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/sched/mm.h>
18 #include <linux/module.h>
19 #include <linux/gfp.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/highmem.h>
25 #include <linux/vmpressure.h>
26 #include <linux/vmstat.h>
27 #include <linux/file.h>
28 #include <linux/writeback.h>
29 #include <linux/blkdev.h>
30 #include <linux/buffer_head.h> /* for try_to_release_page(),
31 buffer_heads_over_limit */
32 #include <linux/mm_inline.h>
33 #include <linux/backing-dev.h>
34 #include <linux/rmap.h>
35 #include <linux/topology.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/compaction.h>
39 #include <linux/notifier.h>
40 #include <linux/rwsem.h>
41 #include <linux/delay.h>
42 #include <linux/kthread.h>
43 #include <linux/freezer.h>
44 #include <linux/memcontrol.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
55 #include <linux/swapops.h>
56 #include <linux/balloon_compaction.h>
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/vmscan.h>
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim;
67 /* This context's GFP mask */
70 /* Allocation order */
74 * Nodemask of nodes allowed by the caller. If NULL, all nodes
80 * The memory cgroup that hit its limit and as a result is the
81 * primary target of this reclaim invocation.
83 struct mem_cgroup *target_mem_cgroup;
85 /* Scan (total_size >> priority) pages at once */
88 /* The highest zone to isolate pages for reclaim from */
89 enum zone_type reclaim_idx;
91 /* Writepage batching in laptop mode; RECLAIM_WRITE */
92 unsigned int may_writepage:1;
94 /* Can mapped pages be reclaimed? */
95 unsigned int may_unmap:1;
97 /* Can pages be swapped as part of reclaim? */
98 unsigned int may_swap:1;
101 * Cgroups are not reclaimed below their configured memory.low,
102 * unless we threaten to OOM. If any cgroups are skipped due to
103 * memory.low and nothing was reclaimed, go back for memory.low.
105 unsigned int memcg_low_reclaim:1;
106 unsigned int memcg_low_skipped:1;
108 unsigned int hibernation_mode:1;
110 /* One of the zones is ready for compaction */
111 unsigned int compaction_ready:1;
113 /* Incremented by the number of inactive pages that were scanned */
114 unsigned long nr_scanned;
116 /* Number of pages freed so far during a call to shrink_zones() */
117 unsigned long nr_reclaimed;
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness = 60;
153 * The total number of pages which are beyond the high watermark within all
156 unsigned long vm_total_pages;
158 static LIST_HEAD(shrinker_list);
159 static DECLARE_RWSEM(shrinker_rwsem);
162 static bool global_reclaim(struct scan_control *sc)
164 return !sc->target_mem_cgroup;
168 * sane_reclaim - is the usual dirty throttling mechanism operational?
169 * @sc: scan_control in question
171 * The normal page dirty throttling mechanism in balance_dirty_pages() is
172 * completely broken with the legacy memcg and direct stalling in
173 * shrink_page_list() is used for throttling instead, which lacks all the
174 * niceties such as fairness, adaptive pausing, bandwidth proportional
175 * allocation and configurability.
177 * This function tests whether the vmscan currently in progress can assume
178 * that the normal dirty throttling mechanism is operational.
180 static bool sane_reclaim(struct scan_control *sc)
182 struct mem_cgroup *memcg = sc->target_mem_cgroup;
186 #ifdef CONFIG_CGROUP_WRITEBACK
187 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
193 static bool global_reclaim(struct scan_control *sc)
198 static bool sane_reclaim(struct scan_control *sc)
205 * This misses isolated pages which are not accounted for to save counters.
206 * As the data only determines if reclaim or compaction continues, it is
207 * not expected that isolated pages will be a dominating factor.
209 unsigned long zone_reclaimable_pages(struct zone *zone)
213 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
214 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
215 if (get_nr_swap_pages() > 0)
216 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
217 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
222 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
226 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
227 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
228 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
230 if (get_nr_swap_pages() > 0)
231 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
232 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
233 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
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) ||
915 (PageAnon(page) && !PageSwapBacked(page))) {
921 /* By default assume that the page flags are accurate */
922 *dirty = PageDirty(page);
923 *writeback = PageWriteback(page);
925 /* Verify dirty/writeback state if the filesystem supports it */
926 if (!page_has_private(page))
929 mapping = page_mapping(page);
930 if (mapping && mapping->a_ops->is_dirty_writeback)
931 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
934 struct reclaim_stat {
936 unsigned nr_unqueued_dirty;
937 unsigned nr_congested;
938 unsigned nr_writeback;
939 unsigned nr_immediate;
940 unsigned nr_activate;
941 unsigned nr_ref_keep;
942 unsigned nr_unmap_fail;
946 * shrink_page_list() returns the number of reclaimed pages
948 static unsigned long shrink_page_list(struct list_head *page_list,
949 struct pglist_data *pgdat,
950 struct scan_control *sc,
951 enum ttu_flags ttu_flags,
952 struct reclaim_stat *stat,
955 LIST_HEAD(ret_pages);
956 LIST_HEAD(free_pages);
958 unsigned nr_unqueued_dirty = 0;
959 unsigned nr_dirty = 0;
960 unsigned nr_congested = 0;
961 unsigned nr_reclaimed = 0;
962 unsigned nr_writeback = 0;
963 unsigned nr_immediate = 0;
964 unsigned nr_ref_keep = 0;
965 unsigned nr_unmap_fail = 0;
969 while (!list_empty(page_list)) {
970 struct address_space *mapping;
973 enum page_references references = PAGEREF_RECLAIM_CLEAN;
974 bool dirty, writeback;
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)) &&
997 !(PageAnon(page) && !PageSwapBacked(page)))
1000 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1001 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1004 * The number of dirty pages determines if a zone is marked
1005 * reclaim_congested which affects wait_iff_congested. kswapd
1006 * will stall and start writing pages if the tail of the LRU
1007 * is all dirty unqueued pages.
1009 page_check_dirty_writeback(page, &dirty, &writeback);
1010 if (dirty || writeback)
1013 if (dirty && !writeback)
1014 nr_unqueued_dirty++;
1017 * Treat this page as congested if the underlying BDI is or if
1018 * pages are cycling through the LRU so quickly that the
1019 * pages marked for immediate reclaim are making it to the
1020 * end of the LRU a second time.
1022 mapping = page_mapping(page);
1023 if (((dirty || writeback) && mapping &&
1024 inode_write_congested(mapping->host)) ||
1025 (writeback && PageReclaim(page)))
1029 * If a page at the tail of the LRU is under writeback, there
1030 * are three cases to consider.
1032 * 1) If reclaim is encountering an excessive number of pages
1033 * under writeback and this page is both under writeback and
1034 * PageReclaim then it indicates that pages are being queued
1035 * for IO but are being recycled through the LRU before the
1036 * IO can complete. Waiting on the page itself risks an
1037 * indefinite stall if it is impossible to writeback the
1038 * page due to IO error or disconnected storage so instead
1039 * note that the LRU is being scanned too quickly and the
1040 * caller can stall after page list has been processed.
1042 * 2) Global or new memcg reclaim encounters a page that is
1043 * not marked for immediate reclaim, or the caller does not
1044 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1045 * not to fs). In this case mark the page for immediate
1046 * reclaim and continue scanning.
1048 * Require may_enter_fs because we would wait on fs, which
1049 * may not have submitted IO yet. And the loop driver might
1050 * enter reclaim, and deadlock if it waits on a page for
1051 * which it is needed to do the write (loop masks off
1052 * __GFP_IO|__GFP_FS for this reason); but more thought
1053 * would probably show more reasons.
1055 * 3) Legacy memcg encounters a page that is already marked
1056 * PageReclaim. memcg does not have any dirty pages
1057 * throttling so we could easily OOM just because too many
1058 * pages are in writeback and there is nothing else to
1059 * reclaim. Wait for the writeback to complete.
1061 * In cases 1) and 2) we activate the pages to get them out of
1062 * the way while we continue scanning for clean pages on the
1063 * inactive list and refilling from the active list. The
1064 * observation here is that waiting for disk writes is more
1065 * expensive than potentially causing reloads down the line.
1066 * Since they're marked for immediate reclaim, they won't put
1067 * memory pressure on the cache working set any longer than it
1068 * takes to write them to disk.
1070 if (PageWriteback(page)) {
1072 if (current_is_kswapd() &&
1073 PageReclaim(page) &&
1074 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1076 goto activate_locked;
1079 } else if (sane_reclaim(sc) ||
1080 !PageReclaim(page) || !may_enter_fs) {
1082 * This is slightly racy - end_page_writeback()
1083 * might have just cleared PageReclaim, then
1084 * setting PageReclaim here end up interpreted
1085 * as PageReadahead - but that does not matter
1086 * enough to care. What we do want is for this
1087 * page to have PageReclaim set next time memcg
1088 * reclaim reaches the tests above, so it will
1089 * then wait_on_page_writeback() to avoid OOM;
1090 * and it's also appropriate in global reclaim.
1092 SetPageReclaim(page);
1094 goto activate_locked;
1099 wait_on_page_writeback(page);
1100 /* then go back and try same page again */
1101 list_add_tail(&page->lru, page_list);
1107 references = page_check_references(page, sc);
1109 switch (references) {
1110 case PAGEREF_ACTIVATE:
1111 goto activate_locked;
1115 case PAGEREF_RECLAIM:
1116 case PAGEREF_RECLAIM_CLEAN:
1117 ; /* try to reclaim the page below */
1121 * Anonymous process memory has backing store?
1122 * Try to allocate it some swap space here.
1123 * Lazyfree page could be freed directly
1125 if (PageAnon(page) && PageSwapBacked(page) &&
1126 !PageSwapCache(page)) {
1127 if (!(sc->gfp_mask & __GFP_IO))
1129 if (!add_to_swap(page, page_list))
1130 goto activate_locked;
1133 /* Adding to swap updated mapping */
1134 mapping = page_mapping(page);
1135 } else if (unlikely(PageTransHuge(page))) {
1136 /* Split file THP */
1137 if (split_huge_page_to_list(page, page_list))
1141 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1144 * The page is mapped into the page tables of one or more
1145 * processes. Try to unmap it here.
1147 if (page_mapped(page)) {
1148 switch (ret = try_to_unmap(page,
1149 ttu_flags | TTU_BATCH_FLUSH)) {
1151 SetPageSwapBacked(page);
1155 goto activate_locked;
1161 ; /* try to free the page below */
1165 if (PageDirty(page)) {
1167 * Only kswapd can writeback filesystem pages
1168 * to avoid risk of stack overflow. But avoid
1169 * injecting inefficient single-page IO into
1170 * flusher writeback as much as possible: only
1171 * write pages when we've encountered many
1172 * dirty pages, and when we've already scanned
1173 * the rest of the LRU for clean pages and see
1174 * the same dirty pages again (PageReclaim).
1176 if (page_is_file_cache(page) &&
1177 (!current_is_kswapd() || !PageReclaim(page) ||
1178 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1180 * Immediately reclaim when written back.
1181 * Similar in principal to deactivate_page()
1182 * except we already have the page isolated
1183 * and know it's dirty
1185 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1186 SetPageReclaim(page);
1188 goto activate_locked;
1191 if (references == PAGEREF_RECLAIM_CLEAN)
1195 if (!sc->may_writepage)
1199 * Page is dirty. Flush the TLB if a writable entry
1200 * potentially exists to avoid CPU writes after IO
1201 * starts and then write it out here.
1203 try_to_unmap_flush_dirty();
1204 switch (pageout(page, mapping, sc)) {
1208 goto activate_locked;
1210 if (PageWriteback(page))
1212 if (PageDirty(page))
1216 * A synchronous write - probably a ramdisk. Go
1217 * ahead and try to reclaim the page.
1219 if (!trylock_page(page))
1221 if (PageDirty(page) || PageWriteback(page))
1223 mapping = page_mapping(page);
1225 ; /* try to free the page below */
1230 * If the page has buffers, try to free the buffer mappings
1231 * associated with this page. If we succeed we try to free
1234 * We do this even if the page is PageDirty().
1235 * try_to_release_page() does not perform I/O, but it is
1236 * possible for a page to have PageDirty set, but it is actually
1237 * clean (all its buffers are clean). This happens if the
1238 * buffers were written out directly, with submit_bh(). ext3
1239 * will do this, as well as the blockdev mapping.
1240 * try_to_release_page() will discover that cleanness and will
1241 * drop the buffers and mark the page clean - it can be freed.
1243 * Rarely, pages can have buffers and no ->mapping. These are
1244 * the pages which were not successfully invalidated in
1245 * truncate_complete_page(). We try to drop those buffers here
1246 * and if that worked, and the page is no longer mapped into
1247 * process address space (page_count == 1) it can be freed.
1248 * Otherwise, leave the page on the LRU so it is swappable.
1250 if (page_has_private(page)) {
1251 if (!try_to_release_page(page, sc->gfp_mask))
1252 goto activate_locked;
1253 if (!mapping && page_count(page) == 1) {
1255 if (put_page_testzero(page))
1259 * rare race with speculative reference.
1260 * the speculative reference will free
1261 * this page shortly, so we may
1262 * increment nr_reclaimed here (and
1263 * leave it off the LRU).
1271 if (PageAnon(page) && !PageSwapBacked(page)) {
1272 /* follow __remove_mapping for reference */
1273 if (!page_ref_freeze(page, 1))
1275 if (PageDirty(page)) {
1276 page_ref_unfreeze(page, 1);
1280 count_vm_event(PGLAZYFREED);
1281 } else if (!mapping || !__remove_mapping(mapping, page, true))
1284 * At this point, we have no other references and there is
1285 * no way to pick any more up (removed from LRU, removed
1286 * from pagecache). Can use non-atomic bitops now (and
1287 * we obviously don't have to worry about waking up a process
1288 * waiting on the page lock, because there are no references.
1290 __ClearPageLocked(page);
1295 * Is there need to periodically free_page_list? It would
1296 * appear not as the counts should be low
1298 list_add(&page->lru, &free_pages);
1302 if (PageSwapCache(page))
1303 try_to_free_swap(page);
1305 list_add(&page->lru, &ret_pages);
1309 /* Not a candidate for swapping, so reclaim swap space. */
1310 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1311 try_to_free_swap(page);
1312 VM_BUG_ON_PAGE(PageActive(page), page);
1313 SetPageActive(page);
1318 list_add(&page->lru, &ret_pages);
1319 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1322 mem_cgroup_uncharge_list(&free_pages);
1323 try_to_unmap_flush();
1324 free_hot_cold_page_list(&free_pages, true);
1326 list_splice(&ret_pages, page_list);
1327 count_vm_events(PGACTIVATE, pgactivate);
1330 stat->nr_dirty = nr_dirty;
1331 stat->nr_congested = nr_congested;
1332 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1333 stat->nr_writeback = nr_writeback;
1334 stat->nr_immediate = nr_immediate;
1335 stat->nr_activate = pgactivate;
1336 stat->nr_ref_keep = nr_ref_keep;
1337 stat->nr_unmap_fail = nr_unmap_fail;
1339 return nr_reclaimed;
1342 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1343 struct list_head *page_list)
1345 struct scan_control sc = {
1346 .gfp_mask = GFP_KERNEL,
1347 .priority = DEF_PRIORITY,
1351 struct page *page, *next;
1352 LIST_HEAD(clean_pages);
1354 list_for_each_entry_safe(page, next, page_list, lru) {
1355 if (page_is_file_cache(page) && !PageDirty(page) &&
1356 !__PageMovable(page)) {
1357 ClearPageActive(page);
1358 list_move(&page->lru, &clean_pages);
1362 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1363 TTU_IGNORE_ACCESS, NULL, true);
1364 list_splice(&clean_pages, page_list);
1365 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1370 * Attempt to remove the specified page from its LRU. Only take this page
1371 * if it is of the appropriate PageActive status. Pages which are being
1372 * freed elsewhere are also ignored.
1374 * page: page to consider
1375 * mode: one of the LRU isolation modes defined above
1377 * returns 0 on success, -ve errno on failure.
1379 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1383 /* Only take pages on the LRU. */
1387 /* Compaction should not handle unevictable pages but CMA can do so */
1388 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1394 * To minimise LRU disruption, the caller can indicate that it only
1395 * wants to isolate pages it will be able to operate on without
1396 * blocking - clean pages for the most part.
1398 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1399 * that it is possible to migrate without blocking
1401 if (mode & ISOLATE_ASYNC_MIGRATE) {
1402 /* All the caller can do on PageWriteback is block */
1403 if (PageWriteback(page))
1406 if (PageDirty(page)) {
1407 struct address_space *mapping;
1410 * Only pages without mappings or that have a
1411 * ->migratepage callback are possible to migrate
1414 mapping = page_mapping(page);
1415 if (mapping && !mapping->a_ops->migratepage)
1420 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1423 if (likely(get_page_unless_zero(page))) {
1425 * Be careful not to clear PageLRU until after we're
1426 * sure the page is not being freed elsewhere -- the
1427 * page release code relies on it.
1438 * Update LRU sizes after isolating pages. The LRU size updates must
1439 * be complete before mem_cgroup_update_lru_size due to a santity check.
1441 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1442 enum lru_list lru, unsigned long *nr_zone_taken)
1446 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1447 if (!nr_zone_taken[zid])
1450 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1452 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1459 * zone_lru_lock is heavily contended. Some of the functions that
1460 * shrink the lists perform better by taking out a batch of pages
1461 * and working on them outside the LRU lock.
1463 * For pagecache intensive workloads, this function is the hottest
1464 * spot in the kernel (apart from copy_*_user functions).
1466 * Appropriate locks must be held before calling this function.
1468 * @nr_to_scan: The number of pages to look through on the list.
1469 * @lruvec: The LRU vector to pull pages from.
1470 * @dst: The temp list to put pages on to.
1471 * @nr_scanned: The number of pages that were scanned.
1472 * @sc: The scan_control struct for this reclaim session
1473 * @mode: One of the LRU isolation modes
1474 * @lru: LRU list id for isolating
1476 * returns how many pages were moved onto *@dst.
1478 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1479 struct lruvec *lruvec, struct list_head *dst,
1480 unsigned long *nr_scanned, struct scan_control *sc,
1481 isolate_mode_t mode, enum lru_list lru)
1483 struct list_head *src = &lruvec->lists[lru];
1484 unsigned long nr_taken = 0;
1485 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1486 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1487 unsigned long skipped = 0;
1488 unsigned long scan, nr_pages;
1489 LIST_HEAD(pages_skipped);
1491 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1492 !list_empty(src); scan++) {
1495 page = lru_to_page(src);
1496 prefetchw_prev_lru_page(page, src, flags);
1498 VM_BUG_ON_PAGE(!PageLRU(page), page);
1500 if (page_zonenum(page) > sc->reclaim_idx) {
1501 list_move(&page->lru, &pages_skipped);
1502 nr_skipped[page_zonenum(page)]++;
1506 switch (__isolate_lru_page(page, mode)) {
1508 nr_pages = hpage_nr_pages(page);
1509 nr_taken += nr_pages;
1510 nr_zone_taken[page_zonenum(page)] += nr_pages;
1511 list_move(&page->lru, dst);
1515 /* else it is being freed elsewhere */
1516 list_move(&page->lru, src);
1525 * Splice any skipped pages to the start of the LRU list. Note that
1526 * this disrupts the LRU order when reclaiming for lower zones but
1527 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1528 * scanning would soon rescan the same pages to skip and put the
1529 * system at risk of premature OOM.
1531 if (!list_empty(&pages_skipped)) {
1534 list_splice(&pages_skipped, src);
1535 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1536 if (!nr_skipped[zid])
1539 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1540 skipped += nr_skipped[zid];
1544 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1545 scan, skipped, nr_taken, mode, lru);
1546 update_lru_sizes(lruvec, lru, nr_zone_taken);
1551 * isolate_lru_page - tries to isolate a page from its LRU list
1552 * @page: page to isolate from its LRU list
1554 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1555 * vmstat statistic corresponding to whatever LRU list the page was on.
1557 * Returns 0 if the page was removed from an LRU list.
1558 * Returns -EBUSY if the page was not on an LRU list.
1560 * The returned page will have PageLRU() cleared. If it was found on
1561 * the active list, it will have PageActive set. If it was found on
1562 * the unevictable list, it will have the PageUnevictable bit set. That flag
1563 * may need to be cleared by the caller before letting the page go.
1565 * The vmstat statistic corresponding to the list on which the page was
1566 * found will be decremented.
1569 * (1) Must be called with an elevated refcount on the page. This is a
1570 * fundamentnal difference from isolate_lru_pages (which is called
1571 * without a stable reference).
1572 * (2) the lru_lock must not be held.
1573 * (3) interrupts must be enabled.
1575 int isolate_lru_page(struct page *page)
1579 VM_BUG_ON_PAGE(!page_count(page), page);
1580 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1582 if (PageLRU(page)) {
1583 struct zone *zone = page_zone(page);
1584 struct lruvec *lruvec;
1586 spin_lock_irq(zone_lru_lock(zone));
1587 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1588 if (PageLRU(page)) {
1589 int lru = page_lru(page);
1592 del_page_from_lru_list(page, lruvec, lru);
1595 spin_unlock_irq(zone_lru_lock(zone));
1601 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1602 * then get resheduled. When there are massive number of tasks doing page
1603 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1604 * the LRU list will go small and be scanned faster than necessary, leading to
1605 * unnecessary swapping, thrashing and OOM.
1607 static int too_many_isolated(struct pglist_data *pgdat, int file,
1608 struct scan_control *sc)
1610 unsigned long inactive, isolated;
1612 if (current_is_kswapd())
1615 if (!sane_reclaim(sc))
1619 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1620 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1622 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1623 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1627 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1628 * won't get blocked by normal direct-reclaimers, forming a circular
1631 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1634 return isolated > inactive;
1637 static noinline_for_stack void
1638 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1640 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1641 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1642 LIST_HEAD(pages_to_free);
1645 * Put back any unfreeable pages.
1647 while (!list_empty(page_list)) {
1648 struct page *page = lru_to_page(page_list);
1651 VM_BUG_ON_PAGE(PageLRU(page), page);
1652 list_del(&page->lru);
1653 if (unlikely(!page_evictable(page))) {
1654 spin_unlock_irq(&pgdat->lru_lock);
1655 putback_lru_page(page);
1656 spin_lock_irq(&pgdat->lru_lock);
1660 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1663 lru = page_lru(page);
1664 add_page_to_lru_list(page, lruvec, lru);
1666 if (is_active_lru(lru)) {
1667 int file = is_file_lru(lru);
1668 int numpages = hpage_nr_pages(page);
1669 reclaim_stat->recent_rotated[file] += numpages;
1671 if (put_page_testzero(page)) {
1672 __ClearPageLRU(page);
1673 __ClearPageActive(page);
1674 del_page_from_lru_list(page, lruvec, lru);
1676 if (unlikely(PageCompound(page))) {
1677 spin_unlock_irq(&pgdat->lru_lock);
1678 mem_cgroup_uncharge(page);
1679 (*get_compound_page_dtor(page))(page);
1680 spin_lock_irq(&pgdat->lru_lock);
1682 list_add(&page->lru, &pages_to_free);
1687 * To save our caller's stack, now use input list for pages to free.
1689 list_splice(&pages_to_free, page_list);
1693 * If a kernel thread (such as nfsd for loop-back mounts) services
1694 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1695 * In that case we should only throttle if the backing device it is
1696 * writing to is congested. In other cases it is safe to throttle.
1698 static int current_may_throttle(void)
1700 return !(current->flags & PF_LESS_THROTTLE) ||
1701 current->backing_dev_info == NULL ||
1702 bdi_write_congested(current->backing_dev_info);
1706 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1707 * of reclaimed pages
1709 static noinline_for_stack unsigned long
1710 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1711 struct scan_control *sc, enum lru_list lru)
1713 LIST_HEAD(page_list);
1714 unsigned long nr_scanned;
1715 unsigned long nr_reclaimed = 0;
1716 unsigned long nr_taken;
1717 struct reclaim_stat stat = {};
1718 isolate_mode_t isolate_mode = 0;
1719 int file = is_file_lru(lru);
1720 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1721 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1723 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1724 congestion_wait(BLK_RW_ASYNC, HZ/10);
1726 /* We are about to die and free our memory. Return now. */
1727 if (fatal_signal_pending(current))
1728 return SWAP_CLUSTER_MAX;
1734 isolate_mode |= ISOLATE_UNMAPPED;
1736 spin_lock_irq(&pgdat->lru_lock);
1738 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1739 &nr_scanned, sc, isolate_mode, lru);
1741 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1742 reclaim_stat->recent_scanned[file] += nr_taken;
1744 if (global_reclaim(sc)) {
1745 if (current_is_kswapd())
1746 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1748 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1750 spin_unlock_irq(&pgdat->lru_lock);
1755 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1758 spin_lock_irq(&pgdat->lru_lock);
1760 if (global_reclaim(sc)) {
1761 if (current_is_kswapd())
1762 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1764 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1767 putback_inactive_pages(lruvec, &page_list);
1769 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1771 spin_unlock_irq(&pgdat->lru_lock);
1773 mem_cgroup_uncharge_list(&page_list);
1774 free_hot_cold_page_list(&page_list, true);
1777 * If reclaim is isolating dirty pages under writeback, it implies
1778 * that the long-lived page allocation rate is exceeding the page
1779 * laundering rate. Either the global limits are not being effective
1780 * at throttling processes due to the page distribution throughout
1781 * zones or there is heavy usage of a slow backing device. The
1782 * only option is to throttle from reclaim context which is not ideal
1783 * as there is no guarantee the dirtying process is throttled in the
1784 * same way balance_dirty_pages() manages.
1786 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1787 * of pages under pages flagged for immediate reclaim and stall if any
1788 * are encountered in the nr_immediate check below.
1790 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1791 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1794 * Legacy memcg will stall in page writeback so avoid forcibly
1797 if (sane_reclaim(sc)) {
1799 * Tag a zone as congested if all the dirty pages scanned were
1800 * backed by a congested BDI and wait_iff_congested will stall.
1802 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1803 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1806 * If dirty pages are scanned that are not queued for IO, it
1807 * implies that flushers are not doing their job. This can
1808 * happen when memory pressure pushes dirty pages to the end of
1809 * the LRU before the dirty limits are breached and the dirty
1810 * data has expired. It can also happen when the proportion of
1811 * dirty pages grows not through writes but through memory
1812 * pressure reclaiming all the clean cache. And in some cases,
1813 * the flushers simply cannot keep up with the allocation
1814 * rate. Nudge the flusher threads in case they are asleep, but
1815 * also allow kswapd to start writing pages during reclaim.
1817 if (stat.nr_unqueued_dirty == nr_taken) {
1818 wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1819 set_bit(PGDAT_DIRTY, &pgdat->flags);
1823 * If kswapd scans pages marked marked for immediate
1824 * reclaim and under writeback (nr_immediate), it implies
1825 * that pages are cycling through the LRU faster than
1826 * they are written so also forcibly stall.
1828 if (stat.nr_immediate && current_may_throttle())
1829 congestion_wait(BLK_RW_ASYNC, HZ/10);
1833 * Stall direct reclaim for IO completions if underlying BDIs or zone
1834 * is congested. Allow kswapd to continue until it starts encountering
1835 * unqueued dirty pages or cycling through the LRU too quickly.
1837 if (!sc->hibernation_mode && !current_is_kswapd() &&
1838 current_may_throttle())
1839 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1841 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1842 nr_scanned, nr_reclaimed,
1843 stat.nr_dirty, stat.nr_writeback,
1844 stat.nr_congested, stat.nr_immediate,
1845 stat.nr_activate, stat.nr_ref_keep,
1847 sc->priority, file);
1848 return nr_reclaimed;
1852 * This moves pages from the active list to the inactive list.
1854 * We move them the other way if the page is referenced by one or more
1855 * processes, from rmap.
1857 * If the pages are mostly unmapped, the processing is fast and it is
1858 * appropriate to hold zone_lru_lock across the whole operation. But if
1859 * the pages are mapped, the processing is slow (page_referenced()) so we
1860 * should drop zone_lru_lock around each page. It's impossible to balance
1861 * this, so instead we remove the pages from the LRU while processing them.
1862 * It is safe to rely on PG_active against the non-LRU pages in here because
1863 * nobody will play with that bit on a non-LRU page.
1865 * The downside is that we have to touch page->_refcount against each page.
1866 * But we had to alter page->flags anyway.
1868 * Returns the number of pages moved to the given lru.
1871 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1872 struct list_head *list,
1873 struct list_head *pages_to_free,
1876 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1881 while (!list_empty(list)) {
1882 page = lru_to_page(list);
1883 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1885 VM_BUG_ON_PAGE(PageLRU(page), page);
1888 nr_pages = hpage_nr_pages(page);
1889 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1890 list_move(&page->lru, &lruvec->lists[lru]);
1892 if (put_page_testzero(page)) {
1893 __ClearPageLRU(page);
1894 __ClearPageActive(page);
1895 del_page_from_lru_list(page, lruvec, lru);
1897 if (unlikely(PageCompound(page))) {
1898 spin_unlock_irq(&pgdat->lru_lock);
1899 mem_cgroup_uncharge(page);
1900 (*get_compound_page_dtor(page))(page);
1901 spin_lock_irq(&pgdat->lru_lock);
1903 list_add(&page->lru, pages_to_free);
1905 nr_moved += nr_pages;
1909 if (!is_active_lru(lru))
1910 __count_vm_events(PGDEACTIVATE, nr_moved);
1915 static void shrink_active_list(unsigned long nr_to_scan,
1916 struct lruvec *lruvec,
1917 struct scan_control *sc,
1920 unsigned long nr_taken;
1921 unsigned long nr_scanned;
1922 unsigned long vm_flags;
1923 LIST_HEAD(l_hold); /* The pages which were snipped off */
1924 LIST_HEAD(l_active);
1925 LIST_HEAD(l_inactive);
1927 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1928 unsigned nr_deactivate, nr_activate;
1929 unsigned nr_rotated = 0;
1930 isolate_mode_t isolate_mode = 0;
1931 int file = is_file_lru(lru);
1932 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1937 isolate_mode |= ISOLATE_UNMAPPED;
1939 spin_lock_irq(&pgdat->lru_lock);
1941 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1942 &nr_scanned, sc, isolate_mode, lru);
1944 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1945 reclaim_stat->recent_scanned[file] += nr_taken;
1947 __count_vm_events(PGREFILL, nr_scanned);
1949 spin_unlock_irq(&pgdat->lru_lock);
1951 while (!list_empty(&l_hold)) {
1953 page = lru_to_page(&l_hold);
1954 list_del(&page->lru);
1956 if (unlikely(!page_evictable(page))) {
1957 putback_lru_page(page);
1961 if (unlikely(buffer_heads_over_limit)) {
1962 if (page_has_private(page) && trylock_page(page)) {
1963 if (page_has_private(page))
1964 try_to_release_page(page, 0);
1969 if (page_referenced(page, 0, sc->target_mem_cgroup,
1971 nr_rotated += hpage_nr_pages(page);
1973 * Identify referenced, file-backed active pages and
1974 * give them one more trip around the active list. So
1975 * that executable code get better chances to stay in
1976 * memory under moderate memory pressure. Anon pages
1977 * are not likely to be evicted by use-once streaming
1978 * IO, plus JVM can create lots of anon VM_EXEC pages,
1979 * so we ignore them here.
1981 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1982 list_add(&page->lru, &l_active);
1987 ClearPageActive(page); /* we are de-activating */
1988 list_add(&page->lru, &l_inactive);
1992 * Move pages back to the lru list.
1994 spin_lock_irq(&pgdat->lru_lock);
1996 * Count referenced pages from currently used mappings as rotated,
1997 * even though only some of them are actually re-activated. This
1998 * helps balance scan pressure between file and anonymous pages in
2001 reclaim_stat->recent_rotated[file] += nr_rotated;
2003 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2004 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2005 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2006 spin_unlock_irq(&pgdat->lru_lock);
2008 mem_cgroup_uncharge_list(&l_hold);
2009 free_hot_cold_page_list(&l_hold, true);
2010 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2011 nr_deactivate, nr_rotated, sc->priority, file);
2015 * The inactive anon list should be small enough that the VM never has
2016 * to do too much work.
2018 * The inactive file list should be small enough to leave most memory
2019 * to the established workingset on the scan-resistant active list,
2020 * but large enough to avoid thrashing the aggregate readahead window.
2022 * Both inactive lists should also be large enough that each inactive
2023 * page has a chance to be referenced again before it is reclaimed.
2025 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2026 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2027 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2030 * memory ratio inactive
2031 * -------------------------------------
2040 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2041 struct scan_control *sc, bool trace)
2043 unsigned long inactive_ratio;
2044 unsigned long inactive, active;
2045 enum lru_list inactive_lru = file * LRU_FILE;
2046 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2050 * If we don't have swap space, anonymous page deactivation
2053 if (!file && !total_swap_pages)
2056 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2057 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2059 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2061 inactive_ratio = int_sqrt(10 * gb);
2066 trace_mm_vmscan_inactive_list_is_low(lruvec_pgdat(lruvec)->node_id,
2068 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2069 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2070 inactive_ratio, file);
2072 return inactive * inactive_ratio < active;
2075 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2076 struct lruvec *lruvec, struct scan_control *sc)
2078 if (is_active_lru(lru)) {
2079 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2080 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2084 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2095 * Determine how aggressively the anon and file LRU lists should be
2096 * scanned. The relative value of each set of LRU lists is determined
2097 * by looking at the fraction of the pages scanned we did rotate back
2098 * onto the active list instead of evict.
2100 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2101 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2103 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2104 struct scan_control *sc, unsigned long *nr,
2105 unsigned long *lru_pages)
2107 int swappiness = mem_cgroup_swappiness(memcg);
2108 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2110 u64 denominator = 0; /* gcc */
2111 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2112 unsigned long anon_prio, file_prio;
2113 enum scan_balance scan_balance;
2114 unsigned long anon, file;
2115 unsigned long ap, fp;
2118 /* If we have no swap space, do not bother scanning anon pages. */
2119 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2120 scan_balance = SCAN_FILE;
2125 * Global reclaim will swap to prevent OOM even with no
2126 * swappiness, but memcg users want to use this knob to
2127 * disable swapping for individual groups completely when
2128 * using the memory controller's swap limit feature would be
2131 if (!global_reclaim(sc) && !swappiness) {
2132 scan_balance = SCAN_FILE;
2137 * Do not apply any pressure balancing cleverness when the
2138 * system is close to OOM, scan both anon and file equally
2139 * (unless the swappiness setting disagrees with swapping).
2141 if (!sc->priority && swappiness) {
2142 scan_balance = SCAN_EQUAL;
2147 * Prevent the reclaimer from falling into the cache trap: as
2148 * cache pages start out inactive, every cache fault will tip
2149 * the scan balance towards the file LRU. And as the file LRU
2150 * shrinks, so does the window for rotation from references.
2151 * This means we have a runaway feedback loop where a tiny
2152 * thrashing file LRU becomes infinitely more attractive than
2153 * anon pages. Try to detect this based on file LRU size.
2155 if (global_reclaim(sc)) {
2156 unsigned long pgdatfile;
2157 unsigned long pgdatfree;
2159 unsigned long total_high_wmark = 0;
2161 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2162 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2163 node_page_state(pgdat, NR_INACTIVE_FILE);
2165 for (z = 0; z < MAX_NR_ZONES; z++) {
2166 struct zone *zone = &pgdat->node_zones[z];
2167 if (!managed_zone(zone))
2170 total_high_wmark += high_wmark_pages(zone);
2173 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2174 scan_balance = SCAN_ANON;
2180 * If there is enough inactive page cache, i.e. if the size of the
2181 * inactive list is greater than that of the active list *and* the
2182 * inactive list actually has some pages to scan on this priority, we
2183 * do not reclaim anything from the anonymous working set right now.
2184 * Without the second condition we could end up never scanning an
2185 * lruvec even if it has plenty of old anonymous pages unless the
2186 * system is under heavy pressure.
2188 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2189 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2190 scan_balance = SCAN_FILE;
2194 scan_balance = SCAN_FRACT;
2197 * With swappiness at 100, anonymous and file have the same priority.
2198 * This scanning priority is essentially the inverse of IO cost.
2200 anon_prio = swappiness;
2201 file_prio = 200 - anon_prio;
2204 * OK, so we have swap space and a fair amount of page cache
2205 * pages. We use the recently rotated / recently scanned
2206 * ratios to determine how valuable each cache is.
2208 * Because workloads change over time (and to avoid overflow)
2209 * we keep these statistics as a floating average, which ends
2210 * up weighing recent references more than old ones.
2212 * anon in [0], file in [1]
2215 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2216 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2217 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2218 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2220 spin_lock_irq(&pgdat->lru_lock);
2221 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2222 reclaim_stat->recent_scanned[0] /= 2;
2223 reclaim_stat->recent_rotated[0] /= 2;
2226 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2227 reclaim_stat->recent_scanned[1] /= 2;
2228 reclaim_stat->recent_rotated[1] /= 2;
2232 * The amount of pressure on anon vs file pages is inversely
2233 * proportional to the fraction of recently scanned pages on
2234 * each list that were recently referenced and in active use.
2236 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2237 ap /= reclaim_stat->recent_rotated[0] + 1;
2239 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2240 fp /= reclaim_stat->recent_rotated[1] + 1;
2241 spin_unlock_irq(&pgdat->lru_lock);
2245 denominator = ap + fp + 1;
2248 for_each_evictable_lru(lru) {
2249 int file = is_file_lru(lru);
2253 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2254 scan = size >> sc->priority;
2256 * If the cgroup's already been deleted, make sure to
2257 * scrape out the remaining cache.
2259 if (!scan && !mem_cgroup_online(memcg))
2260 scan = min(size, SWAP_CLUSTER_MAX);
2262 switch (scan_balance) {
2264 /* Scan lists relative to size */
2268 * Scan types proportional to swappiness and
2269 * their relative recent reclaim efficiency.
2271 scan = div64_u64(scan * fraction[file],
2276 /* Scan one type exclusively */
2277 if ((scan_balance == SCAN_FILE) != file) {
2283 /* Look ma, no brain */
2293 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2295 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2296 struct scan_control *sc, unsigned long *lru_pages)
2298 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2299 unsigned long nr[NR_LRU_LISTS];
2300 unsigned long targets[NR_LRU_LISTS];
2301 unsigned long nr_to_scan;
2303 unsigned long nr_reclaimed = 0;
2304 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2305 struct blk_plug plug;
2308 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2310 /* Record the original scan target for proportional adjustments later */
2311 memcpy(targets, nr, sizeof(nr));
2314 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2315 * event that can occur when there is little memory pressure e.g.
2316 * multiple streaming readers/writers. Hence, we do not abort scanning
2317 * when the requested number of pages are reclaimed when scanning at
2318 * DEF_PRIORITY on the assumption that the fact we are direct
2319 * reclaiming implies that kswapd is not keeping up and it is best to
2320 * do a batch of work at once. For memcg reclaim one check is made to
2321 * abort proportional reclaim if either the file or anon lru has already
2322 * dropped to zero at the first pass.
2324 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2325 sc->priority == DEF_PRIORITY);
2327 blk_start_plug(&plug);
2328 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2329 nr[LRU_INACTIVE_FILE]) {
2330 unsigned long nr_anon, nr_file, percentage;
2331 unsigned long nr_scanned;
2333 for_each_evictable_lru(lru) {
2335 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2336 nr[lru] -= nr_to_scan;
2338 nr_reclaimed += shrink_list(lru, nr_to_scan,
2345 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2349 * For kswapd and memcg, reclaim at least the number of pages
2350 * requested. Ensure that the anon and file LRUs are scanned
2351 * proportionally what was requested by get_scan_count(). We
2352 * stop reclaiming one LRU and reduce the amount scanning
2353 * proportional to the original scan target.
2355 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2356 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2359 * It's just vindictive to attack the larger once the smaller
2360 * has gone to zero. And given the way we stop scanning the
2361 * smaller below, this makes sure that we only make one nudge
2362 * towards proportionality once we've got nr_to_reclaim.
2364 if (!nr_file || !nr_anon)
2367 if (nr_file > nr_anon) {
2368 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2369 targets[LRU_ACTIVE_ANON] + 1;
2371 percentage = nr_anon * 100 / scan_target;
2373 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2374 targets[LRU_ACTIVE_FILE] + 1;
2376 percentage = nr_file * 100 / scan_target;
2379 /* Stop scanning the smaller of the LRU */
2381 nr[lru + LRU_ACTIVE] = 0;
2384 * Recalculate the other LRU scan count based on its original
2385 * scan target and the percentage scanning already complete
2387 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2388 nr_scanned = targets[lru] - nr[lru];
2389 nr[lru] = targets[lru] * (100 - percentage) / 100;
2390 nr[lru] -= min(nr[lru], nr_scanned);
2393 nr_scanned = targets[lru] - nr[lru];
2394 nr[lru] = targets[lru] * (100 - percentage) / 100;
2395 nr[lru] -= min(nr[lru], nr_scanned);
2397 scan_adjusted = true;
2399 blk_finish_plug(&plug);
2400 sc->nr_reclaimed += nr_reclaimed;
2403 * Even if we did not try to evict anon pages at all, we want to
2404 * rebalance the anon lru active/inactive ratio.
2406 if (inactive_list_is_low(lruvec, false, sc, true))
2407 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2408 sc, LRU_ACTIVE_ANON);
2411 /* Use reclaim/compaction for costly allocs or under memory pressure */
2412 static bool in_reclaim_compaction(struct scan_control *sc)
2414 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2415 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2416 sc->priority < DEF_PRIORITY - 2))
2423 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2424 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2425 * true if more pages should be reclaimed such that when the page allocator
2426 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2427 * It will give up earlier than that if there is difficulty reclaiming pages.
2429 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2430 unsigned long nr_reclaimed,
2431 unsigned long nr_scanned,
2432 struct scan_control *sc)
2434 unsigned long pages_for_compaction;
2435 unsigned long inactive_lru_pages;
2438 /* If not in reclaim/compaction mode, stop */
2439 if (!in_reclaim_compaction(sc))
2442 /* Consider stopping depending on scan and reclaim activity */
2443 if (sc->gfp_mask & __GFP_REPEAT) {
2445 * For __GFP_REPEAT allocations, stop reclaiming if the
2446 * full LRU list has been scanned and we are still failing
2447 * to reclaim pages. This full LRU scan is potentially
2448 * expensive but a __GFP_REPEAT caller really wants to succeed
2450 if (!nr_reclaimed && !nr_scanned)
2454 * For non-__GFP_REPEAT allocations which can presumably
2455 * fail without consequence, stop if we failed to reclaim
2456 * any pages from the last SWAP_CLUSTER_MAX number of
2457 * pages that were scanned. This will return to the
2458 * caller faster at the risk reclaim/compaction and
2459 * the resulting allocation attempt fails
2466 * If we have not reclaimed enough pages for compaction and the
2467 * inactive lists are large enough, continue reclaiming
2469 pages_for_compaction = compact_gap(sc->order);
2470 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2471 if (get_nr_swap_pages() > 0)
2472 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2473 if (sc->nr_reclaimed < pages_for_compaction &&
2474 inactive_lru_pages > pages_for_compaction)
2477 /* If compaction would go ahead or the allocation would succeed, stop */
2478 for (z = 0; z <= sc->reclaim_idx; z++) {
2479 struct zone *zone = &pgdat->node_zones[z];
2480 if (!managed_zone(zone))
2483 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2484 case COMPACT_SUCCESS:
2485 case COMPACT_CONTINUE:
2488 /* check next zone */
2495 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2497 struct reclaim_state *reclaim_state = current->reclaim_state;
2498 unsigned long nr_reclaimed, nr_scanned;
2499 bool reclaimable = false;
2502 struct mem_cgroup *root = sc->target_mem_cgroup;
2503 struct mem_cgroup_reclaim_cookie reclaim = {
2505 .priority = sc->priority,
2507 unsigned long node_lru_pages = 0;
2508 struct mem_cgroup *memcg;
2510 nr_reclaimed = sc->nr_reclaimed;
2511 nr_scanned = sc->nr_scanned;
2513 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2515 unsigned long lru_pages;
2516 unsigned long reclaimed;
2517 unsigned long scanned;
2519 if (mem_cgroup_low(root, memcg)) {
2520 if (!sc->memcg_low_reclaim) {
2521 sc->memcg_low_skipped = 1;
2524 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2527 reclaimed = sc->nr_reclaimed;
2528 scanned = sc->nr_scanned;
2530 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2531 node_lru_pages += lru_pages;
2534 shrink_slab(sc->gfp_mask, pgdat->node_id,
2535 memcg, sc->nr_scanned - scanned,
2538 /* Record the group's reclaim efficiency */
2539 vmpressure(sc->gfp_mask, memcg, false,
2540 sc->nr_scanned - scanned,
2541 sc->nr_reclaimed - reclaimed);
2544 * Direct reclaim and kswapd have to scan all memory
2545 * cgroups to fulfill the overall scan target for the
2548 * Limit reclaim, on the other hand, only cares about
2549 * nr_to_reclaim pages to be reclaimed and it will
2550 * retry with decreasing priority if one round over the
2551 * whole hierarchy is not sufficient.
2553 if (!global_reclaim(sc) &&
2554 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2555 mem_cgroup_iter_break(root, memcg);
2558 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2561 * Shrink the slab caches in the same proportion that
2562 * the eligible LRU pages were scanned.
2564 if (global_reclaim(sc))
2565 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2566 sc->nr_scanned - nr_scanned,
2569 if (reclaim_state) {
2570 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2571 reclaim_state->reclaimed_slab = 0;
2574 /* Record the subtree's reclaim efficiency */
2575 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2576 sc->nr_scanned - nr_scanned,
2577 sc->nr_reclaimed - nr_reclaimed);
2579 if (sc->nr_reclaimed - nr_reclaimed)
2582 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2583 sc->nr_scanned - nr_scanned, sc));
2586 * Kswapd gives up on balancing particular nodes after too
2587 * many failures to reclaim anything from them and goes to
2588 * sleep. On reclaim progress, reset the failure counter. A
2589 * successful direct reclaim run will revive a dormant kswapd.
2592 pgdat->kswapd_failures = 0;
2598 * Returns true if compaction should go ahead for a costly-order request, or
2599 * the allocation would already succeed without compaction. Return false if we
2600 * should reclaim first.
2602 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2604 unsigned long watermark;
2605 enum compact_result suitable;
2607 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2608 if (suitable == COMPACT_SUCCESS)
2609 /* Allocation should succeed already. Don't reclaim. */
2611 if (suitable == COMPACT_SKIPPED)
2612 /* Compaction cannot yet proceed. Do reclaim. */
2616 * Compaction is already possible, but it takes time to run and there
2617 * are potentially other callers using the pages just freed. So proceed
2618 * with reclaim to make a buffer of free pages available to give
2619 * compaction a reasonable chance of completing and allocating the page.
2620 * Note that we won't actually reclaim the whole buffer in one attempt
2621 * as the target watermark in should_continue_reclaim() is lower. But if
2622 * we are already above the high+gap watermark, don't reclaim at all.
2624 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2626 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2630 * This is the direct reclaim path, for page-allocating processes. We only
2631 * try to reclaim pages from zones which will satisfy the caller's allocation
2634 * If a zone is deemed to be full of pinned pages then just give it a light
2635 * scan then give up on it.
2637 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2641 unsigned long nr_soft_reclaimed;
2642 unsigned long nr_soft_scanned;
2644 pg_data_t *last_pgdat = NULL;
2647 * If the number of buffer_heads in the machine exceeds the maximum
2648 * allowed level, force direct reclaim to scan the highmem zone as
2649 * highmem pages could be pinning lowmem pages storing buffer_heads
2651 orig_mask = sc->gfp_mask;
2652 if (buffer_heads_over_limit) {
2653 sc->gfp_mask |= __GFP_HIGHMEM;
2654 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2657 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2658 sc->reclaim_idx, sc->nodemask) {
2660 * Take care memory controller reclaiming has small influence
2663 if (global_reclaim(sc)) {
2664 if (!cpuset_zone_allowed(zone,
2665 GFP_KERNEL | __GFP_HARDWALL))
2669 * If we already have plenty of memory free for
2670 * compaction in this zone, don't free any more.
2671 * Even though compaction is invoked for any
2672 * non-zero order, only frequent costly order
2673 * reclamation is disruptive enough to become a
2674 * noticeable problem, like transparent huge
2677 if (IS_ENABLED(CONFIG_COMPACTION) &&
2678 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2679 compaction_ready(zone, sc)) {
2680 sc->compaction_ready = true;
2685 * Shrink each node in the zonelist once. If the
2686 * zonelist is ordered by zone (not the default) then a
2687 * node may be shrunk multiple times but in that case
2688 * the user prefers lower zones being preserved.
2690 if (zone->zone_pgdat == last_pgdat)
2694 * This steals pages from memory cgroups over softlimit
2695 * and returns the number of reclaimed pages and
2696 * scanned pages. This works for global memory pressure
2697 * and balancing, not for a memcg's limit.
2699 nr_soft_scanned = 0;
2700 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2701 sc->order, sc->gfp_mask,
2703 sc->nr_reclaimed += nr_soft_reclaimed;
2704 sc->nr_scanned += nr_soft_scanned;
2705 /* need some check for avoid more shrink_zone() */
2708 /* See comment about same check for global reclaim above */
2709 if (zone->zone_pgdat == last_pgdat)
2711 last_pgdat = zone->zone_pgdat;
2712 shrink_node(zone->zone_pgdat, sc);
2716 * Restore to original mask to avoid the impact on the caller if we
2717 * promoted it to __GFP_HIGHMEM.
2719 sc->gfp_mask = orig_mask;
2723 * This is the main entry point to direct page reclaim.
2725 * If a full scan of the inactive list fails to free enough memory then we
2726 * are "out of memory" and something needs to be killed.
2728 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2729 * high - the zone may be full of dirty or under-writeback pages, which this
2730 * caller can't do much about. We kick the writeback threads and take explicit
2731 * naps in the hope that some of these pages can be written. But if the
2732 * allocating task holds filesystem locks which prevent writeout this might not
2733 * work, and the allocation attempt will fail.
2735 * returns: 0, if no pages reclaimed
2736 * else, the number of pages reclaimed
2738 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2739 struct scan_control *sc)
2741 int initial_priority = sc->priority;
2743 delayacct_freepages_start();
2745 if (global_reclaim(sc))
2746 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2749 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2752 shrink_zones(zonelist, sc);
2754 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2757 if (sc->compaction_ready)
2761 * If we're getting trouble reclaiming, start doing
2762 * writepage even in laptop mode.
2764 if (sc->priority < DEF_PRIORITY - 2)
2765 sc->may_writepage = 1;
2766 } while (--sc->priority >= 0);
2768 delayacct_freepages_end();
2770 if (sc->nr_reclaimed)
2771 return sc->nr_reclaimed;
2773 /* Aborted reclaim to try compaction? don't OOM, then */
2774 if (sc->compaction_ready)
2777 /* Untapped cgroup reserves? Don't OOM, retry. */
2778 if (sc->memcg_low_skipped) {
2779 sc->priority = initial_priority;
2780 sc->memcg_low_reclaim = 1;
2781 sc->memcg_low_skipped = 0;
2788 static bool allow_direct_reclaim(pg_data_t *pgdat)
2791 unsigned long pfmemalloc_reserve = 0;
2792 unsigned long free_pages = 0;
2796 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2799 for (i = 0; i <= ZONE_NORMAL; i++) {
2800 zone = &pgdat->node_zones[i];
2801 if (!managed_zone(zone))
2804 if (!zone_reclaimable_pages(zone))
2807 pfmemalloc_reserve += min_wmark_pages(zone);
2808 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2811 /* If there are no reserves (unexpected config) then do not throttle */
2812 if (!pfmemalloc_reserve)
2815 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2817 /* kswapd must be awake if processes are being throttled */
2818 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2819 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2820 (enum zone_type)ZONE_NORMAL);
2821 wake_up_interruptible(&pgdat->kswapd_wait);
2828 * Throttle direct reclaimers if backing storage is backed by the network
2829 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2830 * depleted. kswapd will continue to make progress and wake the processes
2831 * when the low watermark is reached.
2833 * Returns true if a fatal signal was delivered during throttling. If this
2834 * happens, the page allocator should not consider triggering the OOM killer.
2836 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2837 nodemask_t *nodemask)
2841 pg_data_t *pgdat = NULL;
2844 * Kernel threads should not be throttled as they may be indirectly
2845 * responsible for cleaning pages necessary for reclaim to make forward
2846 * progress. kjournald for example may enter direct reclaim while
2847 * committing a transaction where throttling it could forcing other
2848 * processes to block on log_wait_commit().
2850 if (current->flags & PF_KTHREAD)
2854 * If a fatal signal is pending, this process should not throttle.
2855 * It should return quickly so it can exit and free its memory
2857 if (fatal_signal_pending(current))
2861 * Check if the pfmemalloc reserves are ok by finding the first node
2862 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2863 * GFP_KERNEL will be required for allocating network buffers when
2864 * swapping over the network so ZONE_HIGHMEM is unusable.
2866 * Throttling is based on the first usable node and throttled processes
2867 * wait on a queue until kswapd makes progress and wakes them. There
2868 * is an affinity then between processes waking up and where reclaim
2869 * progress has been made assuming the process wakes on the same node.
2870 * More importantly, processes running on remote nodes will not compete
2871 * for remote pfmemalloc reserves and processes on different nodes
2872 * should make reasonable progress.
2874 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2875 gfp_zone(gfp_mask), nodemask) {
2876 if (zone_idx(zone) > ZONE_NORMAL)
2879 /* Throttle based on the first usable node */
2880 pgdat = zone->zone_pgdat;
2881 if (allow_direct_reclaim(pgdat))
2886 /* If no zone was usable by the allocation flags then do not throttle */
2890 /* Account for the throttling */
2891 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2894 * If the caller cannot enter the filesystem, it's possible that it
2895 * is due to the caller holding an FS lock or performing a journal
2896 * transaction in the case of a filesystem like ext[3|4]. In this case,
2897 * it is not safe to block on pfmemalloc_wait as kswapd could be
2898 * blocked waiting on the same lock. Instead, throttle for up to a
2899 * second before continuing.
2901 if (!(gfp_mask & __GFP_FS)) {
2902 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2903 allow_direct_reclaim(pgdat), HZ);
2908 /* Throttle until kswapd wakes the process */
2909 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2910 allow_direct_reclaim(pgdat));
2913 if (fatal_signal_pending(current))
2920 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2921 gfp_t gfp_mask, nodemask_t *nodemask)
2923 unsigned long nr_reclaimed;
2924 struct scan_control sc = {
2925 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2926 .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)),
2927 .reclaim_idx = gfp_zone(gfp_mask),
2929 .nodemask = nodemask,
2930 .priority = DEF_PRIORITY,
2931 .may_writepage = !laptop_mode,
2937 * Do not enter reclaim if fatal signal was delivered while throttled.
2938 * 1 is returned so that the page allocator does not OOM kill at this
2941 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2944 trace_mm_vmscan_direct_reclaim_begin(order,
2949 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2951 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2953 return nr_reclaimed;
2958 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2959 gfp_t gfp_mask, bool noswap,
2961 unsigned long *nr_scanned)
2963 struct scan_control sc = {
2964 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2965 .target_mem_cgroup = memcg,
2966 .may_writepage = !laptop_mode,
2968 .reclaim_idx = MAX_NR_ZONES - 1,
2969 .may_swap = !noswap,
2971 unsigned long lru_pages;
2973 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2974 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2976 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2982 * NOTE: Although we can get the priority field, using it
2983 * here is not a good idea, since it limits the pages we can scan.
2984 * if we don't reclaim here, the shrink_node from balance_pgdat
2985 * will pick up pages from other mem cgroup's as well. We hack
2986 * the priority and make it zero.
2988 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
2990 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2992 *nr_scanned = sc.nr_scanned;
2993 return sc.nr_reclaimed;
2996 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2997 unsigned long nr_pages,
3001 struct zonelist *zonelist;
3002 unsigned long nr_reclaimed;
3004 struct scan_control sc = {
3005 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3006 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3007 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3008 .reclaim_idx = MAX_NR_ZONES - 1,
3009 .target_mem_cgroup = memcg,
3010 .priority = DEF_PRIORITY,
3011 .may_writepage = !laptop_mode,
3013 .may_swap = may_swap,
3017 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3018 * take care of from where we get pages. So the node where we start the
3019 * scan does not need to be the current node.
3021 nid = mem_cgroup_select_victim_node(memcg);
3023 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3025 trace_mm_vmscan_memcg_reclaim_begin(0,
3030 current->flags |= PF_MEMALLOC;
3031 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3032 current->flags &= ~PF_MEMALLOC;
3034 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3036 return nr_reclaimed;
3040 static void age_active_anon(struct pglist_data *pgdat,
3041 struct scan_control *sc)
3043 struct mem_cgroup *memcg;
3045 if (!total_swap_pages)
3048 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3050 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3052 if (inactive_list_is_low(lruvec, false, sc, true))
3053 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3054 sc, LRU_ACTIVE_ANON);
3056 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3061 * Returns true if there is an eligible zone balanced for the request order
3064 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3067 unsigned long mark = -1;
3070 for (i = 0; i <= classzone_idx; i++) {
3071 zone = pgdat->node_zones + i;
3073 if (!managed_zone(zone))
3076 mark = high_wmark_pages(zone);
3077 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3082 * If a node has no populated zone within classzone_idx, it does not
3083 * need balancing by definition. This can happen if a zone-restricted
3084 * allocation tries to wake a remote kswapd.
3092 /* Clear pgdat state for congested, dirty or under writeback. */
3093 static void clear_pgdat_congested(pg_data_t *pgdat)
3095 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3096 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3097 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3101 * Prepare kswapd for sleeping. This verifies that there are no processes
3102 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3104 * Returns true if kswapd is ready to sleep
3106 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3109 * The throttled processes are normally woken up in balance_pgdat() as
3110 * soon as allow_direct_reclaim() is true. But there is a potential
3111 * race between when kswapd checks the watermarks and a process gets
3112 * throttled. There is also a potential race if processes get
3113 * throttled, kswapd wakes, a large process exits thereby balancing the
3114 * zones, which causes kswapd to exit balance_pgdat() before reaching
3115 * the wake up checks. If kswapd is going to sleep, no process should
3116 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3117 * the wake up is premature, processes will wake kswapd and get
3118 * throttled again. The difference from wake ups in balance_pgdat() is
3119 * that here we are under prepare_to_wait().
3121 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3122 wake_up_all(&pgdat->pfmemalloc_wait);
3124 /* Hopeless node, leave it to direct reclaim */
3125 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3128 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3129 clear_pgdat_congested(pgdat);
3137 * kswapd shrinks a node of pages that are at or below the highest usable
3138 * zone that is currently unbalanced.
3140 * Returns true if kswapd scanned at least the requested number of pages to
3141 * reclaim or if the lack of progress was due to pages under writeback.
3142 * This is used to determine if the scanning priority needs to be raised.
3144 static bool kswapd_shrink_node(pg_data_t *pgdat,
3145 struct scan_control *sc)
3150 /* Reclaim a number of pages proportional to the number of zones */
3151 sc->nr_to_reclaim = 0;
3152 for (z = 0; z <= sc->reclaim_idx; z++) {
3153 zone = pgdat->node_zones + z;
3154 if (!managed_zone(zone))
3157 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3161 * Historically care was taken to put equal pressure on all zones but
3162 * now pressure is applied based on node LRU order.
3164 shrink_node(pgdat, sc);
3167 * Fragmentation may mean that the system cannot be rebalanced for
3168 * high-order allocations. If twice the allocation size has been
3169 * reclaimed then recheck watermarks only at order-0 to prevent
3170 * excessive reclaim. Assume that a process requested a high-order
3171 * can direct reclaim/compact.
3173 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3176 return sc->nr_scanned >= sc->nr_to_reclaim;
3180 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3181 * that are eligible for use by the caller until at least one zone is
3184 * Returns the order kswapd finished reclaiming at.
3186 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3187 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3188 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3189 * or lower is eligible for reclaim until at least one usable zone is
3192 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3195 unsigned long nr_soft_reclaimed;
3196 unsigned long nr_soft_scanned;
3198 struct scan_control sc = {
3199 .gfp_mask = GFP_KERNEL,
3201 .priority = DEF_PRIORITY,
3202 .may_writepage = !laptop_mode,
3206 count_vm_event(PAGEOUTRUN);
3209 unsigned long nr_reclaimed = sc.nr_reclaimed;
3210 bool raise_priority = true;
3212 sc.reclaim_idx = classzone_idx;
3215 * If the number of buffer_heads exceeds the maximum allowed
3216 * then consider reclaiming from all zones. This has a dual
3217 * purpose -- on 64-bit systems it is expected that
3218 * buffer_heads are stripped during active rotation. On 32-bit
3219 * systems, highmem pages can pin lowmem memory and shrinking
3220 * buffers can relieve lowmem pressure. Reclaim may still not
3221 * go ahead if all eligible zones for the original allocation
3222 * request are balanced to avoid excessive reclaim from kswapd.
3224 if (buffer_heads_over_limit) {
3225 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3226 zone = pgdat->node_zones + i;
3227 if (!managed_zone(zone))
3236 * Only reclaim if there are no eligible zones. Note that
3237 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3240 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3244 * Do some background aging of the anon list, to give
3245 * pages a chance to be referenced before reclaiming. All
3246 * pages are rotated regardless of classzone as this is
3247 * about consistent aging.
3249 age_active_anon(pgdat, &sc);
3252 * If we're getting trouble reclaiming, start doing writepage
3253 * even in laptop mode.
3255 if (sc.priority < DEF_PRIORITY - 2)
3256 sc.may_writepage = 1;
3258 /* Call soft limit reclaim before calling shrink_node. */
3260 nr_soft_scanned = 0;
3261 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3262 sc.gfp_mask, &nr_soft_scanned);
3263 sc.nr_reclaimed += nr_soft_reclaimed;
3266 * There should be no need to raise the scanning priority if
3267 * enough pages are already being scanned that that high
3268 * watermark would be met at 100% efficiency.
3270 if (kswapd_shrink_node(pgdat, &sc))
3271 raise_priority = false;
3274 * If the low watermark is met there is no need for processes
3275 * to be throttled on pfmemalloc_wait as they should not be
3276 * able to safely make forward progress. Wake them
3278 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3279 allow_direct_reclaim(pgdat))
3280 wake_up_all(&pgdat->pfmemalloc_wait);
3282 /* Check if kswapd should be suspending */
3283 if (try_to_freeze() || kthread_should_stop())
3287 * Raise priority if scanning rate is too low or there was no
3288 * progress in reclaiming pages
3290 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3291 if (raise_priority || !nr_reclaimed)
3293 } while (sc.priority >= 1);
3295 if (!sc.nr_reclaimed)
3296 pgdat->kswapd_failures++;
3300 * Return the order kswapd stopped reclaiming at as
3301 * prepare_kswapd_sleep() takes it into account. If another caller
3302 * entered the allocator slow path while kswapd was awake, order will
3303 * remain at the higher level.
3309 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3310 * allocation request woke kswapd for. When kswapd has not woken recently,
3311 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3312 * given classzone and returns it or the highest classzone index kswapd
3313 * was recently woke for.
3315 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3316 enum zone_type classzone_idx)
3318 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3319 return classzone_idx;
3321 return max(pgdat->kswapd_classzone_idx, classzone_idx);
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);
3336 * Try to sleep for a short interval. Note that kcompactd will only be
3337 * woken if it is possible to sleep for a short interval. This is
3338 * deliberate on the assumption that if reclaim cannot keep an
3339 * eligible zone balanced that it's also unlikely that compaction will
3342 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3344 * Compaction records what page blocks it recently failed to
3345 * isolate pages from and skips them in the future scanning.
3346 * When kswapd is going to sleep, it is reasonable to assume
3347 * that pages and compaction may succeed so reset the cache.
3349 reset_isolation_suitable(pgdat);
3352 * We have freed the memory, now we should compact it to make
3353 * allocation of the requested order possible.
3355 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3357 remaining = schedule_timeout(HZ/10);
3360 * If woken prematurely then reset kswapd_classzone_idx and
3361 * order. The values will either be from a wakeup request or
3362 * the previous request that slept prematurely.
3365 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3366 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3369 finish_wait(&pgdat->kswapd_wait, &wait);
3370 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3374 * After a short sleep, check if it was a premature sleep. If not, then
3375 * go fully to sleep until explicitly woken up.
3378 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3379 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3382 * vmstat counters are not perfectly accurate and the estimated
3383 * value for counters such as NR_FREE_PAGES can deviate from the
3384 * true value by nr_online_cpus * threshold. To avoid the zone
3385 * watermarks being breached while under pressure, we reduce the
3386 * per-cpu vmstat threshold while kswapd is awake and restore
3387 * them before going back to sleep.
3389 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3391 if (!kthread_should_stop())
3394 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3397 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3399 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3401 finish_wait(&pgdat->kswapd_wait, &wait);
3405 * The background pageout daemon, started as a kernel thread
3406 * from the init process.
3408 * This basically trickles out pages so that we have _some_
3409 * free memory available even if there is no other activity
3410 * that frees anything up. This is needed for things like routing
3411 * etc, where we otherwise might have all activity going on in
3412 * asynchronous contexts that cannot page things out.
3414 * If there are applications that are active memory-allocators
3415 * (most normal use), this basically shouldn't matter.
3417 static int kswapd(void *p)
3419 unsigned int alloc_order, reclaim_order;
3420 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3421 pg_data_t *pgdat = (pg_data_t*)p;
3422 struct task_struct *tsk = current;
3424 struct reclaim_state reclaim_state = {
3425 .reclaimed_slab = 0,
3427 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3429 lockdep_set_current_reclaim_state(GFP_KERNEL);
3431 if (!cpumask_empty(cpumask))
3432 set_cpus_allowed_ptr(tsk, cpumask);
3433 current->reclaim_state = &reclaim_state;
3436 * Tell the memory management that we're a "memory allocator",
3437 * and that if we need more memory we should get access to it
3438 * regardless (see "__alloc_pages()"). "kswapd" should
3439 * never get caught in the normal page freeing logic.
3441 * (Kswapd normally doesn't need memory anyway, but sometimes
3442 * you need a small amount of memory in order to be able to
3443 * page out something else, and this flag essentially protects
3444 * us from recursively trying to free more memory as we're
3445 * trying to free the first piece of memory in the first place).
3447 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3450 pgdat->kswapd_order = 0;
3451 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3455 alloc_order = reclaim_order = pgdat->kswapd_order;
3456 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3459 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3462 /* Read the new order and classzone_idx */
3463 alloc_order = reclaim_order = pgdat->kswapd_order;
3464 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3465 pgdat->kswapd_order = 0;
3466 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3468 ret = try_to_freeze();
3469 if (kthread_should_stop())
3473 * We can speed up thawing tasks if we don't call balance_pgdat
3474 * after returning from the refrigerator
3480 * Reclaim begins at the requested order but if a high-order
3481 * reclaim fails then kswapd falls back to reclaiming for
3482 * order-0. If that happens, kswapd will consider sleeping
3483 * for the order it finished reclaiming at (reclaim_order)
3484 * but kcompactd is woken to compact for the original
3485 * request (alloc_order).
3487 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3489 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3490 if (reclaim_order < alloc_order)
3491 goto kswapd_try_sleep;
3494 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3495 current->reclaim_state = NULL;
3496 lockdep_clear_current_reclaim_state();
3502 * A zone is low on free memory, so wake its kswapd task to service it.
3504 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3508 if (!managed_zone(zone))
3511 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3513 pgdat = zone->zone_pgdat;
3514 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3516 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3517 if (!waitqueue_active(&pgdat->kswapd_wait))
3520 /* Hopeless node, leave it to direct reclaim */
3521 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3524 if (pgdat_balanced(pgdat, order, classzone_idx))
3527 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order);
3528 wake_up_interruptible(&pgdat->kswapd_wait);
3531 #ifdef CONFIG_HIBERNATION
3533 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3536 * Rather than trying to age LRUs the aim is to preserve the overall
3537 * LRU order by reclaiming preferentially
3538 * inactive > active > active referenced > active mapped
3540 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3542 struct reclaim_state reclaim_state;
3543 struct scan_control sc = {
3544 .nr_to_reclaim = nr_to_reclaim,
3545 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3546 .reclaim_idx = MAX_NR_ZONES - 1,
3547 .priority = DEF_PRIORITY,
3551 .hibernation_mode = 1,
3553 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3554 struct task_struct *p = current;
3555 unsigned long nr_reclaimed;
3557 p->flags |= PF_MEMALLOC;
3558 lockdep_set_current_reclaim_state(sc.gfp_mask);
3559 reclaim_state.reclaimed_slab = 0;
3560 p->reclaim_state = &reclaim_state;
3562 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3564 p->reclaim_state = NULL;
3565 lockdep_clear_current_reclaim_state();
3566 p->flags &= ~PF_MEMALLOC;
3568 return nr_reclaimed;
3570 #endif /* CONFIG_HIBERNATION */
3572 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3573 not required for correctness. So if the last cpu in a node goes
3574 away, we get changed to run anywhere: as the first one comes back,
3575 restore their cpu bindings. */
3576 static int kswapd_cpu_online(unsigned int cpu)
3580 for_each_node_state(nid, N_MEMORY) {
3581 pg_data_t *pgdat = NODE_DATA(nid);
3582 const struct cpumask *mask;
3584 mask = cpumask_of_node(pgdat->node_id);
3586 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3587 /* One of our CPUs online: restore mask */
3588 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3594 * This kswapd start function will be called by init and node-hot-add.
3595 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3597 int kswapd_run(int nid)
3599 pg_data_t *pgdat = NODE_DATA(nid);
3605 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3606 if (IS_ERR(pgdat->kswapd)) {
3607 /* failure at boot is fatal */
3608 BUG_ON(system_state == SYSTEM_BOOTING);
3609 pr_err("Failed to start kswapd on node %d\n", nid);
3610 ret = PTR_ERR(pgdat->kswapd);
3611 pgdat->kswapd = NULL;
3617 * Called by memory hotplug when all memory in a node is offlined. Caller must
3618 * hold mem_hotplug_begin/end().
3620 void kswapd_stop(int nid)
3622 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3625 kthread_stop(kswapd);
3626 NODE_DATA(nid)->kswapd = NULL;
3630 static int __init kswapd_init(void)
3635 for_each_node_state(nid, N_MEMORY)
3637 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3638 "mm/vmscan:online", kswapd_cpu_online,
3644 module_init(kswapd_init)
3650 * If non-zero call node_reclaim when the number of free pages falls below
3653 int node_reclaim_mode __read_mostly;
3655 #define RECLAIM_OFF 0
3656 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3657 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3658 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3661 * Priority for NODE_RECLAIM. This determines the fraction of pages
3662 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3665 #define NODE_RECLAIM_PRIORITY 4
3668 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3671 int sysctl_min_unmapped_ratio = 1;
3674 * If the number of slab pages in a zone grows beyond this percentage then
3675 * slab reclaim needs to occur.
3677 int sysctl_min_slab_ratio = 5;
3679 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3681 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3682 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3683 node_page_state(pgdat, NR_ACTIVE_FILE);
3686 * It's possible for there to be more file mapped pages than
3687 * accounted for by the pages on the file LRU lists because
3688 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3690 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3693 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3694 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3696 unsigned long nr_pagecache_reclaimable;
3697 unsigned long delta = 0;
3700 * If RECLAIM_UNMAP is set, then all file pages are considered
3701 * potentially reclaimable. Otherwise, we have to worry about
3702 * pages like swapcache and node_unmapped_file_pages() provides
3705 if (node_reclaim_mode & RECLAIM_UNMAP)
3706 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3708 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3710 /* If we can't clean pages, remove dirty pages from consideration */
3711 if (!(node_reclaim_mode & RECLAIM_WRITE))
3712 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3714 /* Watch for any possible underflows due to delta */
3715 if (unlikely(delta > nr_pagecache_reclaimable))
3716 delta = nr_pagecache_reclaimable;
3718 return nr_pagecache_reclaimable - delta;
3722 * Try to free up some pages from this node through reclaim.
3724 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3726 /* Minimum pages needed in order to stay on node */
3727 const unsigned long nr_pages = 1 << order;
3728 struct task_struct *p = current;
3729 struct reclaim_state reclaim_state;
3730 int classzone_idx = gfp_zone(gfp_mask);
3731 struct scan_control sc = {
3732 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3733 .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)),
3735 .priority = NODE_RECLAIM_PRIORITY,
3736 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3737 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3739 .reclaim_idx = classzone_idx,
3744 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3745 * and we also need to be able to write out pages for RECLAIM_WRITE
3746 * and RECLAIM_UNMAP.
3748 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3749 lockdep_set_current_reclaim_state(gfp_mask);
3750 reclaim_state.reclaimed_slab = 0;
3751 p->reclaim_state = &reclaim_state;
3753 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3755 * Free memory by calling shrink zone with increasing
3756 * priorities until we have enough memory freed.
3759 shrink_node(pgdat, &sc);
3760 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3763 p->reclaim_state = NULL;
3764 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3765 lockdep_clear_current_reclaim_state();
3766 return sc.nr_reclaimed >= nr_pages;
3769 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3774 * Node reclaim reclaims unmapped file backed pages and
3775 * slab pages if we are over the defined limits.
3777 * A small portion of unmapped file backed pages is needed for
3778 * file I/O otherwise pages read by file I/O will be immediately
3779 * thrown out if the node is overallocated. So we do not reclaim
3780 * if less than a specified percentage of the node is used by
3781 * unmapped file backed pages.
3783 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3784 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3785 return NODE_RECLAIM_FULL;
3788 * Do not scan if the allocation should not be delayed.
3790 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3791 return NODE_RECLAIM_NOSCAN;
3794 * Only run node reclaim on the local node or on nodes that do not
3795 * have associated processors. This will favor the local processor
3796 * over remote processors and spread off node memory allocations
3797 * as wide as possible.
3799 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3800 return NODE_RECLAIM_NOSCAN;
3802 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3803 return NODE_RECLAIM_NOSCAN;
3805 ret = __node_reclaim(pgdat, gfp_mask, order);
3806 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3809 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3816 * page_evictable - test whether a page is evictable
3817 * @page: the page to test
3819 * Test whether page is evictable--i.e., should be placed on active/inactive
3820 * lists vs unevictable list.
3822 * Reasons page might not be evictable:
3823 * (1) page's mapping marked unevictable
3824 * (2) page is part of an mlocked VMA
3827 int page_evictable(struct page *page)
3829 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3834 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3835 * @pages: array of pages to check
3836 * @nr_pages: number of pages to check
3838 * Checks pages for evictability and moves them to the appropriate lru list.
3840 * This function is only used for SysV IPC SHM_UNLOCK.
3842 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3844 struct lruvec *lruvec;
3845 struct pglist_data *pgdat = NULL;
3850 for (i = 0; i < nr_pages; i++) {
3851 struct page *page = pages[i];
3852 struct pglist_data *pagepgdat = page_pgdat(page);
3855 if (pagepgdat != pgdat) {
3857 spin_unlock_irq(&pgdat->lru_lock);
3859 spin_lock_irq(&pgdat->lru_lock);
3861 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3863 if (!PageLRU(page) || !PageUnevictable(page))
3866 if (page_evictable(page)) {
3867 enum lru_list lru = page_lru_base_type(page);
3869 VM_BUG_ON_PAGE(PageActive(page), page);
3870 ClearPageUnevictable(page);
3871 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3872 add_page_to_lru_list(page, lruvec, lru);
3878 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3879 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3880 spin_unlock_irq(&pgdat->lru_lock);
3883 #endif /* CONFIG_SHMEM */