback(page))
931 * A synchronous write - probably a ramdisk. Go
932 * ahead and try to reclaim the page.
934 if (!trylock_page(page))
936 if (PageDirty(page) || PageWriteback(page))
938 mapping = page_mapping(page);
940 ; /* try to free the page below */
945 * If the page has buffers, try to free the buffer mappings
946 * associated with this page. If we succeed we try to free
949 * We do this even if the page is PageDirty().
950 * try_to_release_page() does not perform I/O, but it is
951 * possible for a page to have PageDirty set, but it is actually
952 * clean (all its buffers are clean). This happens if the
953 * buffers were written out directly, with submit_bh(). ext3
954 * will do this, as well as the blockdev mapping.
955 * try_to_release_page() will discover that cleanness and will
956 * drop the buffers and mark the page clean - it can be freed.
958 * Rarely, pages can have buffers and no ->mapping. These are
959 * the pages which were not successfully invalidated in
960 * truncate_complete_page(). We try to drop those buffers here
961 * and if that worked, and the page is no longer mapped into
962 * process address space (page_count == 1) it can be freed.
963 * Otherwise, leave the page on the LRU so it is swappable.
965 if (page_has_private(page)) {
966 if (!try_to_release_page(page, sc->gfp_mask))
967 goto activate_locked;
968 if (!mapping && page_count(page) == 1) {
970 if (put_page_testzero(page))
974 * rare race with speculative reference.
975 * the speculative reference will free
976 * this page shortly, so we may
977 * increment nr_reclaimed here (and
978 * leave it off the LRU).
986 if (!mapping || !__remove_mapping(mapping, page))
990 * At this point, we have no other references and there is
991 * no way to pick any more up (removed from LRU, removed
992 * from pagecache). Can use non-atomic bitops now (and
993 * we obviously don't have to worry about waking up a process
994 * waiting on the page lock, because there are no references.
996 __clear_page_locked(page);
1001 * Is there need to periodically free_page_list? It would
1002 * appear not as the counts should be low
1004 list_add(&page->lru, &free_pages);
1008 if (PageSwapCache(page))
1009 try_to_free_swap(page);
1011 putback_lru_page(page);
1015 /* Not a candidate for swapping, so reclaim swap space. */
1016 if (PageSwapCache(page) && vm_swap_full())
1017 try_to_free_swap(page);
1018 VM_BUG_ON(PageActive(page));
1019 SetPageActive(page);
1024 list_add(&page->lru, &ret_pages);
1025 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1029 * Tag a zone as congested if all the dirty pages encountered were
1030 * backed by a congested BDI. In this case, reclaimers should just
1031 * back off and wait for congestion to clear because further reclaim
1032 * will encounter the same problem
1034 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1035 zone_set_flag(zone, ZONE_CONGESTED);
1037 free_hot_cold_page_list(&free_pages, 1);
1039 list_splice(&ret_pages, page_list);
1040 count_vm_events(PGACTIVATE, pgactivate);
1041 mem_cgroup_uncharge_end();
1042 *ret_nr_dirty += nr_dirty;
1043 *ret_nr_writeback += nr_writeback;
1044 return nr_reclaimed;
1047 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1048 struct list_head *page_list)
1050 struct scan_control sc = {
1051 .gfp_mask = GFP_KERNEL,
1052 .priority = DEF_PRIORITY,
1055 unsigned long ret, dummy1, dummy2;
1056 struct page *page, *next;
1057 LIST_HEAD(clean_pages);
1059 list_for_each_entry_safe(page, next, page_list, lru) {
1060 if (page_is_file_cache(page) && !PageDirty(page) &&
1061 !isolated_balloon_page(page)) {
1062 ClearPageActive(page);
1063 list_move(&page->lru, &clean_pages);
1067 ret = shrink_page_list(&clean_pages, zone, &sc,
1068 TTU_UNMAP|TTU_IGNORE_ACCESS,
1069 &dummy1, &dummy2, true);
1070 list_splice(&clean_pages, page_list);
1071 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1076 * Attempt to remove the specified page from its LRU. Only take this page
1077 * if it is of the appropriate PageActive status. Pages which are being
1078 * freed elsewhere are also ignored.
1080 * page: page to consider
1081 * mode: one of the LRU isolation modes defined above
1083 * returns 0 on success, -ve errno on failure.
1085 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1089 /* Only take pages on the LRU. */
1093 /* Compaction should not handle unevictable pages but CMA can do so */
1094 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1100 if ((mode & ISOLATE_NO_CMA) && is_cma_pageblock(page))
1104 * To minimise LRU disruption, the caller can indicate that it only
1105 * wants to isolate pages it will be able to operate on without
1106 * blocking - clean pages for the most part.
1108 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1109 * is used by reclaim when it is cannot write to backing storage
1111 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1112 * that it is possible to migrate without blocking
1114 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1115 /* All the caller can do on PageWriteback is block */
1116 if (PageWriteback(page))
1119 if (PageDirty(page)) {
1120 struct address_space *mapping;
1122 /* ISOLATE_CLEAN means only clean pages */
1123 if (mode & ISOLATE_CLEAN)
1127 * Only pages without mappings or that have a
1128 * ->migratepage callback are possible to migrate
1131 mapping = page_mapping(page);
1132 if (mapping && !mapping->a_ops->migratepage)
1137 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1140 if (likely(get_page_unless_zero(page))) {
1142 * Be careful not to clear PageLRU until after we're
1143 * sure the page is not being freed elsewhere -- the
1144 * page release code relies on it.
1154 * zone->lru_lock is heavily contended. Some of the functions that
1155 * shrink the lists perform better by taking out a batch of pages
1156 * and working on them outside the LRU lock.
1158 * For pagecache intensive workloads, this function is the hottest
1159 * spot in the kernel (apart from copy_*_user functions).
1161 * Appropriate locks must be held before calling this function.
1163 * @nr_to_scan: The number of pages to look through on the list.
1164 * @lruvec: The LRU vector to pull pages from.
1165 * @dst: The temp list to put pages on to.
1166 * @nr_scanned: The number of pages that were scanned.
1167 * @sc: The scan_control struct for this reclaim session
1168 * @mode: One of the LRU isolation modes
1169 * @lru: LRU list id for isolating
1171 * returns how many pages were moved onto *@dst.
1173 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1174 struct lruvec *lruvec, struct list_head *dst,
1175 unsigned long *nr_scanned, struct scan_control *sc,
1176 isolate_mode_t mode, enum lru_list lru)
1178 struct list_head *src = &lruvec->lists[lru];
1179 unsigned long nr_taken = 0;
1182 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1186 page = lru_to_page(src);
1187 prefetchw_prev_lru_page(page, src, flags);
1189 VM_BUG_ON(!PageLRU(page));
1191 switch (__isolate_lru_page(page, mode)) {
1193 nr_pages = hpage_nr_pages(page);
1194 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1195 list_move(&page->lru, dst);
1196 nr_taken += nr_pages;
1200 /* else it is being freed elsewhere */
1201 list_move(&page->lru, src);
1210 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1211 nr_taken, mode, is_file_lru(lru));
1216 * isolate_lru_page - tries to isolate a page from its LRU list
1217 * @page: page to isolate from its LRU list
1219 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1220 * vmstat statistic corresponding to whatever LRU list the page was on.
1222 * Returns 0 if the page was removed from an LRU list.
1223 * Returns -EBUSY if the page was not on an LRU list.
1225 * The returned page will have PageLRU() cleared. If it was found on
1226 * the active list, it will have PageActive set. If it was found on
1227 * the unevictable list, it will have the PageUnevictable bit set. That flag
1228 * may need to be cleared by the caller before letting the page go.
1230 * The vmstat statistic corresponding to the list on which the page was
1231 * found will be decremented.
1234 * (1) Must be called with an elevated refcount on the page. This is a
1235 * fundamentnal difference from isolate_lru_pages (which is called
1236 * without a stable reference).
1237 * (2) the lru_lock must not be held.
1238 * (3) interrupts must be enabled.
1240 int isolate_lru_page(struct page *page)
1244 VM_BUG_ON(!page_count(page));
1246 if (PageLRU(page)) {
1247 struct zone *zone = page_zone(page);
1248 struct lruvec *lruvec;
1250 spin_lock_irq(&zone->lru_lock);
1251 lruvec = mem_cgroup_page_lruvec(page, zone);
1252 if (PageLRU(page)) {
1253 int lru = page_lru(page);
1256 del_page_from_lru_list(page, lruvec, lru);
1259 spin_unlock_irq(&zone->lru_lock);
1265 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1266 * then get resheduled. When there are massive number of tasks doing page
1267 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1268 * the LRU list will go small and be scanned faster than necessary, leading to
1269 * unnecessary swapping, thrashing and OOM.
1271 static int too_many_isolated(struct zone *zone, int file,
1272 struct scan_control *sc)
1274 unsigned long inactive, isolated;
1276 #ifdef CONFIG_RUNTIME_COMPCACHE
1277 if (get_rtcc_status() == 1)
1279 #endif /* CONFIG_RUNTIME_COMPCACHE */
1281 if (current_is_kswapd())
1284 if (!global_reclaim(sc))
1288 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1289 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1291 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1292 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1296 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1297 * won't get blocked by normal direct-reclaimers, forming a circular
1300 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1303 return isolated > inactive;
1306 static noinline_for_stack void
1307 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1309 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1310 struct zone *zone = lruvec_zone(lruvec);
1311 LIST_HEAD(pages_to_free);
1314 * Put back any unfreeable pages.
1316 while (!list_empty(page_list)) {
1317 struct page *page = lru_to_page(page_list);
1320 VM_BUG_ON(PageLRU(page));
1321 list_del(&page->lru);
1322 if (unlikely(!page_evictable(page))) {
1323 spin_unlock_irq(&zone->lru_lock);
1324 putback_lru_page(page);
1325 spin_lock_irq(&zone->lru_lock);
1329 lruvec = mem_cgroup_page_lruvec(page, zone);
1332 lru = page_lru(page);
1333 add_page_to_lru_list(page, lruvec, lru);
1335 if (is_active_lru(lru)) {
1336 int file = is_file_lru(lru);
1337 int numpages = hpage_nr_pages(page);
1338 reclaim_stat->recent_rotated[file] += numpages;
1340 if (put_page_testzero(page)) {
1341 __ClearPageLRU(page);
1342 __ClearPageActive(page);
1343 del_page_from_lru_list(page, lruvec, lru);
1345 if (unlikely(PageCompound(page))) {
1346 spin_unlock_irq(&zone->lru_lock);
1347 (*get_compound_page_dtor(page))(page);
1348 spin_lock_irq(&zone->lru_lock);
1350 list_add(&page->lru, &pages_to_free);
1355 * To save our caller's stack, now use input list for pages to free.
1357 list_splice(&pages_to_free, page_list);
1361 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1362 * of reclaimed pages
1364 static noinline_for_stack unsigned long
1365 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1366 struct scan_control *sc, enum lru_list lru)
1368 LIST_HEAD(page_list);
1369 unsigned long nr_scanned;
1370 unsigned long nr_reclaimed = 0;
1371 unsigned long nr_taken;
1372 unsigned long nr_dirty = 0;
1373 unsigned long nr_writeback = 0;
1374 isolate_mode_t isolate_mode = 0;
1375 int file = is_file_lru(lru);
1376 struct zone *zone = lruvec_zone(lruvec);
1377 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1379 while (unlikely(too_many_isolated(zone, file, sc))) {
1380 congestion_wait(BLK_RW_ASYNC, HZ/10);
1382 /* We are about to die and free our memory. Return now. */
1383 if (fatal_signal_pending(current))
1384 return SWAP_CLUSTER_MAX;
1390 isolate_mode |= ISOLATE_UNMAPPED;
1391 if (!sc->may_writepage)
1392 isolate_mode |= ISOLATE_CLEAN;
1394 if (allocflags_to_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE)
1395 isolate_mode |= ISOLATE_NO_CMA;
1398 spin_lock_irq(&zone->lru_lock);
1400 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1401 &nr_scanned, sc, isolate_mode, lru);
1403 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1404 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1406 if (global_reclaim(sc)) {
1407 zone->pages_scanned += nr_scanned;
1408 if (current_is_kswapd())
1409 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1411 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1413 spin_unlock_irq(&zone->lru_lock);
1418 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1419 &nr_dirty, &nr_writeback, false);
1421 spin_lock_irq(&zone->lru_lock);
1423 reclaim_stat->recent_scanned[file] += nr_taken;
1425 if (global_reclaim(sc)) {
1426 if (current_is_kswapd())
1427 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1430 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1434 putback_inactive_pages(lruvec, &page_list);
1436 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1438 spin_unlock_irq(&zone->lru_lock);
1440 free_hot_cold_page_list(&page_list, 1);
1443 * If reclaim is isolating dirty pages under writeback, it implies
1444 * that the long-lived page allocation rate is exceeding the page
1445 * laundering rate. Either the global limits are not being effective
1446 * at throttling processes due to the page distribution throughout
1447 * zones or there is heavy usage of a slow backing device. The
1448 * only option is to throttle from reclaim context which is not ideal
1449 * as there is no guarantee the dirtying process is throttled in the
1450 * same way balance_dirty_pages() manages.
1452 * This scales the number of dirty pages that must be under writeback
1453 * before throttling depending on priority. It is a simple backoff
1454 * function that has the most effect in the range DEF_PRIORITY to
1455 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1456 * in trouble and reclaim is considered to be in trouble.
1458 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1459 * DEF_PRIORITY-1 50% must be PageWriteback
1460 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1462 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1463 * isolated page is PageWriteback
1465 if (nr_writeback && nr_writeback >=
1466 (nr_taken >> (DEF_PRIORITY - sc->priority)))
1467 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1469 #ifdef CONFIG_RUNTIME_COMPCACHE
1471 if (rtcc_reclaim(sc))
1472 sc->rc->nr_swapped += nr_reclaimed;
1474 nr_kswapd_swapped += nr_reclaimed;
1476 #endif /* CONFIG_RUNTIME_COMPCACHE */
1478 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1480 nr_scanned, nr_reclaimed,
1482 trace_shrink_flags(file));
1483 return nr_reclaimed;
1487 * This moves pages from the active list to the inactive list.
1489 * We move them the other way if the page is referenced by one or more
1490 * processes, from rmap.
1492 * If the pages are mostly unmapped, the processing is fast and it is
1493 * appropriate to hold zone->lru_lock across the whole operation. But if
1494 * the pages are mapped, the processing is slow (page_referenced()) so we
1495 * should drop zone->lru_lock around each page. It's impossible to balance
1496 * this, so instead we remove the pages from the LRU while processing them.
1497 * It is safe to rely on PG_active against the non-LRU pages in here because
1498 * nobody will play with that bit on a non-LRU page.
1500 * The downside is that we have to touch page->_count against each page.
1501 * But we had to alter page->flags anyway.
1504 static void move_active_pages_to_lru(struct lruvec *lruvec,
1505 struct list_head *list,
1506 struct list_head *pages_to_free,
1509 struct zone *zone = lruvec_zone(lruvec);
1510 unsigned long pgmoved = 0;
1514 while (!list_empty(list)) {
1515 page = lru_to_page(list);
1516 lruvec = mem_cgroup_page_lruvec(page, zone);
1518 VM_BUG_ON(PageLRU(page));
1521 nr_pages = hpage_nr_pages(page);
1522 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1523 list_move(&page->lru, &lruvec->lists[lru]);
1524 pgmoved += nr_pages;
1526 if (put_page_testzero(page)) {
1527 __ClearPageLRU(page);
1528 __ClearPageActive(page);
1529 del_page_from_lru_list(page, lruvec, lru);
1531 if (unlikely(PageCompound(page))) {
1532 spin_unlock_irq(&zone->lru_lock);
1533 (*get_compound_page_dtor(page))(page);
1534 spin_lock_irq(&zone->lru_lock);
1536 list_add(&page->lru, pages_to_free);
1539 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1540 if (!is_active_lru(lru))
1541 __count_vm_events(PGDEACTIVATE, pgmoved);
1544 static void shrink_active_list(unsigned long nr_to_scan,
1545 struct lruvec *lruvec,
1546 struct scan_control *sc,
1549 unsigned long nr_taken;
1550 unsigned long nr_scanned;
1551 unsigned long vm_flags;
1552 LIST_HEAD(l_hold); /* The pages which were snipped off */
1553 LIST_HEAD(l_active);
1554 LIST_HEAD(l_inactive);
1556 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1557 unsigned long nr_rotated = 0;
1558 isolate_mode_t isolate_mode = 0;
1559 int file = is_file_lru(lru);
1560 struct zone *zone = lruvec_zone(lruvec);
1565 isolate_mode |= ISOLATE_UNMAPPED;
1566 if (!sc->may_writepage)
1567 isolate_mode |= ISOLATE_CLEAN;
1569 spin_lock_irq(&zone->lru_lock);
1571 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1572 &nr_scanned, sc, isolate_mode, lru);
1573 if (global_reclaim(sc))
1574 zone->pages_scanned += nr_scanned;
1576 reclaim_stat->recent_scanned[file] += nr_taken;
1578 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1579 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1580 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1581 spin_unlock_irq(&zone->lru_lock);
1583 while (!list_empty(&l_hold)) {
1585 page = lru_to_page(&l_hold);
1586 list_del(&page->lru);
1588 if (unlikely(!page_evictable(page))) {
1589 putback_lru_page(page);
1593 if (unlikely(buffer_heads_over_limit)) {
1594 if (page_has_private(page) && trylock_page(page)) {
1595 if (page_has_private(page))
1596 try_to_release_page(page, 0);
1601 if (page_referenced(page, 0, sc->target_mem_cgroup,
1603 nr_rotated += hpage_nr_pages(page);
1605 * Identify referenced, file-backed active pages and
1606 * give them one more trip around the active list. So
1607 * that executable code get better chances to stay in
1608 * memory under moderate memory pressure. Anon pages
1609 * are not likely to be evicted by use-once streaming
1610 * IO, plus JVM can create lots of anon VM_EXEC pages,
1611 * so we ignore them here.
1613 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1614 list_add(&page->lru, &l_active);
1619 ClearPageActive(page); /* we are de-activating */
1620 list_add(&page->lru, &l_inactive);
1624 * Move pages back to the lru list.
1626 spin_lock_irq(&zone->lru_lock);
1628 * Count referenced pages from currently used mappings as rotated,
1629 * even though only some of them are actually re-activated. This
1630 * helps balance scan pressure between file and anonymous pages in
1633 reclaim_stat->recent_rotated[file] += nr_rotated;
1635 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1636 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1637 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1638 spin_unlock_irq(&zone->lru_lock);
1640 free_hot_cold_page_list(&l_hold, 1);
1644 static int inactive_anon_is_low_global(struct zone *zone)
1646 unsigned long active, inactive;
1648 active = zone_page_state(zone, NR_ACTIVE_ANON);
1649 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1651 if (inactive * zone->inactive_ratio < active)
1658 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1659 * @lruvec: LRU vector to check
1661 * Returns true if the zone does not have enough inactive anon pages,
1662 * meaning some active anon pages need to be deactivated.
1664 static int inactive_anon_is_low(struct lruvec *lruvec)
1667 * If we don't have swap space, anonymous page deactivation
1670 if (!total_swap_pages)
1673 if (!mem_cgroup_disabled())
1674 return mem_cgroup_inactive_anon_is_low(lruvec);
1676 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1679 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1686 * inactive_file_is_low - check if file pages need to be deactivated
1687 * @lruvec: LRU vector to check
1689 * When the system is doing streaming IO, memory pressure here
1690 * ensures that active file pages get deactivated, until more
1691 * than half of the file pages are on the inactive list.
1693 * Once we get to that situation, protect the system's working
1694 * set from being evicted by disabling active file page aging.
1696 * This uses a different ratio than the anonymous pages, because
1697 * the page cache uses a use-once replacement algorithm.
1699 static int inactive_file_is_low(struct lruvec *lruvec)
1701 unsigned long inactive;
1702 unsigned long active;
1704 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1705 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1707 return active > inactive;
1710 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1712 if (is_file_lru(lru))
1713 return inactive_file_is_low(lruvec);
1715 return inactive_anon_is_low(lruvec);
1718 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1719 struct lruvec *lruvec, struct scan_control *sc)
1721 if (is_active_lru(lru)) {
1722 if (inactive_list_is_low(lruvec, lru))
1723 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1727 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1730 static int vmscan_swappiness(struct scan_control *sc)
1732 #ifdef CONFIG_RUNTIME_COMPCACHE
1733 if (rtcc_reclaim(sc))
1734 return sc->rc->swappiness;
1735 #endif /* CONFIG_RUNTIME_COMPCACHE */
1736 if (global_reclaim(sc))
1737 return vm_swappiness;
1738 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1749 * Determine how aggressively the anon and file LRU lists should be
1750 * scanned. The relative value of each set of LRU lists is determined
1751 * by looking at the fraction of the pages scanned we did rotate back
1752 * onto the active list instead of evict.
1754 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1755 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1757 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1760 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1762 u64 denominator = 0; /* gcc */
1763 struct zone *zone = lruvec_zone(lruvec);
1764 unsigned long anon_prio, file_prio;
1765 enum scan_balance scan_balance;
1766 unsigned long anon, file, free;
1767 bool force_scan = false;
1768 unsigned long ap, fp;
1772 * If the zone or memcg is small, nr[l] can be 0. This
1773 * results in no scanning on this priority and a potential
1774 * priority drop. Global direct reclaim can go to the next
1775 * zone and tends to have no problems. Global kswapd is for
1776 * zone balancing and it needs to scan a minimum amount. When
1777 * reclaiming for a memcg, a priority drop can cause high
1778 * latencies, so it's better to scan a minimum amount there as
1781 if (current_is_kswapd() && !zone_reclaimable(zone))
1783 if (!global_reclaim(sc))
1786 /* If we have no swap space, do not bother scanning anon pages. */
1787 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1788 scan_balance = SCAN_FILE;
1793 * Global reclaim will swap to prevent OOM even with no
1794 * swappiness, but memcg users want to use this knob to
1795 * disable swapping for individual groups completely when
1796 * using the memory controller's swap limit feature would be
1799 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1800 scan_balance = SCAN_FILE;
1805 * Do not apply any pressure balancing cleverness when the
1806 * system is close to OOM, scan both anon and file equally
1807 * (unless the swappiness setting disagrees with swapping).
1809 if (!sc->priority && vmscan_swappiness(sc)) {
1810 scan_balance = SCAN_EQUAL;
1814 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1815 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1816 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1817 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1820 * If it's foreseeable that reclaiming the file cache won't be
1821 * enough to get the zone back into a desirable shape, we have
1822 * to swap. Better start now and leave the - probably heavily
1823 * thrashing - remaining file pages alone.
1825 if (global_reclaim(sc)) {
1826 free = zone_page_state(zone, NR_FREE_PAGES);
1827 if (unlikely(file + free <= high_wmark_pages(zone))) {
1828 scan_balance = SCAN_ANON;
1834 * There is enough inactive page cache, do not reclaim
1835 * anything from the anonymous working set right now.
1837 if (!IS_ENABLED(CONFIG_ZRAM) &&
1838 !inactive_file_is_low(lruvec)) {
1839 scan_balance = SCAN_FILE;
1843 scan_balance = SCAN_FRACT;
1846 * With swappiness at 100, anonymous and file have the same priority.
1847 * This scanning priority is essentially the inverse of IO cost.
1849 #if defined(CONFIG_ZRAM) && defined(CONFIG_RUNTIME_COMPCACHE)
1850 if (rtcc_reclaim(sc)) {
1851 anon_prio = vmscan_swappiness(sc);
1852 file_prio = 200 - anon_prio;
1854 anon_prio = (vmscan_swappiness(sc) * anon) / (anon + file + 1);
1855 file_prio = (200 - vmscan_swappiness(sc)) * file / (anon + file + 1);
1858 anon_prio = vmscan_swappiness(sc);
1859 file_prio = 200 - anon_prio;
1862 * OK, so we have swap space and a fair amount of page cache
1863 * pages. We use the recently rotated / recently scanned
1864 * ratios to determine how valuable each cache is.
1866 * Because workloads change over time (and to avoid overflow)
1867 * we keep these statistics as a floating average, which ends
1868 * up weighing recent references more than old ones.
1870 * anon in [0], file in [1]
1872 spin_lock_irq(&zone->lru_lock);
1873 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1874 reclaim_stat->recent_scanned[0] /= 2;
1875 reclaim_stat->recent_rotated[0] /= 2;
1878 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1879 reclaim_stat->recent_scanned[1] /= 2;
1880 reclaim_stat->recent_rotated[1] /= 2;
1884 * The amount of pressure on anon vs file pages is inversely
1885 * proportional to the fraction of recently scanned pages on
1886 * each list that were recently referenced and in active use.
1888 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1889 ap /= reclaim_stat->recent_rotated[0] + 1;
1891 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1892 fp /= reclaim_stat->recent_rotated[1] + 1;
1893 spin_unlock_irq(&zone->lru_lock);
1897 denominator = ap + fp + 1;
1899 for_each_evictable_lru(lru) {
1900 int file = is_file_lru(lru);
1904 size = get_lru_size(lruvec, lru);
1905 scan = size >> sc->priority;
1907 if (!scan && force_scan)
1908 scan = min(size, SWAP_CLUSTER_MAX);
1910 switch (scan_balance) {
1912 /* Scan lists relative to size */
1916 * Scan types proportional to swappiness and
1917 * their relative recent reclaim efficiency.
1919 scan = div64_u64(scan * fraction[file], denominator);
1923 /* Scan one type exclusively */
1924 if ((scan_balance == SCAN_FILE) != file)
1928 /* Look ma, no brain */
1935 #ifdef CONFIG_RUNTIME_COMPCACHE
1936 /* reserve 512 pages when allocate memory for swap */
1937 static int swap_reserve = 512;
1938 module_param_named(swap_reserve, swap_reserve, int, S_IRUGO | S_IWUSR);
1942 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1944 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1946 unsigned long nr[NR_LRU_LISTS];
1947 unsigned long nr_to_scan;
1949 unsigned long nr_reclaimed = 0;
1950 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1951 struct blk_plug plug;
1952 #ifdef CONFIG_RUNTIME_COMPCACHE
1953 struct rtcc_control *rc = sc->rc;
1954 unsigned long mem_available;
1955 struct zone *zone = lruvec_zone(lruvec);
1956 unsigned long file_pages;
1957 #endif /* CONFIG_RUNTIME_COMPCACHE */
1959 get_scan_count(lruvec, sc, nr);
1961 blk_start_plug(&plug);
1962 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1963 nr[LRU_INACTIVE_FILE]) {
1964 #ifdef CONFIG_RUNTIME_COMPCACHE
1965 if (rtcc_reclaim(sc)) {
1966 if (rc->nr_swapped >= rc->nr_anon)
1967 nr[LRU_INACTIVE_ANON] = nr[LRU_ACTIVE_ANON] = 0;
1969 if ((sc->nr_reclaimed + nr_reclaimed - rc->nr_swapped) >= rc->nr_file)
1970 nr[LRU_INACTIVE_FILE] = nr[LRU_ACTIVE_FILE] = 0;
1972 /* Stop swap out when there are not enough available memory */
1973 mem_available = global_page_state(NR_FREE_PAGES) - global_page_state(NR_FREE_CMA_PAGES);
1974 if(mem_available < swap_reserve)
1975 nr[LRU_INACTIVE_ANON] = nr[LRU_ACTIVE_ANON] = 0;
1977 /* Stop dropping file caches when there are too little left */
1978 file_pages = global_page_state(NR_ACTIVE_FILE) + global_page_state(NR_INACTIVE_FILE);
1979 if(file_pages < min_wmark_pages(zone))
1980 nr[LRU_INACTIVE_FILE] = nr[LRU_ACTIVE_FILE] = 0;
1981 #endif /* CONFIG_RUNTIME_COMPCACHE */
1983 for_each_evictable_lru(lru) {
1985 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1986 nr[lru] -= nr_to_scan;
1988 nr_reclaimed += shrink_list(lru, nr_to_scan,
1993 * On large memory systems, scan >> priority can become
1994 * really large. This is fine for the starting priority;
1995 * we want to put equal scanning pressure on each zone.
1996 * However, if the VM has a harder time of freeing pages,
1997 * with multiple processes reclaiming pages, the total
1998 * freeing target can get unreasonably large.
2000 if (nr_reclaimed >= nr_to_reclaim &&
2001 sc->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(lruvec))
2012 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2013 sc, LRU_ACTIVE_ANON);
2015 throttle_vm_writeout(sc->gfp_mask);
2018 /* Use reclaim/compaction for costly allocs or under memory pressure */
2019 static bool in_reclaim_compaction(struct scan_control *sc)
2021 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2022 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2023 sc->priority < DEF_PRIORITY - 2))
2030 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2031 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2032 * true if more pages should be reclaimed such that when the page allocator
2033 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2034 * It will give up earlier than that if there is difficulty reclaiming pages.
2036 static inline bool should_continue_reclaim(struct zone *zone,
2037 unsigned long nr_reclaimed,
2038 unsigned long nr_scanned,
2039 struct scan_control *sc)
2041 unsigned long pages_for_compaction;
2042 unsigned long inactive_lru_pages;
2044 /* If not in reclaim/compaction mode, stop */
2045 if (!in_reclaim_compaction(sc))
2048 /* Consider stopping depending on scan and reclaim activity */
2049 if (sc->gfp_mask & __GFP_REPEAT) {
2051 * For __GFP_REPEAT allocations, stop reclaiming if the
2052 * full LRU list has been scanned and we are still failing
2053 * to reclaim pages. This full LRU scan is potentially
2054 * expensive but a __GFP_REPEAT caller really wants to succeed
2056 if (!nr_reclaimed && !nr_scanned)
2060 * For non-__GFP_REPEAT allocations which can presumably
2061 * fail without consequence, stop if we failed to reclaim
2062 * any pages from the last SWAP_CLUSTER_MAX number of
2063 * pages that were scanned. This will return to the
2064 * caller faster at the risk reclaim/compaction and
2065 * the resulting allocation attempt fails
2072 * If we have not reclaimed enough pages for compaction and the
2073 * inactive lists are large enough, continue reclaiming
2075 pages_for_compaction = (2UL << sc->order);
2076 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2077 if (get_nr_swap_pages() > 0)
2078 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2079 if (sc->nr_reclaimed < pages_for_compaction &&
2080 inactive_lru_pages > pages_for_compaction)
2083 /* If compaction would go ahead or the allocation would succeed, stop */
2084 switch (compaction_suitable(zone, sc->order)) {
2085 case COMPACT_PARTIAL:
2086 case COMPACT_CONTINUE:
2093 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2095 unsigned long nr_reclaimed, nr_scanned;
2098 struct mem_cgroup *root = sc->target_mem_cgroup;
2099 struct mem_cgroup_reclaim_cookie reclaim = {
2101 .priority = sc->priority,
2103 struct mem_cgroup *memcg;
2105 nr_reclaimed = sc->nr_reclaimed;
2106 nr_scanned = sc->nr_scanned;
2108 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2110 struct lruvec *lruvec;
2112 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2114 shrink_lruvec(lruvec, sc);
2117 * Direct reclaim and kswapd have to scan all memory
2118 * cgroups to fulfill the overall scan target for the
2121 * Limit reclaim, on the other hand, only cares about
2122 * nr_to_reclaim pages to be reclaimed and it will
2123 * retry with decreasing priority if one round over the
2124 * whole hierarchy is not sufficient.
2126 if (!global_reclaim(sc) &&
2127 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2128 mem_cgroup_iter_break(root, memcg);
2131 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2134 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2135 sc->nr_scanned - nr_scanned,
2136 sc->nr_reclaimed - nr_reclaimed);
2138 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2139 sc->nr_scanned - nr_scanned, sc));
2142 /* Returns true if compaction should go ahead for a high-order request */
2143 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2145 unsigned long balance_gap, watermark;
2148 /* Do not consider compaction for orders reclaim is meant to satisfy */
2149 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2153 * Compaction takes time to run and there are potentially other
2154 * callers using the pages just freed. Continue reclaiming until
2155 * there is a buffer of free pages available to give compaction
2156 * a reasonable chance of completing and allocating the page
2158 balance_gap = min(low_wmark_pages(zone),
2159 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2160 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2161 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2162 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2165 * If compaction is deferred, reclaim up to a point where
2166 * compaction will have a chance of success when re-enabled
2168 if (compaction_deferred(zone, sc->order))
2169 return watermark_ok;
2171 /* If compaction is not ready to start, keep reclaiming */
2172 if (!compaction_suitable(zone, sc->order))
2175 return watermark_ok;
2179 * This is the direct reclaim path, for page-allocating processes. We only
2180 * try to reclaim pages from zones which will satisfy the caller's allocation
2183 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2185 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2187 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2188 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2189 * zone defense algorithm.
2191 * If a zone is deemed to be full of pinned pages then just give it a light
2192 * scan then give up on it.
2194 * This function returns true if a zone is being reclaimed for a costly
2195 * high-order allocation and compaction is ready to begin. This indicates to
2196 * the caller that it should consider retrying the allocation instead of
2199 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2203 unsigned long nr_soft_reclaimed;
2204 unsigned long nr_soft_scanned;
2205 bool aborted_reclaim = false;
2208 * If the number of buffer_heads in the machine exceeds the maximum
2209 * allowed level, force direct reclaim to scan the highmem zone as
2210 * highmem pages could be pinning lowmem pages storing buffer_heads
2212 if (buffer_heads_over_limit)
2213 sc->gfp_mask |= __GFP_HIGHMEM;
2215 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2216 gfp_zone(sc->gfp_mask), sc->nodemask) {
2217 if (!populated_zone(zone))
2220 * Take care memory controller reclaiming has small influence
2223 if (global_reclaim(sc)) {
2224 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2226 if (sc->priority != DEF_PRIORITY &&
2227 !zone_reclaimable(zone))
2228 continue; /* Let kswapd poll it */
2229 if (IS_ENABLED(CONFIG_COMPACTION)) {
2231 * If we already have plenty of memory free for
2232 * compaction in this zone, don't free any more.
2233 * Even though compaction is invoked for any
2234 * non-zero order, only frequent costly order
2235 * reclamation is disruptive enough to become a
2236 * noticeable problem, like transparent huge
2239 if (compaction_ready(zone, sc)) {
2240 aborted_reclaim = true;
2245 * This steals pages from memory cgroups over softlimit
2246 * and returns the number of reclaimed pages and
2247 * scanned pages. This works for global memory pressure
2248 * and balancing, not for a memcg's limit.
2250 nr_soft_scanned = 0;
2251 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2252 sc->order, sc->gfp_mask,
2254 sc->nr_reclaimed += nr_soft_reclaimed;
2255 sc->nr_scanned += nr_soft_scanned;
2256 /* need some check for avoid more shrink_zone() */
2259 shrink_zone(zone, sc);
2262 return aborted_reclaim;
2265 /* All zones in zonelist are unreclaimable? */
2266 static bool all_unreclaimable(struct zonelist *zonelist,
2267 struct scan_control *sc)
2272 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2273 gfp_zone(sc->gfp_mask), sc->nodemask) {
2274 if (!populated_zone(zone))
2276 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2278 if (zone_reclaimable(zone))
2286 * This is the main entry point to direct page reclaim.
2288 * If a full scan of the inactive list fails to free enough memory then we
2289 * are "out of memory" and something needs to be killed.
2291 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2292 * high - the zone may be full of dirty or under-writeback pages, which this
2293 * caller can't do much about. We kick the writeback threads and take explicit
2294 * naps in the hope that some of these pages can be written. But if the
2295 * allocating task holds filesystem locks which prevent writeout this might not
2296 * work, and the allocation attempt will fail.
2298 * returns: 0, if no pages reclaimed
2299 * else, the number of pages reclaimed
2301 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2302 struct scan_control *sc,
2303 struct shrink_control *shrink)
2305 unsigned long total_scanned = 0;
2306 struct reclaim_state *reclaim_state = current->reclaim_state;
2309 unsigned long writeback_threshold;
2310 bool aborted_reclaim;
2312 delayacct_freepages_start();
2314 if (global_reclaim(sc))
2315 count_vm_event(ALLOCSTALL);
2318 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2321 aborted_reclaim = shrink_zones(zonelist, sc);
2324 * Don't shrink slabs when reclaiming memory from
2325 * over limit cgroups
2327 if (global_reclaim(sc)) {
2328 unsigned long lru_pages = 0;
2329 for_each_zone_zonelist(zone, z, zonelist,
2330 gfp_zone(sc->gfp_mask)) {
2331 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2334 lru_pages += zone_reclaimable_pages(zone);
2337 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2338 if (reclaim_state) {
2339 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2340 reclaim_state->reclaimed_slab = 0;
2343 total_scanned += sc->nr_scanned;
2344 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2348 * If we're getting trouble reclaiming, start doing
2349 * writepage even in laptop mode.
2351 if (sc->priority < DEF_PRIORITY - 2)
2352 sc->may_writepage = 1;
2355 * Try to write back as many pages as we just scanned. This
2356 * tends to cause slow streaming writers to write data to the
2357 * disk smoothly, at the dirtying rate, which is nice. But
2358 * that's undesirable in laptop mode, where we *want* lumpy
2359 * writeout. So in laptop mode, write out the whole world.
2361 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2362 if (total_scanned > writeback_threshold) {
2363 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2364 WB_REASON_TRY_TO_FREE_PAGES);
2365 sc->may_writepage = 1;
2368 /* Take a nap, wait for some writeback to complete */
2369 if (!sc->hibernation_mode && sc->nr_scanned &&
2370 sc->priority < DEF_PRIORITY - 2) {
2371 struct zone *preferred_zone;
2373 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2374 &cpuset_current_mems_allowed,
2376 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2378 } while (--sc->priority >= 0);
2381 delayacct_freepages_end();
2383 if (sc->nr_reclaimed)
2384 return sc->nr_reclaimed;
2387 * As hibernation is going on, kswapd is freezed so that it can't mark
2388 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2391 if (oom_killer_disabled)
2394 /* Aborted reclaim to try compaction? don't OOM, then */
2395 if (aborted_reclaim)
2398 /* top priority shrink_zones still had more to do? don't OOM, then */
2399 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2405 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2408 unsigned long pfmemalloc_reserve = 0;
2409 unsigned long free_pages = 0;
2413 for (i = 0; i <= ZONE_NORMAL; i++) {
2414 zone = &pgdat->node_zones[i];
2415 if (!populated_zone(zone))
2418 pfmemalloc_reserve += min_wmark_pages(zone);
2419 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2422 /* If there are no reserves (unexpected config) then do not throttle */
2423 if (!pfmemalloc_reserve)
2426 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2428 /* kswapd must be awake if processes are being throttled */
2429 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2430 pgdat->classzone_idx = min(pgdat->classzone_idx,
2431 (enum zone_type)ZONE_NORMAL);
2432 wake_up_interruptible(&pgdat->kswapd_wait);
2439 * Throttle direct reclaimers if backing storage is backed by the network
2440 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2441 * depleted. kswapd will continue to make progress and wake the processes
2442 * when the low watermark is reached.
2444 * Returns true if a fatal signal was delivered during throttling. If this
2445 * happens, the page allocator should not consider triggering the OOM killer.
2447 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2448 nodemask_t *nodemask)
2452 pg_data_t *pgdat = NULL;
2455 * Kernel threads should not be throttled as they may be indirectly
2456 * responsible for cleaning pages necessary for reclaim to make forward
2457 * progress. kjournald for example may enter direct reclaim while
2458 * committing a transaction where throttling it could forcing other
2459 * processes to block on log_wait_commit().
2461 if (current->flags & PF_KTHREAD)
2465 * If a fatal signal is pending, this process should not throttle.
2466 * It should return quickly so it can exit and free its memory
2468 if (fatal_signal_pending(current))
2472 * Check if the pfmemalloc reserves are ok by finding the first node
2473 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2474 * GFP_KERNEL will be required for allocating network buffers when
2475 * swapping over the network so ZONE_HIGHMEM is unusable.
2477 * Throttling is based on the first usable node and throttled processes
2478 * wait on a queue until kswapd makes progress and wakes them. There
2479 * is an affinity then between processes waking up and where reclaim
2480 * progress has been made assuming the process wakes on the same node.
2481 * More importantly, processes running on remote nodes will not compete
2482 * for remote pfmemalloc reserves and processes on different nodes
2483 * should make reasonable progress.
2485 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2486 gfp_mask, nodemask) {
2487 if (zone_idx(zone) > ZONE_NORMAL)
2490 /* Throttle based on the first usable node */
2491 pgdat = zone->zone_pgdat;
2492 if (pfmemalloc_watermark_ok(pgdat))
2497 /* If no zone was usable by the allocation flags then do not throttle */
2501 /* Account for the throttling */
2502 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2505 * If the caller cannot enter the filesystem, it's possible that it
2506 * is due to the caller holding an FS lock or performing a journal
2507 * transaction in the case of a filesystem like ext[3|4]. In this case,
2508 * it is not safe to block on pfmemalloc_wait as kswapd could be
2509 * blocked waiting on the same lock. Instead, throttle for up to a
2510 * second before continuing.
2512 if (!(gfp_mask & __GFP_FS)) {
2513 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2514 pfmemalloc_watermark_ok(pgdat), HZ);
2519 /* Throttle until kswapd wakes the process */
2520 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2521 pfmemalloc_watermark_ok(pgdat));
2524 if (fatal_signal_pending(current))
2531 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2532 gfp_t gfp_mask, nodemask_t *nodemask)
2534 unsigned long nr_reclaimed;
2535 struct scan_control sc = {
2536 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2537 .may_writepage = !laptop_mode,
2538 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2540 #ifdef CONFIG_RUNTIME_COMPCACHE
2544 #endif /* CONFIG_RUNTIME_COMPCACHE */
2546 .priority = DEF_PRIORITY,
2547 .target_mem_cgroup = NULL,
2548 .nodemask = nodemask,
2550 struct shrink_control shrink = {
2551 .gfp_mask = sc.gfp_mask,
2555 * Do not enter reclaim if fatal signal was delivered while throttled.
2556 * 1 is returned so that the page allocator does not OOM kill at this
2559 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2562 trace_mm_vmscan_direct_reclaim_begin(order,
2566 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2568 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2570 return nr_reclaimed;
2575 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2576 gfp_t gfp_mask, bool noswap,
2578 unsigned long *nr_scanned)
2580 struct scan_control sc = {
2582 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2583 .may_writepage = !laptop_mode,
2585 .may_swap = !noswap,
2588 .target_mem_cgroup = memcg,
2590 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2592 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2593 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2595 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2600 * NOTE: Although we can get the priority field, using it
2601 * here is not a good idea, since it limits the pages we can scan.
2602 * if we don't reclaim here, the shrink_zone from balance_pgdat
2603 * will pick up pages from other mem cgroup's as well. We hack
2604 * the priority and make it zero.
2606 shrink_lruvec(lruvec, &sc);
2608 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2610 *nr_scanned = sc.nr_scanned;
2611 return sc.nr_reclaimed;
2614 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2618 struct zonelist *zonelist;
2619 unsigned long nr_reclaimed;
2621 struct scan_control sc = {
2622 .may_writepage = !laptop_mode,
2624 .may_swap = !noswap,
2625 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2627 .priority = DEF_PRIORITY,
2628 .target_mem_cgroup = memcg,
2629 .nodemask = NULL, /* we don't care the placement */
2630 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2631 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2633 struct shrink_control shrink = {
2634 .gfp_mask = sc.gfp_mask,
2638 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2639 * take care of from where we get pages. So the node where we start the
2640 * scan does not need to be the current node.
2642 nid = mem_cgroup_select_victim_node(memcg);
2644 zonelist = NODE_DATA(nid)->node_zonelists;
2646 trace_mm_vmscan_memcg_reclaim_begin(0,
2650 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2652 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2654 return nr_reclaimed;
2658 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2660 struct mem_cgroup *memcg;
2662 if (!total_swap_pages)
2665 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2667 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2669 if (inactive_anon_is_low(lruvec))
2670 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2671 sc, LRU_ACTIVE_ANON);
2673 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2677 static bool zone_balanced(struct zone *zone, int order,
2678 unsigned long balance_gap, int classzone_idx)
2680 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2681 balance_gap, classzone_idx, 0))
2684 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2685 !compaction_suitable(zone, order))
2692 * pgdat_balanced() is used when checking if a node is balanced.
2694 * For order-0, all zones must be balanced!
2696 * For high-order allocations only zones that meet watermarks and are in a
2697 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2698 * total of balanced pages must be at least 25% of the zones allowed by
2699 * classzone_idx for the node to be considered balanced. Forcing all zones to
2700 * be balanced for high orders can cause excessive reclaim when there are
2702 * The choice of 25% is due to
2703 * o a 16M DMA zone that is balanced will not balance a zone on any
2704 * reasonable sized machine
2705 * o On all other machines, the top zone must be at least a reasonable
2706 * percentage of the middle zones. For example, on 32-bit x86, highmem
2707 * would need to be at least 256M for it to be balance a whole node.
2708 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2709 * to balance a node on its own. These seemed like reasonable ratios.
2711 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2713 unsigned long managed_pages = 0;
2714 unsigned long balanced_pages = 0;
2717 /* Check the watermark levels */
2718 for (i = 0; i <= classzone_idx; i++) {
2719 struct zone *zone = pgdat->node_zones + i;
2721 if (!populated_zone(zone))
2724 managed_pages += zone->managed_pages;
2727 * A special case here:
2729 * balance_pgdat() skips over all_unreclaimable after
2730 * DEF_PRIORITY. Effectively, it considers them balanced so
2731 * they must be considered balanced here as well!
2733 if (!zone_reclaimable(zone)) {
2734 balanced_pages += zone->managed_pages;
2738 if (zone_balanced(zone, order, 0, i))
2739 balanced_pages += zone->managed_pages;
2745 return balanced_pages >= (managed_pages >> 2);
2751 * Prepare kswapd for sleeping. This verifies that there are no processes
2752 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2754 * Returns true if kswapd is ready to sleep
2756 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2759 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2764 * The throttled processes are normally woken up in balance_pgdat() as
2765 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2766 * race between when kswapd checks the watermarks and a process gets
2767 * throttled. There is also a potential race if processes get
2768 * throttled, kswapd wakes, a large process exits thereby balancing the
2769 * zones, which causes kswapd to exit balance_pgdat() before reaching
2770 * the wake up checks. If kswapd is going to sleep, no process should
2771 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2772 * the wake up is premature, processes will wake kswapd and get
2773 * throttled again. The difference from wake ups in balance_pgdat() is
2774 * that here we are under prepare_to_wait().
2776 if (waitqueue_active(&pgdat->pfmemalloc_wait))
2777 wake_up_all(&pgdat->pfmemalloc_wait);
2779 return pgdat_balanced(pgdat, order, classzone_idx);
2783 * For kswapd, balance_pgdat() will work across all this node's zones until
2784 * they are all at high_wmark_pages(zone).
2786 * Returns the final order kswapd was reclaiming at
2788 * There is special handling here for zones which are full of pinned pages.
2789 * This can happen if the pages are all mlocked, or if they are all used by
2790 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2791 * What we do is to detect the case where all pages in the zone have been
2792 * scanned twice and there has been zero successful reclaim. Mark the zone as
2793 * dead and from now on, only perform a short scan. Basically we're polling
2794 * the zone for when the problem goes away.
2796 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2797 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2798 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2799 * lower zones regardless of the number of free pages in the lower zones. This
2800 * interoperates with the page allocator fallback scheme to ensure that aging
2801 * of pages is balanced across the zones.
2803 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2806 bool pgdat_is_balanced = false;
2808 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2809 struct reclaim_state *reclaim_state = current->reclaim_state;
2810 unsigned long nr_soft_reclaimed;
2811 unsigned long nr_soft_scanned;
2812 struct scan_control sc = {
2813 .gfp_mask = GFP_KERNEL,
2815 #ifndef CONFIG_KSWAPD_NOSWAP
2819 #endif /* CONFIG_KSWAPD_NOSWAP */
2821 * kswapd doesn't want to be bailed out while reclaim. because
2822 * we want to put equal scanning pressure on each zone.
2824 .nr_to_reclaim = ULONG_MAX,
2826 .target_mem_cgroup = NULL,
2828 struct shrink_control shrink = {
2829 .gfp_mask = sc.gfp_mask,
2832 sc.priority = DEF_PRIORITY;
2833 sc.nr_reclaimed = 0;
2834 sc.may_writepage = !laptop_mode;
2835 count_vm_event(PAGEOUTRUN);
2838 unsigned long lru_pages = 0;
2841 * Scan in the highmem->dma direction for the highest
2842 * zone which needs scanning
2844 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2845 struct zone *zone = pgdat->node_zones + i;
2847 if (!populated_zone(zone))
2850 if (sc.priority != DEF_PRIORITY &&
2851 !zone_reclaimable(zone))
2855 * Do some background aging of the anon list, to give
2856 * pages a chance to be referenced before reclaiming.
2858 age_active_anon(zone, &sc);
2861 * If the number of buffer_heads in the machine
2862 * exceeds the maximum allowed level and this node
2863 * has a highmem zone, force kswapd to reclaim from
2864 * it to relieve lowmem pressure.
2866 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2871 if (!zone_balanced(zone, order, 0, 0)) {
2875 /* If balanced, clear the congested flag */
2876 zone_clear_flag(zone, ZONE_CONGESTED);
2881 pgdat_is_balanced = true;
2885 for (i = 0; i <= end_zone; i++) {
2886 struct zone *zone = pgdat->node_zones + i;
2888 lru_pages += zone_reclaimable_pages(zone);
2892 * Now scan the zone in the dma->highmem direction, stopping
2893 * at the last zone which needs scanning.
2895 * We do this because the page allocator works in the opposite
2896 * direction. This prevents the page allocator from allocating
2897 * pages behind kswapd's direction of progress, which would
2898 * cause too much scanning of the lower zones.
2900 for (i = 0; i <= end_zone; i++) {
2901 struct zone *zone = pgdat->node_zones + i;
2903 unsigned long balance_gap;
2905 if (!populated_zone(zone))
2908 if (sc.priority != DEF_PRIORITY &&
2909 !zone_reclaimable(zone))
2914 nr_soft_scanned = 0;
2916 * Call soft limit reclaim before calling shrink_zone.
2918 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2921 sc.nr_reclaimed += nr_soft_reclaimed;
2924 * We put equal pressure on every zone, unless
2925 * one zone has way too many pages free
2926 * already. The "too many pages" is defined
2927 * as the high wmark plus a "gap" where the
2928 * gap is either the low watermark or 1%
2929 * of the zone, whichever is smaller.
2931 balance_gap = min(low_wmark_pages(zone),
2932 (zone->managed_pages +
2933 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2934 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2936 * Kswapd reclaims only single pages with compaction
2937 * enabled. Trying too hard to reclaim until contiguous
2938 * free pages have become available can hurt performance
2939 * by evicting too much useful data from memory.
2940 * Do not reclaim more than needed for compaction.
2943 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2944 compaction_suitable(zone, order) !=
2948 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2949 !zone_balanced(zone, testorder,
2950 balance_gap, end_zone)) {
2951 shrink_zone(zone, &sc);
2953 reclaim_state->reclaimed_slab = 0;
2954 shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2955 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2959 * If we're getting trouble reclaiming, start doing
2960 * writepage even in laptop mode.
2962 if (sc.priority < DEF_PRIORITY - 2)
2963 sc.may_writepage = 1;
2965 if (!zone_reclaimable(zone)) {
2966 if (end_zone && end_zone == i)
2971 if (zone_balanced(zone, testorder, 0, end_zone))
2973 * If a zone reaches its high watermark,
2974 * consider it to be no longer congested. It's
2975 * possible there are dirty pages backed by
2976 * congested BDIs but as pressure is relieved,
2977 * speculatively avoid congestion waits
2979 zone_clear_flag(zone, ZONE_CONGESTED);
2983 * If the low watermark is met there is no need for processes
2984 * to be throttled on pfmemalloc_wait as they should not be
2985 * able to safely make forward progress. Wake them
2987 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2988 pfmemalloc_watermark_ok(pgdat))
2989 wake_up(&pgdat->pfmemalloc_wait);
2991 if (pgdat_balanced(pgdat, order, *classzone_idx)) {
2992 pgdat_is_balanced = true;
2993 break; /* kswapd: all done */
2997 * We do this so kswapd doesn't build up large priorities for
2998 * example when it is freeing in parallel with allocators. It
2999 * matches the direct reclaim path behaviour in terms of impact
3000 * on zone->*_priority.
3002 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
3004 } while (--sc.priority >= 0);
3007 if (!pgdat_is_balanced) {
3013 * Fragmentation may mean that the system cannot be
3014 * rebalanced for high-order allocations in all zones.
3015 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
3016 * it means the zones have been fully scanned and are still
3017 * not balanced. For high-order allocations, there is
3018 * little point trying all over again as kswapd may
3021 * Instead, recheck all watermarks at order-0 as they
3022 * are the most important. If watermarks are ok, kswapd will go
3023 * back to sleep. High-order users can still perform direct
3024 * reclaim if they wish.
3026 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
3027 order = sc.order = 0;
3033 * If kswapd was reclaiming at a higher order, it has the option of
3034 * sleeping without all zones being balanced. Before it does, it must
3035 * ensure that the watermarks for order-0 on *all* zones are met and
3036 * that the congestion flags are cleared. The congestion flag must
3037 * be cleared as kswapd is the only mechanism that clears the flag
3038 * and it is potentially going to sleep here.
3041 int zones_need_compaction = 1;
3043 for (i = 0; i <= end_zone; i++) {
3044 struct zone *zone = pgdat->node_zones + i;
3046 if (!populated_zone(zone))
3049 /* Check if the memory needs to be defragmented. */
3050 if (zone_watermark_ok(zone, order,
3051 low_wmark_pages(zone), *classzone_idx, 0))
3052 zones_need_compaction = 0;
3055 if (zones_need_compaction)
3056 compact_pgdat(pgdat, order);
3060 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3061 * makes a decision on the order we were last reclaiming at. However,
3062 * if another caller entered the allocator slow path while kswapd
3063 * was awake, order will remain at the higher level
3065 *classzone_idx = end_zone;
3069 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3074 if (freezing(current) || kthread_should_stop())
3077 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3079 /* Try to sleep for a short interval */
3080 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3081 remaining = schedule_timeout(HZ/10);
3082 finish_wait(&pgdat->kswapd_wait, &wait);
3083 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3087 * After a short sleep, check if it was a premature sleep. If not, then
3088 * go fully to sleep until explicitly woken up.
3090 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3091 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3094 * vmstat counters are not perfectly accurate and the estimated
3095 * value for counters such as NR_FREE_PAGES can deviate from the
3096 * true value by nr_online_cpus * threshold. To avoid the zone
3097 * watermarks being breached while under pressure, we reduce the
3098 * per-cpu vmstat threshold while kswapd is awake and restore
3099 * them before going back to sleep.
3101 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3104 * Compaction records what page blocks it recently failed to
3105 * isolate pages from and skips them in the future scanning.
3106 * When kswapd is going to sleep, it is reasonable to assume
3107 * that pages and compaction may succeed so reset the cache.
3109 reset_isolation_suitable(pgdat);
3111 if (!kthread_should_stop())
3114 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3117 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3119 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3121 finish_wait(&pgdat->kswapd_wait, &wait);
3125 * The background pageout daemon, started as a kernel thread
3126 * from the init process.
3128 * This basically trickles out pages so that we have _some_
3129 * free memory available even if there is no other activity
3130 * that frees anything up. This is needed for things like routing
3131 * etc, where we otherwise might have all activity going on in
3132 * asynchronous contexts that cannot page things out.
3134 * If there are applications that are active memory-allocators
3135 * (most normal use), this basically shouldn't matter.
3137 static int kswapd(void *p)
3139 unsigned long order, new_order;
3140 unsigned balanced_order;
3141 int classzone_idx, new_classzone_idx;
3142 int balanced_classzone_idx;
3143 pg_data_t *pgdat = (pg_data_t*)p;
3144 struct task_struct *tsk = current;
3146 struct reclaim_state reclaim_state = {
3147 .reclaimed_slab = 0,
3149 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3151 lockdep_set_current_reclaim_state(GFP_KERNEL);
3153 if (!cpumask_empty(cpumask))
3154 set_cpus_allowed_ptr(tsk, cpumask);
3155 current->reclaim_state = &reclaim_state;
3158 * Tell the memory management that we're a "memory allocator",
3159 * and that if we need more memory we should get access to it
3160 * regardless (see "__alloc_pages()"). "kswapd" should
3161 * never get caught in the normal page freeing logic.
3163 * (Kswapd normally doesn't need memory anyway, but sometimes
3164 * you need a small amount of memory in order to be able to
3165 * page out something else, and this flag essentially protects
3166 * us from recursively trying to free more memory as we're
3167 * trying to free the first piece of memory in the first place).
3169 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3172 order = new_order = 0;
3174 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3175 balanced_classzone_idx = classzone_idx;
3180 * If the last balance_pgdat was unsuccessful it's unlikely a
3181 * new request of a similar or harder type will succeed soon
3182 * so consider going to sleep on the basis we reclaimed at
3184 if (balanced_classzone_idx >= new_classzone_idx &&
3185 balanced_order == new_order) {
3186 new_order = pgdat->kswapd_max_order;
3187 new_classzone_idx = pgdat->classzone_idx;
3188 pgdat->kswapd_max_order = 0;
3189 pgdat->classzone_idx = pgdat->nr_zones - 1;
3192 if (order < new_order || classzone_idx > new_classzone_idx) {
3194 * Don't sleep if someone wants a larger 'order'
3195 * allocation or has tigher zone constraints
3198 classzone_idx = new_classzone_idx;
3200 kswapd_try_to_sleep(pgdat, balanced_order,
3201 balanced_classzone_idx);
3202 order = pgdat->kswapd_max_order;
3203 classzone_idx = pgdat->classzone_idx;
3205 new_classzone_idx = classzone_idx;
3206 pgdat->kswapd_max_order = 0;
3207 pgdat->classzone_idx = pgdat->nr_zones - 1;
3210 ret = try_to_freeze();
3211 if (kthread_should_stop())
3215 * We can speed up thawing tasks if we don't call balance_pgdat
3216 * after returning from the refrigerator
3219 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3220 balanced_classzone_idx = classzone_idx;
3221 balanced_order = balance_pgdat(pgdat, order,
3222 &balanced_classzone_idx);
3226 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3227 current->reclaim_state = NULL;
3228 lockdep_clear_current_reclaim_state();
3235 static uint debug_kswapd_wakeup = 0;
3237 module_param_named(debug_kswapd_wakeup, debug_kswapd_wakeup, uint, S_IRUGO | S_IWUSR);
3241 * A zone is low on free memory, so wake its kswapd task to service it.
3243 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3247 if (!populated_zone(zone))
3250 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3252 pgdat = zone->zone_pgdat;
3253 if (pgdat->kswapd_max_order < order) {
3254 pgdat->kswapd_max_order = order;
3255 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3257 if (!waitqueue_active(&pgdat->kswapd_wait))
3259 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3262 if(debug_kswapd_wakeup &&
3263 zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + min_wmark_pages(zone), 0, 0))
3265 printk("%s(pages): free:%d, free_cma:%d, high:%d, low:%d, min:%d, order:%d\r\n",
3266 __func__, global_page_state(NR_FREE_PAGES), global_page_state(NR_FREE_CMA_PAGES), high_wmark_pages(zone), low_wmark_pages(zone) , min_wmark_pages(zone) ,order);
3270 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3271 wake_up_interruptible(&pgdat->kswapd_wait);
3275 * The reclaimable count would be mostly accurate.
3276 * The less reclaimable pages may be
3277 * - mlocked pages, which will be moved to unevictable list when encountered
3278 * - mapped pages, which may require several travels to be reclaimed
3279 * - dirty pages, which is not "instantly" reclaimable
3281 unsigned long global_reclaimable_pages(void)
3285 nr = global_page_state(NR_ACTIVE_FILE) +
3286 global_page_state(NR_INACTIVE_FILE);
3288 #ifndef CONFIG_KSWAPD_NOSWAP
3289 if (get_nr_swap_pages() > 0)
3290 nr += global_page_state(NR_ACTIVE_ANON) +
3291 global_page_state(NR_INACTIVE_ANON);
3292 #endif /* CONFIG_KSWAPD_NOSWAP */
3297 unsigned long zone_reclaimable_pages(struct zone *zone)
3301 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3302 zone_page_state(zone, NR_INACTIVE_FILE);
3304 #ifndef CONFIG_KSWAPD_NOSWAP
3305 if (get_nr_swap_pages() > 0)
3306 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3307 zone_page_state(zone, NR_INACTIVE_ANON);
3308 #endif /* CONFIG_KSWAPD_NOSWAP */
3313 #ifdef CONFIG_RUNTIME_COMPCACHE
3315 * This is the main entry point to direct page reclaim for RTCC.
3317 * If a full scan of the inactive list fails to free enough memory then we
3318 * are "out of memory" and something needs to be killed.
3320 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3321 * high - the zone may be full of dirty or under-writeback pages, which this
3322 * caller can't do much about. We kick the writeback threads and take explicit
3323 * naps in the hope that some of these pages can be written. But if the
3324 * allocating task holds filesystem locks which prevent writeout this might not
3325 * work, and the allocation attempt will fail.
3327 * returns: 0, if no pages reclaimed
3328 * else, the number of pages reclaimed
3330 static unsigned long rtcc_do_try_to_free_pages(struct zonelist *zonelist,
3331 struct scan_control *sc,
3332 struct shrink_control *shrink)
3334 unsigned long total_scanned = 0;
3335 unsigned long writeback_threshold;
3336 bool aborted_reclaim;
3338 delayacct_freepages_start();
3340 if (global_reclaim(sc))
3341 count_vm_event(ALLOCSTALL);
3344 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3347 aborted_reclaim = shrink_zones(zonelist, sc);
3349 total_scanned += sc->nr_scanned;
3350 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3354 * If we're getting trouble reclaiming, start doing
3355 * writepage even in laptop mode.
3357 if (sc->priority < DEF_PRIORITY - 2)
3358 sc->may_writepage = 1;
3361 * Try to write back as many pages as we just scanned. This
3362 * tends to cause slow streaming writers to write data to the
3363 * disk smoothly, at the dirtying rate, which is nice. But
3364 * that's undesirable in laptop mode, where we *want* lumpy
3365 * writeout. So in laptop mode, write out the whole world.
3367 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
3368 if (total_scanned > writeback_threshold) {
3369 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
3370 WB_REASON_TRY_TO_FREE_PAGES);
3371 sc->may_writepage = 1;
3374 /* Take a nap, wait for some writeback to complete */
3375 if (!sc->hibernation_mode && sc->nr_scanned &&
3376 sc->priority < DEF_PRIORITY - 2) {
3377 struct zone *preferred_zone;
3379 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
3380 &cpuset_current_mems_allowed,
3382 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
3384 } while (--sc->priority >= 0);
3387 delayacct_freepages_end();
3389 if (sc->nr_reclaimed)
3390 return sc->nr_reclaimed;
3393 * As hibernation is going on, kswapd is freezed so that it can't mark
3394 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
3397 if (oom_killer_disabled)
3400 /* Aborted reclaim to try compaction? don't OOM, then */
3401 if (aborted_reclaim)
3404 /* top priority shrink_zones still had more to do? don't OOM, then */
3405 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
3411 unsigned long rtcc_reclaim_pages(unsigned long nr_to_reclaim, int swappiness, unsigned long *nr_swapped)
3413 struct reclaim_state reclaim_state;
3414 struct scan_control sc = {
3415 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3419 .nr_to_reclaim = nr_to_reclaim,
3421 .priority = DEF_PRIORITY/2,
3423 struct shrink_control shrink = {
3424 .gfp_mask = sc.gfp_mask,
3426 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3427 struct task_struct *p = current;
3428 unsigned long nr_reclaimed;
3429 struct rtcc_control rc;
3431 rc.swappiness = swappiness;
3432 rc.nr_anon = nr_to_reclaim * swappiness / 200;
3433 rc.nr_file = nr_to_reclaim - rc.nr_anon;
3437 if (swappiness <= 1)
3440 p->flags |= PF_MEMALLOC;
3441 lockdep_set_current_reclaim_state(sc.gfp_mask);
3442 reclaim_state.reclaimed_slab = 0;
3443 p->reclaim_state = &reclaim_state;
3445 nr_reclaimed = rtcc_do_try_to_free_pages(zonelist, &sc, &shrink);
3446 *nr_swapped = rc.nr_swapped;
3448 p->reclaim_state = NULL;
3449 lockdep_clear_current_reclaim_state();
3450 p->flags &= ~PF_MEMALLOC;
3452 return nr_reclaimed;
3454 #endif /* CONFIG_RUNTIME_COMPCACHE */
3456 #if defined CONFIG_HIBERNATION || CONFIG_SHRINK_MEMORY
3458 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3461 * Rather than trying to age LRUs the aim is to preserve the overall
3462 * LRU order by reclaiming preferentially
3463 * inactive > active > active referenced > active mapped
3465 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3467 struct reclaim_state reclaim_state;
3468 struct scan_control sc = {
3469 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3473 .nr_to_reclaim = nr_to_reclaim,
3474 .hibernation_mode = 1,
3476 .priority = DEF_PRIORITY,
3478 struct shrink_control shrink = {
3479 .gfp_mask = sc.gfp_mask,
3481 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3482 struct task_struct *p = current;
3483 unsigned long nr_reclaimed;
3485 if (system_entering_hibernation())
3486 sc.hibernation_mode = 1;
3488 sc.hibernation_mode = 0;
3490 p->flags |= PF_MEMALLOC;
3491 lockdep_set_current_reclaim_state(sc.gfp_mask);
3492 reclaim_state.reclaimed_slab = 0;
3493 p->reclaim_state = &reclaim_state;
3495 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3497 p->reclaim_state = NULL;
3498 lockdep_clear_current_reclaim_state();
3499 p->flags &= ~PF_MEMALLOC;
3501 return nr_reclaimed;
3503 #endif /* CONFIG_HIBERNATION */
3505 #ifdef CONFIG_SHRINK_MEMORY
3506 int sysctl_shrink_memory;
3507 /* This is the entry point for system-wide shrink memory
3508 via /proc/sys/vm/shrink_memory */
3509 int sysctl_shrinkmem_handler(struct ctl_table *table, int write,
3510 void __user *buffer, size_t *length, loff_t *ppos)
3513 shrink_all_memory(totalram_pages);
3519 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3520 not required for correctness. So if the last cpu in a node goes
3521 away, we get changed to run anywhere: as the first one comes back,
3522 restore their cpu bindings. */
3523 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3528 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3529 for_each_node_state(nid, N_MEMORY) {
3530 pg_data_t *pgdat = NODE_DATA(nid);
3531 const struct cpumask *mask;
3533 mask = cpumask_of_node(pgdat->node_id);
3535 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3536 /* One of our CPUs online: restore mask */
3537 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3544 * This kswapd start function will be called by init and node-hot-add.
3545 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3547 int kswapd_run(int nid)
3549 pg_data_t *pgdat = NODE_DATA(nid);
3555 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3556 if (IS_ERR(pgdat->kswapd)) {
3557 /* failure at boot is fatal */
3558 BUG_ON(system_state == SYSTEM_BOOTING);
3559 pr_err("Failed to start kswapd on node %d\n", nid);
3560 ret = PTR_ERR(pgdat->kswapd);
3561 pgdat->kswapd = NULL;
3567 * Called by memory hotplug when all memory in a node is offlined. Caller must
3568 * hold lock_memory_hotplug().
3570 void kswapd_stop(int nid)
3572 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3575 kthread_stop(kswapd);
3576 NODE_DATA(nid)->kswapd = NULL;
3580 static int __init kswapd_init(void)
3585 for_each_node_state(nid, N_MEMORY)
3587 hotcpu_notifier(cpu_callback, 0);
3591 module_init(kswapd_init)
3597 * If non-zero call zone_reclaim when the number of free pages falls below
3600 int zone_reclaim_mode __read_mostly;
3602 #define RECLAIM_OFF 0
3603 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3604 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3605 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3608 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3609 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3612 #define ZONE_RECLAIM_PRIORITY 4
3615 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3618 int sysctl_min_unmapped_ratio = 1;
3621 * If the number of slab pages in a zone grows beyond this percentage then
3622 * slab reclaim needs to occur.
3624 int sysctl_min_slab_ratio = 5;
3626 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3628 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3629 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3630 zone_page_state(zone, NR_ACTIVE_FILE);
3633 * It's possible for there to be more file mapped pages than
3634 * accounted for by the pages on the file LRU lists because
3635 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3637 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3640 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3641 static long zone_pagecache_reclaimable(struct zone *zone)
3643 long nr_pagecache_reclaimable;
3647 * If RECLAIM_SWAP is set, then all file pages are considered
3648 * potentially reclaimable. Otherwise, we have to worry about
3649 * pages like swapcache and zone_unmapped_file_pages() provides
3652 if (zone_reclaim_mode & RECLAIM_SWAP)
3653 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3655 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3657 /* If we can't clean pages, remove dirty pages from consideration */
3658 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3659 delta += zone_page_state(zone, NR_FILE_DIRTY);
3661 /* Watch for any possible underflows due to delta */
3662 if (unlikely(delta > nr_pagecache_reclaimable))
3663 delta = nr_pagecache_reclaimable;
3665 return nr_pagecache_reclaimable - delta;
3669 * Try to free up some pages from this zone through reclaim.
3671 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3673 /* Minimum pages needed in order to stay on node */
3674 const unsigned long nr_pages = 1 << order;
3675 struct task_struct *p = current;
3676 struct reclaim_state reclaim_state;
3677 struct scan_control sc = {
3678 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3679 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3680 #ifdef CONFIG_RUNTIME_COMPCACHE
3684 #endif /* CONFIG_RUNTIME_COMPCACHE */
3685 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3686 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3688 .priority = ZONE_RECLAIM_PRIORITY,
3690 struct shrink_control shrink = {
3691 .gfp_mask = sc.gfp_mask,
3693 unsigned long nr_slab_pages0, nr_slab_pages1;
3697 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3698 * and we also need to be able to write out pages for RECLAIM_WRITE
3701 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3702 lockdep_set_current_reclaim_state(gfp_mask);
3703 reclaim_state.reclaimed_slab = 0;
3704 p->reclaim_state = &reclaim_state;
3706 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3708 * Free memory by calling shrink zone with increasing
3709 * priorities until we have enough memory freed.
3712 shrink_zone(zone, &sc);
3713 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3716 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3717 if (nr_slab_pages0 > zone->min_slab_pages) {
3719 * shrink_slab() does not currently allow us to determine how
3720 * many pages were freed in this zone. So we take the current
3721 * number of slab pages and shake the slab until it is reduced
3722 * by the same nr_pages that we used for reclaiming unmapped
3725 * Note that shrink_slab will free memory on all zones and may
3729 unsigned long lru_pages = zone_reclaimable_pages(zone);
3731 /* No reclaimable slab or very low memory pressure */
3732 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3735 /* Freed enough memory */
3736 nr_slab_pages1 = zone_page_state(zone,
3737 NR_SLAB_RECLAIMABLE);
3738 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3743 * Update nr_reclaimed by the number of slab pages we
3744 * reclaimed from this zone.
3746 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3747 if (nr_slab_pages1 < nr_slab_pages0)
3748 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3751 p->reclaim_state = NULL;
3752 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3753 lockdep_clear_current_reclaim_state();
3754 return sc.nr_reclaimed >= nr_pages;
3757 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3763 * Zone reclaim reclaims unmapped file backed pages and
3764 * slab pages if we are over the defined limits.
3766 * A small portion of unmapped file backed pages is needed for
3767 * file I/O otherwise pages read by file I/O will be immediately
3768 * thrown out if the zone is overallocated. So we do not reclaim
3769 * if less than a specified percentage of the zone is used by
3770 * unmapped file backed pages.
3772 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3773 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3774 return ZONE_RECLAIM_FULL;
3776 if (!zone_reclaimable(zone))
3777 return ZONE_RECLAIM_FULL;
3780 * Do not scan if the allocation should not be delayed.
3782 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3783 return ZONE_RECLAIM_NOSCAN;
3786 * Only run zone reclaim on the local zone or on zones that do not
3787 * have associated processors. This will favor the local processor
3788 * over remote processors and spread off node memory allocations
3789 * as wide as possible.
3791 node_id = zone_to_nid(zone);
3792 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3793 return ZONE_RECLAIM_NOSCAN;
3795 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3796 return ZONE_RECLAIM_NOSCAN;
3798 ret = __zone_reclaim(zone, gfp_mask, order);
3799 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3802 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3809 * page_evictable - test whether a page is evictable
3810 * @page: the page to test
3812 * Test whether page is evictable--i.e., should be placed on active/inactive
3813 * lists vs unevictable list.
3815 * Reasons page might not be evictable:
3816 * (1) page's mapping marked unevictable
3817 * (2) page is part of an mlocked VMA
3820 int page_evictable(struct page *page)
3822 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3827 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3828 * @pages: array of pages to check
3829 * @nr_pages: number of pages to check
3831 * Checks pages for evictability and moves them to the appropriate lru list.
3833 * This function is only used for SysV IPC SHM_UNLOCK.
3835 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3837 struct lruvec *lruvec;
3838 struct zone *zone = NULL;
3843 for (i = 0; i < nr_pages; i++) {
3844 struct page *page = pages[i];
3845 struct zone *pagezone;
3848 pagezone = page_zone(page);
3849 if (pagezone != zone) {
3851 spin_unlock_irq(&zone->lru_lock);
3853 spin_lock_irq(&zone->lru_lock);
3855 lruvec = mem_cgroup_page_lruvec(page, zone);
3857 if (!PageLRU(page) || !PageUnevictable(page))
3860 if (page_evictable(page)) {
3861 enum lru_list lru = page_lru_base_type(page);
3863 VM_BUG_ON(PageActive(page));
3864 ClearPageUnevictable(page);
3865 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3866 add_page_to_lru_list(page, lruvec, lru);
3872 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3873 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3874 spin_unlock_irq(&zone->lru_lock);
3877 #endif /* CONFIG_SHMEM */
3879 static void warn_scan_unevictable_pages(void)
3881 printk_once(KERN_WARNING
3882 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3883 "disabled for lack of a legitimate use case. If you have "
3884 "one, please send an email to linux-mm@kvack.org.\n",
3889 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3890 * all nodes' unevictable lists for evictable pages
3892 unsigned long scan_unevictable_pages;
3894 int scan_unevictable_handler(struct ctl_table *table, int write,
3895 void __user *buffer,
3896 size_t *length, loff_t *ppos)
3898 warn_scan_unevictable_pages();
3899 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3900 scan_unevictable_pages = 0;
3906 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3907 * a specified node's per zone unevictable lists for evictable pages.
3910 static ssize_t read_scan_unevictable_node(struct device *dev,
3911 struct device_attribute *attr,
3914 warn_scan_unevictable_pages();
3915 return sprintf(buf, "0\n"); /* always zero; should fit... */
3918 static ssize_t write_scan_unevictable_node(struct device *dev,
3919 struct device_attribute *attr,
3920 const char *buf, size_t count)
3922 warn_scan_unevictable_pages();
3927 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3928 read_scan_unevictable_node,
3929 write_scan_unevictable_node);
3931 int scan_unevictable_register_node(struct node *node)
3933 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3936 void scan_unevictable_unregister_node(struct node *node)
3938 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);