4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup *target_mem_cgroup;
84 /* Scan (total_size >> priority) pages at once */
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
90 unsigned int may_writepage:1;
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
101 unsigned int hibernation_mode:1;
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
113 #ifdef ARCH_HAS_PREFETCH
114 #define prefetch_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
124 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
127 #ifdef ARCH_HAS_PREFETCHW
128 #define prefetchw_prev_lru_page(_page, _base, _field) \
130 if ((_page)->lru.prev != _base) { \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
138 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
142 * From 0 .. 100. Higher means more swappy.
144 int vm_swappiness = 60;
146 * The total number of pages which are beyond the high watermark within all
149 unsigned long vm_total_pages;
151 static LIST_HEAD(shrinker_list);
152 static DECLARE_RWSEM(shrinker_rwsem);
155 static bool global_reclaim(struct scan_control *sc)
157 return !sc->target_mem_cgroup;
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
173 static bool sane_reclaim(struct scan_control *sc)
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
179 #ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
186 static bool global_reclaim(struct scan_control *sc)
191 static bool sane_reclaim(struct scan_control *sc)
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
202 unsigned long zone_reclaimable_pages(struct zone *zone)
206 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
207 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
208 if (get_nr_swap_pages() > 0)
209 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
210 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
215 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
219 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
220 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
223 if (get_nr_swap_pages() > 0)
224 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
225 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
231 bool pgdat_reclaimable(struct pglist_data *pgdat)
233 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
234 pgdat_reclaimable_pages(pgdat) * 6;
238 * lruvec_lru_size - Returns the number of pages on the given LRU list.
239 * @lruvec: lru vector
241 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
243 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
245 unsigned long lru_size;
248 if (!mem_cgroup_disabled())
249 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
251 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
253 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
254 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
257 if (!managed_zone(zone))
260 if (!mem_cgroup_disabled())
261 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
263 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
264 NR_ZONE_LRU_BASE + lru);
265 lru_size -= min(size, lru_size);
273 * Add a shrinker callback to be called from the vm.
275 int register_shrinker(struct shrinker *shrinker)
277 size_t size = sizeof(*shrinker->nr_deferred);
279 if (shrinker->flags & SHRINKER_NUMA_AWARE)
282 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
283 if (!shrinker->nr_deferred)
286 down_write(&shrinker_rwsem);
287 list_add_tail(&shrinker->list, &shrinker_list);
288 up_write(&shrinker_rwsem);
291 EXPORT_SYMBOL(register_shrinker);
296 void unregister_shrinker(struct shrinker *shrinker)
298 down_write(&shrinker_rwsem);
299 list_del(&shrinker->list);
300 up_write(&shrinker_rwsem);
301 kfree(shrinker->nr_deferred);
303 EXPORT_SYMBOL(unregister_shrinker);
305 #define SHRINK_BATCH 128
307 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
308 struct shrinker *shrinker,
309 unsigned long nr_scanned,
310 unsigned long nr_eligible)
312 unsigned long freed = 0;
313 unsigned long long delta;
318 int nid = shrinkctl->nid;
319 long batch_size = shrinker->batch ? shrinker->batch
321 long scanned = 0, next_deferred;
323 freeable = shrinker->count_objects(shrinker, shrinkctl);
328 * copy the current shrinker scan count into a local variable
329 * and zero it so that other concurrent shrinker invocations
330 * don't also do this scanning work.
332 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
335 delta = (4 * nr_scanned) / shrinker->seeks;
337 do_div(delta, nr_eligible + 1);
339 if (total_scan < 0) {
340 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
341 shrinker->scan_objects, total_scan);
342 total_scan = freeable;
345 next_deferred = total_scan;
348 * We need to avoid excessive windup on filesystem shrinkers
349 * due to large numbers of GFP_NOFS allocations causing the
350 * shrinkers to return -1 all the time. This results in a large
351 * nr being built up so when a shrink that can do some work
352 * comes along it empties the entire cache due to nr >>>
353 * freeable. This is bad for sustaining a working set in
356 * Hence only allow the shrinker to scan the entire cache when
357 * a large delta change is calculated directly.
359 if (delta < freeable / 4)
360 total_scan = min(total_scan, freeable / 2);
363 * Avoid risking looping forever due to too large nr value:
364 * never try to free more than twice the estimate number of
367 if (total_scan > freeable * 2)
368 total_scan = freeable * 2;
370 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
371 nr_scanned, nr_eligible,
372 freeable, delta, total_scan);
375 * Normally, we should not scan less than batch_size objects in one
376 * pass to avoid too frequent shrinker calls, but if the slab has less
377 * than batch_size objects in total and we are really tight on memory,
378 * we will try to reclaim all available objects, otherwise we can end
379 * up failing allocations although there are plenty of reclaimable
380 * objects spread over several slabs with usage less than the
383 * We detect the "tight on memory" situations by looking at the total
384 * number of objects we want to scan (total_scan). If it is greater
385 * than the total number of objects on slab (freeable), we must be
386 * scanning at high prio and therefore should try to reclaim as much as
389 while (total_scan >= batch_size ||
390 total_scan >= freeable) {
392 unsigned long nr_to_scan = min(batch_size, total_scan);
394 shrinkctl->nr_to_scan = nr_to_scan;
395 ret = shrinker->scan_objects(shrinker, shrinkctl);
396 if (ret == SHRINK_STOP)
400 count_vm_events(SLABS_SCANNED, nr_to_scan);
401 total_scan -= nr_to_scan;
402 scanned += nr_to_scan;
407 if (next_deferred >= scanned)
408 next_deferred -= scanned;
412 * move the unused scan count back into the shrinker in a
413 * manner that handles concurrent updates. If we exhausted the
414 * scan, there is no need to do an update.
416 if (next_deferred > 0)
417 new_nr = atomic_long_add_return(next_deferred,
418 &shrinker->nr_deferred[nid]);
420 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
422 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
427 * shrink_slab - shrink slab caches
428 * @gfp_mask: allocation context
429 * @nid: node whose slab caches to target
430 * @memcg: memory cgroup whose slab caches to target
431 * @nr_scanned: pressure numerator
432 * @nr_eligible: pressure denominator
434 * Call the shrink functions to age shrinkable caches.
436 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
437 * unaware shrinkers will receive a node id of 0 instead.
439 * @memcg specifies the memory cgroup to target. If it is not NULL,
440 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
441 * objects from the memory cgroup specified. Otherwise, only unaware
442 * shrinkers are called.
444 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
445 * the available objects should be scanned. Page reclaim for example
446 * passes the number of pages scanned and the number of pages on the
447 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
448 * when it encountered mapped pages. The ratio is further biased by
449 * the ->seeks setting of the shrink function, which indicates the
450 * cost to recreate an object relative to that of an LRU page.
452 * Returns the number of reclaimed slab objects.
454 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
455 struct mem_cgroup *memcg,
456 unsigned long nr_scanned,
457 unsigned long nr_eligible)
459 struct shrinker *shrinker;
460 unsigned long freed = 0;
462 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
466 nr_scanned = SWAP_CLUSTER_MAX;
468 if (!down_read_trylock(&shrinker_rwsem)) {
470 * If we would return 0, our callers would understand that we
471 * have nothing else to shrink and give up trying. By returning
472 * 1 we keep it going and assume we'll be able to shrink next
479 list_for_each_entry(shrinker, &shrinker_list, list) {
480 struct shrink_control sc = {
481 .gfp_mask = gfp_mask,
487 * If kernel memory accounting is disabled, we ignore
488 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
489 * passing NULL for memcg.
491 if (memcg_kmem_enabled() &&
492 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
495 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
498 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
501 up_read(&shrinker_rwsem);
507 void drop_slab_node(int nid)
512 struct mem_cgroup *memcg = NULL;
516 freed += shrink_slab(GFP_KERNEL, nid, memcg,
518 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
519 } while (freed > 10);
526 for_each_online_node(nid)
530 static inline int is_page_cache_freeable(struct page *page)
533 * A freeable page cache page is referenced only by the caller
534 * that isolated the page, the page cache radix tree and
535 * optional buffer heads at page->private.
537 return page_count(page) - page_has_private(page) == 2;
540 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
542 if (current->flags & PF_SWAPWRITE)
544 if (!inode_write_congested(inode))
546 if (inode_to_bdi(inode) == current->backing_dev_info)
552 * We detected a synchronous write error writing a page out. Probably
553 * -ENOSPC. We need to propagate that into the address_space for a subsequent
554 * fsync(), msync() or close().
556 * The tricky part is that after writepage we cannot touch the mapping: nothing
557 * prevents it from being freed up. But we have a ref on the page and once
558 * that page is locked, the mapping is pinned.
560 * We're allowed to run sleeping lock_page() here because we know the caller has
563 static void handle_write_error(struct address_space *mapping,
564 struct page *page, int error)
567 if (page_mapping(page) == mapping)
568 mapping_set_error(mapping, error);
572 /* possible outcome of pageout() */
574 /* failed to write page out, page is locked */
576 /* move page to the active list, page is locked */
578 /* page has been sent to the disk successfully, page is unlocked */
580 /* page is clean and locked */
585 * pageout is called by shrink_page_list() for each dirty page.
586 * Calls ->writepage().
588 static pageout_t pageout(struct page *page, struct address_space *mapping,
589 struct scan_control *sc)
592 * If the page is dirty, only perform writeback if that write
593 * will be non-blocking. To prevent this allocation from being
594 * stalled by pagecache activity. But note that there may be
595 * stalls if we need to run get_block(). We could test
596 * PagePrivate for that.
598 * If this process is currently in __generic_file_write_iter() against
599 * this page's queue, we can perform writeback even if that
602 * If the page is swapcache, write it back even if that would
603 * block, for some throttling. This happens by accident, because
604 * swap_backing_dev_info is bust: it doesn't reflect the
605 * congestion state of the swapdevs. Easy to fix, if needed.
607 if (!is_page_cache_freeable(page))
611 * Some data journaling orphaned pages can have
612 * page->mapping == NULL while being dirty with clean buffers.
614 if (page_has_private(page)) {
615 if (try_to_free_buffers(page)) {
616 ClearPageDirty(page);
617 pr_info("%s: orphaned page\n", __func__);
623 if (mapping->a_ops->writepage == NULL)
624 return PAGE_ACTIVATE;
625 if (!may_write_to_inode(mapping->host, sc))
628 if (clear_page_dirty_for_io(page)) {
630 struct writeback_control wbc = {
631 .sync_mode = WB_SYNC_NONE,
632 .nr_to_write = SWAP_CLUSTER_MAX,
634 .range_end = LLONG_MAX,
638 SetPageReclaim(page);
639 res = mapping->a_ops->writepage(page, &wbc);
641 handle_write_error(mapping, page, res);
642 if (res == AOP_WRITEPAGE_ACTIVATE) {
643 ClearPageReclaim(page);
644 return PAGE_ACTIVATE;
647 if (!PageWriteback(page)) {
648 /* synchronous write or broken a_ops? */
649 ClearPageReclaim(page);
651 trace_mm_vmscan_writepage(page);
652 inc_node_page_state(page, NR_VMSCAN_WRITE);
660 * Same as remove_mapping, but if the page is removed from the mapping, it
661 * gets returned with a refcount of 0.
663 static int __remove_mapping(struct address_space *mapping, struct page *page,
668 BUG_ON(!PageLocked(page));
669 BUG_ON(mapping != page_mapping(page));
671 spin_lock_irqsave(&mapping->tree_lock, flags);
673 * The non racy check for a busy page.
675 * Must be careful with the order of the tests. When someone has
676 * a ref to the page, it may be possible that they dirty it then
677 * drop the reference. So if PageDirty is tested before page_count
678 * here, then the following race may occur:
680 * get_user_pages(&page);
681 * [user mapping goes away]
683 * !PageDirty(page) [good]
684 * SetPageDirty(page);
686 * !page_count(page) [good, discard it]
688 * [oops, our write_to data is lost]
690 * Reversing the order of the tests ensures such a situation cannot
691 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
692 * load is not satisfied before that of page->_refcount.
694 * Note that if SetPageDirty is always performed via set_page_dirty,
695 * and thus under tree_lock, then this ordering is not required.
697 if (!page_ref_freeze(page, 2))
699 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
700 if (unlikely(PageDirty(page))) {
701 page_ref_unfreeze(page, 2);
705 if (PageSwapCache(page)) {
706 swp_entry_t swap = { .val = page_private(page) };
707 mem_cgroup_swapout(page, swap);
708 __delete_from_swap_cache(page);
709 spin_unlock_irqrestore(&mapping->tree_lock, flags);
710 swapcache_free(swap);
712 void (*freepage)(struct page *);
715 freepage = mapping->a_ops->freepage;
717 * Remember a shadow entry for reclaimed file cache in
718 * order to detect refaults, thus thrashing, later on.
720 * But don't store shadows in an address space that is
721 * already exiting. This is not just an optizimation,
722 * inode reclaim needs to empty out the radix tree or
723 * the nodes are lost. Don't plant shadows behind its
726 * We also don't store shadows for DAX mappings because the
727 * only page cache pages found in these are zero pages
728 * covering holes, and because we don't want to mix DAX
729 * exceptional entries and shadow exceptional entries in the
732 if (reclaimed && page_is_file_cache(page) &&
733 !mapping_exiting(mapping) && !dax_mapping(mapping))
734 shadow = workingset_eviction(mapping, page);
735 __delete_from_page_cache(page, shadow);
736 spin_unlock_irqrestore(&mapping->tree_lock, flags);
738 if (freepage != NULL)
745 spin_unlock_irqrestore(&mapping->tree_lock, flags);
750 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
751 * someone else has a ref on the page, abort and return 0. If it was
752 * successfully detached, return 1. Assumes the caller has a single ref on
755 int remove_mapping(struct address_space *mapping, struct page *page)
757 if (__remove_mapping(mapping, page, false)) {
759 * Unfreezing the refcount with 1 rather than 2 effectively
760 * drops the pagecache ref for us without requiring another
763 page_ref_unfreeze(page, 1);
770 * putback_lru_page - put previously isolated page onto appropriate LRU list
771 * @page: page to be put back to appropriate lru list
773 * Add previously isolated @page to appropriate LRU list.
774 * Page may still be unevictable for other reasons.
776 * lru_lock must not be held, interrupts must be enabled.
778 void putback_lru_page(struct page *page)
781 int was_unevictable = PageUnevictable(page);
783 VM_BUG_ON_PAGE(PageLRU(page), page);
786 ClearPageUnevictable(page);
788 if (page_evictable(page)) {
790 * For evictable pages, we can use the cache.
791 * In event of a race, worst case is we end up with an
792 * unevictable page on [in]active list.
793 * We know how to handle that.
795 is_unevictable = false;
799 * Put unevictable pages directly on zone's unevictable
802 is_unevictable = true;
803 add_page_to_unevictable_list(page);
805 * When racing with an mlock or AS_UNEVICTABLE clearing
806 * (page is unlocked) make sure that if the other thread
807 * does not observe our setting of PG_lru and fails
808 * isolation/check_move_unevictable_pages,
809 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
810 * the page back to the evictable list.
812 * The other side is TestClearPageMlocked() or shmem_lock().
818 * page's status can change while we move it among lru. If an evictable
819 * page is on unevictable list, it never be freed. To avoid that,
820 * check after we added it to the list, again.
822 if (is_unevictable && page_evictable(page)) {
823 if (!isolate_lru_page(page)) {
827 /* This means someone else dropped this page from LRU
828 * So, it will be freed or putback to LRU again. There is
829 * nothing to do here.
833 if (was_unevictable && !is_unevictable)
834 count_vm_event(UNEVICTABLE_PGRESCUED);
835 else if (!was_unevictable && is_unevictable)
836 count_vm_event(UNEVICTABLE_PGCULLED);
838 put_page(page); /* drop ref from isolate */
841 enum page_references {
843 PAGEREF_RECLAIM_CLEAN,
848 static enum page_references page_check_references(struct page *page,
849 struct scan_control *sc)
851 int referenced_ptes, referenced_page;
852 unsigned long vm_flags;
854 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
856 referenced_page = TestClearPageReferenced(page);
859 * Mlock lost the isolation race with us. Let try_to_unmap()
860 * move the page to the unevictable list.
862 if (vm_flags & VM_LOCKED)
863 return PAGEREF_RECLAIM;
865 if (referenced_ptes) {
866 if (PageSwapBacked(page))
867 return PAGEREF_ACTIVATE;
869 * All mapped pages start out with page table
870 * references from the instantiating fault, so we need
871 * to look twice if a mapped file page is used more
874 * Mark it and spare it for another trip around the
875 * inactive list. Another page table reference will
876 * lead to its activation.
878 * Note: the mark is set for activated pages as well
879 * so that recently deactivated but used pages are
882 SetPageReferenced(page);
884 if (referenced_page || referenced_ptes > 1)
885 return PAGEREF_ACTIVATE;
888 * Activate file-backed executable pages after first usage.
890 if (vm_flags & VM_EXEC)
891 return PAGEREF_ACTIVATE;
896 /* Reclaim if clean, defer dirty pages to writeback */
897 if (referenced_page && !PageSwapBacked(page))
898 return PAGEREF_RECLAIM_CLEAN;
900 return PAGEREF_RECLAIM;
903 /* Check if a page is dirty or under writeback */
904 static void page_check_dirty_writeback(struct page *page,
905 bool *dirty, bool *writeback)
907 struct address_space *mapping;
910 * Anonymous pages are not handled by flushers and must be written
911 * from reclaim context. Do not stall reclaim based on them
913 if (!page_is_file_cache(page)) {
919 /* By default assume that the page flags are accurate */
920 *dirty = PageDirty(page);
921 *writeback = PageWriteback(page);
923 /* Verify dirty/writeback state if the filesystem supports it */
924 if (!page_has_private(page))
927 mapping = page_mapping(page);
928 if (mapping && mapping->a_ops->is_dirty_writeback)
929 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
932 struct reclaim_stat {
934 unsigned nr_unqueued_dirty;
935 unsigned nr_congested;
936 unsigned nr_writeback;
937 unsigned nr_immediate;
938 unsigned nr_activate;
939 unsigned nr_ref_keep;
940 unsigned nr_unmap_fail;
944 * shrink_page_list() returns the number of reclaimed pages
946 static unsigned long shrink_page_list(struct list_head *page_list,
947 struct pglist_data *pgdat,
948 struct scan_control *sc,
949 enum ttu_flags ttu_flags,
950 struct reclaim_stat *stat,
953 LIST_HEAD(ret_pages);
954 LIST_HEAD(free_pages);
956 unsigned nr_unqueued_dirty = 0;
957 unsigned nr_dirty = 0;
958 unsigned nr_congested = 0;
959 unsigned nr_reclaimed = 0;
960 unsigned nr_writeback = 0;
961 unsigned nr_immediate = 0;
962 unsigned nr_ref_keep = 0;
963 unsigned nr_unmap_fail = 0;
967 while (!list_empty(page_list)) {
968 struct address_space *mapping;
971 enum page_references references = PAGEREF_RECLAIM_CLEAN;
972 bool dirty, writeback;
973 bool lazyfree = false;
974 int ret = SWAP_SUCCESS;
978 page = lru_to_page(page_list);
979 list_del(&page->lru);
981 if (!trylock_page(page))
984 VM_BUG_ON_PAGE(PageActive(page), page);
988 if (unlikely(!page_evictable(page)))
991 if (!sc->may_unmap && page_mapped(page))
994 /* Double the slab pressure for mapped and swapcache pages */
995 if (page_mapped(page) || PageSwapCache(page))
998 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
999 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1002 * The number of dirty pages determines if a zone is marked
1003 * reclaim_congested which affects wait_iff_congested. kswapd
1004 * will stall and start writing pages if the tail of the LRU
1005 * is all dirty unqueued pages.
1007 page_check_dirty_writeback(page, &dirty, &writeback);
1008 if (dirty || writeback)
1011 if (dirty && !writeback)
1012 nr_unqueued_dirty++;
1015 * Treat this page as congested if the underlying BDI is or if
1016 * pages are cycling through the LRU so quickly that the
1017 * pages marked for immediate reclaim are making it to the
1018 * end of the LRU a second time.
1020 mapping = page_mapping(page);
1021 if (((dirty || writeback) && mapping &&
1022 inode_write_congested(mapping->host)) ||
1023 (writeback && PageReclaim(page)))
1027 * If a page at the tail of the LRU is under writeback, there
1028 * are three cases to consider.
1030 * 1) If reclaim is encountering an excessive number of pages
1031 * under writeback and this page is both under writeback and
1032 * PageReclaim then it indicates that pages are being queued
1033 * for IO but are being recycled through the LRU before the
1034 * IO can complete. Waiting on the page itself risks an
1035 * indefinite stall if it is impossible to writeback the
1036 * page due to IO error or disconnected storage so instead
1037 * note that the LRU is being scanned too quickly and the
1038 * caller can stall after page list has been processed.
1040 * 2) Global or new memcg reclaim encounters a page that is
1041 * not marked for immediate reclaim, or the caller does not
1042 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1043 * not to fs). In this case mark the page for immediate
1044 * reclaim and continue scanning.
1046 * Require may_enter_fs because we would wait on fs, which
1047 * may not have submitted IO yet. And the loop driver might
1048 * enter reclaim, and deadlock if it waits on a page for
1049 * which it is needed to do the write (loop masks off
1050 * __GFP_IO|__GFP_FS for this reason); but more thought
1051 * would probably show more reasons.
1053 * 3) Legacy memcg encounters a page that is already marked
1054 * PageReclaim. memcg does not have any dirty pages
1055 * throttling so we could easily OOM just because too many
1056 * pages are in writeback and there is nothing else to
1057 * reclaim. Wait for the writeback to complete.
1059 if (PageWriteback(page)) {
1061 if (current_is_kswapd() &&
1062 PageReclaim(page) &&
1063 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1068 } else if (sane_reclaim(sc) ||
1069 !PageReclaim(page) || !may_enter_fs) {
1071 * This is slightly racy - end_page_writeback()
1072 * might have just cleared PageReclaim, then
1073 * setting PageReclaim here end up interpreted
1074 * as PageReadahead - but that does not matter
1075 * enough to care. What we do want is for this
1076 * page to have PageReclaim set next time memcg
1077 * reclaim reaches the tests above, so it will
1078 * then wait_on_page_writeback() to avoid OOM;
1079 * and it's also appropriate in global reclaim.
1081 SetPageReclaim(page);
1088 wait_on_page_writeback(page);
1089 /* then go back and try same page again */
1090 list_add_tail(&page->lru, page_list);
1096 references = page_check_references(page, sc);
1098 switch (references) {
1099 case PAGEREF_ACTIVATE:
1100 goto activate_locked;
1104 case PAGEREF_RECLAIM:
1105 case PAGEREF_RECLAIM_CLEAN:
1106 ; /* try to reclaim the page below */
1110 * Anonymous process memory has backing store?
1111 * Try to allocate it some swap space here.
1113 if (PageAnon(page) && !PageSwapCache(page)) {
1114 if (!(sc->gfp_mask & __GFP_IO))
1116 if (!add_to_swap(page, page_list))
1117 goto activate_locked;
1121 /* Adding to swap updated mapping */
1122 mapping = page_mapping(page);
1123 } else if (unlikely(PageTransHuge(page))) {
1124 /* Split file THP */
1125 if (split_huge_page_to_list(page, page_list))
1129 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1132 * The page is mapped into the page tables of one or more
1133 * processes. Try to unmap it here.
1135 if (page_mapped(page) && mapping) {
1136 switch (ret = try_to_unmap(page, lazyfree ?
1137 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1138 (ttu_flags | TTU_BATCH_FLUSH))) {
1141 goto activate_locked;
1149 ; /* try to free the page below */
1153 if (PageDirty(page)) {
1155 * Only kswapd can writeback filesystem pages to
1156 * avoid risk of stack overflow but only writeback
1157 * if many dirty pages have been encountered.
1159 if (page_is_file_cache(page) &&
1160 (!current_is_kswapd() ||
1161 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1163 * Immediately reclaim when written back.
1164 * Similar in principal to deactivate_page()
1165 * except we already have the page isolated
1166 * and know it's dirty
1168 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1169 SetPageReclaim(page);
1174 if (references == PAGEREF_RECLAIM_CLEAN)
1178 if (!sc->may_writepage)
1182 * Page is dirty. Flush the TLB if a writable entry
1183 * potentially exists to avoid CPU writes after IO
1184 * starts and then write it out here.
1186 try_to_unmap_flush_dirty();
1187 switch (pageout(page, mapping, sc)) {
1191 goto activate_locked;
1193 if (PageWriteback(page))
1195 if (PageDirty(page))
1199 * A synchronous write - probably a ramdisk. Go
1200 * ahead and try to reclaim the page.
1202 if (!trylock_page(page))
1204 if (PageDirty(page) || PageWriteback(page))
1206 mapping = page_mapping(page);
1208 ; /* try to free the page below */
1213 * If the page has buffers, try to free the buffer mappings
1214 * associated with this page. If we succeed we try to free
1217 * We do this even if the page is PageDirty().
1218 * try_to_release_page() does not perform I/O, but it is
1219 * possible for a page to have PageDirty set, but it is actually
1220 * clean (all its buffers are clean). This happens if the
1221 * buffers were written out directly, with submit_bh(). ext3
1222 * will do this, as well as the blockdev mapping.
1223 * try_to_release_page() will discover that cleanness and will
1224 * drop the buffers and mark the page clean - it can be freed.
1226 * Rarely, pages can have buffers and no ->mapping. These are
1227 * the pages which were not successfully invalidated in
1228 * truncate_complete_page(). We try to drop those buffers here
1229 * and if that worked, and the page is no longer mapped into
1230 * process address space (page_count == 1) it can be freed.
1231 * Otherwise, leave the page on the LRU so it is swappable.
1233 if (page_has_private(page)) {
1234 if (!try_to_release_page(page, sc->gfp_mask))
1235 goto activate_locked;
1236 if (!mapping && page_count(page) == 1) {
1238 if (put_page_testzero(page))
1242 * rare race with speculative reference.
1243 * the speculative reference will free
1244 * this page shortly, so we may
1245 * increment nr_reclaimed here (and
1246 * leave it off the LRU).
1255 if (!mapping || !__remove_mapping(mapping, page, true))
1259 * At this point, we have no other references and there is
1260 * no way to pick any more up (removed from LRU, removed
1261 * from pagecache). Can use non-atomic bitops now (and
1262 * we obviously don't have to worry about waking up a process
1263 * waiting on the page lock, because there are no references.
1265 __ClearPageLocked(page);
1267 if (ret == SWAP_LZFREE)
1268 count_vm_event(PGLAZYFREED);
1273 * Is there need to periodically free_page_list? It would
1274 * appear not as the counts should be low
1276 list_add(&page->lru, &free_pages);
1280 if (PageSwapCache(page))
1281 try_to_free_swap(page);
1283 list_add(&page->lru, &ret_pages);
1287 /* Not a candidate for swapping, so reclaim swap space. */
1288 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1289 try_to_free_swap(page);
1290 VM_BUG_ON_PAGE(PageActive(page), page);
1291 SetPageActive(page);
1296 list_add(&page->lru, &ret_pages);
1297 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1300 mem_cgroup_uncharge_list(&free_pages);
1301 try_to_unmap_flush();
1302 free_hot_cold_page_list(&free_pages, true);
1304 list_splice(&ret_pages, page_list);
1305 count_vm_events(PGACTIVATE, pgactivate);
1308 stat->nr_dirty = nr_dirty;
1309 stat->nr_congested = nr_congested;
1310 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1311 stat->nr_writeback = nr_writeback;
1312 stat->nr_immediate = nr_immediate;
1313 stat->nr_activate = pgactivate;
1314 stat->nr_ref_keep = nr_ref_keep;
1315 stat->nr_unmap_fail = nr_unmap_fail;
1317 return nr_reclaimed;
1320 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1321 struct list_head *page_list)
1323 struct scan_control sc = {
1324 .gfp_mask = GFP_KERNEL,
1325 .priority = DEF_PRIORITY,
1329 struct page *page, *next;
1330 LIST_HEAD(clean_pages);
1332 list_for_each_entry_safe(page, next, page_list, lru) {
1333 if (page_is_file_cache(page) && !PageDirty(page) &&
1334 !__PageMovable(page)) {
1335 ClearPageActive(page);
1336 list_move(&page->lru, &clean_pages);
1340 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1341 TTU_UNMAP|TTU_IGNORE_ACCESS, NULL, true);
1342 list_splice(&clean_pages, page_list);
1343 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1348 * Attempt to remove the specified page from its LRU. Only take this page
1349 * if it is of the appropriate PageActive status. Pages which are being
1350 * freed elsewhere are also ignored.
1352 * page: page to consider
1353 * mode: one of the LRU isolation modes defined above
1355 * returns 0 on success, -ve errno on failure.
1357 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1361 /* Only take pages on the LRU. */
1365 /* Compaction should not handle unevictable pages but CMA can do so */
1366 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1372 * To minimise LRU disruption, the caller can indicate that it only
1373 * wants to isolate pages it will be able to operate on without
1374 * blocking - clean pages for the most part.
1376 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1377 * is used by reclaim when it is cannot write to backing storage
1379 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1380 * that it is possible to migrate without blocking
1382 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1383 /* All the caller can do on PageWriteback is block */
1384 if (PageWriteback(page))
1387 if (PageDirty(page)) {
1388 struct address_space *mapping;
1390 /* ISOLATE_CLEAN means only clean pages */
1391 if (mode & ISOLATE_CLEAN)
1395 * Only pages without mappings or that have a
1396 * ->migratepage callback are possible to migrate
1399 mapping = page_mapping(page);
1400 if (mapping && !mapping->a_ops->migratepage)
1405 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1408 if (likely(get_page_unless_zero(page))) {
1410 * Be careful not to clear PageLRU until after we're
1411 * sure the page is not being freed elsewhere -- the
1412 * page release code relies on it.
1423 * Update LRU sizes after isolating pages. The LRU size updates must
1424 * be complete before mem_cgroup_update_lru_size due to a santity check.
1426 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1427 enum lru_list lru, unsigned long *nr_zone_taken)
1431 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1432 if (!nr_zone_taken[zid])
1435 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1437 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1444 * zone_lru_lock is heavily contended. Some of the functions that
1445 * shrink the lists perform better by taking out a batch of pages
1446 * and working on them outside the LRU lock.
1448 * For pagecache intensive workloads, this function is the hottest
1449 * spot in the kernel (apart from copy_*_user functions).
1451 * Appropriate locks must be held before calling this function.
1453 * @nr_to_scan: The number of pages to look through on the list.
1454 * @lruvec: The LRU vector to pull pages from.
1455 * @dst: The temp list to put pages on to.
1456 * @nr_scanned: The number of pages that were scanned.
1457 * @sc: The scan_control struct for this reclaim session
1458 * @mode: One of the LRU isolation modes
1459 * @lru: LRU list id for isolating
1461 * returns how many pages were moved onto *@dst.
1463 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1464 struct lruvec *lruvec, struct list_head *dst,
1465 unsigned long *nr_scanned, struct scan_control *sc,
1466 isolate_mode_t mode, enum lru_list lru)
1468 struct list_head *src = &lruvec->lists[lru];
1469 unsigned long nr_taken = 0;
1470 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1471 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1472 unsigned long skipped = 0, total_skipped = 0;
1473 unsigned long scan, nr_pages;
1474 LIST_HEAD(pages_skipped);
1476 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1477 !list_empty(src);) {
1480 page = lru_to_page(src);
1481 prefetchw_prev_lru_page(page, src, flags);
1483 VM_BUG_ON_PAGE(!PageLRU(page), page);
1485 if (page_zonenum(page) > sc->reclaim_idx) {
1486 list_move(&page->lru, &pages_skipped);
1487 nr_skipped[page_zonenum(page)]++;
1492 * Account for scanned and skipped separetly to avoid the pgdat
1493 * being prematurely marked unreclaimable by pgdat_reclaimable.
1497 switch (__isolate_lru_page(page, mode)) {
1499 nr_pages = hpage_nr_pages(page);
1500 nr_taken += nr_pages;
1501 nr_zone_taken[page_zonenum(page)] += nr_pages;
1502 list_move(&page->lru, dst);
1506 /* else it is being freed elsewhere */
1507 list_move(&page->lru, src);
1516 * Splice any skipped pages to the start of the LRU list. Note that
1517 * this disrupts the LRU order when reclaiming for lower zones but
1518 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1519 * scanning would soon rescan the same pages to skip and put the
1520 * system at risk of premature OOM.
1522 if (!list_empty(&pages_skipped)) {
1525 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1526 if (!nr_skipped[zid])
1529 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1530 skipped += nr_skipped[zid];
1534 * Account skipped pages as a partial scan as the pgdat may be
1535 * close to unreclaimable. If the LRU list is empty, account
1536 * skipped pages as a full scan.
1538 total_skipped = list_empty(src) ? skipped : skipped >> 2;
1540 list_splice(&pages_skipped, src);
1542 *nr_scanned = scan + total_skipped;
1543 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1544 scan, skipped, nr_taken, mode, lru);
1545 update_lru_sizes(lruvec, lru, nr_zone_taken);
1550 * isolate_lru_page - tries to isolate a page from its LRU list
1551 * @page: page to isolate from its LRU list
1553 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1554 * vmstat statistic corresponding to whatever LRU list the page was on.
1556 * Returns 0 if the page was removed from an LRU list.
1557 * Returns -EBUSY if the page was not on an LRU list.
1559 * The returned page will have PageLRU() cleared. If it was found on
1560 * the active list, it will have PageActive set. If it was found on
1561 * the unevictable list, it will have the PageUnevictable bit set. That flag
1562 * may need to be cleared by the caller before letting the page go.
1564 * The vmstat statistic corresponding to the list on which the page was
1565 * found will be decremented.
1568 * (1) Must be called with an elevated refcount on the page. This is a
1569 * fundamentnal difference from isolate_lru_pages (which is called
1570 * without a stable reference).
1571 * (2) the lru_lock must not be held.
1572 * (3) interrupts must be enabled.
1574 int isolate_lru_page(struct page *page)
1578 VM_BUG_ON_PAGE(!page_count(page), page);
1579 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1581 if (PageLRU(page)) {
1582 struct zone *zone = page_zone(page);
1583 struct lruvec *lruvec;
1585 spin_lock_irq(zone_lru_lock(zone));
1586 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1587 if (PageLRU(page)) {
1588 int lru = page_lru(page);
1591 del_page_from_lru_list(page, lruvec, lru);
1594 spin_unlock_irq(zone_lru_lock(zone));
1600 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1601 * then get resheduled. When there are massive number of tasks doing page
1602 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1603 * the LRU list will go small and be scanned faster than necessary, leading to
1604 * unnecessary swapping, thrashing and OOM.
1606 static int too_many_isolated(struct pglist_data *pgdat, int file,
1607 struct scan_control *sc)
1609 unsigned long inactive, isolated;
1611 if (current_is_kswapd())
1614 if (!sane_reclaim(sc))
1618 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1619 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1621 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1622 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1626 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1627 * won't get blocked by normal direct-reclaimers, forming a circular
1630 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1633 return isolated > inactive;
1636 static noinline_for_stack void
1637 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1639 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1640 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1641 LIST_HEAD(pages_to_free);
1644 * Put back any unfreeable pages.
1646 while (!list_empty(page_list)) {
1647 struct page *page = lru_to_page(page_list);
1650 VM_BUG_ON_PAGE(PageLRU(page), page);
1651 list_del(&page->lru);
1652 if (unlikely(!page_evictable(page))) {
1653 spin_unlock_irq(&pgdat->lru_lock);
1654 putback_lru_page(page);
1655 spin_lock_irq(&pgdat->lru_lock);
1659 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1662 lru = page_lru(page);
1663 add_page_to_lru_list(page, lruvec, lru);
1665 if (is_active_lru(lru)) {
1666 int file = is_file_lru(lru);
1667 int numpages = hpage_nr_pages(page);
1668 reclaim_stat->recent_rotated[file] += numpages;
1670 if (put_page_testzero(page)) {
1671 __ClearPageLRU(page);
1672 __ClearPageActive(page);
1673 del_page_from_lru_list(page, lruvec, lru);
1675 if (unlikely(PageCompound(page))) {
1676 spin_unlock_irq(&pgdat->lru_lock);
1677 mem_cgroup_uncharge(page);
1678 (*get_compound_page_dtor(page))(page);
1679 spin_lock_irq(&pgdat->lru_lock);
1681 list_add(&page->lru, &pages_to_free);
1686 * To save our caller's stack, now use input list for pages to free.
1688 list_splice(&pages_to_free, page_list);
1692 * If a kernel thread (such as nfsd for loop-back mounts) services
1693 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1694 * In that case we should only throttle if the backing device it is
1695 * writing to is congested. In other cases it is safe to throttle.
1697 static int current_may_throttle(void)
1699 return !(current->flags & PF_LESS_THROTTLE) ||
1700 current->backing_dev_info == NULL ||
1701 bdi_write_congested(current->backing_dev_info);
1705 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1706 * of reclaimed pages
1708 static noinline_for_stack unsigned long
1709 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1710 struct scan_control *sc, enum lru_list lru)
1712 LIST_HEAD(page_list);
1713 unsigned long nr_scanned;
1714 unsigned long nr_reclaimed = 0;
1715 unsigned long nr_taken;
1716 struct reclaim_stat stat = {};
1717 isolate_mode_t isolate_mode = 0;
1718 int file = is_file_lru(lru);
1719 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1720 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1722 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1723 congestion_wait(BLK_RW_ASYNC, HZ/10);
1725 /* We are about to die and free our memory. Return now. */
1726 if (fatal_signal_pending(current))
1727 return SWAP_CLUSTER_MAX;
1733 isolate_mode |= ISOLATE_UNMAPPED;
1734 if (!sc->may_writepage)
1735 isolate_mode |= ISOLATE_CLEAN;
1737 spin_lock_irq(&pgdat->lru_lock);
1739 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1740 &nr_scanned, sc, isolate_mode, lru);
1742 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1743 reclaim_stat->recent_scanned[file] += nr_taken;
1745 if (global_reclaim(sc)) {
1746 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1747 if (current_is_kswapd())
1748 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1750 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1752 spin_unlock_irq(&pgdat->lru_lock);
1757 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1760 spin_lock_irq(&pgdat->lru_lock);
1762 if (global_reclaim(sc)) {
1763 if (current_is_kswapd())
1764 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1766 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1769 putback_inactive_pages(lruvec, &page_list);
1771 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1773 spin_unlock_irq(&pgdat->lru_lock);
1775 mem_cgroup_uncharge_list(&page_list);
1776 free_hot_cold_page_list(&page_list, true);
1779 * If reclaim is isolating dirty pages under writeback, it implies
1780 * that the long-lived page allocation rate is exceeding the page
1781 * laundering rate. Either the global limits are not being effective
1782 * at throttling processes due to the page distribution throughout
1783 * zones or there is heavy usage of a slow backing device. The
1784 * only option is to throttle from reclaim context which is not ideal
1785 * as there is no guarantee the dirtying process is throttled in the
1786 * same way balance_dirty_pages() manages.
1788 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1789 * of pages under pages flagged for immediate reclaim and stall if any
1790 * are encountered in the nr_immediate check below.
1792 if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1793 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1796 * Legacy memcg will stall in page writeback so avoid forcibly
1799 if (sane_reclaim(sc)) {
1801 * Tag a zone as congested if all the dirty pages scanned were
1802 * backed by a congested BDI and wait_iff_congested will stall.
1804 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1805 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1808 * If dirty pages are scanned that are not queued for IO, it
1809 * implies that flushers are not keeping up. In this case, flag
1810 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1813 if (stat.nr_unqueued_dirty == nr_taken)
1814 set_bit(PGDAT_DIRTY, &pgdat->flags);
1817 * If kswapd scans pages marked marked for immediate
1818 * reclaim and under writeback (nr_immediate), it implies
1819 * that pages are cycling through the LRU faster than
1820 * they are written so also forcibly stall.
1822 if (stat.nr_immediate && current_may_throttle())
1823 congestion_wait(BLK_RW_ASYNC, HZ/10);
1827 * Stall direct reclaim for IO completions if underlying BDIs or zone
1828 * is congested. Allow kswapd to continue until it starts encountering
1829 * unqueued dirty pages or cycling through the LRU too quickly.
1831 if (!sc->hibernation_mode && !current_is_kswapd() &&
1832 current_may_throttle())
1833 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1835 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1836 nr_scanned, nr_reclaimed,
1837 stat.nr_dirty, stat.nr_writeback,
1838 stat.nr_congested, stat.nr_immediate,
1839 stat.nr_activate, stat.nr_ref_keep,
1841 sc->priority, file);
1842 return nr_reclaimed;
1846 * This moves pages from the active list to the inactive list.
1848 * We move them the other way if the page is referenced by one or more
1849 * processes, from rmap.
1851 * If the pages are mostly unmapped, the processing is fast and it is
1852 * appropriate to hold zone_lru_lock across the whole operation. But if
1853 * the pages are mapped, the processing is slow (page_referenced()) so we
1854 * should drop zone_lru_lock around each page. It's impossible to balance
1855 * this, so instead we remove the pages from the LRU while processing them.
1856 * It is safe to rely on PG_active against the non-LRU pages in here because
1857 * nobody will play with that bit on a non-LRU page.
1859 * The downside is that we have to touch page->_refcount against each page.
1860 * But we had to alter page->flags anyway.
1862 * Returns the number of pages moved to the given lru.
1865 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1866 struct list_head *list,
1867 struct list_head *pages_to_free,
1870 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1875 while (!list_empty(list)) {
1876 page = lru_to_page(list);
1877 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1879 VM_BUG_ON_PAGE(PageLRU(page), page);
1882 nr_pages = hpage_nr_pages(page);
1883 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1884 list_move(&page->lru, &lruvec->lists[lru]);
1886 if (put_page_testzero(page)) {
1887 __ClearPageLRU(page);
1888 __ClearPageActive(page);
1889 del_page_from_lru_list(page, lruvec, lru);
1891 if (unlikely(PageCompound(page))) {
1892 spin_unlock_irq(&pgdat->lru_lock);
1893 mem_cgroup_uncharge(page);
1894 (*get_compound_page_dtor(page))(page);
1895 spin_lock_irq(&pgdat->lru_lock);
1897 list_add(&page->lru, pages_to_free);
1899 nr_moved += nr_pages;
1903 if (!is_active_lru(lru))
1904 __count_vm_events(PGDEACTIVATE, nr_moved);
1909 static void shrink_active_list(unsigned long nr_to_scan,
1910 struct lruvec *lruvec,
1911 struct scan_control *sc,
1914 unsigned long nr_taken;
1915 unsigned long nr_scanned;
1916 unsigned long vm_flags;
1917 LIST_HEAD(l_hold); /* The pages which were snipped off */
1918 LIST_HEAD(l_active);
1919 LIST_HEAD(l_inactive);
1921 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1922 unsigned nr_deactivate, nr_activate;
1923 unsigned nr_rotated = 0;
1924 isolate_mode_t isolate_mode = 0;
1925 int file = is_file_lru(lru);
1926 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1931 isolate_mode |= ISOLATE_UNMAPPED;
1932 if (!sc->may_writepage)
1933 isolate_mode |= ISOLATE_CLEAN;
1935 spin_lock_irq(&pgdat->lru_lock);
1937 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1938 &nr_scanned, sc, isolate_mode, lru);
1940 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1941 reclaim_stat->recent_scanned[file] += nr_taken;
1943 if (global_reclaim(sc))
1944 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1945 __count_vm_events(PGREFILL, nr_scanned);
1947 spin_unlock_irq(&pgdat->lru_lock);
1949 while (!list_empty(&l_hold)) {
1951 page = lru_to_page(&l_hold);
1952 list_del(&page->lru);
1954 if (unlikely(!page_evictable(page))) {
1955 putback_lru_page(page);
1959 if (unlikely(buffer_heads_over_limit)) {
1960 if (page_has_private(page) && trylock_page(page)) {
1961 if (page_has_private(page))
1962 try_to_release_page(page, 0);
1967 if (page_referenced(page, 0, sc->target_mem_cgroup,
1969 nr_rotated += hpage_nr_pages(page);
1971 * Identify referenced, file-backed active pages and
1972 * give them one more trip around the active list. So
1973 * that executable code get better chances to stay in
1974 * memory under moderate memory pressure. Anon pages
1975 * are not likely to be evicted by use-once streaming
1976 * IO, plus JVM can create lots of anon VM_EXEC pages,
1977 * so we ignore them here.
1979 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1980 list_add(&page->lru, &l_active);
1985 ClearPageActive(page); /* we are de-activating */
1986 list_add(&page->lru, &l_inactive);
1990 * Move pages back to the lru list.
1992 spin_lock_irq(&pgdat->lru_lock);
1994 * Count referenced pages from currently used mappings as rotated,
1995 * even though only some of them are actually re-activated. This
1996 * helps balance scan pressure between file and anonymous pages in
1999 reclaim_stat->recent_rotated[file] += nr_rotated;
2001 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2002 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2003 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2004 spin_unlock_irq(&pgdat->lru_lock);
2006 mem_cgroup_uncharge_list(&l_hold);
2007 free_hot_cold_page_list(&l_hold, true);
2008 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2009 nr_deactivate, nr_rotated, sc->priority, file);
2013 * The inactive anon list should be small enough that the VM never has
2014 * to do too much work.
2016 * The inactive file list should be small enough to leave most memory
2017 * to the established workingset on the scan-resistant active list,
2018 * but large enough to avoid thrashing the aggregate readahead window.
2020 * Both inactive lists should also be large enough that each inactive
2021 * page has a chance to be referenced again before it is reclaimed.
2023 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2024 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2025 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2028 * memory ratio inactive
2029 * -------------------------------------
2038 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2039 struct scan_control *sc, bool trace)
2041 unsigned long inactive_ratio;
2042 unsigned long inactive, active;
2043 enum lru_list inactive_lru = file * LRU_FILE;
2044 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2048 * If we don't have swap space, anonymous page deactivation
2051 if (!file && !total_swap_pages)
2054 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2055 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2057 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2059 inactive_ratio = int_sqrt(10 * gb);
2064 trace_mm_vmscan_inactive_list_is_low(lruvec_pgdat(lruvec)->node_id,
2066 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2067 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2068 inactive_ratio, file);
2070 return inactive * inactive_ratio < active;
2073 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2074 struct lruvec *lruvec, struct scan_control *sc)
2076 if (is_active_lru(lru)) {
2077 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2078 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2082 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2093 * Determine how aggressively the anon and file LRU lists should be
2094 * scanned. The relative value of each set of LRU lists is determined
2095 * by looking at the fraction of the pages scanned we did rotate back
2096 * onto the active list instead of evict.
2098 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2099 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2101 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2102 struct scan_control *sc, unsigned long *nr,
2103 unsigned long *lru_pages)
2105 int swappiness = mem_cgroup_swappiness(memcg);
2106 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2108 u64 denominator = 0; /* gcc */
2109 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2110 unsigned long anon_prio, file_prio;
2111 enum scan_balance scan_balance;
2112 unsigned long anon, file;
2113 bool force_scan = false;
2114 unsigned long ap, fp;
2120 * If the zone or memcg is small, nr[l] can be 0. This
2121 * results in no scanning on this priority and a potential
2122 * priority drop. Global direct reclaim can go to the next
2123 * zone and tends to have no problems. Global kswapd is for
2124 * zone balancing and it needs to scan a minimum amount. When
2125 * reclaiming for a memcg, a priority drop can cause high
2126 * latencies, so it's better to scan a minimum amount there as
2129 if (current_is_kswapd()) {
2130 if (!pgdat_reclaimable(pgdat))
2132 if (!mem_cgroup_online(memcg))
2135 if (!global_reclaim(sc))
2138 /* If we have no swap space, do not bother scanning anon pages. */
2139 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2140 scan_balance = SCAN_FILE;
2145 * Global reclaim will swap to prevent OOM even with no
2146 * swappiness, but memcg users want to use this knob to
2147 * disable swapping for individual groups completely when
2148 * using the memory controller's swap limit feature would be
2151 if (!global_reclaim(sc) && !swappiness) {
2152 scan_balance = SCAN_FILE;
2157 * Do not apply any pressure balancing cleverness when the
2158 * system is close to OOM, scan both anon and file equally
2159 * (unless the swappiness setting disagrees with swapping).
2161 if (!sc->priority && swappiness) {
2162 scan_balance = SCAN_EQUAL;
2167 * Prevent the reclaimer from falling into the cache trap: as
2168 * cache pages start out inactive, every cache fault will tip
2169 * the scan balance towards the file LRU. And as the file LRU
2170 * shrinks, so does the window for rotation from references.
2171 * This means we have a runaway feedback loop where a tiny
2172 * thrashing file LRU becomes infinitely more attractive than
2173 * anon pages. Try to detect this based on file LRU size.
2175 if (global_reclaim(sc)) {
2176 unsigned long pgdatfile;
2177 unsigned long pgdatfree;
2179 unsigned long total_high_wmark = 0;
2181 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2182 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2183 node_page_state(pgdat, NR_INACTIVE_FILE);
2185 for (z = 0; z < MAX_NR_ZONES; z++) {
2186 struct zone *zone = &pgdat->node_zones[z];
2187 if (!managed_zone(zone))
2190 total_high_wmark += high_wmark_pages(zone);
2193 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2194 scan_balance = SCAN_ANON;
2200 * If there is enough inactive page cache, i.e. if the size of the
2201 * inactive list is greater than that of the active list *and* the
2202 * inactive list actually has some pages to scan on this priority, we
2203 * do not reclaim anything from the anonymous working set right now.
2204 * Without the second condition we could end up never scanning an
2205 * lruvec even if it has plenty of old anonymous pages unless the
2206 * system is under heavy pressure.
2208 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2209 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2210 scan_balance = SCAN_FILE;
2214 scan_balance = SCAN_FRACT;
2217 * With swappiness at 100, anonymous and file have the same priority.
2218 * This scanning priority is essentially the inverse of IO cost.
2220 anon_prio = swappiness;
2221 file_prio = 200 - anon_prio;
2224 * OK, so we have swap space and a fair amount of page cache
2225 * pages. We use the recently rotated / recently scanned
2226 * ratios to determine how valuable each cache is.
2228 * Because workloads change over time (and to avoid overflow)
2229 * we keep these statistics as a floating average, which ends
2230 * up weighing recent references more than old ones.
2232 * anon in [0], file in [1]
2235 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2236 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2237 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2238 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2240 spin_lock_irq(&pgdat->lru_lock);
2241 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2242 reclaim_stat->recent_scanned[0] /= 2;
2243 reclaim_stat->recent_rotated[0] /= 2;
2246 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2247 reclaim_stat->recent_scanned[1] /= 2;
2248 reclaim_stat->recent_rotated[1] /= 2;
2252 * The amount of pressure on anon vs file pages is inversely
2253 * proportional to the fraction of recently scanned pages on
2254 * each list that were recently referenced and in active use.
2256 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2257 ap /= reclaim_stat->recent_rotated[0] + 1;
2259 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2260 fp /= reclaim_stat->recent_rotated[1] + 1;
2261 spin_unlock_irq(&pgdat->lru_lock);
2265 denominator = ap + fp + 1;
2267 some_scanned = false;
2268 /* Only use force_scan on second pass. */
2269 for (pass = 0; !some_scanned && pass < 2; pass++) {
2271 for_each_evictable_lru(lru) {
2272 int file = is_file_lru(lru);
2276 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2277 scan = size >> sc->priority;
2279 if (!scan && pass && force_scan)
2280 scan = min(size, SWAP_CLUSTER_MAX);
2282 switch (scan_balance) {
2284 /* Scan lists relative to size */
2288 * Scan types proportional to swappiness and
2289 * their relative recent reclaim efficiency.
2291 scan = div64_u64(scan * fraction[file],
2296 /* Scan one type exclusively */
2297 if ((scan_balance == SCAN_FILE) != file) {
2303 /* Look ma, no brain */
2311 * Skip the second pass and don't force_scan,
2312 * if we found something to scan.
2314 some_scanned |= !!scan;
2320 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2322 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2323 struct scan_control *sc, unsigned long *lru_pages)
2325 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2326 unsigned long nr[NR_LRU_LISTS];
2327 unsigned long targets[NR_LRU_LISTS];
2328 unsigned long nr_to_scan;
2330 unsigned long nr_reclaimed = 0;
2331 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2332 struct blk_plug plug;
2335 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2337 /* Record the original scan target for proportional adjustments later */
2338 memcpy(targets, nr, sizeof(nr));
2341 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2342 * event that can occur when there is little memory pressure e.g.
2343 * multiple streaming readers/writers. Hence, we do not abort scanning
2344 * when the requested number of pages are reclaimed when scanning at
2345 * DEF_PRIORITY on the assumption that the fact we are direct
2346 * reclaiming implies that kswapd is not keeping up and it is best to
2347 * do a batch of work at once. For memcg reclaim one check is made to
2348 * abort proportional reclaim if either the file or anon lru has already
2349 * dropped to zero at the first pass.
2351 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2352 sc->priority == DEF_PRIORITY);
2354 blk_start_plug(&plug);
2355 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2356 nr[LRU_INACTIVE_FILE]) {
2357 unsigned long nr_anon, nr_file, percentage;
2358 unsigned long nr_scanned;
2360 for_each_evictable_lru(lru) {
2362 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2363 nr[lru] -= nr_to_scan;
2365 nr_reclaimed += shrink_list(lru, nr_to_scan,
2372 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2376 * For kswapd and memcg, reclaim at least the number of pages
2377 * requested. Ensure that the anon and file LRUs are scanned
2378 * proportionally what was requested by get_scan_count(). We
2379 * stop reclaiming one LRU and reduce the amount scanning
2380 * proportional to the original scan target.
2382 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2383 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2386 * It's just vindictive to attack the larger once the smaller
2387 * has gone to zero. And given the way we stop scanning the
2388 * smaller below, this makes sure that we only make one nudge
2389 * towards proportionality once we've got nr_to_reclaim.
2391 if (!nr_file || !nr_anon)
2394 if (nr_file > nr_anon) {
2395 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2396 targets[LRU_ACTIVE_ANON] + 1;
2398 percentage = nr_anon * 100 / scan_target;
2400 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2401 targets[LRU_ACTIVE_FILE] + 1;
2403 percentage = nr_file * 100 / scan_target;
2406 /* Stop scanning the smaller of the LRU */
2408 nr[lru + LRU_ACTIVE] = 0;
2411 * Recalculate the other LRU scan count based on its original
2412 * scan target and the percentage scanning already complete
2414 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2415 nr_scanned = targets[lru] - nr[lru];
2416 nr[lru] = targets[lru] * (100 - percentage) / 100;
2417 nr[lru] -= min(nr[lru], nr_scanned);
2420 nr_scanned = targets[lru] - nr[lru];
2421 nr[lru] = targets[lru] * (100 - percentage) / 100;
2422 nr[lru] -= min(nr[lru], nr_scanned);
2424 scan_adjusted = true;
2426 blk_finish_plug(&plug);
2427 sc->nr_reclaimed += nr_reclaimed;
2430 * Even if we did not try to evict anon pages at all, we want to
2431 * rebalance the anon lru active/inactive ratio.
2433 if (inactive_list_is_low(lruvec, false, sc, true))
2434 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2435 sc, LRU_ACTIVE_ANON);
2438 /* Use reclaim/compaction for costly allocs or under memory pressure */
2439 static bool in_reclaim_compaction(struct scan_control *sc)
2441 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2442 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2443 sc->priority < DEF_PRIORITY - 2))
2450 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2451 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2452 * true if more pages should be reclaimed such that when the page allocator
2453 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2454 * It will give up earlier than that if there is difficulty reclaiming pages.
2456 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2457 unsigned long nr_reclaimed,
2458 unsigned long nr_scanned,
2459 struct scan_control *sc)
2461 unsigned long pages_for_compaction;
2462 unsigned long inactive_lru_pages;
2465 /* If not in reclaim/compaction mode, stop */
2466 if (!in_reclaim_compaction(sc))
2469 /* Consider stopping depending on scan and reclaim activity */
2470 if (sc->gfp_mask & __GFP_REPEAT) {
2472 * For __GFP_REPEAT allocations, stop reclaiming if the
2473 * full LRU list has been scanned and we are still failing
2474 * to reclaim pages. This full LRU scan is potentially
2475 * expensive but a __GFP_REPEAT caller really wants to succeed
2477 if (!nr_reclaimed && !nr_scanned)
2481 * For non-__GFP_REPEAT allocations which can presumably
2482 * fail without consequence, stop if we failed to reclaim
2483 * any pages from the last SWAP_CLUSTER_MAX number of
2484 * pages that were scanned. This will return to the
2485 * caller faster at the risk reclaim/compaction and
2486 * the resulting allocation attempt fails
2493 * If we have not reclaimed enough pages for compaction and the
2494 * inactive lists are large enough, continue reclaiming
2496 pages_for_compaction = compact_gap(sc->order);
2497 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2498 if (get_nr_swap_pages() > 0)
2499 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2500 if (sc->nr_reclaimed < pages_for_compaction &&
2501 inactive_lru_pages > pages_for_compaction)
2504 /* If compaction would go ahead or the allocation would succeed, stop */
2505 for (z = 0; z <= sc->reclaim_idx; z++) {
2506 struct zone *zone = &pgdat->node_zones[z];
2507 if (!managed_zone(zone))
2510 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2511 case COMPACT_SUCCESS:
2512 case COMPACT_CONTINUE:
2515 /* check next zone */
2522 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2524 struct reclaim_state *reclaim_state = current->reclaim_state;
2525 unsigned long nr_reclaimed, nr_scanned;
2526 bool reclaimable = false;
2529 struct mem_cgroup *root = sc->target_mem_cgroup;
2530 struct mem_cgroup_reclaim_cookie reclaim = {
2532 .priority = sc->priority,
2534 unsigned long node_lru_pages = 0;
2535 struct mem_cgroup *memcg;
2537 nr_reclaimed = sc->nr_reclaimed;
2538 nr_scanned = sc->nr_scanned;
2540 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2542 unsigned long lru_pages;
2543 unsigned long reclaimed;
2544 unsigned long scanned;
2546 if (mem_cgroup_low(root, memcg)) {
2547 if (!sc->may_thrash)
2549 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2552 reclaimed = sc->nr_reclaimed;
2553 scanned = sc->nr_scanned;
2555 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2556 node_lru_pages += lru_pages;
2559 shrink_slab(sc->gfp_mask, pgdat->node_id,
2560 memcg, sc->nr_scanned - scanned,
2563 /* Record the group's reclaim efficiency */
2564 vmpressure(sc->gfp_mask, memcg, false,
2565 sc->nr_scanned - scanned,
2566 sc->nr_reclaimed - reclaimed);
2569 * Direct reclaim and kswapd have to scan all memory
2570 * cgroups to fulfill the overall scan target for the
2573 * Limit reclaim, on the other hand, only cares about
2574 * nr_to_reclaim pages to be reclaimed and it will
2575 * retry with decreasing priority if one round over the
2576 * whole hierarchy is not sufficient.
2578 if (!global_reclaim(sc) &&
2579 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2580 mem_cgroup_iter_break(root, memcg);
2583 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2586 * Shrink the slab caches in the same proportion that
2587 * the eligible LRU pages were scanned.
2589 if (global_reclaim(sc))
2590 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2591 sc->nr_scanned - nr_scanned,
2594 if (reclaim_state) {
2595 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2596 reclaim_state->reclaimed_slab = 0;
2599 /* Record the subtree's reclaim efficiency */
2600 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2601 sc->nr_scanned - nr_scanned,
2602 sc->nr_reclaimed - nr_reclaimed);
2604 if (sc->nr_reclaimed - nr_reclaimed)
2607 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2608 sc->nr_scanned - nr_scanned, sc));
2614 * Returns true if compaction should go ahead for a costly-order request, or
2615 * the allocation would already succeed without compaction. Return false if we
2616 * should reclaim first.
2618 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2620 unsigned long watermark;
2621 enum compact_result suitable;
2623 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2624 if (suitable == COMPACT_SUCCESS)
2625 /* Allocation should succeed already. Don't reclaim. */
2627 if (suitable == COMPACT_SKIPPED)
2628 /* Compaction cannot yet proceed. Do reclaim. */
2632 * Compaction is already possible, but it takes time to run and there
2633 * are potentially other callers using the pages just freed. So proceed
2634 * with reclaim to make a buffer of free pages available to give
2635 * compaction a reasonable chance of completing and allocating the page.
2636 * Note that we won't actually reclaim the whole buffer in one attempt
2637 * as the target watermark in should_continue_reclaim() is lower. But if
2638 * we are already above the high+gap watermark, don't reclaim at all.
2640 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2642 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2646 * This is the direct reclaim path, for page-allocating processes. We only
2647 * try to reclaim pages from zones which will satisfy the caller's allocation
2650 * If a zone is deemed to be full of pinned pages then just give it a light
2651 * scan then give up on it.
2653 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2657 unsigned long nr_soft_reclaimed;
2658 unsigned long nr_soft_scanned;
2660 pg_data_t *last_pgdat = NULL;
2663 * If the number of buffer_heads in the machine exceeds the maximum
2664 * allowed level, force direct reclaim to scan the highmem zone as
2665 * highmem pages could be pinning lowmem pages storing buffer_heads
2667 orig_mask = sc->gfp_mask;
2668 if (buffer_heads_over_limit) {
2669 sc->gfp_mask |= __GFP_HIGHMEM;
2670 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2673 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2674 sc->reclaim_idx, sc->nodemask) {
2676 * Take care memory controller reclaiming has small influence
2679 if (global_reclaim(sc)) {
2680 if (!cpuset_zone_allowed(zone,
2681 GFP_KERNEL | __GFP_HARDWALL))
2684 if (sc->priority != DEF_PRIORITY &&
2685 !pgdat_reclaimable(zone->zone_pgdat))
2686 continue; /* Let kswapd poll it */
2689 * If we already have plenty of memory free for
2690 * compaction in this zone, don't free any more.
2691 * Even though compaction is invoked for any
2692 * non-zero order, only frequent costly order
2693 * reclamation is disruptive enough to become a
2694 * noticeable problem, like transparent huge
2697 if (IS_ENABLED(CONFIG_COMPACTION) &&
2698 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2699 compaction_ready(zone, sc)) {
2700 sc->compaction_ready = true;
2705 * Shrink each node in the zonelist once. If the
2706 * zonelist is ordered by zone (not the default) then a
2707 * node may be shrunk multiple times but in that case
2708 * the user prefers lower zones being preserved.
2710 if (zone->zone_pgdat == last_pgdat)
2714 * This steals pages from memory cgroups over softlimit
2715 * and returns the number of reclaimed pages and
2716 * scanned pages. This works for global memory pressure
2717 * and balancing, not for a memcg's limit.
2719 nr_soft_scanned = 0;
2720 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2721 sc->order, sc->gfp_mask,
2723 sc->nr_reclaimed += nr_soft_reclaimed;
2724 sc->nr_scanned += nr_soft_scanned;
2725 /* need some check for avoid more shrink_zone() */
2728 /* See comment about same check for global reclaim above */
2729 if (zone->zone_pgdat == last_pgdat)
2731 last_pgdat = zone->zone_pgdat;
2732 shrink_node(zone->zone_pgdat, sc);
2736 * Restore to original mask to avoid the impact on the caller if we
2737 * promoted it to __GFP_HIGHMEM.
2739 sc->gfp_mask = orig_mask;
2743 * This is the main entry point to direct page reclaim.
2745 * If a full scan of the inactive list fails to free enough memory then we
2746 * are "out of memory" and something needs to be killed.
2748 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2749 * high - the zone may be full of dirty or under-writeback pages, which this
2750 * caller can't do much about. We kick the writeback threads and take explicit
2751 * naps in the hope that some of these pages can be written. But if the
2752 * allocating task holds filesystem locks which prevent writeout this might not
2753 * work, and the allocation attempt will fail.
2755 * returns: 0, if no pages reclaimed
2756 * else, the number of pages reclaimed
2758 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2759 struct scan_control *sc)
2761 int initial_priority = sc->priority;
2762 unsigned long total_scanned = 0;
2763 unsigned long writeback_threshold;
2765 delayacct_freepages_start();
2767 if (global_reclaim(sc))
2768 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2771 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2774 shrink_zones(zonelist, sc);
2776 total_scanned += sc->nr_scanned;
2777 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2780 if (sc->compaction_ready)
2784 * If we're getting trouble reclaiming, start doing
2785 * writepage even in laptop mode.
2787 if (sc->priority < DEF_PRIORITY - 2)
2788 sc->may_writepage = 1;
2791 * Try to write back as many pages as we just scanned. This
2792 * tends to cause slow streaming writers to write data to the
2793 * disk smoothly, at the dirtying rate, which is nice. But
2794 * that's undesirable in laptop mode, where we *want* lumpy
2795 * writeout. So in laptop mode, write out the whole world.
2797 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2798 if (total_scanned > writeback_threshold) {
2799 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2800 WB_REASON_TRY_TO_FREE_PAGES);
2801 sc->may_writepage = 1;
2803 } while (--sc->priority >= 0);
2805 delayacct_freepages_end();
2807 if (sc->nr_reclaimed)
2808 return sc->nr_reclaimed;
2810 /* Aborted reclaim to try compaction? don't OOM, then */
2811 if (sc->compaction_ready)
2814 /* Untapped cgroup reserves? Don't OOM, retry. */
2815 if (!sc->may_thrash) {
2816 sc->priority = initial_priority;
2824 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2827 unsigned long pfmemalloc_reserve = 0;
2828 unsigned long free_pages = 0;
2832 for (i = 0; i <= ZONE_NORMAL; i++) {
2833 zone = &pgdat->node_zones[i];
2834 if (!managed_zone(zone) ||
2835 pgdat_reclaimable_pages(pgdat) == 0)
2838 pfmemalloc_reserve += min_wmark_pages(zone);
2839 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2842 /* If there are no reserves (unexpected config) then do not throttle */
2843 if (!pfmemalloc_reserve)
2846 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2848 /* kswapd must be awake if processes are being throttled */
2849 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2850 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2851 (enum zone_type)ZONE_NORMAL);
2852 wake_up_interruptible(&pgdat->kswapd_wait);
2859 * Throttle direct reclaimers if backing storage is backed by the network
2860 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2861 * depleted. kswapd will continue to make progress and wake the processes
2862 * when the low watermark is reached.
2864 * Returns true if a fatal signal was delivered during throttling. If this
2865 * happens, the page allocator should not consider triggering the OOM killer.
2867 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2868 nodemask_t *nodemask)
2872 pg_data_t *pgdat = NULL;
2875 * Kernel threads should not be throttled as they may be indirectly
2876 * responsible for cleaning pages necessary for reclaim to make forward
2877 * progress. kjournald for example may enter direct reclaim while
2878 * committing a transaction where throttling it could forcing other
2879 * processes to block on log_wait_commit().
2881 if (current->flags & PF_KTHREAD)
2885 * If a fatal signal is pending, this process should not throttle.
2886 * It should return quickly so it can exit and free its memory
2888 if (fatal_signal_pending(current))
2892 * Check if the pfmemalloc reserves are ok by finding the first node
2893 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2894 * GFP_KERNEL will be required for allocating network buffers when
2895 * swapping over the network so ZONE_HIGHMEM is unusable.
2897 * Throttling is based on the first usable node and throttled processes
2898 * wait on a queue until kswapd makes progress and wakes them. There
2899 * is an affinity then between processes waking up and where reclaim
2900 * progress has been made assuming the process wakes on the same node.
2901 * More importantly, processes running on remote nodes will not compete
2902 * for remote pfmemalloc reserves and processes on different nodes
2903 * should make reasonable progress.
2905 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2906 gfp_zone(gfp_mask), nodemask) {
2907 if (zone_idx(zone) > ZONE_NORMAL)
2910 /* Throttle based on the first usable node */
2911 pgdat = zone->zone_pgdat;
2912 if (pfmemalloc_watermark_ok(pgdat))
2917 /* If no zone was usable by the allocation flags then do not throttle */
2921 /* Account for the throttling */
2922 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2925 * If the caller cannot enter the filesystem, it's possible that it
2926 * is due to the caller holding an FS lock or performing a journal
2927 * transaction in the case of a filesystem like ext[3|4]. In this case,
2928 * it is not safe to block on pfmemalloc_wait as kswapd could be
2929 * blocked waiting on the same lock. Instead, throttle for up to a
2930 * second before continuing.
2932 if (!(gfp_mask & __GFP_FS)) {
2933 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2934 pfmemalloc_watermark_ok(pgdat), HZ);
2939 /* Throttle until kswapd wakes the process */
2940 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2941 pfmemalloc_watermark_ok(pgdat));
2944 if (fatal_signal_pending(current))
2951 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2952 gfp_t gfp_mask, nodemask_t *nodemask)
2954 unsigned long nr_reclaimed;
2955 struct scan_control sc = {
2956 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2957 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2958 .reclaim_idx = gfp_zone(gfp_mask),
2960 .nodemask = nodemask,
2961 .priority = DEF_PRIORITY,
2962 .may_writepage = !laptop_mode,
2968 * Do not enter reclaim if fatal signal was delivered while throttled.
2969 * 1 is returned so that the page allocator does not OOM kill at this
2972 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2975 trace_mm_vmscan_direct_reclaim_begin(order,
2980 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2982 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2984 return nr_reclaimed;
2989 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2990 gfp_t gfp_mask, bool noswap,
2992 unsigned long *nr_scanned)
2994 struct scan_control sc = {
2995 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2996 .target_mem_cgroup = memcg,
2997 .may_writepage = !laptop_mode,
2999 .reclaim_idx = MAX_NR_ZONES - 1,
3000 .may_swap = !noswap,
3002 unsigned long lru_pages;
3004 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3005 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3007 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3013 * NOTE: Although we can get the priority field, using it
3014 * here is not a good idea, since it limits the pages we can scan.
3015 * if we don't reclaim here, the shrink_node from balance_pgdat
3016 * will pick up pages from other mem cgroup's as well. We hack
3017 * the priority and make it zero.
3019 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3021 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3023 *nr_scanned = sc.nr_scanned;
3024 return sc.nr_reclaimed;
3027 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3028 unsigned long nr_pages,
3032 struct zonelist *zonelist;
3033 unsigned long nr_reclaimed;
3035 struct scan_control sc = {
3036 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3037 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3038 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3039 .reclaim_idx = MAX_NR_ZONES - 1,
3040 .target_mem_cgroup = memcg,
3041 .priority = DEF_PRIORITY,
3042 .may_writepage = !laptop_mode,
3044 .may_swap = may_swap,
3048 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3049 * take care of from where we get pages. So the node where we start the
3050 * scan does not need to be the current node.
3052 nid = mem_cgroup_select_victim_node(memcg);
3054 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3056 trace_mm_vmscan_memcg_reclaim_begin(0,
3061 current->flags |= PF_MEMALLOC;
3062 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3063 current->flags &= ~PF_MEMALLOC;
3065 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3067 return nr_reclaimed;
3071 static void age_active_anon(struct pglist_data *pgdat,
3072 struct scan_control *sc)
3074 struct mem_cgroup *memcg;
3076 if (!total_swap_pages)
3079 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3081 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3083 if (inactive_list_is_low(lruvec, false, sc, true))
3084 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3085 sc, LRU_ACTIVE_ANON);
3087 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3091 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3093 unsigned long mark = high_wmark_pages(zone);
3095 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3099 * If any eligible zone is balanced then the node is not considered
3100 * to be congested or dirty
3102 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3103 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3109 * Prepare kswapd for sleeping. This verifies that there are no processes
3110 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3112 * Returns true if kswapd is ready to sleep
3114 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3119 * The throttled processes are normally woken up in balance_pgdat() as
3120 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3121 * race between when kswapd checks the watermarks and a process gets
3122 * throttled. There is also a potential race if processes get
3123 * throttled, kswapd wakes, a large process exits thereby balancing the
3124 * zones, which causes kswapd to exit balance_pgdat() before reaching
3125 * the wake up checks. If kswapd is going to sleep, no process should
3126 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3127 * the wake up is premature, processes will wake kswapd and get
3128 * throttled again. The difference from wake ups in balance_pgdat() is
3129 * that here we are under prepare_to_wait().
3131 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3132 wake_up_all(&pgdat->pfmemalloc_wait);
3134 for (i = 0; i <= classzone_idx; i++) {
3135 struct zone *zone = pgdat->node_zones + i;
3137 if (!managed_zone(zone))
3140 if (!zone_balanced(zone, order, classzone_idx))
3148 * kswapd shrinks a node of pages that are at or below the highest usable
3149 * zone that is currently unbalanced.
3151 * Returns true if kswapd scanned at least the requested number of pages to
3152 * reclaim or if the lack of progress was due to pages under writeback.
3153 * This is used to determine if the scanning priority needs to be raised.
3155 static bool kswapd_shrink_node(pg_data_t *pgdat,
3156 struct scan_control *sc)
3161 /* Reclaim a number of pages proportional to the number of zones */
3162 sc->nr_to_reclaim = 0;
3163 for (z = 0; z <= sc->reclaim_idx; z++) {
3164 zone = pgdat->node_zones + z;
3165 if (!managed_zone(zone))
3168 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3172 * Historically care was taken to put equal pressure on all zones but
3173 * now pressure is applied based on node LRU order.
3175 shrink_node(pgdat, sc);
3178 * Fragmentation may mean that the system cannot be rebalanced for
3179 * high-order allocations. If twice the allocation size has been
3180 * reclaimed then recheck watermarks only at order-0 to prevent
3181 * excessive reclaim. Assume that a process requested a high-order
3182 * can direct reclaim/compact.
3184 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3187 return sc->nr_scanned >= sc->nr_to_reclaim;
3191 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3192 * that are eligible for use by the caller until at least one zone is
3195 * Returns the order kswapd finished reclaiming at.
3197 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3198 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3199 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3200 * or lower is eligible for reclaim until at least one usable zone is
3203 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3206 unsigned long nr_soft_reclaimed;
3207 unsigned long nr_soft_scanned;
3209 struct scan_control sc = {
3210 .gfp_mask = GFP_KERNEL,
3212 .priority = DEF_PRIORITY,
3213 .may_writepage = !laptop_mode,
3217 count_vm_event(PAGEOUTRUN);
3220 bool raise_priority = true;
3222 sc.nr_reclaimed = 0;
3223 sc.reclaim_idx = classzone_idx;
3226 * If the number of buffer_heads exceeds the maximum allowed
3227 * then consider reclaiming from all zones. This has a dual
3228 * purpose -- on 64-bit systems it is expected that
3229 * buffer_heads are stripped during active rotation. On 32-bit
3230 * systems, highmem pages can pin lowmem memory and shrinking
3231 * buffers can relieve lowmem pressure. Reclaim may still not
3232 * go ahead if all eligible zones for the original allocation
3233 * request are balanced to avoid excessive reclaim from kswapd.
3235 if (buffer_heads_over_limit) {
3236 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3237 zone = pgdat->node_zones + i;
3238 if (!managed_zone(zone))
3247 * Only reclaim if there are no eligible zones. Check from
3248 * high to low zone as allocations prefer higher zones.
3249 * Scanning from low to high zone would allow congestion to be
3250 * cleared during a very small window when a small low
3251 * zone was balanced even under extreme pressure when the
3252 * overall node may be congested. Note that sc.reclaim_idx
3253 * is not used as buffer_heads_over_limit may have adjusted
3256 for (i = classzone_idx; i >= 0; i--) {
3257 zone = pgdat->node_zones + i;
3258 if (!managed_zone(zone))
3261 if (zone_balanced(zone, sc.order, classzone_idx))
3266 * Do some background aging of the anon list, to give
3267 * pages a chance to be referenced before reclaiming. All
3268 * pages are rotated regardless of classzone as this is
3269 * about consistent aging.
3271 age_active_anon(pgdat, &sc);
3274 * If we're getting trouble reclaiming, start doing writepage
3275 * even in laptop mode.
3277 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3278 sc.may_writepage = 1;
3280 /* Call soft limit reclaim before calling shrink_node. */
3282 nr_soft_scanned = 0;
3283 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3284 sc.gfp_mask, &nr_soft_scanned);
3285 sc.nr_reclaimed += nr_soft_reclaimed;
3288 * There should be no need to raise the scanning priority if
3289 * enough pages are already being scanned that that high
3290 * watermark would be met at 100% efficiency.
3292 if (kswapd_shrink_node(pgdat, &sc))
3293 raise_priority = false;
3296 * If the low watermark is met there is no need for processes
3297 * to be throttled on pfmemalloc_wait as they should not be
3298 * able to safely make forward progress. Wake them
3300 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3301 pfmemalloc_watermark_ok(pgdat))
3302 wake_up_all(&pgdat->pfmemalloc_wait);
3304 /* Check if kswapd should be suspending */
3305 if (try_to_freeze() || kthread_should_stop())
3309 * Raise priority if scanning rate is too low or there was no
3310 * progress in reclaiming pages
3312 if (raise_priority || !sc.nr_reclaimed)
3314 } while (sc.priority >= 1);
3318 * Return the order kswapd stopped reclaiming at as
3319 * prepare_kswapd_sleep() takes it into account. If another caller
3320 * entered the allocator slow path while kswapd was awake, order will
3321 * remain at the higher level.
3326 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3327 unsigned int classzone_idx)
3332 if (freezing(current) || kthread_should_stop())
3335 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3337 /* Try to sleep for a short interval */
3338 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3340 * Compaction records what page blocks it recently failed to
3341 * isolate pages from and skips them in the future scanning.
3342 * When kswapd is going to sleep, it is reasonable to assume
3343 * that pages and compaction may succeed so reset the cache.
3345 reset_isolation_suitable(pgdat);
3348 * We have freed the memory, now we should compact it to make
3349 * allocation of the requested order possible.
3351 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3353 remaining = schedule_timeout(HZ/10);
3356 * If woken prematurely then reset kswapd_classzone_idx and
3357 * order. The values will either be from a wakeup request or
3358 * the previous request that slept prematurely.
3361 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3362 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3365 finish_wait(&pgdat->kswapd_wait, &wait);
3366 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3370 * After a short sleep, check if it was a premature sleep. If not, then
3371 * go fully to sleep until explicitly woken up.
3374 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3375 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3378 * vmstat counters are not perfectly accurate and the estimated
3379 * value for counters such as NR_FREE_PAGES can deviate from the
3380 * true value by nr_online_cpus * threshold. To avoid the zone
3381 * watermarks being breached while under pressure, we reduce the
3382 * per-cpu vmstat threshold while kswapd is awake and restore
3383 * them before going back to sleep.
3385 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3387 if (!kthread_should_stop())
3390 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3393 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3395 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3397 finish_wait(&pgdat->kswapd_wait, &wait);
3401 * The background pageout daemon, started as a kernel thread
3402 * from the init process.
3404 * This basically trickles out pages so that we have _some_
3405 * free memory available even if there is no other activity
3406 * that frees anything up. This is needed for things like routing
3407 * etc, where we otherwise might have all activity going on in
3408 * asynchronous contexts that cannot page things out.
3410 * If there are applications that are active memory-allocators
3411 * (most normal use), this basically shouldn't matter.
3413 static int kswapd(void *p)
3415 unsigned int alloc_order, reclaim_order, classzone_idx;
3416 pg_data_t *pgdat = (pg_data_t*)p;
3417 struct task_struct *tsk = current;
3419 struct reclaim_state reclaim_state = {
3420 .reclaimed_slab = 0,
3422 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3424 lockdep_set_current_reclaim_state(GFP_KERNEL);
3426 if (!cpumask_empty(cpumask))
3427 set_cpus_allowed_ptr(tsk, cpumask);
3428 current->reclaim_state = &reclaim_state;
3431 * Tell the memory management that we're a "memory allocator",
3432 * and that if we need more memory we should get access to it
3433 * regardless (see "__alloc_pages()"). "kswapd" should
3434 * never get caught in the normal page freeing logic.
3436 * (Kswapd normally doesn't need memory anyway, but sometimes
3437 * you need a small amount of memory in order to be able to
3438 * page out something else, and this flag essentially protects
3439 * us from recursively trying to free more memory as we're
3440 * trying to free the first piece of memory in the first place).
3442 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3445 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3446 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3451 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3454 /* Read the new order and classzone_idx */
3455 alloc_order = reclaim_order = pgdat->kswapd_order;
3456 classzone_idx = pgdat->kswapd_classzone_idx;
3457 pgdat->kswapd_order = 0;
3458 pgdat->kswapd_classzone_idx = 0;
3460 ret = try_to_freeze();
3461 if (kthread_should_stop())
3465 * We can speed up thawing tasks if we don't call balance_pgdat
3466 * after returning from the refrigerator
3472 * Reclaim begins at the requested order but if a high-order
3473 * reclaim fails then kswapd falls back to reclaiming for
3474 * order-0. If that happens, kswapd will consider sleeping
3475 * for the order it finished reclaiming at (reclaim_order)
3476 * but kcompactd is woken to compact for the original
3477 * request (alloc_order).
3479 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3481 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3482 if (reclaim_order < alloc_order)
3483 goto kswapd_try_sleep;
3485 alloc_order = reclaim_order = pgdat->kswapd_order;
3486 classzone_idx = pgdat->kswapd_classzone_idx;
3489 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3490 current->reclaim_state = NULL;
3491 lockdep_clear_current_reclaim_state();
3497 * A zone is low on free memory, so wake its kswapd task to service it.
3499 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3504 if (!managed_zone(zone))
3507 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3509 pgdat = zone->zone_pgdat;
3510 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3511 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3512 if (!waitqueue_active(&pgdat->kswapd_wait))
3515 /* Only wake kswapd if all zones are unbalanced */
3516 for (z = 0; z <= classzone_idx; z++) {
3517 zone = pgdat->node_zones + z;
3518 if (!managed_zone(zone))
3521 if (zone_balanced(zone, order, classzone_idx))
3525 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3526 wake_up_interruptible(&pgdat->kswapd_wait);
3529 #ifdef CONFIG_HIBERNATION
3531 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3534 * Rather than trying to age LRUs the aim is to preserve the overall
3535 * LRU order by reclaiming preferentially
3536 * inactive > active > active referenced > active mapped
3538 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3540 struct reclaim_state reclaim_state;
3541 struct scan_control sc = {
3542 .nr_to_reclaim = nr_to_reclaim,
3543 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3544 .reclaim_idx = MAX_NR_ZONES - 1,
3545 .priority = DEF_PRIORITY,
3549 .hibernation_mode = 1,
3551 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3552 struct task_struct *p = current;
3553 unsigned long nr_reclaimed;
3555 p->flags |= PF_MEMALLOC;
3556 lockdep_set_current_reclaim_state(sc.gfp_mask);
3557 reclaim_state.reclaimed_slab = 0;
3558 p->reclaim_state = &reclaim_state;
3560 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3562 p->reclaim_state = NULL;
3563 lockdep_clear_current_reclaim_state();
3564 p->flags &= ~PF_MEMALLOC;
3566 return nr_reclaimed;
3568 #endif /* CONFIG_HIBERNATION */
3570 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3571 not required for correctness. So if the last cpu in a node goes
3572 away, we get changed to run anywhere: as the first one comes back,
3573 restore their cpu bindings. */
3574 static int kswapd_cpu_online(unsigned int cpu)
3578 for_each_node_state(nid, N_MEMORY) {
3579 pg_data_t *pgdat = NODE_DATA(nid);
3580 const struct cpumask *mask;
3582 mask = cpumask_of_node(pgdat->node_id);
3584 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3585 /* One of our CPUs online: restore mask */
3586 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3592 * This kswapd start function will be called by init and node-hot-add.
3593 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3595 int kswapd_run(int nid)
3597 pg_data_t *pgdat = NODE_DATA(nid);
3603 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3604 if (IS_ERR(pgdat->kswapd)) {
3605 /* failure at boot is fatal */
3606 BUG_ON(system_state == SYSTEM_BOOTING);
3607 pr_err("Failed to start kswapd on node %d\n", nid);
3608 ret = PTR_ERR(pgdat->kswapd);
3609 pgdat->kswapd = NULL;
3615 * Called by memory hotplug when all memory in a node is offlined. Caller must
3616 * hold mem_hotplug_begin/end().
3618 void kswapd_stop(int nid)
3620 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3623 kthread_stop(kswapd);
3624 NODE_DATA(nid)->kswapd = NULL;
3628 static int __init kswapd_init(void)
3633 for_each_node_state(nid, N_MEMORY)
3635 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3636 "mm/vmscan:online", kswapd_cpu_online,
3642 module_init(kswapd_init)
3648 * If non-zero call node_reclaim when the number of free pages falls below
3651 int node_reclaim_mode __read_mostly;
3653 #define RECLAIM_OFF 0
3654 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3655 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3656 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3659 * Priority for NODE_RECLAIM. This determines the fraction of pages
3660 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3663 #define NODE_RECLAIM_PRIORITY 4
3666 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3669 int sysctl_min_unmapped_ratio = 1;
3672 * If the number of slab pages in a zone grows beyond this percentage then
3673 * slab reclaim needs to occur.
3675 int sysctl_min_slab_ratio = 5;
3677 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3679 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3680 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3681 node_page_state(pgdat, NR_ACTIVE_FILE);
3684 * It's possible for there to be more file mapped pages than
3685 * accounted for by the pages on the file LRU lists because
3686 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3688 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3691 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3692 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3694 unsigned long nr_pagecache_reclaimable;
3695 unsigned long delta = 0;
3698 * If RECLAIM_UNMAP is set, then all file pages are considered
3699 * potentially reclaimable. Otherwise, we have to worry about
3700 * pages like swapcache and node_unmapped_file_pages() provides
3703 if (node_reclaim_mode & RECLAIM_UNMAP)
3704 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3706 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3708 /* If we can't clean pages, remove dirty pages from consideration */
3709 if (!(node_reclaim_mode & RECLAIM_WRITE))
3710 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3712 /* Watch for any possible underflows due to delta */
3713 if (unlikely(delta > nr_pagecache_reclaimable))
3714 delta = nr_pagecache_reclaimable;
3716 return nr_pagecache_reclaimable - delta;
3720 * Try to free up some pages from this node through reclaim.
3722 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3724 /* Minimum pages needed in order to stay on node */
3725 const unsigned long nr_pages = 1 << order;
3726 struct task_struct *p = current;
3727 struct reclaim_state reclaim_state;
3728 int classzone_idx = gfp_zone(gfp_mask);
3729 struct scan_control sc = {
3730 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3731 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3733 .priority = NODE_RECLAIM_PRIORITY,
3734 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3735 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3737 .reclaim_idx = classzone_idx,
3742 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3743 * and we also need to be able to write out pages for RECLAIM_WRITE
3744 * and RECLAIM_UNMAP.
3746 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3747 lockdep_set_current_reclaim_state(gfp_mask);
3748 reclaim_state.reclaimed_slab = 0;
3749 p->reclaim_state = &reclaim_state;
3751 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3753 * Free memory by calling shrink zone with increasing
3754 * priorities until we have enough memory freed.
3757 shrink_node(pgdat, &sc);
3758 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3761 p->reclaim_state = NULL;
3762 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3763 lockdep_clear_current_reclaim_state();
3764 return sc.nr_reclaimed >= nr_pages;
3767 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3772 * Node reclaim reclaims unmapped file backed pages and
3773 * slab pages if we are over the defined limits.
3775 * A small portion of unmapped file backed pages is needed for
3776 * file I/O otherwise pages read by file I/O will be immediately
3777 * thrown out if the node is overallocated. So we do not reclaim
3778 * if less than a specified percentage of the node is used by
3779 * unmapped file backed pages.
3781 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3782 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3783 return NODE_RECLAIM_FULL;
3785 if (!pgdat_reclaimable(pgdat))
3786 return NODE_RECLAIM_FULL;
3789 * Do not scan if the allocation should not be delayed.
3791 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3792 return NODE_RECLAIM_NOSCAN;
3795 * Only run node reclaim on the local node or on nodes that do not
3796 * have associated processors. This will favor the local processor
3797 * over remote processors and spread off node memory allocations
3798 * as wide as possible.
3800 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3801 return NODE_RECLAIM_NOSCAN;
3803 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3804 return NODE_RECLAIM_NOSCAN;
3806 ret = __node_reclaim(pgdat, gfp_mask, order);
3807 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3810 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3817 * page_evictable - test whether a page is evictable
3818 * @page: the page to test
3820 * Test whether page is evictable--i.e., should be placed on active/inactive
3821 * lists vs unevictable list.
3823 * Reasons page might not be evictable:
3824 * (1) page's mapping marked unevictable
3825 * (2) page is part of an mlocked VMA
3828 int page_evictable(struct page *page)
3830 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3835 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3836 * @pages: array of pages to check
3837 * @nr_pages: number of pages to check
3839 * Checks pages for evictability and moves them to the appropriate lru list.
3841 * This function is only used for SysV IPC SHM_UNLOCK.
3843 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3845 struct lruvec *lruvec;
3846 struct pglist_data *pgdat = NULL;
3851 for (i = 0; i < nr_pages; i++) {
3852 struct page *page = pages[i];
3853 struct pglist_data *pagepgdat = page_pgdat(page);
3856 if (pagepgdat != pgdat) {
3858 spin_unlock_irq(&pgdat->lru_lock);
3860 spin_lock_irq(&pgdat->lru_lock);
3862 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3864 if (!PageLRU(page) || !PageUnevictable(page))
3867 if (page_evictable(page)) {
3868 enum lru_list lru = page_lru_base_type(page);
3870 VM_BUG_ON_PAGE(PageActive(page), page);
3871 ClearPageUnevictable(page);
3872 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3873 add_page_to_lru_list(page, lruvec, lru);
3879 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3880 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3881 spin_unlock_irq(&pgdat->lru_lock);
3884 #endif /* CONFIG_SHMEM */