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.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup *target_mem_cgroup;
84 /* Scan (total_size >> priority) pages at once */
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
90 /* Writepage batching in laptop mode; RECLAIM_WRITE */
91 unsigned int may_writepage:1;
93 /* Can mapped pages be reclaimed? */
94 unsigned int may_unmap:1;
96 /* Can pages be swapped as part of reclaim? */
97 unsigned int may_swap:1;
99 /* Can cgroups be reclaimed below their normal consumption range? */
100 unsigned int may_thrash:1;
102 unsigned int hibernation_mode:1;
104 /* One of the zones is ready for compaction */
105 unsigned int compaction_ready:1;
107 /* Incremented by the number of inactive pages that were scanned */
108 unsigned long nr_scanned;
110 /* Number of pages freed so far during a call to shrink_zones() */
111 unsigned long nr_reclaimed;
114 #ifdef ARCH_HAS_PREFETCH
115 #define prefetch_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetch(&prev->_field); \
125 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
128 #ifdef ARCH_HAS_PREFETCHW
129 #define prefetchw_prev_lru_page(_page, _base, _field) \
131 if ((_page)->lru.prev != _base) { \
134 prev = lru_to_page(&(_page->lru)); \
135 prefetchw(&prev->_field); \
139 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
143 * From 0 .. 100. Higher means more swappy.
145 int vm_swappiness = 60;
147 * The total number of pages which are beyond the high watermark within all
150 unsigned long vm_total_pages;
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
156 static bool global_reclaim(struct scan_control *sc)
158 return !sc->target_mem_cgroup;
162 * sane_reclaim - is the usual dirty throttling mechanism operational?
163 * @sc: scan_control in question
165 * The normal page dirty throttling mechanism in balance_dirty_pages() is
166 * completely broken with the legacy memcg and direct stalling in
167 * shrink_page_list() is used for throttling instead, which lacks all the
168 * niceties such as fairness, adaptive pausing, bandwidth proportional
169 * allocation and configurability.
171 * This function tests whether the vmscan currently in progress can assume
172 * that the normal dirty throttling mechanism is operational.
174 static bool sane_reclaim(struct scan_control *sc)
176 struct mem_cgroup *memcg = sc->target_mem_cgroup;
180 #ifdef CONFIG_CGROUP_WRITEBACK
181 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
187 static bool global_reclaim(struct scan_control *sc)
192 static bool sane_reclaim(struct scan_control *sc)
199 * This misses isolated pages which are not accounted for to save counters.
200 * As the data only determines if reclaim or compaction continues, it is
201 * not expected that isolated pages will be a dominating factor.
203 unsigned long zone_reclaimable_pages(struct zone *zone)
207 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
208 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
209 if (get_nr_swap_pages() > 0)
210 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
211 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
216 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
220 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
222 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
224 if (get_nr_swap_pages() > 0)
225 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
227 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
232 bool pgdat_reclaimable(struct pglist_data *pgdat)
234 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
235 pgdat_reclaimable_pages(pgdat) * 6;
239 * lruvec_lru_size - Returns the number of pages on the given LRU list.
240 * @lruvec: lru vector
242 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
244 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
246 unsigned long lru_size;
249 if (!mem_cgroup_disabled())
250 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
252 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
254 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
255 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
258 if (!managed_zone(zone))
261 if (!mem_cgroup_disabled())
262 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
264 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
265 NR_ZONE_LRU_BASE + lru);
266 lru_size -= min(size, lru_size);
274 * Add a shrinker callback to be called from the vm.
276 int register_shrinker(struct shrinker *shrinker)
278 size_t size = sizeof(*shrinker->nr_deferred);
280 if (shrinker->flags & SHRINKER_NUMA_AWARE)
283 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
284 if (!shrinker->nr_deferred)
287 down_write(&shrinker_rwsem);
288 list_add_tail(&shrinker->list, &shrinker_list);
289 up_write(&shrinker_rwsem);
292 EXPORT_SYMBOL(register_shrinker);
297 void unregister_shrinker(struct shrinker *shrinker)
299 down_write(&shrinker_rwsem);
300 list_del(&shrinker->list);
301 up_write(&shrinker_rwsem);
302 kfree(shrinker->nr_deferred);
304 EXPORT_SYMBOL(unregister_shrinker);
306 #define SHRINK_BATCH 128
308 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
309 struct shrinker *shrinker,
310 unsigned long nr_scanned,
311 unsigned long nr_eligible)
313 unsigned long freed = 0;
314 unsigned long long delta;
319 int nid = shrinkctl->nid;
320 long batch_size = shrinker->batch ? shrinker->batch
322 long scanned = 0, next_deferred;
324 freeable = shrinker->count_objects(shrinker, shrinkctl);
329 * copy the current shrinker scan count into a local variable
330 * and zero it so that other concurrent shrinker invocations
331 * don't also do this scanning work.
333 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
336 delta = (4 * nr_scanned) / shrinker->seeks;
338 do_div(delta, nr_eligible + 1);
340 if (total_scan < 0) {
341 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
342 shrinker->scan_objects, total_scan);
343 total_scan = freeable;
346 next_deferred = total_scan;
349 * We need to avoid excessive windup on filesystem shrinkers
350 * due to large numbers of GFP_NOFS allocations causing the
351 * shrinkers to return -1 all the time. This results in a large
352 * nr being built up so when a shrink that can do some work
353 * comes along it empties the entire cache due to nr >>>
354 * freeable. This is bad for sustaining a working set in
357 * Hence only allow the shrinker to scan the entire cache when
358 * a large delta change is calculated directly.
360 if (delta < freeable / 4)
361 total_scan = min(total_scan, freeable / 2);
364 * Avoid risking looping forever due to too large nr value:
365 * never try to free more than twice the estimate number of
368 if (total_scan > freeable * 2)
369 total_scan = freeable * 2;
371 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
372 nr_scanned, nr_eligible,
373 freeable, delta, total_scan);
376 * Normally, we should not scan less than batch_size objects in one
377 * pass to avoid too frequent shrinker calls, but if the slab has less
378 * than batch_size objects in total and we are really tight on memory,
379 * we will try to reclaim all available objects, otherwise we can end
380 * up failing allocations although there are plenty of reclaimable
381 * objects spread over several slabs with usage less than the
384 * We detect the "tight on memory" situations by looking at the total
385 * number of objects we want to scan (total_scan). If it is greater
386 * than the total number of objects on slab (freeable), we must be
387 * scanning at high prio and therefore should try to reclaim as much as
390 while (total_scan >= batch_size ||
391 total_scan >= freeable) {
393 unsigned long nr_to_scan = min(batch_size, total_scan);
395 shrinkctl->nr_to_scan = nr_to_scan;
396 ret = shrinker->scan_objects(shrinker, shrinkctl);
397 if (ret == SHRINK_STOP)
401 count_vm_events(SLABS_SCANNED, nr_to_scan);
402 total_scan -= nr_to_scan;
403 scanned += nr_to_scan;
408 if (next_deferred >= scanned)
409 next_deferred -= scanned;
413 * move the unused scan count back into the shrinker in a
414 * manner that handles concurrent updates. If we exhausted the
415 * scan, there is no need to do an update.
417 if (next_deferred > 0)
418 new_nr = atomic_long_add_return(next_deferred,
419 &shrinker->nr_deferred[nid]);
421 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
423 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
428 * shrink_slab - shrink slab caches
429 * @gfp_mask: allocation context
430 * @nid: node whose slab caches to target
431 * @memcg: memory cgroup whose slab caches to target
432 * @nr_scanned: pressure numerator
433 * @nr_eligible: pressure denominator
435 * Call the shrink functions to age shrinkable caches.
437 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
438 * unaware shrinkers will receive a node id of 0 instead.
440 * @memcg specifies the memory cgroup to target. If it is not NULL,
441 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
442 * objects from the memory cgroup specified. Otherwise, only unaware
443 * shrinkers are called.
445 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
446 * the available objects should be scanned. Page reclaim for example
447 * passes the number of pages scanned and the number of pages on the
448 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
449 * when it encountered mapped pages. The ratio is further biased by
450 * the ->seeks setting of the shrink function, which indicates the
451 * cost to recreate an object relative to that of an LRU page.
453 * Returns the number of reclaimed slab objects.
455 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
456 struct mem_cgroup *memcg,
457 unsigned long nr_scanned,
458 unsigned long nr_eligible)
460 struct shrinker *shrinker;
461 unsigned long freed = 0;
463 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
467 nr_scanned = SWAP_CLUSTER_MAX;
469 if (!down_read_trylock(&shrinker_rwsem)) {
471 * If we would return 0, our callers would understand that we
472 * have nothing else to shrink and give up trying. By returning
473 * 1 we keep it going and assume we'll be able to shrink next
480 list_for_each_entry(shrinker, &shrinker_list, list) {
481 struct shrink_control sc = {
482 .gfp_mask = gfp_mask,
488 * If kernel memory accounting is disabled, we ignore
489 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
490 * passing NULL for memcg.
492 if (memcg_kmem_enabled() &&
493 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
496 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
499 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
502 up_read(&shrinker_rwsem);
508 void drop_slab_node(int nid)
513 struct mem_cgroup *memcg = NULL;
517 freed += shrink_slab(GFP_KERNEL, nid, memcg,
519 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
520 } while (freed > 10);
527 for_each_online_node(nid)
531 static inline int is_page_cache_freeable(struct page *page)
534 * A freeable page cache page is referenced only by the caller
535 * that isolated the page, the page cache radix tree and
536 * optional buffer heads at page->private.
538 return page_count(page) - page_has_private(page) == 2;
541 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
543 if (current->flags & PF_SWAPWRITE)
545 if (!inode_write_congested(inode))
547 if (inode_to_bdi(inode) == current->backing_dev_info)
553 * We detected a synchronous write error writing a page out. Probably
554 * -ENOSPC. We need to propagate that into the address_space for a subsequent
555 * fsync(), msync() or close().
557 * The tricky part is that after writepage we cannot touch the mapping: nothing
558 * prevents it from being freed up. But we have a ref on the page and once
559 * that page is locked, the mapping is pinned.
561 * We're allowed to run sleeping lock_page() here because we know the caller has
564 static void handle_write_error(struct address_space *mapping,
565 struct page *page, int error)
568 if (page_mapping(page) == mapping)
569 mapping_set_error(mapping, error);
573 /* possible outcome of pageout() */
575 /* failed to write page out, page is locked */
577 /* move page to the active list, page is locked */
579 /* page has been sent to the disk successfully, page is unlocked */
581 /* page is clean and locked */
586 * pageout is called by shrink_page_list() for each dirty page.
587 * Calls ->writepage().
589 static pageout_t pageout(struct page *page, struct address_space *mapping,
590 struct scan_control *sc)
593 * If the page is dirty, only perform writeback if that write
594 * will be non-blocking. To prevent this allocation from being
595 * stalled by pagecache activity. But note that there may be
596 * stalls if we need to run get_block(). We could test
597 * PagePrivate for that.
599 * If this process is currently in __generic_file_write_iter() against
600 * this page's queue, we can perform writeback even if that
603 * If the page is swapcache, write it back even if that would
604 * block, for some throttling. This happens by accident, because
605 * swap_backing_dev_info is bust: it doesn't reflect the
606 * congestion state of the swapdevs. Easy to fix, if needed.
608 if (!is_page_cache_freeable(page))
612 * Some data journaling orphaned pages can have
613 * page->mapping == NULL while being dirty with clean buffers.
615 if (page_has_private(page)) {
616 if (try_to_free_buffers(page)) {
617 ClearPageDirty(page);
618 pr_info("%s: orphaned page\n", __func__);
624 if (mapping->a_ops->writepage == NULL)
625 return PAGE_ACTIVATE;
626 if (!may_write_to_inode(mapping->host, sc))
629 if (clear_page_dirty_for_io(page)) {
631 struct writeback_control wbc = {
632 .sync_mode = WB_SYNC_NONE,
633 .nr_to_write = SWAP_CLUSTER_MAX,
635 .range_end = LLONG_MAX,
639 SetPageReclaim(page);
640 res = mapping->a_ops->writepage(page, &wbc);
642 handle_write_error(mapping, page, res);
643 if (res == AOP_WRITEPAGE_ACTIVATE) {
644 ClearPageReclaim(page);
645 return PAGE_ACTIVATE;
648 if (!PageWriteback(page)) {
649 /* synchronous write or broken a_ops? */
650 ClearPageReclaim(page);
652 trace_mm_vmscan_writepage(page);
653 inc_node_page_state(page, NR_VMSCAN_WRITE);
661 * Same as remove_mapping, but if the page is removed from the mapping, it
662 * gets returned with a refcount of 0.
664 static int __remove_mapping(struct address_space *mapping, struct page *page,
669 BUG_ON(!PageLocked(page));
670 BUG_ON(mapping != page_mapping(page));
672 spin_lock_irqsave(&mapping->tree_lock, flags);
674 * The non racy check for a busy page.
676 * Must be careful with the order of the tests. When someone has
677 * a ref to the page, it may be possible that they dirty it then
678 * drop the reference. So if PageDirty is tested before page_count
679 * here, then the following race may occur:
681 * get_user_pages(&page);
682 * [user mapping goes away]
684 * !PageDirty(page) [good]
685 * SetPageDirty(page);
687 * !page_count(page) [good, discard it]
689 * [oops, our write_to data is lost]
691 * Reversing the order of the tests ensures such a situation cannot
692 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
693 * load is not satisfied before that of page->_refcount.
695 * Note that if SetPageDirty is always performed via set_page_dirty,
696 * and thus under tree_lock, then this ordering is not required.
698 if (!page_ref_freeze(page, 2))
700 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
701 if (unlikely(PageDirty(page))) {
702 page_ref_unfreeze(page, 2);
706 if (PageSwapCache(page)) {
707 swp_entry_t swap = { .val = page_private(page) };
708 mem_cgroup_swapout(page, swap);
709 __delete_from_swap_cache(page);
710 spin_unlock_irqrestore(&mapping->tree_lock, flags);
711 swapcache_free(swap);
713 void (*freepage)(struct page *);
716 freepage = mapping->a_ops->freepage;
718 * Remember a shadow entry for reclaimed file cache in
719 * order to detect refaults, thus thrashing, later on.
721 * But don't store shadows in an address space that is
722 * already exiting. This is not just an optizimation,
723 * inode reclaim needs to empty out the radix tree or
724 * the nodes are lost. Don't plant shadows behind its
727 * We also don't store shadows for DAX mappings because the
728 * only page cache pages found in these are zero pages
729 * covering holes, and because we don't want to mix DAX
730 * exceptional entries and shadow exceptional entries in the
733 if (reclaimed && page_is_file_cache(page) &&
734 !mapping_exiting(mapping) && !dax_mapping(mapping))
735 shadow = workingset_eviction(mapping, page);
736 __delete_from_page_cache(page, shadow);
737 spin_unlock_irqrestore(&mapping->tree_lock, flags);
739 if (freepage != NULL)
746 spin_unlock_irqrestore(&mapping->tree_lock, flags);
751 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
752 * someone else has a ref on the page, abort and return 0. If it was
753 * successfully detached, return 1. Assumes the caller has a single ref on
756 int remove_mapping(struct address_space *mapping, struct page *page)
758 if (__remove_mapping(mapping, page, false)) {
760 * Unfreezing the refcount with 1 rather than 2 effectively
761 * drops the pagecache ref for us without requiring another
764 page_ref_unfreeze(page, 1);
771 * putback_lru_page - put previously isolated page onto appropriate LRU list
772 * @page: page to be put back to appropriate lru list
774 * Add previously isolated @page to appropriate LRU list.
775 * Page may still be unevictable for other reasons.
777 * lru_lock must not be held, interrupts must be enabled.
779 void putback_lru_page(struct page *page)
782 int was_unevictable = PageUnevictable(page);
784 VM_BUG_ON_PAGE(PageLRU(page), page);
787 ClearPageUnevictable(page);
789 if (page_evictable(page)) {
791 * For evictable pages, we can use the cache.
792 * In event of a race, worst case is we end up with an
793 * unevictable page on [in]active list.
794 * We know how to handle that.
796 is_unevictable = false;
800 * Put unevictable pages directly on zone's unevictable
803 is_unevictable = true;
804 add_page_to_unevictable_list(page);
806 * When racing with an mlock or AS_UNEVICTABLE clearing
807 * (page is unlocked) make sure that if the other thread
808 * does not observe our setting of PG_lru and fails
809 * isolation/check_move_unevictable_pages,
810 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
811 * the page back to the evictable list.
813 * The other side is TestClearPageMlocked() or shmem_lock().
819 * page's status can change while we move it among lru. If an evictable
820 * page is on unevictable list, it never be freed. To avoid that,
821 * check after we added it to the list, again.
823 if (is_unevictable && page_evictable(page)) {
824 if (!isolate_lru_page(page)) {
828 /* This means someone else dropped this page from LRU
829 * So, it will be freed or putback to LRU again. There is
830 * nothing to do here.
834 if (was_unevictable && !is_unevictable)
835 count_vm_event(UNEVICTABLE_PGRESCUED);
836 else if (!was_unevictable && is_unevictable)
837 count_vm_event(UNEVICTABLE_PGCULLED);
839 put_page(page); /* drop ref from isolate */
842 enum page_references {
844 PAGEREF_RECLAIM_CLEAN,
849 static enum page_references page_check_references(struct page *page,
850 struct scan_control *sc)
852 int referenced_ptes, referenced_page;
853 unsigned long vm_flags;
855 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
857 referenced_page = TestClearPageReferenced(page);
860 * Mlock lost the isolation race with us. Let try_to_unmap()
861 * move the page to the unevictable list.
863 if (vm_flags & VM_LOCKED)
864 return PAGEREF_RECLAIM;
866 if (referenced_ptes) {
867 if (PageSwapBacked(page))
868 return PAGEREF_ACTIVATE;
870 * All mapped pages start out with page table
871 * references from the instantiating fault, so we need
872 * to look twice if a mapped file page is used more
875 * Mark it and spare it for another trip around the
876 * inactive list. Another page table reference will
877 * lead to its activation.
879 * Note: the mark is set for activated pages as well
880 * so that recently deactivated but used pages are
883 SetPageReferenced(page);
885 if (referenced_page || referenced_ptes > 1)
886 return PAGEREF_ACTIVATE;
889 * Activate file-backed executable pages after first usage.
891 if (vm_flags & VM_EXEC)
892 return PAGEREF_ACTIVATE;
897 /* Reclaim if clean, defer dirty pages to writeback */
898 if (referenced_page && !PageSwapBacked(page))
899 return PAGEREF_RECLAIM_CLEAN;
901 return PAGEREF_RECLAIM;
904 /* Check if a page is dirty or under writeback */
905 static void page_check_dirty_writeback(struct page *page,
906 bool *dirty, bool *writeback)
908 struct address_space *mapping;
911 * Anonymous pages are not handled by flushers and must be written
912 * from reclaim context. Do not stall reclaim based on them
914 if (!page_is_file_cache(page)) {
920 /* By default assume that the page flags are accurate */
921 *dirty = PageDirty(page);
922 *writeback = PageWriteback(page);
924 /* Verify dirty/writeback state if the filesystem supports it */
925 if (!page_has_private(page))
928 mapping = page_mapping(page);
929 if (mapping && mapping->a_ops->is_dirty_writeback)
930 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
933 struct reclaim_stat {
935 unsigned nr_unqueued_dirty;
936 unsigned nr_congested;
937 unsigned nr_writeback;
938 unsigned nr_immediate;
939 unsigned nr_activate;
940 unsigned nr_ref_keep;
941 unsigned nr_unmap_fail;
945 * shrink_page_list() returns the number of reclaimed pages
947 static unsigned long shrink_page_list(struct list_head *page_list,
948 struct pglist_data *pgdat,
949 struct scan_control *sc,
950 enum ttu_flags ttu_flags,
951 struct reclaim_stat *stat,
954 LIST_HEAD(ret_pages);
955 LIST_HEAD(free_pages);
957 unsigned nr_unqueued_dirty = 0;
958 unsigned nr_dirty = 0;
959 unsigned nr_congested = 0;
960 unsigned nr_reclaimed = 0;
961 unsigned nr_writeback = 0;
962 unsigned nr_immediate = 0;
963 unsigned nr_ref_keep = 0;
964 unsigned nr_unmap_fail = 0;
968 while (!list_empty(page_list)) {
969 struct address_space *mapping;
972 enum page_references references = PAGEREF_RECLAIM_CLEAN;
973 bool dirty, writeback;
974 bool lazyfree = false;
975 int ret = SWAP_SUCCESS;
979 page = lru_to_page(page_list);
980 list_del(&page->lru);
982 if (!trylock_page(page))
985 VM_BUG_ON_PAGE(PageActive(page), page);
989 if (unlikely(!page_evictable(page)))
992 if (!sc->may_unmap && page_mapped(page))
995 /* Double the slab pressure for mapped and swapcache pages */
996 if (page_mapped(page) || PageSwapCache(page))
999 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1000 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1003 * The number of dirty pages determines if a zone is marked
1004 * reclaim_congested which affects wait_iff_congested. kswapd
1005 * will stall and start writing pages if the tail of the LRU
1006 * is all dirty unqueued pages.
1008 page_check_dirty_writeback(page, &dirty, &writeback);
1009 if (dirty || writeback)
1012 if (dirty && !writeback)
1013 nr_unqueued_dirty++;
1016 * Treat this page as congested if the underlying BDI is or if
1017 * pages are cycling through the LRU so quickly that the
1018 * pages marked for immediate reclaim are making it to the
1019 * end of the LRU a second time.
1021 mapping = page_mapping(page);
1022 if (((dirty || writeback) && mapping &&
1023 inode_write_congested(mapping->host)) ||
1024 (writeback && PageReclaim(page)))
1028 * If a page at the tail of the LRU is under writeback, there
1029 * are three cases to consider.
1031 * 1) If reclaim is encountering an excessive number of pages
1032 * under writeback and this page is both under writeback and
1033 * PageReclaim then it indicates that pages are being queued
1034 * for IO but are being recycled through the LRU before the
1035 * IO can complete. Waiting on the page itself risks an
1036 * indefinite stall if it is impossible to writeback the
1037 * page due to IO error or disconnected storage so instead
1038 * note that the LRU is being scanned too quickly and the
1039 * caller can stall after page list has been processed.
1041 * 2) Global or new memcg reclaim encounters a page that is
1042 * not marked for immediate reclaim, or the caller does not
1043 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1044 * not to fs). In this case mark the page for immediate
1045 * reclaim and continue scanning.
1047 * Require may_enter_fs because we would wait on fs, which
1048 * may not have submitted IO yet. And the loop driver might
1049 * enter reclaim, and deadlock if it waits on a page for
1050 * which it is needed to do the write (loop masks off
1051 * __GFP_IO|__GFP_FS for this reason); but more thought
1052 * would probably show more reasons.
1054 * 3) Legacy memcg encounters a page that is already marked
1055 * PageReclaim. memcg does not have any dirty pages
1056 * throttling so we could easily OOM just because too many
1057 * pages are in writeback and there is nothing else to
1058 * reclaim. Wait for the writeback to complete.
1060 if (PageWriteback(page)) {
1062 if (current_is_kswapd() &&
1063 PageReclaim(page) &&
1064 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1069 } else if (sane_reclaim(sc) ||
1070 !PageReclaim(page) || !may_enter_fs) {
1072 * This is slightly racy - end_page_writeback()
1073 * might have just cleared PageReclaim, then
1074 * setting PageReclaim here end up interpreted
1075 * as PageReadahead - but that does not matter
1076 * enough to care. What we do want is for this
1077 * page to have PageReclaim set next time memcg
1078 * reclaim reaches the tests above, so it will
1079 * then wait_on_page_writeback() to avoid OOM;
1080 * and it's also appropriate in global reclaim.
1082 SetPageReclaim(page);
1089 wait_on_page_writeback(page);
1090 /* then go back and try same page again */
1091 list_add_tail(&page->lru, page_list);
1097 references = page_check_references(page, sc);
1099 switch (references) {
1100 case PAGEREF_ACTIVATE:
1101 goto activate_locked;
1105 case PAGEREF_RECLAIM:
1106 case PAGEREF_RECLAIM_CLEAN:
1107 ; /* try to reclaim the page below */
1111 * Anonymous process memory has backing store?
1112 * Try to allocate it some swap space here.
1114 if (PageAnon(page) && !PageSwapCache(page)) {
1115 if (!(sc->gfp_mask & __GFP_IO))
1117 if (!add_to_swap(page, page_list))
1118 goto activate_locked;
1122 /* Adding to swap updated mapping */
1123 mapping = page_mapping(page);
1124 } else if (unlikely(PageTransHuge(page))) {
1125 /* Split file THP */
1126 if (split_huge_page_to_list(page, page_list))
1130 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1133 * The page is mapped into the page tables of one or more
1134 * processes. Try to unmap it here.
1136 if (page_mapped(page) && mapping) {
1137 switch (ret = try_to_unmap(page, lazyfree ?
1138 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1139 (ttu_flags | TTU_BATCH_FLUSH))) {
1142 goto activate_locked;
1150 ; /* try to free the page below */
1154 if (PageDirty(page)) {
1156 * Only kswapd can writeback filesystem pages to
1157 * avoid risk of stack overflow but only writeback
1158 * if many dirty pages have been encountered.
1160 if (page_is_file_cache(page) &&
1161 (!current_is_kswapd() ||
1162 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1164 * Immediately reclaim when written back.
1165 * Similar in principal to deactivate_page()
1166 * except we already have the page isolated
1167 * and know it's dirty
1169 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1170 SetPageReclaim(page);
1175 if (references == PAGEREF_RECLAIM_CLEAN)
1179 if (!sc->may_writepage)
1183 * Page is dirty. Flush the TLB if a writable entry
1184 * potentially exists to avoid CPU writes after IO
1185 * starts and then write it out here.
1187 try_to_unmap_flush_dirty();
1188 switch (pageout(page, mapping, sc)) {
1192 goto activate_locked;
1194 if (PageWriteback(page))
1196 if (PageDirty(page))
1200 * A synchronous write - probably a ramdisk. Go
1201 * ahead and try to reclaim the page.
1203 if (!trylock_page(page))
1205 if (PageDirty(page) || PageWriteback(page))
1207 mapping = page_mapping(page);
1209 ; /* try to free the page below */
1214 * If the page has buffers, try to free the buffer mappings
1215 * associated with this page. If we succeed we try to free
1218 * We do this even if the page is PageDirty().
1219 * try_to_release_page() does not perform I/O, but it is
1220 * possible for a page to have PageDirty set, but it is actually
1221 * clean (all its buffers are clean). This happens if the
1222 * buffers were written out directly, with submit_bh(). ext3
1223 * will do this, as well as the blockdev mapping.
1224 * try_to_release_page() will discover that cleanness and will
1225 * drop the buffers and mark the page clean - it can be freed.
1227 * Rarely, pages can have buffers and no ->mapping. These are
1228 * the pages which were not successfully invalidated in
1229 * truncate_complete_page(). We try to drop those buffers here
1230 * and if that worked, and the page is no longer mapped into
1231 * process address space (page_count == 1) it can be freed.
1232 * Otherwise, leave the page on the LRU so it is swappable.
1234 if (page_has_private(page)) {
1235 if (!try_to_release_page(page, sc->gfp_mask))
1236 goto activate_locked;
1237 if (!mapping && page_count(page) == 1) {
1239 if (put_page_testzero(page))
1243 * rare race with speculative reference.
1244 * the speculative reference will free
1245 * this page shortly, so we may
1246 * increment nr_reclaimed here (and
1247 * leave it off the LRU).
1256 if (!mapping || !__remove_mapping(mapping, page, true))
1260 * At this point, we have no other references and there is
1261 * no way to pick any more up (removed from LRU, removed
1262 * from pagecache). Can use non-atomic bitops now (and
1263 * we obviously don't have to worry about waking up a process
1264 * waiting on the page lock, because there are no references.
1266 __ClearPageLocked(page);
1268 if (ret == SWAP_LZFREE)
1269 count_vm_event(PGLAZYFREED);
1274 * Is there need to periodically free_page_list? It would
1275 * appear not as the counts should be low
1277 list_add(&page->lru, &free_pages);
1281 if (PageSwapCache(page))
1282 try_to_free_swap(page);
1284 list_add(&page->lru, &ret_pages);
1288 /* Not a candidate for swapping, so reclaim swap space. */
1289 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1290 try_to_free_swap(page);
1291 VM_BUG_ON_PAGE(PageActive(page), page);
1292 SetPageActive(page);
1297 list_add(&page->lru, &ret_pages);
1298 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1301 mem_cgroup_uncharge_list(&free_pages);
1302 try_to_unmap_flush();
1303 free_hot_cold_page_list(&free_pages, true);
1305 list_splice(&ret_pages, page_list);
1306 count_vm_events(PGACTIVATE, pgactivate);
1309 stat->nr_dirty = nr_dirty;
1310 stat->nr_congested = nr_congested;
1311 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1312 stat->nr_writeback = nr_writeback;
1313 stat->nr_immediate = nr_immediate;
1314 stat->nr_activate = pgactivate;
1315 stat->nr_ref_keep = nr_ref_keep;
1316 stat->nr_unmap_fail = nr_unmap_fail;
1318 return nr_reclaimed;
1321 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1322 struct list_head *page_list)
1324 struct scan_control sc = {
1325 .gfp_mask = GFP_KERNEL,
1326 .priority = DEF_PRIORITY,
1330 struct page *page, *next;
1331 LIST_HEAD(clean_pages);
1333 list_for_each_entry_safe(page, next, page_list, lru) {
1334 if (page_is_file_cache(page) && !PageDirty(page) &&
1335 !__PageMovable(page)) {
1336 ClearPageActive(page);
1337 list_move(&page->lru, &clean_pages);
1341 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1342 TTU_UNMAP|TTU_IGNORE_ACCESS, NULL, true);
1343 list_splice(&clean_pages, page_list);
1344 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1349 * Attempt to remove the specified page from its LRU. Only take this page
1350 * if it is of the appropriate PageActive status. Pages which are being
1351 * freed elsewhere are also ignored.
1353 * page: page to consider
1354 * mode: one of the LRU isolation modes defined above
1356 * returns 0 on success, -ve errno on failure.
1358 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1362 /* Only take pages on the LRU. */
1366 /* Compaction should not handle unevictable pages but CMA can do so */
1367 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1373 * To minimise LRU disruption, the caller can indicate that it only
1374 * wants to isolate pages it will be able to operate on without
1375 * blocking - clean pages for the most part.
1377 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1378 * that it is possible to migrate without blocking
1380 if (mode & ISOLATE_ASYNC_MIGRATE) {
1381 /* All the caller can do on PageWriteback is block */
1382 if (PageWriteback(page))
1385 if (PageDirty(page)) {
1386 struct address_space *mapping;
1389 * Only pages without mappings or that have a
1390 * ->migratepage callback are possible to migrate
1393 mapping = page_mapping(page);
1394 if (mapping && !mapping->a_ops->migratepage)
1399 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1402 if (likely(get_page_unless_zero(page))) {
1404 * Be careful not to clear PageLRU until after we're
1405 * sure the page is not being freed elsewhere -- the
1406 * page release code relies on it.
1417 * Update LRU sizes after isolating pages. The LRU size updates must
1418 * be complete before mem_cgroup_update_lru_size due to a santity check.
1420 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1421 enum lru_list lru, unsigned long *nr_zone_taken)
1425 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1426 if (!nr_zone_taken[zid])
1429 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1431 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1438 * zone_lru_lock is heavily contended. Some of the functions that
1439 * shrink the lists perform better by taking out a batch of pages
1440 * and working on them outside the LRU lock.
1442 * For pagecache intensive workloads, this function is the hottest
1443 * spot in the kernel (apart from copy_*_user functions).
1445 * Appropriate locks must be held before calling this function.
1447 * @nr_to_scan: The number of pages to look through on the list.
1448 * @lruvec: The LRU vector to pull pages from.
1449 * @dst: The temp list to put pages on to.
1450 * @nr_scanned: The number of pages that were scanned.
1451 * @sc: The scan_control struct for this reclaim session
1452 * @mode: One of the LRU isolation modes
1453 * @lru: LRU list id for isolating
1455 * returns how many pages were moved onto *@dst.
1457 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1458 struct lruvec *lruvec, struct list_head *dst,
1459 unsigned long *nr_scanned, struct scan_control *sc,
1460 isolate_mode_t mode, enum lru_list lru)
1462 struct list_head *src = &lruvec->lists[lru];
1463 unsigned long nr_taken = 0;
1464 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1465 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1466 unsigned long skipped = 0, total_skipped = 0;
1467 unsigned long scan, nr_pages;
1468 LIST_HEAD(pages_skipped);
1470 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1471 !list_empty(src);) {
1474 page = lru_to_page(src);
1475 prefetchw_prev_lru_page(page, src, flags);
1477 VM_BUG_ON_PAGE(!PageLRU(page), page);
1479 if (page_zonenum(page) > sc->reclaim_idx) {
1480 list_move(&page->lru, &pages_skipped);
1481 nr_skipped[page_zonenum(page)]++;
1486 * Account for scanned and skipped separetly to avoid the pgdat
1487 * being prematurely marked unreclaimable by pgdat_reclaimable.
1491 switch (__isolate_lru_page(page, mode)) {
1493 nr_pages = hpage_nr_pages(page);
1494 nr_taken += nr_pages;
1495 nr_zone_taken[page_zonenum(page)] += nr_pages;
1496 list_move(&page->lru, dst);
1500 /* else it is being freed elsewhere */
1501 list_move(&page->lru, src);
1510 * Splice any skipped pages to the start of the LRU list. Note that
1511 * this disrupts the LRU order when reclaiming for lower zones but
1512 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1513 * scanning would soon rescan the same pages to skip and put the
1514 * system at risk of premature OOM.
1516 if (!list_empty(&pages_skipped)) {
1519 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1520 if (!nr_skipped[zid])
1523 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1524 skipped += nr_skipped[zid];
1528 * Account skipped pages as a partial scan as the pgdat may be
1529 * close to unreclaimable. If the LRU list is empty, account
1530 * skipped pages as a full scan.
1532 total_skipped = list_empty(src) ? skipped : skipped >> 2;
1534 list_splice(&pages_skipped, src);
1536 *nr_scanned = scan + total_skipped;
1537 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1538 scan, skipped, nr_taken, mode, lru);
1539 update_lru_sizes(lruvec, lru, nr_zone_taken);
1544 * isolate_lru_page - tries to isolate a page from its LRU list
1545 * @page: page to isolate from its LRU list
1547 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1548 * vmstat statistic corresponding to whatever LRU list the page was on.
1550 * Returns 0 if the page was removed from an LRU list.
1551 * Returns -EBUSY if the page was not on an LRU list.
1553 * The returned page will have PageLRU() cleared. If it was found on
1554 * the active list, it will have PageActive set. If it was found on
1555 * the unevictable list, it will have the PageUnevictable bit set. That flag
1556 * may need to be cleared by the caller before letting the page go.
1558 * The vmstat statistic corresponding to the list on which the page was
1559 * found will be decremented.
1562 * (1) Must be called with an elevated refcount on the page. This is a
1563 * fundamentnal difference from isolate_lru_pages (which is called
1564 * without a stable reference).
1565 * (2) the lru_lock must not be held.
1566 * (3) interrupts must be enabled.
1568 int isolate_lru_page(struct page *page)
1572 VM_BUG_ON_PAGE(!page_count(page), page);
1573 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1575 if (PageLRU(page)) {
1576 struct zone *zone = page_zone(page);
1577 struct lruvec *lruvec;
1579 spin_lock_irq(zone_lru_lock(zone));
1580 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1581 if (PageLRU(page)) {
1582 int lru = page_lru(page);
1585 del_page_from_lru_list(page, lruvec, lru);
1588 spin_unlock_irq(zone_lru_lock(zone));
1594 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1595 * then get resheduled. When there are massive number of tasks doing page
1596 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1597 * the LRU list will go small and be scanned faster than necessary, leading to
1598 * unnecessary swapping, thrashing and OOM.
1600 static int too_many_isolated(struct pglist_data *pgdat, int file,
1601 struct scan_control *sc)
1603 unsigned long inactive, isolated;
1605 if (current_is_kswapd())
1608 if (!sane_reclaim(sc))
1612 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1613 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1615 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1616 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1620 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1621 * won't get blocked by normal direct-reclaimers, forming a circular
1624 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1627 return isolated > inactive;
1630 static noinline_for_stack void
1631 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1633 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1634 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1635 LIST_HEAD(pages_to_free);
1638 * Put back any unfreeable pages.
1640 while (!list_empty(page_list)) {
1641 struct page *page = lru_to_page(page_list);
1644 VM_BUG_ON_PAGE(PageLRU(page), page);
1645 list_del(&page->lru);
1646 if (unlikely(!page_evictable(page))) {
1647 spin_unlock_irq(&pgdat->lru_lock);
1648 putback_lru_page(page);
1649 spin_lock_irq(&pgdat->lru_lock);
1653 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1656 lru = page_lru(page);
1657 add_page_to_lru_list(page, lruvec, lru);
1659 if (is_active_lru(lru)) {
1660 int file = is_file_lru(lru);
1661 int numpages = hpage_nr_pages(page);
1662 reclaim_stat->recent_rotated[file] += numpages;
1664 if (put_page_testzero(page)) {
1665 __ClearPageLRU(page);
1666 __ClearPageActive(page);
1667 del_page_from_lru_list(page, lruvec, lru);
1669 if (unlikely(PageCompound(page))) {
1670 spin_unlock_irq(&pgdat->lru_lock);
1671 mem_cgroup_uncharge(page);
1672 (*get_compound_page_dtor(page))(page);
1673 spin_lock_irq(&pgdat->lru_lock);
1675 list_add(&page->lru, &pages_to_free);
1680 * To save our caller's stack, now use input list for pages to free.
1682 list_splice(&pages_to_free, page_list);
1686 * If a kernel thread (such as nfsd for loop-back mounts) services
1687 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1688 * In that case we should only throttle if the backing device it is
1689 * writing to is congested. In other cases it is safe to throttle.
1691 static int current_may_throttle(void)
1693 return !(current->flags & PF_LESS_THROTTLE) ||
1694 current->backing_dev_info == NULL ||
1695 bdi_write_congested(current->backing_dev_info);
1699 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1700 * of reclaimed pages
1702 static noinline_for_stack unsigned long
1703 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1704 struct scan_control *sc, enum lru_list lru)
1706 LIST_HEAD(page_list);
1707 unsigned long nr_scanned;
1708 unsigned long nr_reclaimed = 0;
1709 unsigned long nr_taken;
1710 struct reclaim_stat stat = {};
1711 isolate_mode_t isolate_mode = 0;
1712 int file = is_file_lru(lru);
1713 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1714 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1716 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1717 congestion_wait(BLK_RW_ASYNC, HZ/10);
1719 /* We are about to die and free our memory. Return now. */
1720 if (fatal_signal_pending(current))
1721 return SWAP_CLUSTER_MAX;
1727 isolate_mode |= ISOLATE_UNMAPPED;
1729 spin_lock_irq(&pgdat->lru_lock);
1731 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1732 &nr_scanned, sc, isolate_mode, lru);
1734 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1735 reclaim_stat->recent_scanned[file] += nr_taken;
1737 if (global_reclaim(sc)) {
1738 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1739 if (current_is_kswapd())
1740 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1742 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1744 spin_unlock_irq(&pgdat->lru_lock);
1749 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1752 spin_lock_irq(&pgdat->lru_lock);
1754 if (global_reclaim(sc)) {
1755 if (current_is_kswapd())
1756 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1758 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1761 putback_inactive_pages(lruvec, &page_list);
1763 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1765 spin_unlock_irq(&pgdat->lru_lock);
1767 mem_cgroup_uncharge_list(&page_list);
1768 free_hot_cold_page_list(&page_list, true);
1771 * If reclaim is isolating dirty pages under writeback, it implies
1772 * that the long-lived page allocation rate is exceeding the page
1773 * laundering rate. Either the global limits are not being effective
1774 * at throttling processes due to the page distribution throughout
1775 * zones or there is heavy usage of a slow backing device. The
1776 * only option is to throttle from reclaim context which is not ideal
1777 * as there is no guarantee the dirtying process is throttled in the
1778 * same way balance_dirty_pages() manages.
1780 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1781 * of pages under pages flagged for immediate reclaim and stall if any
1782 * are encountered in the nr_immediate check below.
1784 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1785 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1788 * Legacy memcg will stall in page writeback so avoid forcibly
1791 if (sane_reclaim(sc)) {
1793 * Tag a zone as congested if all the dirty pages scanned were
1794 * backed by a congested BDI and wait_iff_congested will stall.
1796 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1797 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1800 * If dirty pages are scanned that are not queued for IO, it
1801 * implies that flushers are not doing their job. This can
1802 * happen when memory pressure pushes dirty pages to the end of
1803 * the LRU before the dirty limits are breached and the dirty
1804 * data has expired. It can also happen when the proportion of
1805 * dirty pages grows not through writes but through memory
1806 * pressure reclaiming all the clean cache. And in some cases,
1807 * the flushers simply cannot keep up with the allocation
1808 * rate. Nudge the flusher threads in case they are asleep, but
1809 * also allow kswapd to start writing pages during reclaim.
1811 if (stat.nr_unqueued_dirty == nr_taken) {
1812 wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1813 set_bit(PGDAT_DIRTY, &pgdat->flags);
1817 * If kswapd scans pages marked marked for immediate
1818 * reclaim and under writeback (nr_immediate), it implies
1819 * that pages are cycling through the LRU faster than
1820 * they are written so also forcibly stall.
1822 if (stat.nr_immediate && current_may_throttle())
1823 congestion_wait(BLK_RW_ASYNC, HZ/10);
1827 * Stall direct reclaim for IO completions if underlying BDIs or zone
1828 * is congested. Allow kswapd to continue until it starts encountering
1829 * unqueued dirty pages or cycling through the LRU too quickly.
1831 if (!sc->hibernation_mode && !current_is_kswapd() &&
1832 current_may_throttle())
1833 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1835 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1836 nr_scanned, nr_reclaimed,
1837 stat.nr_dirty, stat.nr_writeback,
1838 stat.nr_congested, stat.nr_immediate,
1839 stat.nr_activate, stat.nr_ref_keep,
1841 sc->priority, file);
1842 return nr_reclaimed;
1846 * This moves pages from the active list to the inactive list.
1848 * We move them the other way if the page is referenced by one or more
1849 * processes, from rmap.
1851 * If the pages are mostly unmapped, the processing is fast and it is
1852 * appropriate to hold zone_lru_lock across the whole operation. But if
1853 * the pages are mapped, the processing is slow (page_referenced()) so we
1854 * should drop zone_lru_lock around each page. It's impossible to balance
1855 * this, so instead we remove the pages from the LRU while processing them.
1856 * It is safe to rely on PG_active against the non-LRU pages in here because
1857 * nobody will play with that bit on a non-LRU page.
1859 * The downside is that we have to touch page->_refcount against each page.
1860 * But we had to alter page->flags anyway.
1862 * Returns the number of pages moved to the given lru.
1865 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1866 struct list_head *list,
1867 struct list_head *pages_to_free,
1870 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1875 while (!list_empty(list)) {
1876 page = lru_to_page(list);
1877 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1879 VM_BUG_ON_PAGE(PageLRU(page), page);
1882 nr_pages = hpage_nr_pages(page);
1883 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1884 list_move(&page->lru, &lruvec->lists[lru]);
1886 if (put_page_testzero(page)) {
1887 __ClearPageLRU(page);
1888 __ClearPageActive(page);
1889 del_page_from_lru_list(page, lruvec, lru);
1891 if (unlikely(PageCompound(page))) {
1892 spin_unlock_irq(&pgdat->lru_lock);
1893 mem_cgroup_uncharge(page);
1894 (*get_compound_page_dtor(page))(page);
1895 spin_lock_irq(&pgdat->lru_lock);
1897 list_add(&page->lru, pages_to_free);
1899 nr_moved += nr_pages;
1903 if (!is_active_lru(lru))
1904 __count_vm_events(PGDEACTIVATE, nr_moved);
1909 static void shrink_active_list(unsigned long nr_to_scan,
1910 struct lruvec *lruvec,
1911 struct scan_control *sc,
1914 unsigned long nr_taken;
1915 unsigned long nr_scanned;
1916 unsigned long vm_flags;
1917 LIST_HEAD(l_hold); /* The pages which were snipped off */
1918 LIST_HEAD(l_active);
1919 LIST_HEAD(l_inactive);
1921 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1922 unsigned nr_deactivate, nr_activate;
1923 unsigned nr_rotated = 0;
1924 isolate_mode_t isolate_mode = 0;
1925 int file = is_file_lru(lru);
1926 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1931 isolate_mode |= ISOLATE_UNMAPPED;
1933 spin_lock_irq(&pgdat->lru_lock);
1935 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1936 &nr_scanned, sc, isolate_mode, lru);
1938 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1939 reclaim_stat->recent_scanned[file] += nr_taken;
1941 if (global_reclaim(sc))
1942 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1943 __count_vm_events(PGREFILL, nr_scanned);
1945 spin_unlock_irq(&pgdat->lru_lock);
1947 while (!list_empty(&l_hold)) {
1949 page = lru_to_page(&l_hold);
1950 list_del(&page->lru);
1952 if (unlikely(!page_evictable(page))) {
1953 putback_lru_page(page);
1957 if (unlikely(buffer_heads_over_limit)) {
1958 if (page_has_private(page) && trylock_page(page)) {
1959 if (page_has_private(page))
1960 try_to_release_page(page, 0);
1965 if (page_referenced(page, 0, sc->target_mem_cgroup,
1967 nr_rotated += hpage_nr_pages(page);
1969 * Identify referenced, file-backed active pages and
1970 * give them one more trip around the active list. So
1971 * that executable code get better chances to stay in
1972 * memory under moderate memory pressure. Anon pages
1973 * are not likely to be evicted by use-once streaming
1974 * IO, plus JVM can create lots of anon VM_EXEC pages,
1975 * so we ignore them here.
1977 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1978 list_add(&page->lru, &l_active);
1983 ClearPageActive(page); /* we are de-activating */
1984 list_add(&page->lru, &l_inactive);
1988 * Move pages back to the lru list.
1990 spin_lock_irq(&pgdat->lru_lock);
1992 * Count referenced pages from currently used mappings as rotated,
1993 * even though only some of them are actually re-activated. This
1994 * helps balance scan pressure between file and anonymous pages in
1997 reclaim_stat->recent_rotated[file] += nr_rotated;
1999 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2000 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2001 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2002 spin_unlock_irq(&pgdat->lru_lock);
2004 mem_cgroup_uncharge_list(&l_hold);
2005 free_hot_cold_page_list(&l_hold, true);
2006 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2007 nr_deactivate, nr_rotated, sc->priority, file);
2011 * The inactive anon list should be small enough that the VM never has
2012 * to do too much work.
2014 * The inactive file list should be small enough to leave most memory
2015 * to the established workingset on the scan-resistant active list,
2016 * but large enough to avoid thrashing the aggregate readahead window.
2018 * Both inactive lists should also be large enough that each inactive
2019 * page has a chance to be referenced again before it is reclaimed.
2021 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2022 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2023 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2026 * memory ratio inactive
2027 * -------------------------------------
2036 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2037 struct scan_control *sc, bool trace)
2039 unsigned long inactive_ratio;
2040 unsigned long inactive, active;
2041 enum lru_list inactive_lru = file * LRU_FILE;
2042 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2046 * If we don't have swap space, anonymous page deactivation
2049 if (!file && !total_swap_pages)
2052 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2053 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2055 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2057 inactive_ratio = int_sqrt(10 * gb);
2062 trace_mm_vmscan_inactive_list_is_low(lruvec_pgdat(lruvec)->node_id,
2064 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2065 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2066 inactive_ratio, file);
2068 return inactive * inactive_ratio < active;
2071 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2072 struct lruvec *lruvec, struct scan_control *sc)
2074 if (is_active_lru(lru)) {
2075 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2076 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2080 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2091 * Determine how aggressively the anon and file LRU lists should be
2092 * scanned. The relative value of each set of LRU lists is determined
2093 * by looking at the fraction of the pages scanned we did rotate back
2094 * onto the active list instead of evict.
2096 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2097 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2099 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2100 struct scan_control *sc, unsigned long *nr,
2101 unsigned long *lru_pages)
2103 int swappiness = mem_cgroup_swappiness(memcg);
2104 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2106 u64 denominator = 0; /* gcc */
2107 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2108 unsigned long anon_prio, file_prio;
2109 enum scan_balance scan_balance;
2110 unsigned long anon, file;
2111 bool force_scan = false;
2112 unsigned long ap, fp;
2118 * If the zone or memcg is small, nr[l] can be 0. This
2119 * results in no scanning on this priority and a potential
2120 * priority drop. Global direct reclaim can go to the next
2121 * zone and tends to have no problems. Global kswapd is for
2122 * zone balancing and it needs to scan a minimum amount. When
2123 * reclaiming for a memcg, a priority drop can cause high
2124 * latencies, so it's better to scan a minimum amount there as
2127 if (current_is_kswapd()) {
2128 if (!pgdat_reclaimable(pgdat))
2130 if (!mem_cgroup_online(memcg))
2133 if (!global_reclaim(sc))
2136 /* If we have no swap space, do not bother scanning anon pages. */
2137 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2138 scan_balance = SCAN_FILE;
2143 * Global reclaim will swap to prevent OOM even with no
2144 * swappiness, but memcg users want to use this knob to
2145 * disable swapping for individual groups completely when
2146 * using the memory controller's swap limit feature would be
2149 if (!global_reclaim(sc) && !swappiness) {
2150 scan_balance = SCAN_FILE;
2155 * Do not apply any pressure balancing cleverness when the
2156 * system is close to OOM, scan both anon and file equally
2157 * (unless the swappiness setting disagrees with swapping).
2159 if (!sc->priority && swappiness) {
2160 scan_balance = SCAN_EQUAL;
2165 * Prevent the reclaimer from falling into the cache trap: as
2166 * cache pages start out inactive, every cache fault will tip
2167 * the scan balance towards the file LRU. And as the file LRU
2168 * shrinks, so does the window for rotation from references.
2169 * This means we have a runaway feedback loop where a tiny
2170 * thrashing file LRU becomes infinitely more attractive than
2171 * anon pages. Try to detect this based on file LRU size.
2173 if (global_reclaim(sc)) {
2174 unsigned long pgdatfile;
2175 unsigned long pgdatfree;
2177 unsigned long total_high_wmark = 0;
2179 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2180 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2181 node_page_state(pgdat, NR_INACTIVE_FILE);
2183 for (z = 0; z < MAX_NR_ZONES; z++) {
2184 struct zone *zone = &pgdat->node_zones[z];
2185 if (!managed_zone(zone))
2188 total_high_wmark += high_wmark_pages(zone);
2191 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2192 scan_balance = SCAN_ANON;
2198 * If there is enough inactive page cache, i.e. if the size of the
2199 * inactive list is greater than that of the active list *and* the
2200 * inactive list actually has some pages to scan on this priority, we
2201 * do not reclaim anything from the anonymous working set right now.
2202 * Without the second condition we could end up never scanning an
2203 * lruvec even if it has plenty of old anonymous pages unless the
2204 * system is under heavy pressure.
2206 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2207 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2208 scan_balance = SCAN_FILE;
2212 scan_balance = SCAN_FRACT;
2215 * With swappiness at 100, anonymous and file have the same priority.
2216 * This scanning priority is essentially the inverse of IO cost.
2218 anon_prio = swappiness;
2219 file_prio = 200 - anon_prio;
2222 * OK, so we have swap space and a fair amount of page cache
2223 * pages. We use the recently rotated / recently scanned
2224 * ratios to determine how valuable each cache is.
2226 * Because workloads change over time (and to avoid overflow)
2227 * we keep these statistics as a floating average, which ends
2228 * up weighing recent references more than old ones.
2230 * anon in [0], file in [1]
2233 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2234 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2235 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2236 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2238 spin_lock_irq(&pgdat->lru_lock);
2239 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2240 reclaim_stat->recent_scanned[0] /= 2;
2241 reclaim_stat->recent_rotated[0] /= 2;
2244 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2245 reclaim_stat->recent_scanned[1] /= 2;
2246 reclaim_stat->recent_rotated[1] /= 2;
2250 * The amount of pressure on anon vs file pages is inversely
2251 * proportional to the fraction of recently scanned pages on
2252 * each list that were recently referenced and in active use.
2254 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2255 ap /= reclaim_stat->recent_rotated[0] + 1;
2257 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2258 fp /= reclaim_stat->recent_rotated[1] + 1;
2259 spin_unlock_irq(&pgdat->lru_lock);
2263 denominator = ap + fp + 1;
2265 some_scanned = false;
2266 /* Only use force_scan on second pass. */
2267 for (pass = 0; !some_scanned && pass < 2; pass++) {
2269 for_each_evictable_lru(lru) {
2270 int file = is_file_lru(lru);
2274 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2275 scan = size >> sc->priority;
2277 if (!scan && pass && force_scan)
2278 scan = min(size, SWAP_CLUSTER_MAX);
2280 switch (scan_balance) {
2282 /* Scan lists relative to size */
2286 * Scan types proportional to swappiness and
2287 * their relative recent reclaim efficiency.
2289 scan = div64_u64(scan * fraction[file],
2294 /* Scan one type exclusively */
2295 if ((scan_balance == SCAN_FILE) != file) {
2301 /* Look ma, no brain */
2309 * Skip the second pass and don't force_scan,
2310 * if we found something to scan.
2312 some_scanned |= !!scan;
2318 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2320 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2321 struct scan_control *sc, unsigned long *lru_pages)
2323 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2324 unsigned long nr[NR_LRU_LISTS];
2325 unsigned long targets[NR_LRU_LISTS];
2326 unsigned long nr_to_scan;
2328 unsigned long nr_reclaimed = 0;
2329 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2330 struct blk_plug plug;
2333 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2335 /* Record the original scan target for proportional adjustments later */
2336 memcpy(targets, nr, sizeof(nr));
2339 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2340 * event that can occur when there is little memory pressure e.g.
2341 * multiple streaming readers/writers. Hence, we do not abort scanning
2342 * when the requested number of pages are reclaimed when scanning at
2343 * DEF_PRIORITY on the assumption that the fact we are direct
2344 * reclaiming implies that kswapd is not keeping up and it is best to
2345 * do a batch of work at once. For memcg reclaim one check is made to
2346 * abort proportional reclaim if either the file or anon lru has already
2347 * dropped to zero at the first pass.
2349 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2350 sc->priority == DEF_PRIORITY);
2352 blk_start_plug(&plug);
2353 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2354 nr[LRU_INACTIVE_FILE]) {
2355 unsigned long nr_anon, nr_file, percentage;
2356 unsigned long nr_scanned;
2358 for_each_evictable_lru(lru) {
2360 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2361 nr[lru] -= nr_to_scan;
2363 nr_reclaimed += shrink_list(lru, nr_to_scan,
2370 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2374 * For kswapd and memcg, reclaim at least the number of pages
2375 * requested. Ensure that the anon and file LRUs are scanned
2376 * proportionally what was requested by get_scan_count(). We
2377 * stop reclaiming one LRU and reduce the amount scanning
2378 * proportional to the original scan target.
2380 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2381 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2384 * It's just vindictive to attack the larger once the smaller
2385 * has gone to zero. And given the way we stop scanning the
2386 * smaller below, this makes sure that we only make one nudge
2387 * towards proportionality once we've got nr_to_reclaim.
2389 if (!nr_file || !nr_anon)
2392 if (nr_file > nr_anon) {
2393 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2394 targets[LRU_ACTIVE_ANON] + 1;
2396 percentage = nr_anon * 100 / scan_target;
2398 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2399 targets[LRU_ACTIVE_FILE] + 1;
2401 percentage = nr_file * 100 / scan_target;
2404 /* Stop scanning the smaller of the LRU */
2406 nr[lru + LRU_ACTIVE] = 0;
2409 * Recalculate the other LRU scan count based on its original
2410 * scan target and the percentage scanning already complete
2412 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2413 nr_scanned = targets[lru] - nr[lru];
2414 nr[lru] = targets[lru] * (100 - percentage) / 100;
2415 nr[lru] -= min(nr[lru], nr_scanned);
2418 nr_scanned = targets[lru] - nr[lru];
2419 nr[lru] = targets[lru] * (100 - percentage) / 100;
2420 nr[lru] -= min(nr[lru], nr_scanned);
2422 scan_adjusted = true;
2424 blk_finish_plug(&plug);
2425 sc->nr_reclaimed += nr_reclaimed;
2428 * Even if we did not try to evict anon pages at all, we want to
2429 * rebalance the anon lru active/inactive ratio.
2431 if (inactive_list_is_low(lruvec, false, sc, true))
2432 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2433 sc, LRU_ACTIVE_ANON);
2436 /* Use reclaim/compaction for costly allocs or under memory pressure */
2437 static bool in_reclaim_compaction(struct scan_control *sc)
2439 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2440 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2441 sc->priority < DEF_PRIORITY - 2))
2448 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2449 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2450 * true if more pages should be reclaimed such that when the page allocator
2451 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2452 * It will give up earlier than that if there is difficulty reclaiming pages.
2454 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2455 unsigned long nr_reclaimed,
2456 unsigned long nr_scanned,
2457 struct scan_control *sc)
2459 unsigned long pages_for_compaction;
2460 unsigned long inactive_lru_pages;
2463 /* If not in reclaim/compaction mode, stop */
2464 if (!in_reclaim_compaction(sc))
2467 /* Consider stopping depending on scan and reclaim activity */
2468 if (sc->gfp_mask & __GFP_REPEAT) {
2470 * For __GFP_REPEAT allocations, stop reclaiming if the
2471 * full LRU list has been scanned and we are still failing
2472 * to reclaim pages. This full LRU scan is potentially
2473 * expensive but a __GFP_REPEAT caller really wants to succeed
2475 if (!nr_reclaimed && !nr_scanned)
2479 * For non-__GFP_REPEAT allocations which can presumably
2480 * fail without consequence, stop if we failed to reclaim
2481 * any pages from the last SWAP_CLUSTER_MAX number of
2482 * pages that were scanned. This will return to the
2483 * caller faster at the risk reclaim/compaction and
2484 * the resulting allocation attempt fails
2491 * If we have not reclaimed enough pages for compaction and the
2492 * inactive lists are large enough, continue reclaiming
2494 pages_for_compaction = compact_gap(sc->order);
2495 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2496 if (get_nr_swap_pages() > 0)
2497 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2498 if (sc->nr_reclaimed < pages_for_compaction &&
2499 inactive_lru_pages > pages_for_compaction)
2502 /* If compaction would go ahead or the allocation would succeed, stop */
2503 for (z = 0; z <= sc->reclaim_idx; z++) {
2504 struct zone *zone = &pgdat->node_zones[z];
2505 if (!managed_zone(zone))
2508 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2509 case COMPACT_SUCCESS:
2510 case COMPACT_CONTINUE:
2513 /* check next zone */
2520 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2522 struct reclaim_state *reclaim_state = current->reclaim_state;
2523 unsigned long nr_reclaimed, nr_scanned;
2524 bool reclaimable = false;
2527 struct mem_cgroup *root = sc->target_mem_cgroup;
2528 struct mem_cgroup_reclaim_cookie reclaim = {
2530 .priority = sc->priority,
2532 unsigned long node_lru_pages = 0;
2533 struct mem_cgroup *memcg;
2535 nr_reclaimed = sc->nr_reclaimed;
2536 nr_scanned = sc->nr_scanned;
2538 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2540 unsigned long lru_pages;
2541 unsigned long reclaimed;
2542 unsigned long scanned;
2544 if (mem_cgroup_low(root, memcg)) {
2545 if (!sc->may_thrash)
2547 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2550 reclaimed = sc->nr_reclaimed;
2551 scanned = sc->nr_scanned;
2553 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2554 node_lru_pages += lru_pages;
2557 shrink_slab(sc->gfp_mask, pgdat->node_id,
2558 memcg, sc->nr_scanned - scanned,
2561 /* Record the group's reclaim efficiency */
2562 vmpressure(sc->gfp_mask, memcg, false,
2563 sc->nr_scanned - scanned,
2564 sc->nr_reclaimed - reclaimed);
2567 * Direct reclaim and kswapd have to scan all memory
2568 * cgroups to fulfill the overall scan target for the
2571 * Limit reclaim, on the other hand, only cares about
2572 * nr_to_reclaim pages to be reclaimed and it will
2573 * retry with decreasing priority if one round over the
2574 * whole hierarchy is not sufficient.
2576 if (!global_reclaim(sc) &&
2577 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2578 mem_cgroup_iter_break(root, memcg);
2581 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2584 * Shrink the slab caches in the same proportion that
2585 * the eligible LRU pages were scanned.
2587 if (global_reclaim(sc))
2588 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2589 sc->nr_scanned - nr_scanned,
2592 if (reclaim_state) {
2593 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2594 reclaim_state->reclaimed_slab = 0;
2597 /* Record the subtree's reclaim efficiency */
2598 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2599 sc->nr_scanned - nr_scanned,
2600 sc->nr_reclaimed - nr_reclaimed);
2602 if (sc->nr_reclaimed - nr_reclaimed)
2605 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2606 sc->nr_scanned - nr_scanned, sc));
2612 * Returns true if compaction should go ahead for a costly-order request, or
2613 * the allocation would already succeed without compaction. Return false if we
2614 * should reclaim first.
2616 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2618 unsigned long watermark;
2619 enum compact_result suitable;
2621 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2622 if (suitable == COMPACT_SUCCESS)
2623 /* Allocation should succeed already. Don't reclaim. */
2625 if (suitable == COMPACT_SKIPPED)
2626 /* Compaction cannot yet proceed. Do reclaim. */
2630 * Compaction is already possible, but it takes time to run and there
2631 * are potentially other callers using the pages just freed. So proceed
2632 * with reclaim to make a buffer of free pages available to give
2633 * compaction a reasonable chance of completing and allocating the page.
2634 * Note that we won't actually reclaim the whole buffer in one attempt
2635 * as the target watermark in should_continue_reclaim() is lower. But if
2636 * we are already above the high+gap watermark, don't reclaim at all.
2638 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2640 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2644 * This is the direct reclaim path, for page-allocating processes. We only
2645 * try to reclaim pages from zones which will satisfy the caller's allocation
2648 * If a zone is deemed to be full of pinned pages then just give it a light
2649 * scan then give up on it.
2651 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2655 unsigned long nr_soft_reclaimed;
2656 unsigned long nr_soft_scanned;
2658 pg_data_t *last_pgdat = NULL;
2661 * If the number of buffer_heads in the machine exceeds the maximum
2662 * allowed level, force direct reclaim to scan the highmem zone as
2663 * highmem pages could be pinning lowmem pages storing buffer_heads
2665 orig_mask = sc->gfp_mask;
2666 if (buffer_heads_over_limit) {
2667 sc->gfp_mask |= __GFP_HIGHMEM;
2668 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2671 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2672 sc->reclaim_idx, sc->nodemask) {
2674 * Take care memory controller reclaiming has small influence
2677 if (global_reclaim(sc)) {
2678 if (!cpuset_zone_allowed(zone,
2679 GFP_KERNEL | __GFP_HARDWALL))
2682 if (sc->priority != DEF_PRIORITY &&
2683 !pgdat_reclaimable(zone->zone_pgdat))
2684 continue; /* Let kswapd poll it */
2687 * If we already have plenty of memory free for
2688 * compaction in this zone, don't free any more.
2689 * Even though compaction is invoked for any
2690 * non-zero order, only frequent costly order
2691 * reclamation is disruptive enough to become a
2692 * noticeable problem, like transparent huge
2695 if (IS_ENABLED(CONFIG_COMPACTION) &&
2696 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2697 compaction_ready(zone, sc)) {
2698 sc->compaction_ready = true;
2703 * Shrink each node in the zonelist once. If the
2704 * zonelist is ordered by zone (not the default) then a
2705 * node may be shrunk multiple times but in that case
2706 * the user prefers lower zones being preserved.
2708 if (zone->zone_pgdat == last_pgdat)
2712 * This steals pages from memory cgroups over softlimit
2713 * and returns the number of reclaimed pages and
2714 * scanned pages. This works for global memory pressure
2715 * and balancing, not for a memcg's limit.
2717 nr_soft_scanned = 0;
2718 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2719 sc->order, sc->gfp_mask,
2721 sc->nr_reclaimed += nr_soft_reclaimed;
2722 sc->nr_scanned += nr_soft_scanned;
2723 /* need some check for avoid more shrink_zone() */
2726 /* See comment about same check for global reclaim above */
2727 if (zone->zone_pgdat == last_pgdat)
2729 last_pgdat = zone->zone_pgdat;
2730 shrink_node(zone->zone_pgdat, sc);
2734 * Restore to original mask to avoid the impact on the caller if we
2735 * promoted it to __GFP_HIGHMEM.
2737 sc->gfp_mask = orig_mask;
2741 * This is the main entry point to direct page reclaim.
2743 * If a full scan of the inactive list fails to free enough memory then we
2744 * are "out of memory" and something needs to be killed.
2746 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2747 * high - the zone may be full of dirty or under-writeback pages, which this
2748 * caller can't do much about. We kick the writeback threads and take explicit
2749 * naps in the hope that some of these pages can be written. But if the
2750 * allocating task holds filesystem locks which prevent writeout this might not
2751 * work, and the allocation attempt will fail.
2753 * returns: 0, if no pages reclaimed
2754 * else, the number of pages reclaimed
2756 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2757 struct scan_control *sc)
2759 int initial_priority = sc->priority;
2760 unsigned long total_scanned = 0;
2761 unsigned long writeback_threshold;
2763 delayacct_freepages_start();
2765 if (global_reclaim(sc))
2766 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2769 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2772 shrink_zones(zonelist, sc);
2774 total_scanned += sc->nr_scanned;
2775 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2778 if (sc->compaction_ready)
2782 * If we're getting trouble reclaiming, start doing
2783 * writepage even in laptop mode.
2785 if (sc->priority < DEF_PRIORITY - 2)
2786 sc->may_writepage = 1;
2789 * Try to write back as many pages as we just scanned. This
2790 * tends to cause slow streaming writers to write data to the
2791 * disk smoothly, at the dirtying rate, which is nice. But
2792 * that's undesirable in laptop mode, where we *want* lumpy
2793 * writeout. So in laptop mode, write out the whole world.
2795 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2796 if (total_scanned > writeback_threshold) {
2797 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2799 sc->may_writepage = 1;
2801 } while (--sc->priority >= 0);
2803 delayacct_freepages_end();
2805 if (sc->nr_reclaimed)
2806 return sc->nr_reclaimed;
2808 /* Aborted reclaim to try compaction? don't OOM, then */
2809 if (sc->compaction_ready)
2812 /* Untapped cgroup reserves? Don't OOM, retry. */
2813 if (!sc->may_thrash) {
2814 sc->priority = initial_priority;
2822 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2825 unsigned long pfmemalloc_reserve = 0;
2826 unsigned long free_pages = 0;
2830 for (i = 0; i <= ZONE_NORMAL; i++) {
2831 zone = &pgdat->node_zones[i];
2832 if (!managed_zone(zone) ||
2833 pgdat_reclaimable_pages(pgdat) == 0)
2836 pfmemalloc_reserve += min_wmark_pages(zone);
2837 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2840 /* If there are no reserves (unexpected config) then do not throttle */
2841 if (!pfmemalloc_reserve)
2844 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2846 /* kswapd must be awake if processes are being throttled */
2847 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2848 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2849 (enum zone_type)ZONE_NORMAL);
2850 wake_up_interruptible(&pgdat->kswapd_wait);
2857 * Throttle direct reclaimers if backing storage is backed by the network
2858 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2859 * depleted. kswapd will continue to make progress and wake the processes
2860 * when the low watermark is reached.
2862 * Returns true if a fatal signal was delivered during throttling. If this
2863 * happens, the page allocator should not consider triggering the OOM killer.
2865 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2866 nodemask_t *nodemask)
2870 pg_data_t *pgdat = NULL;
2873 * Kernel threads should not be throttled as they may be indirectly
2874 * responsible for cleaning pages necessary for reclaim to make forward
2875 * progress. kjournald for example may enter direct reclaim while
2876 * committing a transaction where throttling it could forcing other
2877 * processes to block on log_wait_commit().
2879 if (current->flags & PF_KTHREAD)
2883 * If a fatal signal is pending, this process should not throttle.
2884 * It should return quickly so it can exit and free its memory
2886 if (fatal_signal_pending(current))
2890 * Check if the pfmemalloc reserves are ok by finding the first node
2891 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2892 * GFP_KERNEL will be required for allocating network buffers when
2893 * swapping over the network so ZONE_HIGHMEM is unusable.
2895 * Throttling is based on the first usable node and throttled processes
2896 * wait on a queue until kswapd makes progress and wakes them. There
2897 * is an affinity then between processes waking up and where reclaim
2898 * progress has been made assuming the process wakes on the same node.
2899 * More importantly, processes running on remote nodes will not compete
2900 * for remote pfmemalloc reserves and processes on different nodes
2901 * should make reasonable progress.
2903 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2904 gfp_zone(gfp_mask), nodemask) {
2905 if (zone_idx(zone) > ZONE_NORMAL)
2908 /* Throttle based on the first usable node */
2909 pgdat = zone->zone_pgdat;
2910 if (pfmemalloc_watermark_ok(pgdat))
2915 /* If no zone was usable by the allocation flags then do not throttle */
2919 /* Account for the throttling */
2920 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2923 * If the caller cannot enter the filesystem, it's possible that it
2924 * is due to the caller holding an FS lock or performing a journal
2925 * transaction in the case of a filesystem like ext[3|4]. In this case,
2926 * it is not safe to block on pfmemalloc_wait as kswapd could be
2927 * blocked waiting on the same lock. Instead, throttle for up to a
2928 * second before continuing.
2930 if (!(gfp_mask & __GFP_FS)) {
2931 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2932 pfmemalloc_watermark_ok(pgdat), HZ);
2937 /* Throttle until kswapd wakes the process */
2938 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2939 pfmemalloc_watermark_ok(pgdat));
2942 if (fatal_signal_pending(current))
2949 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2950 gfp_t gfp_mask, nodemask_t *nodemask)
2952 unsigned long nr_reclaimed;
2953 struct scan_control sc = {
2954 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2955 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2956 .reclaim_idx = gfp_zone(gfp_mask),
2958 .nodemask = nodemask,
2959 .priority = DEF_PRIORITY,
2960 .may_writepage = !laptop_mode,
2966 * Do not enter reclaim if fatal signal was delivered while throttled.
2967 * 1 is returned so that the page allocator does not OOM kill at this
2970 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2973 trace_mm_vmscan_direct_reclaim_begin(order,
2978 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2980 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2982 return nr_reclaimed;
2987 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2988 gfp_t gfp_mask, bool noswap,
2990 unsigned long *nr_scanned)
2992 struct scan_control sc = {
2993 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2994 .target_mem_cgroup = memcg,
2995 .may_writepage = !laptop_mode,
2997 .reclaim_idx = MAX_NR_ZONES - 1,
2998 .may_swap = !noswap,
3000 unsigned long lru_pages;
3002 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3003 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3005 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3011 * NOTE: Although we can get the priority field, using it
3012 * here is not a good idea, since it limits the pages we can scan.
3013 * if we don't reclaim here, the shrink_node from balance_pgdat
3014 * will pick up pages from other mem cgroup's as well. We hack
3015 * the priority and make it zero.
3017 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3019 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3021 *nr_scanned = sc.nr_scanned;
3022 return sc.nr_reclaimed;
3025 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3026 unsigned long nr_pages,
3030 struct zonelist *zonelist;
3031 unsigned long nr_reclaimed;
3033 struct scan_control sc = {
3034 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3035 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3036 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3037 .reclaim_idx = MAX_NR_ZONES - 1,
3038 .target_mem_cgroup = memcg,
3039 .priority = DEF_PRIORITY,
3040 .may_writepage = !laptop_mode,
3042 .may_swap = may_swap,
3046 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3047 * take care of from where we get pages. So the node where we start the
3048 * scan does not need to be the current node.
3050 nid = mem_cgroup_select_victim_node(memcg);
3052 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3054 trace_mm_vmscan_memcg_reclaim_begin(0,
3059 current->flags |= PF_MEMALLOC;
3060 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3061 current->flags &= ~PF_MEMALLOC;
3063 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3065 return nr_reclaimed;
3069 static void age_active_anon(struct pglist_data *pgdat,
3070 struct scan_control *sc)
3072 struct mem_cgroup *memcg;
3074 if (!total_swap_pages)
3077 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3079 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3081 if (inactive_list_is_low(lruvec, false, sc, true))
3082 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3083 sc, LRU_ACTIVE_ANON);
3085 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3089 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3091 unsigned long mark = high_wmark_pages(zone);
3093 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3097 * If any eligible zone is balanced then the node is not considered
3098 * to be congested or dirty
3100 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3101 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3107 * Prepare kswapd for sleeping. This verifies that there are no processes
3108 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3110 * Returns true if kswapd is ready to sleep
3112 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3117 * The throttled processes are normally woken up in balance_pgdat() as
3118 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3119 * race between when kswapd checks the watermarks and a process gets
3120 * throttled. There is also a potential race if processes get
3121 * throttled, kswapd wakes, a large process exits thereby balancing the
3122 * zones, which causes kswapd to exit balance_pgdat() before reaching
3123 * the wake up checks. If kswapd is going to sleep, no process should
3124 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3125 * the wake up is premature, processes will wake kswapd and get
3126 * throttled again. The difference from wake ups in balance_pgdat() is
3127 * that here we are under prepare_to_wait().
3129 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3130 wake_up_all(&pgdat->pfmemalloc_wait);
3132 for (i = 0; i <= classzone_idx; i++) {
3133 struct zone *zone = pgdat->node_zones + i;
3135 if (!managed_zone(zone))
3138 if (!zone_balanced(zone, order, classzone_idx))
3146 * kswapd shrinks a node of pages that are at or below the highest usable
3147 * zone that is currently unbalanced.
3149 * Returns true if kswapd scanned at least the requested number of pages to
3150 * reclaim or if the lack of progress was due to pages under writeback.
3151 * This is used to determine if the scanning priority needs to be raised.
3153 static bool kswapd_shrink_node(pg_data_t *pgdat,
3154 struct scan_control *sc)
3159 /* Reclaim a number of pages proportional to the number of zones */
3160 sc->nr_to_reclaim = 0;
3161 for (z = 0; z <= sc->reclaim_idx; z++) {
3162 zone = pgdat->node_zones + z;
3163 if (!managed_zone(zone))
3166 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3170 * Historically care was taken to put equal pressure on all zones but
3171 * now pressure is applied based on node LRU order.
3173 shrink_node(pgdat, sc);
3176 * Fragmentation may mean that the system cannot be rebalanced for
3177 * high-order allocations. If twice the allocation size has been
3178 * reclaimed then recheck watermarks only at order-0 to prevent
3179 * excessive reclaim. Assume that a process requested a high-order
3180 * can direct reclaim/compact.
3182 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3185 return sc->nr_scanned >= sc->nr_to_reclaim;
3189 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3190 * that are eligible for use by the caller until at least one zone is
3193 * Returns the order kswapd finished reclaiming at.
3195 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3196 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3197 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3198 * or lower is eligible for reclaim until at least one usable zone is
3201 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3204 unsigned long nr_soft_reclaimed;
3205 unsigned long nr_soft_scanned;
3207 struct scan_control sc = {
3208 .gfp_mask = GFP_KERNEL,
3210 .priority = DEF_PRIORITY,
3211 .may_writepage = !laptop_mode,
3215 count_vm_event(PAGEOUTRUN);
3218 bool raise_priority = true;
3220 sc.nr_reclaimed = 0;
3221 sc.reclaim_idx = classzone_idx;
3224 * If the number of buffer_heads exceeds the maximum allowed
3225 * then consider reclaiming from all zones. This has a dual
3226 * purpose -- on 64-bit systems it is expected that
3227 * buffer_heads are stripped during active rotation. On 32-bit
3228 * systems, highmem pages can pin lowmem memory and shrinking
3229 * buffers can relieve lowmem pressure. Reclaim may still not
3230 * go ahead if all eligible zones for the original allocation
3231 * request are balanced to avoid excessive reclaim from kswapd.
3233 if (buffer_heads_over_limit) {
3234 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3235 zone = pgdat->node_zones + i;
3236 if (!managed_zone(zone))
3245 * Only reclaim if there are no eligible zones. Check from
3246 * high to low zone as allocations prefer higher zones.
3247 * Scanning from low to high zone would allow congestion to be
3248 * cleared during a very small window when a small low
3249 * zone was balanced even under extreme pressure when the
3250 * overall node may be congested. Note that sc.reclaim_idx
3251 * is not used as buffer_heads_over_limit may have adjusted
3254 for (i = classzone_idx; i >= 0; i--) {
3255 zone = pgdat->node_zones + i;
3256 if (!managed_zone(zone))
3259 if (zone_balanced(zone, sc.order, classzone_idx))
3264 * Do some background aging of the anon list, to give
3265 * pages a chance to be referenced before reclaiming. All
3266 * pages are rotated regardless of classzone as this is
3267 * about consistent aging.
3269 age_active_anon(pgdat, &sc);
3272 * If we're getting trouble reclaiming, start doing writepage
3273 * even in laptop mode.
3275 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3276 sc.may_writepage = 1;
3278 /* Call soft limit reclaim before calling shrink_node. */
3280 nr_soft_scanned = 0;
3281 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3282 sc.gfp_mask, &nr_soft_scanned);
3283 sc.nr_reclaimed += nr_soft_reclaimed;
3286 * There should be no need to raise the scanning priority if
3287 * enough pages are already being scanned that that high
3288 * watermark would be met at 100% efficiency.
3290 if (kswapd_shrink_node(pgdat, &sc))
3291 raise_priority = false;
3294 * If the low watermark is met there is no need for processes
3295 * to be throttled on pfmemalloc_wait as they should not be
3296 * able to safely make forward progress. Wake them
3298 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3299 pfmemalloc_watermark_ok(pgdat))
3300 wake_up_all(&pgdat->pfmemalloc_wait);
3302 /* Check if kswapd should be suspending */
3303 if (try_to_freeze() || kthread_should_stop())
3307 * Raise priority if scanning rate is too low or there was no
3308 * progress in reclaiming pages
3310 if (raise_priority || !sc.nr_reclaimed)
3312 } while (sc.priority >= 1);
3316 * Return the order kswapd stopped reclaiming at as
3317 * prepare_kswapd_sleep() takes it into account. If another caller
3318 * entered the allocator slow path while kswapd was awake, order will
3319 * remain at the higher level.
3324 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3325 unsigned int classzone_idx)
3330 if (freezing(current) || kthread_should_stop())
3333 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3335 /* Try to sleep for a short interval */
3336 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3338 * Compaction records what page blocks it recently failed to
3339 * isolate pages from and skips them in the future scanning.
3340 * When kswapd is going to sleep, it is reasonable to assume
3341 * that pages and compaction may succeed so reset the cache.
3343 reset_isolation_suitable(pgdat);
3346 * We have freed the memory, now we should compact it to make
3347 * allocation of the requested order possible.
3349 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3351 remaining = schedule_timeout(HZ/10);
3354 * If woken prematurely then reset kswapd_classzone_idx and
3355 * order. The values will either be from a wakeup request or
3356 * the previous request that slept prematurely.
3359 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3360 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3363 finish_wait(&pgdat->kswapd_wait, &wait);
3364 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3368 * After a short sleep, check if it was a premature sleep. If not, then
3369 * go fully to sleep until explicitly woken up.
3372 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3373 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3376 * vmstat counters are not perfectly accurate and the estimated
3377 * value for counters such as NR_FREE_PAGES can deviate from the
3378 * true value by nr_online_cpus * threshold. To avoid the zone
3379 * watermarks being breached while under pressure, we reduce the
3380 * per-cpu vmstat threshold while kswapd is awake and restore
3381 * them before going back to sleep.
3383 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3385 if (!kthread_should_stop())
3388 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3391 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3393 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3395 finish_wait(&pgdat->kswapd_wait, &wait);
3399 * The background pageout daemon, started as a kernel thread
3400 * from the init process.
3402 * This basically trickles out pages so that we have _some_
3403 * free memory available even if there is no other activity
3404 * that frees anything up. This is needed for things like routing
3405 * etc, where we otherwise might have all activity going on in
3406 * asynchronous contexts that cannot page things out.
3408 * If there are applications that are active memory-allocators
3409 * (most normal use), this basically shouldn't matter.
3411 static int kswapd(void *p)
3413 unsigned int alloc_order, reclaim_order, classzone_idx;
3414 pg_data_t *pgdat = (pg_data_t*)p;
3415 struct task_struct *tsk = current;
3417 struct reclaim_state reclaim_state = {
3418 .reclaimed_slab = 0,
3420 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3422 lockdep_set_current_reclaim_state(GFP_KERNEL);
3424 if (!cpumask_empty(cpumask))
3425 set_cpus_allowed_ptr(tsk, cpumask);
3426 current->reclaim_state = &reclaim_state;
3429 * Tell the memory management that we're a "memory allocator",
3430 * and that if we need more memory we should get access to it
3431 * regardless (see "__alloc_pages()"). "kswapd" should
3432 * never get caught in the normal page freeing logic.
3434 * (Kswapd normally doesn't need memory anyway, but sometimes
3435 * you need a small amount of memory in order to be able to
3436 * page out something else, and this flag essentially protects
3437 * us from recursively trying to free more memory as we're
3438 * trying to free the first piece of memory in the first place).
3440 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3443 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3444 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3449 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3452 /* Read the new order and classzone_idx */
3453 alloc_order = reclaim_order = pgdat->kswapd_order;
3454 classzone_idx = pgdat->kswapd_classzone_idx;
3455 pgdat->kswapd_order = 0;
3456 pgdat->kswapd_classzone_idx = 0;
3458 ret = try_to_freeze();
3459 if (kthread_should_stop())
3463 * We can speed up thawing tasks if we don't call balance_pgdat
3464 * after returning from the refrigerator
3470 * Reclaim begins at the requested order but if a high-order
3471 * reclaim fails then kswapd falls back to reclaiming for
3472 * order-0. If that happens, kswapd will consider sleeping
3473 * for the order it finished reclaiming at (reclaim_order)
3474 * but kcompactd is woken to compact for the original
3475 * request (alloc_order).
3477 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3479 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3480 if (reclaim_order < alloc_order)
3481 goto kswapd_try_sleep;
3483 alloc_order = reclaim_order = pgdat->kswapd_order;
3484 classzone_idx = pgdat->kswapd_classzone_idx;
3487 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3488 current->reclaim_state = NULL;
3489 lockdep_clear_current_reclaim_state();
3495 * A zone is low on free memory, so wake its kswapd task to service it.
3497 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3502 if (!managed_zone(zone))
3505 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3507 pgdat = zone->zone_pgdat;
3508 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3509 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3510 if (!waitqueue_active(&pgdat->kswapd_wait))
3513 /* Only wake kswapd if all zones are unbalanced */
3514 for (z = 0; z <= classzone_idx; z++) {
3515 zone = pgdat->node_zones + z;
3516 if (!managed_zone(zone))
3519 if (zone_balanced(zone, order, classzone_idx))
3523 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3524 wake_up_interruptible(&pgdat->kswapd_wait);
3527 #ifdef CONFIG_HIBERNATION
3529 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3532 * Rather than trying to age LRUs the aim is to preserve the overall
3533 * LRU order by reclaiming preferentially
3534 * inactive > active > active referenced > active mapped
3536 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3538 struct reclaim_state reclaim_state;
3539 struct scan_control sc = {
3540 .nr_to_reclaim = nr_to_reclaim,
3541 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3542 .reclaim_idx = MAX_NR_ZONES - 1,
3543 .priority = DEF_PRIORITY,
3547 .hibernation_mode = 1,
3549 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3550 struct task_struct *p = current;
3551 unsigned long nr_reclaimed;
3553 p->flags |= PF_MEMALLOC;
3554 lockdep_set_current_reclaim_state(sc.gfp_mask);
3555 reclaim_state.reclaimed_slab = 0;
3556 p->reclaim_state = &reclaim_state;
3558 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3560 p->reclaim_state = NULL;
3561 lockdep_clear_current_reclaim_state();
3562 p->flags &= ~PF_MEMALLOC;
3564 return nr_reclaimed;
3566 #endif /* CONFIG_HIBERNATION */
3568 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3569 not required for correctness. So if the last cpu in a node goes
3570 away, we get changed to run anywhere: as the first one comes back,
3571 restore their cpu bindings. */
3572 static int kswapd_cpu_online(unsigned int cpu)
3576 for_each_node_state(nid, N_MEMORY) {
3577 pg_data_t *pgdat = NODE_DATA(nid);
3578 const struct cpumask *mask;
3580 mask = cpumask_of_node(pgdat->node_id);
3582 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3583 /* One of our CPUs online: restore mask */
3584 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3590 * This kswapd start function will be called by init and node-hot-add.
3591 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3593 int kswapd_run(int nid)
3595 pg_data_t *pgdat = NODE_DATA(nid);
3601 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3602 if (IS_ERR(pgdat->kswapd)) {
3603 /* failure at boot is fatal */
3604 BUG_ON(system_state == SYSTEM_BOOTING);
3605 pr_err("Failed to start kswapd on node %d\n", nid);
3606 ret = PTR_ERR(pgdat->kswapd);
3607 pgdat->kswapd = NULL;
3613 * Called by memory hotplug when all memory in a node is offlined. Caller must
3614 * hold mem_hotplug_begin/end().
3616 void kswapd_stop(int nid)
3618 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3621 kthread_stop(kswapd);
3622 NODE_DATA(nid)->kswapd = NULL;
3626 static int __init kswapd_init(void)
3631 for_each_node_state(nid, N_MEMORY)
3633 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3634 "mm/vmscan:online", kswapd_cpu_online,
3640 module_init(kswapd_init)
3646 * If non-zero call node_reclaim when the number of free pages falls below
3649 int node_reclaim_mode __read_mostly;
3651 #define RECLAIM_OFF 0
3652 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3653 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3654 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3657 * Priority for NODE_RECLAIM. This determines the fraction of pages
3658 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3661 #define NODE_RECLAIM_PRIORITY 4
3664 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3667 int sysctl_min_unmapped_ratio = 1;
3670 * If the number of slab pages in a zone grows beyond this percentage then
3671 * slab reclaim needs to occur.
3673 int sysctl_min_slab_ratio = 5;
3675 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3677 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3678 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3679 node_page_state(pgdat, NR_ACTIVE_FILE);
3682 * It's possible for there to be more file mapped pages than
3683 * accounted for by the pages on the file LRU lists because
3684 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3686 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3689 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3690 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3692 unsigned long nr_pagecache_reclaimable;
3693 unsigned long delta = 0;
3696 * If RECLAIM_UNMAP is set, then all file pages are considered
3697 * potentially reclaimable. Otherwise, we have to worry about
3698 * pages like swapcache and node_unmapped_file_pages() provides
3701 if (node_reclaim_mode & RECLAIM_UNMAP)
3702 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3704 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3706 /* If we can't clean pages, remove dirty pages from consideration */
3707 if (!(node_reclaim_mode & RECLAIM_WRITE))
3708 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3710 /* Watch for any possible underflows due to delta */
3711 if (unlikely(delta > nr_pagecache_reclaimable))
3712 delta = nr_pagecache_reclaimable;
3714 return nr_pagecache_reclaimable - delta;
3718 * Try to free up some pages from this node through reclaim.
3720 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3722 /* Minimum pages needed in order to stay on node */
3723 const unsigned long nr_pages = 1 << order;
3724 struct task_struct *p = current;
3725 struct reclaim_state reclaim_state;
3726 int classzone_idx = gfp_zone(gfp_mask);
3727 struct scan_control sc = {
3728 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3729 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3731 .priority = NODE_RECLAIM_PRIORITY,
3732 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3733 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3735 .reclaim_idx = classzone_idx,
3740 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3741 * and we also need to be able to write out pages for RECLAIM_WRITE
3742 * and RECLAIM_UNMAP.
3744 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3745 lockdep_set_current_reclaim_state(gfp_mask);
3746 reclaim_state.reclaimed_slab = 0;
3747 p->reclaim_state = &reclaim_state;
3749 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3751 * Free memory by calling shrink zone with increasing
3752 * priorities until we have enough memory freed.
3755 shrink_node(pgdat, &sc);
3756 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3759 p->reclaim_state = NULL;
3760 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3761 lockdep_clear_current_reclaim_state();
3762 return sc.nr_reclaimed >= nr_pages;
3765 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3770 * Node reclaim reclaims unmapped file backed pages and
3771 * slab pages if we are over the defined limits.
3773 * A small portion of unmapped file backed pages is needed for
3774 * file I/O otherwise pages read by file I/O will be immediately
3775 * thrown out if the node is overallocated. So we do not reclaim
3776 * if less than a specified percentage of the node is used by
3777 * unmapped file backed pages.
3779 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3780 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3781 return NODE_RECLAIM_FULL;
3783 if (!pgdat_reclaimable(pgdat))
3784 return NODE_RECLAIM_FULL;
3787 * Do not scan if the allocation should not be delayed.
3789 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3790 return NODE_RECLAIM_NOSCAN;
3793 * Only run node reclaim on the local node or on nodes that do not
3794 * have associated processors. This will favor the local processor
3795 * over remote processors and spread off node memory allocations
3796 * as wide as possible.
3798 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3799 return NODE_RECLAIM_NOSCAN;
3801 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3802 return NODE_RECLAIM_NOSCAN;
3804 ret = __node_reclaim(pgdat, gfp_mask, order);
3805 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3808 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3815 * page_evictable - test whether a page is evictable
3816 * @page: the page to test
3818 * Test whether page is evictable--i.e., should be placed on active/inactive
3819 * lists vs unevictable list.
3821 * Reasons page might not be evictable:
3822 * (1) page's mapping marked unevictable
3823 * (2) page is part of an mlocked VMA
3826 int page_evictable(struct page *page)
3828 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3833 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3834 * @pages: array of pages to check
3835 * @nr_pages: number of pages to check
3837 * Checks pages for evictability and moves them to the appropriate lru list.
3839 * This function is only used for SysV IPC SHM_UNLOCK.
3841 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3843 struct lruvec *lruvec;
3844 struct pglist_data *pgdat = NULL;
3849 for (i = 0; i < nr_pages; i++) {
3850 struct page *page = pages[i];
3851 struct pglist_data *pagepgdat = page_pgdat(page);
3854 if (pagepgdat != pgdat) {
3856 spin_unlock_irq(&pgdat->lru_lock);
3858 spin_lock_irq(&pgdat->lru_lock);
3860 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3862 if (!PageLRU(page) || !PageUnevictable(page))
3865 if (page_evictable(page)) {
3866 enum lru_list lru = page_lru_base_type(page);
3868 VM_BUG_ON_PAGE(PageActive(page), page);
3869 ClearPageUnevictable(page);
3870 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3871 add_page_to_lru_list(page, lruvec, lru);
3877 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3878 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3879 spin_unlock_irq(&pgdat->lru_lock);
3882 #endif /* CONFIG_SHMEM */