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;
237 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
239 if (!mem_cgroup_disabled())
240 return mem_cgroup_get_lru_size(lruvec, lru);
242 return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
246 * Add a shrinker callback to be called from the vm.
248 int register_shrinker(struct shrinker *shrinker)
250 size_t size = sizeof(*shrinker->nr_deferred);
252 if (shrinker->flags & SHRINKER_NUMA_AWARE)
255 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
256 if (!shrinker->nr_deferred)
259 down_write(&shrinker_rwsem);
260 list_add_tail(&shrinker->list, &shrinker_list);
261 up_write(&shrinker_rwsem);
264 EXPORT_SYMBOL(register_shrinker);
269 void unregister_shrinker(struct shrinker *shrinker)
271 down_write(&shrinker_rwsem);
272 list_del(&shrinker->list);
273 up_write(&shrinker_rwsem);
274 kfree(shrinker->nr_deferred);
276 EXPORT_SYMBOL(unregister_shrinker);
278 #define SHRINK_BATCH 128
280 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
281 struct shrinker *shrinker,
282 unsigned long nr_scanned,
283 unsigned long nr_eligible)
285 unsigned long freed = 0;
286 unsigned long long delta;
291 int nid = shrinkctl->nid;
292 long batch_size = shrinker->batch ? shrinker->batch
295 freeable = shrinker->count_objects(shrinker, shrinkctl);
300 * copy the current shrinker scan count into a local variable
301 * and zero it so that other concurrent shrinker invocations
302 * don't also do this scanning work.
304 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
307 delta = (4 * nr_scanned) / shrinker->seeks;
309 do_div(delta, nr_eligible + 1);
311 if (total_scan < 0) {
312 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
313 shrinker->scan_objects, total_scan);
314 total_scan = freeable;
318 * We need to avoid excessive windup on filesystem shrinkers
319 * due to large numbers of GFP_NOFS allocations causing the
320 * shrinkers to return -1 all the time. This results in a large
321 * nr being built up so when a shrink that can do some work
322 * comes along it empties the entire cache due to nr >>>
323 * freeable. This is bad for sustaining a working set in
326 * Hence only allow the shrinker to scan the entire cache when
327 * a large delta change is calculated directly.
329 if (delta < freeable / 4)
330 total_scan = min(total_scan, freeable / 2);
333 * Avoid risking looping forever due to too large nr value:
334 * never try to free more than twice the estimate number of
337 if (total_scan > freeable * 2)
338 total_scan = freeable * 2;
340 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
341 nr_scanned, nr_eligible,
342 freeable, delta, total_scan);
345 * Normally, we should not scan less than batch_size objects in one
346 * pass to avoid too frequent shrinker calls, but if the slab has less
347 * than batch_size objects in total and we are really tight on memory,
348 * we will try to reclaim all available objects, otherwise we can end
349 * up failing allocations although there are plenty of reclaimable
350 * objects spread over several slabs with usage less than the
353 * We detect the "tight on memory" situations by looking at the total
354 * number of objects we want to scan (total_scan). If it is greater
355 * than the total number of objects on slab (freeable), we must be
356 * scanning at high prio and therefore should try to reclaim as much as
359 while (total_scan >= batch_size ||
360 total_scan >= freeable) {
362 unsigned long nr_to_scan = min(batch_size, total_scan);
364 shrinkctl->nr_to_scan = nr_to_scan;
365 ret = shrinker->scan_objects(shrinker, shrinkctl);
366 if (ret == SHRINK_STOP)
370 count_vm_events(SLABS_SCANNED, nr_to_scan);
371 total_scan -= nr_to_scan;
377 * move the unused scan count back into the shrinker in a
378 * manner that handles concurrent updates. If we exhausted the
379 * scan, there is no need to do an update.
382 new_nr = atomic_long_add_return(total_scan,
383 &shrinker->nr_deferred[nid]);
385 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
387 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
392 * shrink_slab - shrink slab caches
393 * @gfp_mask: allocation context
394 * @nid: node whose slab caches to target
395 * @memcg: memory cgroup whose slab caches to target
396 * @nr_scanned: pressure numerator
397 * @nr_eligible: pressure denominator
399 * Call the shrink functions to age shrinkable caches.
401 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
402 * unaware shrinkers will receive a node id of 0 instead.
404 * @memcg specifies the memory cgroup to target. If it is not NULL,
405 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
406 * objects from the memory cgroup specified. Otherwise, only unaware
407 * shrinkers are called.
409 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
410 * the available objects should be scanned. Page reclaim for example
411 * passes the number of pages scanned and the number of pages on the
412 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
413 * when it encountered mapped pages. The ratio is further biased by
414 * the ->seeks setting of the shrink function, which indicates the
415 * cost to recreate an object relative to that of an LRU page.
417 * Returns the number of reclaimed slab objects.
419 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
420 struct mem_cgroup *memcg,
421 unsigned long nr_scanned,
422 unsigned long nr_eligible)
424 struct shrinker *shrinker;
425 unsigned long freed = 0;
427 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
431 nr_scanned = SWAP_CLUSTER_MAX;
433 if (!down_read_trylock(&shrinker_rwsem)) {
435 * If we would return 0, our callers would understand that we
436 * have nothing else to shrink and give up trying. By returning
437 * 1 we keep it going and assume we'll be able to shrink next
444 list_for_each_entry(shrinker, &shrinker_list, list) {
445 struct shrink_control sc = {
446 .gfp_mask = gfp_mask,
452 * If kernel memory accounting is disabled, we ignore
453 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
454 * passing NULL for memcg.
456 if (memcg_kmem_enabled() &&
457 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
460 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
463 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
466 up_read(&shrinker_rwsem);
472 void drop_slab_node(int nid)
477 struct mem_cgroup *memcg = NULL;
481 freed += shrink_slab(GFP_KERNEL, nid, memcg,
483 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
484 } while (freed > 10);
491 for_each_online_node(nid)
495 static inline int is_page_cache_freeable(struct page *page)
498 * A freeable page cache page is referenced only by the caller
499 * that isolated the page, the page cache radix tree and
500 * optional buffer heads at page->private.
502 return page_count(page) - page_has_private(page) == 2;
505 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
507 if (current->flags & PF_SWAPWRITE)
509 if (!inode_write_congested(inode))
511 if (inode_to_bdi(inode) == current->backing_dev_info)
517 * We detected a synchronous write error writing a page out. Probably
518 * -ENOSPC. We need to propagate that into the address_space for a subsequent
519 * fsync(), msync() or close().
521 * The tricky part is that after writepage we cannot touch the mapping: nothing
522 * prevents it from being freed up. But we have a ref on the page and once
523 * that page is locked, the mapping is pinned.
525 * We're allowed to run sleeping lock_page() here because we know the caller has
528 static void handle_write_error(struct address_space *mapping,
529 struct page *page, int error)
532 if (page_mapping(page) == mapping)
533 mapping_set_error(mapping, error);
537 /* possible outcome of pageout() */
539 /* failed to write page out, page is locked */
541 /* move page to the active list, page is locked */
543 /* page has been sent to the disk successfully, page is unlocked */
545 /* page is clean and locked */
550 * pageout is called by shrink_page_list() for each dirty page.
551 * Calls ->writepage().
553 static pageout_t pageout(struct page *page, struct address_space *mapping,
554 struct scan_control *sc)
557 * If the page is dirty, only perform writeback if that write
558 * will be non-blocking. To prevent this allocation from being
559 * stalled by pagecache activity. But note that there may be
560 * stalls if we need to run get_block(). We could test
561 * PagePrivate for that.
563 * If this process is currently in __generic_file_write_iter() against
564 * this page's queue, we can perform writeback even if that
567 * If the page is swapcache, write it back even if that would
568 * block, for some throttling. This happens by accident, because
569 * swap_backing_dev_info is bust: it doesn't reflect the
570 * congestion state of the swapdevs. Easy to fix, if needed.
572 if (!is_page_cache_freeable(page))
576 * Some data journaling orphaned pages can have
577 * page->mapping == NULL while being dirty with clean buffers.
579 if (page_has_private(page)) {
580 if (try_to_free_buffers(page)) {
581 ClearPageDirty(page);
582 pr_info("%s: orphaned page\n", __func__);
588 if (mapping->a_ops->writepage == NULL)
589 return PAGE_ACTIVATE;
590 if (!may_write_to_inode(mapping->host, sc))
593 if (clear_page_dirty_for_io(page)) {
595 struct writeback_control wbc = {
596 .sync_mode = WB_SYNC_NONE,
597 .nr_to_write = SWAP_CLUSTER_MAX,
599 .range_end = LLONG_MAX,
603 SetPageReclaim(page);
604 res = mapping->a_ops->writepage(page, &wbc);
606 handle_write_error(mapping, page, res);
607 if (res == AOP_WRITEPAGE_ACTIVATE) {
608 ClearPageReclaim(page);
609 return PAGE_ACTIVATE;
612 if (!PageWriteback(page)) {
613 /* synchronous write or broken a_ops? */
614 ClearPageReclaim(page);
616 trace_mm_vmscan_writepage(page);
617 inc_node_page_state(page, NR_VMSCAN_WRITE);
625 * Same as remove_mapping, but if the page is removed from the mapping, it
626 * gets returned with a refcount of 0.
628 static int __remove_mapping(struct address_space *mapping, struct page *page,
633 BUG_ON(!PageLocked(page));
634 BUG_ON(mapping != page_mapping(page));
636 spin_lock_irqsave(&mapping->tree_lock, flags);
638 * The non racy check for a busy page.
640 * Must be careful with the order of the tests. When someone has
641 * a ref to the page, it may be possible that they dirty it then
642 * drop the reference. So if PageDirty is tested before page_count
643 * here, then the following race may occur:
645 * get_user_pages(&page);
646 * [user mapping goes away]
648 * !PageDirty(page) [good]
649 * SetPageDirty(page);
651 * !page_count(page) [good, discard it]
653 * [oops, our write_to data is lost]
655 * Reversing the order of the tests ensures such a situation cannot
656 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
657 * load is not satisfied before that of page->_refcount.
659 * Note that if SetPageDirty is always performed via set_page_dirty,
660 * and thus under tree_lock, then this ordering is not required.
662 if (!page_ref_freeze(page, 2))
664 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
665 if (unlikely(PageDirty(page))) {
666 page_ref_unfreeze(page, 2);
670 if (PageSwapCache(page)) {
671 swp_entry_t swap = { .val = page_private(page) };
672 mem_cgroup_swapout(page, swap);
673 __delete_from_swap_cache(page);
674 spin_unlock_irqrestore(&mapping->tree_lock, flags);
675 swapcache_free(swap);
677 void (*freepage)(struct page *);
680 freepage = mapping->a_ops->freepage;
682 * Remember a shadow entry for reclaimed file cache in
683 * order to detect refaults, thus thrashing, later on.
685 * But don't store shadows in an address space that is
686 * already exiting. This is not just an optizimation,
687 * inode reclaim needs to empty out the radix tree or
688 * the nodes are lost. Don't plant shadows behind its
691 * We also don't store shadows for DAX mappings because the
692 * only page cache pages found in these are zero pages
693 * covering holes, and because we don't want to mix DAX
694 * exceptional entries and shadow exceptional entries in the
697 if (reclaimed && page_is_file_cache(page) &&
698 !mapping_exiting(mapping) && !dax_mapping(mapping))
699 shadow = workingset_eviction(mapping, page);
700 __delete_from_page_cache(page, shadow);
701 spin_unlock_irqrestore(&mapping->tree_lock, flags);
703 if (freepage != NULL)
710 spin_unlock_irqrestore(&mapping->tree_lock, flags);
715 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
716 * someone else has a ref on the page, abort and return 0. If it was
717 * successfully detached, return 1. Assumes the caller has a single ref on
720 int remove_mapping(struct address_space *mapping, struct page *page)
722 if (__remove_mapping(mapping, page, false)) {
724 * Unfreezing the refcount with 1 rather than 2 effectively
725 * drops the pagecache ref for us without requiring another
728 page_ref_unfreeze(page, 1);
735 * putback_lru_page - put previously isolated page onto appropriate LRU list
736 * @page: page to be put back to appropriate lru list
738 * Add previously isolated @page to appropriate LRU list.
739 * Page may still be unevictable for other reasons.
741 * lru_lock must not be held, interrupts must be enabled.
743 void putback_lru_page(struct page *page)
746 int was_unevictable = PageUnevictable(page);
748 VM_BUG_ON_PAGE(PageLRU(page), page);
751 ClearPageUnevictable(page);
753 if (page_evictable(page)) {
755 * For evictable pages, we can use the cache.
756 * In event of a race, worst case is we end up with an
757 * unevictable page on [in]active list.
758 * We know how to handle that.
760 is_unevictable = false;
764 * Put unevictable pages directly on zone's unevictable
767 is_unevictable = true;
768 add_page_to_unevictable_list(page);
770 * When racing with an mlock or AS_UNEVICTABLE clearing
771 * (page is unlocked) make sure that if the other thread
772 * does not observe our setting of PG_lru and fails
773 * isolation/check_move_unevictable_pages,
774 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
775 * the page back to the evictable list.
777 * The other side is TestClearPageMlocked() or shmem_lock().
783 * page's status can change while we move it among lru. If an evictable
784 * page is on unevictable list, it never be freed. To avoid that,
785 * check after we added it to the list, again.
787 if (is_unevictable && page_evictable(page)) {
788 if (!isolate_lru_page(page)) {
792 /* This means someone else dropped this page from LRU
793 * So, it will be freed or putback to LRU again. There is
794 * nothing to do here.
798 if (was_unevictable && !is_unevictable)
799 count_vm_event(UNEVICTABLE_PGRESCUED);
800 else if (!was_unevictable && is_unevictable)
801 count_vm_event(UNEVICTABLE_PGCULLED);
803 put_page(page); /* drop ref from isolate */
806 enum page_references {
808 PAGEREF_RECLAIM_CLEAN,
813 static enum page_references page_check_references(struct page *page,
814 struct scan_control *sc)
816 int referenced_ptes, referenced_page;
817 unsigned long vm_flags;
819 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
821 referenced_page = TestClearPageReferenced(page);
824 * Mlock lost the isolation race with us. Let try_to_unmap()
825 * move the page to the unevictable list.
827 if (vm_flags & VM_LOCKED)
828 return PAGEREF_RECLAIM;
830 if (referenced_ptes) {
831 if (PageSwapBacked(page))
832 return PAGEREF_ACTIVATE;
834 * All mapped pages start out with page table
835 * references from the instantiating fault, so we need
836 * to look twice if a mapped file page is used more
839 * Mark it and spare it for another trip around the
840 * inactive list. Another page table reference will
841 * lead to its activation.
843 * Note: the mark is set for activated pages as well
844 * so that recently deactivated but used pages are
847 SetPageReferenced(page);
849 if (referenced_page || referenced_ptes > 1)
850 return PAGEREF_ACTIVATE;
853 * Activate file-backed executable pages after first usage.
855 if (vm_flags & VM_EXEC)
856 return PAGEREF_ACTIVATE;
861 /* Reclaim if clean, defer dirty pages to writeback */
862 if (referenced_page && !PageSwapBacked(page))
863 return PAGEREF_RECLAIM_CLEAN;
865 return PAGEREF_RECLAIM;
868 /* Check if a page is dirty or under writeback */
869 static void page_check_dirty_writeback(struct page *page,
870 bool *dirty, bool *writeback)
872 struct address_space *mapping;
875 * Anonymous pages are not handled by flushers and must be written
876 * from reclaim context. Do not stall reclaim based on them
878 if (!page_is_file_cache(page)) {
884 /* By default assume that the page flags are accurate */
885 *dirty = PageDirty(page);
886 *writeback = PageWriteback(page);
888 /* Verify dirty/writeback state if the filesystem supports it */
889 if (!page_has_private(page))
892 mapping = page_mapping(page);
893 if (mapping && mapping->a_ops->is_dirty_writeback)
894 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
898 * shrink_page_list() returns the number of reclaimed pages
900 static unsigned long shrink_page_list(struct list_head *page_list,
901 struct pglist_data *pgdat,
902 struct scan_control *sc,
903 enum ttu_flags ttu_flags,
904 unsigned long *ret_nr_dirty,
905 unsigned long *ret_nr_unqueued_dirty,
906 unsigned long *ret_nr_congested,
907 unsigned long *ret_nr_writeback,
908 unsigned long *ret_nr_immediate,
911 LIST_HEAD(ret_pages);
912 LIST_HEAD(free_pages);
914 unsigned long nr_unqueued_dirty = 0;
915 unsigned long nr_dirty = 0;
916 unsigned long nr_congested = 0;
917 unsigned long nr_reclaimed = 0;
918 unsigned long nr_writeback = 0;
919 unsigned long nr_immediate = 0;
923 while (!list_empty(page_list)) {
924 struct address_space *mapping;
927 enum page_references references = PAGEREF_RECLAIM_CLEAN;
928 bool dirty, writeback;
929 bool lazyfree = false;
930 int ret = SWAP_SUCCESS;
934 page = lru_to_page(page_list);
935 list_del(&page->lru);
937 if (!trylock_page(page))
940 VM_BUG_ON_PAGE(PageActive(page), page);
944 if (unlikely(!page_evictable(page)))
947 if (!sc->may_unmap && page_mapped(page))
950 /* Double the slab pressure for mapped and swapcache pages */
951 if (page_mapped(page) || PageSwapCache(page))
954 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
955 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
958 * The number of dirty pages determines if a zone is marked
959 * reclaim_congested which affects wait_iff_congested. kswapd
960 * will stall and start writing pages if the tail of the LRU
961 * is all dirty unqueued pages.
963 page_check_dirty_writeback(page, &dirty, &writeback);
964 if (dirty || writeback)
967 if (dirty && !writeback)
971 * Treat this page as congested if the underlying BDI is or if
972 * pages are cycling through the LRU so quickly that the
973 * pages marked for immediate reclaim are making it to the
974 * end of the LRU a second time.
976 mapping = page_mapping(page);
977 if (((dirty || writeback) && mapping &&
978 inode_write_congested(mapping->host)) ||
979 (writeback && PageReclaim(page)))
983 * If a page at the tail of the LRU is under writeback, there
984 * are three cases to consider.
986 * 1) If reclaim is encountering an excessive number of pages
987 * under writeback and this page is both under writeback and
988 * PageReclaim then it indicates that pages are being queued
989 * for IO but are being recycled through the LRU before the
990 * IO can complete. Waiting on the page itself risks an
991 * indefinite stall if it is impossible to writeback the
992 * page due to IO error or disconnected storage so instead
993 * note that the LRU is being scanned too quickly and the
994 * caller can stall after page list has been processed.
996 * 2) Global or new memcg reclaim encounters a page that is
997 * not marked for immediate reclaim, or the caller does not
998 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
999 * not to fs). In this case mark the page for immediate
1000 * reclaim and continue scanning.
1002 * Require may_enter_fs because we would wait on fs, which
1003 * may not have submitted IO yet. And the loop driver might
1004 * enter reclaim, and deadlock if it waits on a page for
1005 * which it is needed to do the write (loop masks off
1006 * __GFP_IO|__GFP_FS for this reason); but more thought
1007 * would probably show more reasons.
1009 * 3) Legacy memcg encounters a page that is already marked
1010 * PageReclaim. memcg does not have any dirty pages
1011 * throttling so we could easily OOM just because too many
1012 * pages are in writeback and there is nothing else to
1013 * reclaim. Wait for the writeback to complete.
1015 if (PageWriteback(page)) {
1017 if (current_is_kswapd() &&
1018 PageReclaim(page) &&
1019 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1024 } else if (sane_reclaim(sc) ||
1025 !PageReclaim(page) || !may_enter_fs) {
1027 * This is slightly racy - end_page_writeback()
1028 * might have just cleared PageReclaim, then
1029 * setting PageReclaim here end up interpreted
1030 * as PageReadahead - but that does not matter
1031 * enough to care. What we do want is for this
1032 * page to have PageReclaim set next time memcg
1033 * reclaim reaches the tests above, so it will
1034 * then wait_on_page_writeback() to avoid OOM;
1035 * and it's also appropriate in global reclaim.
1037 SetPageReclaim(page);
1044 wait_on_page_writeback(page);
1045 /* then go back and try same page again */
1046 list_add_tail(&page->lru, page_list);
1052 references = page_check_references(page, sc);
1054 switch (references) {
1055 case PAGEREF_ACTIVATE:
1056 goto activate_locked;
1059 case PAGEREF_RECLAIM:
1060 case PAGEREF_RECLAIM_CLEAN:
1061 ; /* try to reclaim the page below */
1065 * Anonymous process memory has backing store?
1066 * Try to allocate it some swap space here.
1068 if (PageAnon(page) && !PageSwapCache(page)) {
1069 if (!(sc->gfp_mask & __GFP_IO))
1071 if (!add_to_swap(page, page_list))
1072 goto activate_locked;
1076 /* Adding to swap updated mapping */
1077 mapping = page_mapping(page);
1078 } else if (unlikely(PageTransHuge(page))) {
1079 /* Split file THP */
1080 if (split_huge_page_to_list(page, page_list))
1084 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1087 * The page is mapped into the page tables of one or more
1088 * processes. Try to unmap it here.
1090 if (page_mapped(page) && mapping) {
1091 switch (ret = try_to_unmap(page, lazyfree ?
1092 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1093 (ttu_flags | TTU_BATCH_FLUSH))) {
1095 goto activate_locked;
1103 ; /* try to free the page below */
1107 if (PageDirty(page)) {
1109 * Only kswapd can writeback filesystem pages to
1110 * avoid risk of stack overflow but only writeback
1111 * if many dirty pages have been encountered.
1113 if (page_is_file_cache(page) &&
1114 (!current_is_kswapd() ||
1115 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1117 * Immediately reclaim when written back.
1118 * Similar in principal to deactivate_page()
1119 * except we already have the page isolated
1120 * and know it's dirty
1122 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1123 SetPageReclaim(page);
1128 if (references == PAGEREF_RECLAIM_CLEAN)
1132 if (!sc->may_writepage)
1136 * Page is dirty. Flush the TLB if a writable entry
1137 * potentially exists to avoid CPU writes after IO
1138 * starts and then write it out here.
1140 try_to_unmap_flush_dirty();
1141 switch (pageout(page, mapping, sc)) {
1145 goto activate_locked;
1147 if (PageWriteback(page))
1149 if (PageDirty(page))
1153 * A synchronous write - probably a ramdisk. Go
1154 * ahead and try to reclaim the page.
1156 if (!trylock_page(page))
1158 if (PageDirty(page) || PageWriteback(page))
1160 mapping = page_mapping(page);
1162 ; /* try to free the page below */
1167 * If the page has buffers, try to free the buffer mappings
1168 * associated with this page. If we succeed we try to free
1171 * We do this even if the page is PageDirty().
1172 * try_to_release_page() does not perform I/O, but it is
1173 * possible for a page to have PageDirty set, but it is actually
1174 * clean (all its buffers are clean). This happens if the
1175 * buffers were written out directly, with submit_bh(). ext3
1176 * will do this, as well as the blockdev mapping.
1177 * try_to_release_page() will discover that cleanness and will
1178 * drop the buffers and mark the page clean - it can be freed.
1180 * Rarely, pages can have buffers and no ->mapping. These are
1181 * the pages which were not successfully invalidated in
1182 * truncate_complete_page(). We try to drop those buffers here
1183 * and if that worked, and the page is no longer mapped into
1184 * process address space (page_count == 1) it can be freed.
1185 * Otherwise, leave the page on the LRU so it is swappable.
1187 if (page_has_private(page)) {
1188 if (!try_to_release_page(page, sc->gfp_mask))
1189 goto activate_locked;
1190 if (!mapping && page_count(page) == 1) {
1192 if (put_page_testzero(page))
1196 * rare race with speculative reference.
1197 * the speculative reference will free
1198 * this page shortly, so we may
1199 * increment nr_reclaimed here (and
1200 * leave it off the LRU).
1209 if (!mapping || !__remove_mapping(mapping, page, true))
1213 * At this point, we have no other references and there is
1214 * no way to pick any more up (removed from LRU, removed
1215 * from pagecache). Can use non-atomic bitops now (and
1216 * we obviously don't have to worry about waking up a process
1217 * waiting on the page lock, because there are no references.
1219 __ClearPageLocked(page);
1221 if (ret == SWAP_LZFREE)
1222 count_vm_event(PGLAZYFREED);
1227 * Is there need to periodically free_page_list? It would
1228 * appear not as the counts should be low
1230 list_add(&page->lru, &free_pages);
1234 if (PageSwapCache(page))
1235 try_to_free_swap(page);
1237 list_add(&page->lru, &ret_pages);
1241 /* Not a candidate for swapping, so reclaim swap space. */
1242 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1243 try_to_free_swap(page);
1244 VM_BUG_ON_PAGE(PageActive(page), page);
1245 SetPageActive(page);
1250 list_add(&page->lru, &ret_pages);
1251 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1254 mem_cgroup_uncharge_list(&free_pages);
1255 try_to_unmap_flush();
1256 free_hot_cold_page_list(&free_pages, true);
1258 list_splice(&ret_pages, page_list);
1259 count_vm_events(PGACTIVATE, pgactivate);
1261 *ret_nr_dirty += nr_dirty;
1262 *ret_nr_congested += nr_congested;
1263 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1264 *ret_nr_writeback += nr_writeback;
1265 *ret_nr_immediate += nr_immediate;
1266 return nr_reclaimed;
1269 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1270 struct list_head *page_list)
1272 struct scan_control sc = {
1273 .gfp_mask = GFP_KERNEL,
1274 .priority = DEF_PRIORITY,
1277 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1278 struct page *page, *next;
1279 LIST_HEAD(clean_pages);
1281 list_for_each_entry_safe(page, next, page_list, lru) {
1282 if (page_is_file_cache(page) && !PageDirty(page) &&
1283 !__PageMovable(page)) {
1284 ClearPageActive(page);
1285 list_move(&page->lru, &clean_pages);
1289 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1290 TTU_UNMAP|TTU_IGNORE_ACCESS,
1291 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1292 list_splice(&clean_pages, page_list);
1293 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1298 * Attempt to remove the specified page from its LRU. Only take this page
1299 * if it is of the appropriate PageActive status. Pages which are being
1300 * freed elsewhere are also ignored.
1302 * page: page to consider
1303 * mode: one of the LRU isolation modes defined above
1305 * returns 0 on success, -ve errno on failure.
1307 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1311 /* Only take pages on the LRU. */
1315 /* Compaction should not handle unevictable pages but CMA can do so */
1316 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1322 * To minimise LRU disruption, the caller can indicate that it only
1323 * wants to isolate pages it will be able to operate on without
1324 * blocking - clean pages for the most part.
1326 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1327 * is used by reclaim when it is cannot write to backing storage
1329 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1330 * that it is possible to migrate without blocking
1332 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1333 /* All the caller can do on PageWriteback is block */
1334 if (PageWriteback(page))
1337 if (PageDirty(page)) {
1338 struct address_space *mapping;
1340 /* ISOLATE_CLEAN means only clean pages */
1341 if (mode & ISOLATE_CLEAN)
1345 * Only pages without mappings or that have a
1346 * ->migratepage callback are possible to migrate
1349 mapping = page_mapping(page);
1350 if (mapping && !mapping->a_ops->migratepage)
1355 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1358 if (likely(get_page_unless_zero(page))) {
1360 * Be careful not to clear PageLRU until after we're
1361 * sure the page is not being freed elsewhere -- the
1362 * page release code relies on it.
1373 * Update LRU sizes after isolating pages. The LRU size updates must
1374 * be complete before mem_cgroup_update_lru_size due to a santity check.
1376 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1377 enum lru_list lru, unsigned long *nr_zone_taken,
1378 unsigned long nr_taken)
1382 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1383 if (!nr_zone_taken[zid])
1386 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1390 mem_cgroup_update_lru_size(lruvec, lru, -nr_taken);
1395 * zone_lru_lock is heavily contended. Some of the functions that
1396 * shrink the lists perform better by taking out a batch of pages
1397 * and working on them outside the LRU lock.
1399 * For pagecache intensive workloads, this function is the hottest
1400 * spot in the kernel (apart from copy_*_user functions).
1402 * Appropriate locks must be held before calling this function.
1404 * @nr_to_scan: The number of pages to look through on the list.
1405 * @lruvec: The LRU vector to pull pages from.
1406 * @dst: The temp list to put pages on to.
1407 * @nr_scanned: The number of pages that were scanned.
1408 * @sc: The scan_control struct for this reclaim session
1409 * @mode: One of the LRU isolation modes
1410 * @lru: LRU list id for isolating
1412 * returns how many pages were moved onto *@dst.
1414 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1415 struct lruvec *lruvec, struct list_head *dst,
1416 unsigned long *nr_scanned, struct scan_control *sc,
1417 isolate_mode_t mode, enum lru_list lru)
1419 struct list_head *src = &lruvec->lists[lru];
1420 unsigned long nr_taken = 0;
1421 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1422 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1423 unsigned long scan, nr_pages;
1424 LIST_HEAD(pages_skipped);
1426 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1427 !list_empty(src);) {
1430 page = lru_to_page(src);
1431 prefetchw_prev_lru_page(page, src, flags);
1433 VM_BUG_ON_PAGE(!PageLRU(page), page);
1435 if (page_zonenum(page) > sc->reclaim_idx) {
1436 list_move(&page->lru, &pages_skipped);
1437 nr_skipped[page_zonenum(page)]++;
1442 * Account for scanned and skipped separetly to avoid the pgdat
1443 * being prematurely marked unreclaimable by pgdat_reclaimable.
1447 switch (__isolate_lru_page(page, mode)) {
1449 nr_pages = hpage_nr_pages(page);
1450 nr_taken += nr_pages;
1451 nr_zone_taken[page_zonenum(page)] += nr_pages;
1452 list_move(&page->lru, dst);
1456 /* else it is being freed elsewhere */
1457 list_move(&page->lru, src);
1466 * Splice any skipped pages to the start of the LRU list. Note that
1467 * this disrupts the LRU order when reclaiming for lower zones but
1468 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1469 * scanning would soon rescan the same pages to skip and put the
1470 * system at risk of premature OOM.
1472 if (!list_empty(&pages_skipped)) {
1474 unsigned long total_skipped = 0;
1476 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1477 if (!nr_skipped[zid])
1480 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1481 total_skipped += nr_skipped[zid];
1485 * Account skipped pages as a partial scan as the pgdat may be
1486 * close to unreclaimable. If the LRU list is empty, account
1487 * skipped pages as a full scan.
1489 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1491 list_splice(&pages_skipped, src);
1494 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
1495 nr_taken, mode, is_file_lru(lru));
1496 update_lru_sizes(lruvec, lru, nr_zone_taken, nr_taken);
1501 * isolate_lru_page - tries to isolate a page from its LRU list
1502 * @page: page to isolate from its LRU list
1504 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1505 * vmstat statistic corresponding to whatever LRU list the page was on.
1507 * Returns 0 if the page was removed from an LRU list.
1508 * Returns -EBUSY if the page was not on an LRU list.
1510 * The returned page will have PageLRU() cleared. If it was found on
1511 * the active list, it will have PageActive set. If it was found on
1512 * the unevictable list, it will have the PageUnevictable bit set. That flag
1513 * may need to be cleared by the caller before letting the page go.
1515 * The vmstat statistic corresponding to the list on which the page was
1516 * found will be decremented.
1519 * (1) Must be called with an elevated refcount on the page. This is a
1520 * fundamentnal difference from isolate_lru_pages (which is called
1521 * without a stable reference).
1522 * (2) the lru_lock must not be held.
1523 * (3) interrupts must be enabled.
1525 int isolate_lru_page(struct page *page)
1529 VM_BUG_ON_PAGE(!page_count(page), page);
1530 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1532 if (PageLRU(page)) {
1533 struct zone *zone = page_zone(page);
1534 struct lruvec *lruvec;
1536 spin_lock_irq(zone_lru_lock(zone));
1537 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1538 if (PageLRU(page)) {
1539 int lru = page_lru(page);
1542 del_page_from_lru_list(page, lruvec, lru);
1545 spin_unlock_irq(zone_lru_lock(zone));
1551 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1552 * then get resheduled. When there are massive number of tasks doing page
1553 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1554 * the LRU list will go small and be scanned faster than necessary, leading to
1555 * unnecessary swapping, thrashing and OOM.
1557 static int too_many_isolated(struct pglist_data *pgdat, int file,
1558 struct scan_control *sc)
1560 unsigned long inactive, isolated;
1562 if (current_is_kswapd())
1565 if (!sane_reclaim(sc))
1569 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1570 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1572 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1573 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1577 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1578 * won't get blocked by normal direct-reclaimers, forming a circular
1581 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1584 return isolated > inactive;
1587 static noinline_for_stack void
1588 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1590 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1591 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1592 LIST_HEAD(pages_to_free);
1595 * Put back any unfreeable pages.
1597 while (!list_empty(page_list)) {
1598 struct page *page = lru_to_page(page_list);
1601 VM_BUG_ON_PAGE(PageLRU(page), page);
1602 list_del(&page->lru);
1603 if (unlikely(!page_evictable(page))) {
1604 spin_unlock_irq(&pgdat->lru_lock);
1605 putback_lru_page(page);
1606 spin_lock_irq(&pgdat->lru_lock);
1610 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1613 lru = page_lru(page);
1614 add_page_to_lru_list(page, lruvec, lru);
1616 if (is_active_lru(lru)) {
1617 int file = is_file_lru(lru);
1618 int numpages = hpage_nr_pages(page);
1619 reclaim_stat->recent_rotated[file] += numpages;
1621 if (put_page_testzero(page)) {
1622 __ClearPageLRU(page);
1623 __ClearPageActive(page);
1624 del_page_from_lru_list(page, lruvec, lru);
1626 if (unlikely(PageCompound(page))) {
1627 spin_unlock_irq(&pgdat->lru_lock);
1628 mem_cgroup_uncharge(page);
1629 (*get_compound_page_dtor(page))(page);
1630 spin_lock_irq(&pgdat->lru_lock);
1632 list_add(&page->lru, &pages_to_free);
1637 * To save our caller's stack, now use input list for pages to free.
1639 list_splice(&pages_to_free, page_list);
1643 * If a kernel thread (such as nfsd for loop-back mounts) services
1644 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1645 * In that case we should only throttle if the backing device it is
1646 * writing to is congested. In other cases it is safe to throttle.
1648 static int current_may_throttle(void)
1650 return !(current->flags & PF_LESS_THROTTLE) ||
1651 current->backing_dev_info == NULL ||
1652 bdi_write_congested(current->backing_dev_info);
1656 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1657 * of reclaimed pages
1659 static noinline_for_stack unsigned long
1660 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1661 struct scan_control *sc, enum lru_list lru)
1663 LIST_HEAD(page_list);
1664 unsigned long nr_scanned;
1665 unsigned long nr_reclaimed = 0;
1666 unsigned long nr_taken;
1667 unsigned long nr_dirty = 0;
1668 unsigned long nr_congested = 0;
1669 unsigned long nr_unqueued_dirty = 0;
1670 unsigned long nr_writeback = 0;
1671 unsigned long nr_immediate = 0;
1672 isolate_mode_t isolate_mode = 0;
1673 int file = is_file_lru(lru);
1674 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1675 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1677 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1678 congestion_wait(BLK_RW_ASYNC, HZ/10);
1680 /* We are about to die and free our memory. Return now. */
1681 if (fatal_signal_pending(current))
1682 return SWAP_CLUSTER_MAX;
1688 isolate_mode |= ISOLATE_UNMAPPED;
1689 if (!sc->may_writepage)
1690 isolate_mode |= ISOLATE_CLEAN;
1692 spin_lock_irq(&pgdat->lru_lock);
1694 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1695 &nr_scanned, sc, isolate_mode, lru);
1697 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1698 reclaim_stat->recent_scanned[file] += nr_taken;
1700 if (global_reclaim(sc)) {
1701 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1702 if (current_is_kswapd())
1703 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1705 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1707 spin_unlock_irq(&pgdat->lru_lock);
1712 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1713 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1714 &nr_writeback, &nr_immediate,
1717 spin_lock_irq(&pgdat->lru_lock);
1719 if (global_reclaim(sc)) {
1720 if (current_is_kswapd())
1721 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1723 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1726 putback_inactive_pages(lruvec, &page_list);
1728 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1730 spin_unlock_irq(&pgdat->lru_lock);
1732 mem_cgroup_uncharge_list(&page_list);
1733 free_hot_cold_page_list(&page_list, true);
1736 * If reclaim is isolating dirty pages under writeback, it implies
1737 * that the long-lived page allocation rate is exceeding the page
1738 * laundering rate. Either the global limits are not being effective
1739 * at throttling processes due to the page distribution throughout
1740 * zones or there is heavy usage of a slow backing device. The
1741 * only option is to throttle from reclaim context which is not ideal
1742 * as there is no guarantee the dirtying process is throttled in the
1743 * same way balance_dirty_pages() manages.
1745 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1746 * of pages under pages flagged for immediate reclaim and stall if any
1747 * are encountered in the nr_immediate check below.
1749 if (nr_writeback && nr_writeback == nr_taken)
1750 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1753 * Legacy memcg will stall in page writeback so avoid forcibly
1756 if (sane_reclaim(sc)) {
1758 * Tag a zone as congested if all the dirty pages scanned were
1759 * backed by a congested BDI and wait_iff_congested will stall.
1761 if (nr_dirty && nr_dirty == nr_congested)
1762 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1765 * If dirty pages are scanned that are not queued for IO, it
1766 * implies that flushers are not keeping up. In this case, flag
1767 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1770 if (nr_unqueued_dirty == nr_taken)
1771 set_bit(PGDAT_DIRTY, &pgdat->flags);
1774 * If kswapd scans pages marked marked for immediate
1775 * reclaim and under writeback (nr_immediate), it implies
1776 * that pages are cycling through the LRU faster than
1777 * they are written so also forcibly stall.
1779 if (nr_immediate && current_may_throttle())
1780 congestion_wait(BLK_RW_ASYNC, HZ/10);
1784 * Stall direct reclaim for IO completions if underlying BDIs or zone
1785 * is congested. Allow kswapd to continue until it starts encountering
1786 * unqueued dirty pages or cycling through the LRU too quickly.
1788 if (!sc->hibernation_mode && !current_is_kswapd() &&
1789 current_may_throttle())
1790 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1792 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1793 nr_scanned, nr_reclaimed,
1794 sc->priority, file);
1795 return nr_reclaimed;
1799 * This moves pages from the active list to the inactive list.
1801 * We move them the other way if the page is referenced by one or more
1802 * processes, from rmap.
1804 * If the pages are mostly unmapped, the processing is fast and it is
1805 * appropriate to hold zone_lru_lock across the whole operation. But if
1806 * the pages are mapped, the processing is slow (page_referenced()) so we
1807 * should drop zone_lru_lock around each page. It's impossible to balance
1808 * this, so instead we remove the pages from the LRU while processing them.
1809 * It is safe to rely on PG_active against the non-LRU pages in here because
1810 * nobody will play with that bit on a non-LRU page.
1812 * The downside is that we have to touch page->_refcount against each page.
1813 * But we had to alter page->flags anyway.
1816 static void move_active_pages_to_lru(struct lruvec *lruvec,
1817 struct list_head *list,
1818 struct list_head *pages_to_free,
1821 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1822 unsigned long pgmoved = 0;
1826 while (!list_empty(list)) {
1827 page = lru_to_page(list);
1828 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1830 VM_BUG_ON_PAGE(PageLRU(page), page);
1833 nr_pages = hpage_nr_pages(page);
1834 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1835 list_move(&page->lru, &lruvec->lists[lru]);
1836 pgmoved += nr_pages;
1838 if (put_page_testzero(page)) {
1839 __ClearPageLRU(page);
1840 __ClearPageActive(page);
1841 del_page_from_lru_list(page, lruvec, lru);
1843 if (unlikely(PageCompound(page))) {
1844 spin_unlock_irq(&pgdat->lru_lock);
1845 mem_cgroup_uncharge(page);
1846 (*get_compound_page_dtor(page))(page);
1847 spin_lock_irq(&pgdat->lru_lock);
1849 list_add(&page->lru, pages_to_free);
1853 if (!is_active_lru(lru))
1854 __count_vm_events(PGDEACTIVATE, pgmoved);
1857 static void shrink_active_list(unsigned long nr_to_scan,
1858 struct lruvec *lruvec,
1859 struct scan_control *sc,
1862 unsigned long nr_taken;
1863 unsigned long nr_scanned;
1864 unsigned long vm_flags;
1865 LIST_HEAD(l_hold); /* The pages which were snipped off */
1866 LIST_HEAD(l_active);
1867 LIST_HEAD(l_inactive);
1869 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1870 unsigned long nr_rotated = 0;
1871 isolate_mode_t isolate_mode = 0;
1872 int file = is_file_lru(lru);
1873 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1878 isolate_mode |= ISOLATE_UNMAPPED;
1879 if (!sc->may_writepage)
1880 isolate_mode |= ISOLATE_CLEAN;
1882 spin_lock_irq(&pgdat->lru_lock);
1884 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1885 &nr_scanned, sc, isolate_mode, lru);
1887 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1888 reclaim_stat->recent_scanned[file] += nr_taken;
1890 if (global_reclaim(sc))
1891 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1892 __count_vm_events(PGREFILL, nr_scanned);
1894 spin_unlock_irq(&pgdat->lru_lock);
1896 while (!list_empty(&l_hold)) {
1898 page = lru_to_page(&l_hold);
1899 list_del(&page->lru);
1901 if (unlikely(!page_evictable(page))) {
1902 putback_lru_page(page);
1906 if (unlikely(buffer_heads_over_limit)) {
1907 if (page_has_private(page) && trylock_page(page)) {
1908 if (page_has_private(page))
1909 try_to_release_page(page, 0);
1914 if (page_referenced(page, 0, sc->target_mem_cgroup,
1916 nr_rotated += hpage_nr_pages(page);
1918 * Identify referenced, file-backed active pages and
1919 * give them one more trip around the active list. So
1920 * that executable code get better chances to stay in
1921 * memory under moderate memory pressure. Anon pages
1922 * are not likely to be evicted by use-once streaming
1923 * IO, plus JVM can create lots of anon VM_EXEC pages,
1924 * so we ignore them here.
1926 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1927 list_add(&page->lru, &l_active);
1932 ClearPageActive(page); /* we are de-activating */
1933 list_add(&page->lru, &l_inactive);
1937 * Move pages back to the lru list.
1939 spin_lock_irq(&pgdat->lru_lock);
1941 * Count referenced pages from currently used mappings as rotated,
1942 * even though only some of them are actually re-activated. This
1943 * helps balance scan pressure between file and anonymous pages in
1946 reclaim_stat->recent_rotated[file] += nr_rotated;
1948 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1949 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1950 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1951 spin_unlock_irq(&pgdat->lru_lock);
1953 mem_cgroup_uncharge_list(&l_hold);
1954 free_hot_cold_page_list(&l_hold, true);
1958 * The inactive anon list should be small enough that the VM never has
1959 * to do too much work.
1961 * The inactive file list should be small enough to leave most memory
1962 * to the established workingset on the scan-resistant active list,
1963 * but large enough to avoid thrashing the aggregate readahead window.
1965 * Both inactive lists should also be large enough that each inactive
1966 * page has a chance to be referenced again before it is reclaimed.
1968 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1969 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1970 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1973 * memory ratio inactive
1974 * -------------------------------------
1983 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
1984 struct scan_control *sc)
1986 unsigned long inactive_ratio;
1987 unsigned long inactive;
1988 unsigned long active;
1990 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1994 * If we don't have swap space, anonymous page deactivation
1997 if (!file && !total_swap_pages)
2000 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
2001 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
2004 * For zone-constrained allocations, it is necessary to check if
2005 * deactivations are required for lowmem to be reclaimed. This
2006 * calculates the inactive/active pages available in eligible zones.
2008 for (zid = sc->reclaim_idx + 1; zid < MAX_NR_ZONES; zid++) {
2009 struct zone *zone = &pgdat->node_zones[zid];
2010 unsigned long inactive_zone, active_zone;
2012 if (!populated_zone(zone))
2015 inactive_zone = zone_page_state(zone,
2016 NR_ZONE_LRU_BASE + (file * LRU_FILE));
2017 active_zone = zone_page_state(zone,
2018 NR_ZONE_LRU_BASE + (file * LRU_FILE) + LRU_ACTIVE);
2020 inactive -= min(inactive, inactive_zone);
2021 active -= min(active, active_zone);
2024 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2026 inactive_ratio = int_sqrt(10 * gb);
2030 return inactive * inactive_ratio < active;
2033 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2034 struct lruvec *lruvec, struct scan_control *sc)
2036 if (is_active_lru(lru)) {
2037 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
2038 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2042 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2053 * Determine how aggressively the anon and file LRU lists should be
2054 * scanned. The relative value of each set of LRU lists is determined
2055 * by looking at the fraction of the pages scanned we did rotate back
2056 * onto the active list instead of evict.
2058 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2059 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2061 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2062 struct scan_control *sc, unsigned long *nr,
2063 unsigned long *lru_pages)
2065 int swappiness = mem_cgroup_swappiness(memcg);
2066 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2068 u64 denominator = 0; /* gcc */
2069 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2070 unsigned long anon_prio, file_prio;
2071 enum scan_balance scan_balance;
2072 unsigned long anon, file;
2073 bool force_scan = false;
2074 unsigned long ap, fp;
2080 * If the zone or memcg is small, nr[l] can be 0. This
2081 * results in no scanning on this priority and a potential
2082 * priority drop. Global direct reclaim can go to the next
2083 * zone and tends to have no problems. Global kswapd is for
2084 * zone balancing and it needs to scan a minimum amount. When
2085 * reclaiming for a memcg, a priority drop can cause high
2086 * latencies, so it's better to scan a minimum amount there as
2089 if (current_is_kswapd()) {
2090 if (!pgdat_reclaimable(pgdat))
2092 if (!mem_cgroup_online(memcg))
2095 if (!global_reclaim(sc))
2098 /* If we have no swap space, do not bother scanning anon pages. */
2099 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2100 scan_balance = SCAN_FILE;
2105 * Global reclaim will swap to prevent OOM even with no
2106 * swappiness, but memcg users want to use this knob to
2107 * disable swapping for individual groups completely when
2108 * using the memory controller's swap limit feature would be
2111 if (!global_reclaim(sc) && !swappiness) {
2112 scan_balance = SCAN_FILE;
2117 * Do not apply any pressure balancing cleverness when the
2118 * system is close to OOM, scan both anon and file equally
2119 * (unless the swappiness setting disagrees with swapping).
2121 if (!sc->priority && swappiness) {
2122 scan_balance = SCAN_EQUAL;
2127 * Prevent the reclaimer from falling into the cache trap: as
2128 * cache pages start out inactive, every cache fault will tip
2129 * the scan balance towards the file LRU. And as the file LRU
2130 * shrinks, so does the window for rotation from references.
2131 * This means we have a runaway feedback loop where a tiny
2132 * thrashing file LRU becomes infinitely more attractive than
2133 * anon pages. Try to detect this based on file LRU size.
2135 if (global_reclaim(sc)) {
2136 unsigned long pgdatfile;
2137 unsigned long pgdatfree;
2139 unsigned long total_high_wmark = 0;
2141 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2142 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2143 node_page_state(pgdat, NR_INACTIVE_FILE);
2145 for (z = 0; z < MAX_NR_ZONES; z++) {
2146 struct zone *zone = &pgdat->node_zones[z];
2147 if (!populated_zone(zone))
2150 total_high_wmark += high_wmark_pages(zone);
2153 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2154 scan_balance = SCAN_ANON;
2160 * If there is enough inactive page cache, i.e. if the size of the
2161 * inactive list is greater than that of the active list *and* the
2162 * inactive list actually has some pages to scan on this priority, we
2163 * do not reclaim anything from the anonymous working set right now.
2164 * Without the second condition we could end up never scanning an
2165 * lruvec even if it has plenty of old anonymous pages unless the
2166 * system is under heavy pressure.
2168 if (!inactive_list_is_low(lruvec, true, sc) &&
2169 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2170 scan_balance = SCAN_FILE;
2174 scan_balance = SCAN_FRACT;
2177 * With swappiness at 100, anonymous and file have the same priority.
2178 * This scanning priority is essentially the inverse of IO cost.
2180 anon_prio = swappiness;
2181 file_prio = 200 - anon_prio;
2184 * OK, so we have swap space and a fair amount of page cache
2185 * pages. We use the recently rotated / recently scanned
2186 * ratios to determine how valuable each cache is.
2188 * Because workloads change over time (and to avoid overflow)
2189 * we keep these statistics as a floating average, which ends
2190 * up weighing recent references more than old ones.
2192 * anon in [0], file in [1]
2195 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2196 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2197 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2198 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2200 spin_lock_irq(&pgdat->lru_lock);
2201 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2202 reclaim_stat->recent_scanned[0] /= 2;
2203 reclaim_stat->recent_rotated[0] /= 2;
2206 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2207 reclaim_stat->recent_scanned[1] /= 2;
2208 reclaim_stat->recent_rotated[1] /= 2;
2212 * The amount of pressure on anon vs file pages is inversely
2213 * proportional to the fraction of recently scanned pages on
2214 * each list that were recently referenced and in active use.
2216 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2217 ap /= reclaim_stat->recent_rotated[0] + 1;
2219 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2220 fp /= reclaim_stat->recent_rotated[1] + 1;
2221 spin_unlock_irq(&pgdat->lru_lock);
2225 denominator = ap + fp + 1;
2227 some_scanned = false;
2228 /* Only use force_scan on second pass. */
2229 for (pass = 0; !some_scanned && pass < 2; pass++) {
2231 for_each_evictable_lru(lru) {
2232 int file = is_file_lru(lru);
2236 size = lruvec_lru_size(lruvec, lru);
2237 scan = size >> sc->priority;
2239 if (!scan && pass && force_scan)
2240 scan = min(size, SWAP_CLUSTER_MAX);
2242 switch (scan_balance) {
2244 /* Scan lists relative to size */
2248 * Scan types proportional to swappiness and
2249 * their relative recent reclaim efficiency.
2251 scan = div64_u64(scan * fraction[file],
2256 /* Scan one type exclusively */
2257 if ((scan_balance == SCAN_FILE) != file) {
2263 /* Look ma, no brain */
2271 * Skip the second pass and don't force_scan,
2272 * if we found something to scan.
2274 some_scanned |= !!scan;
2279 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2280 static void init_tlb_ubc(void)
2283 * This deliberately does not clear the cpumask as it's expensive
2284 * and unnecessary. If there happens to be data in there then the
2285 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2286 * then will be cleared.
2288 current->tlb_ubc.flush_required = false;
2291 static inline void init_tlb_ubc(void)
2294 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2297 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2299 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2300 struct scan_control *sc, unsigned long *lru_pages)
2302 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2303 unsigned long nr[NR_LRU_LISTS];
2304 unsigned long targets[NR_LRU_LISTS];
2305 unsigned long nr_to_scan;
2307 unsigned long nr_reclaimed = 0;
2308 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2309 struct blk_plug plug;
2312 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2314 /* Record the original scan target for proportional adjustments later */
2315 memcpy(targets, nr, sizeof(nr));
2318 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2319 * event that can occur when there is little memory pressure e.g.
2320 * multiple streaming readers/writers. Hence, we do not abort scanning
2321 * when the requested number of pages are reclaimed when scanning at
2322 * DEF_PRIORITY on the assumption that the fact we are direct
2323 * reclaiming implies that kswapd is not keeping up and it is best to
2324 * do a batch of work at once. For memcg reclaim one check is made to
2325 * abort proportional reclaim if either the file or anon lru has already
2326 * dropped to zero at the first pass.
2328 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2329 sc->priority == DEF_PRIORITY);
2333 blk_start_plug(&plug);
2334 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2335 nr[LRU_INACTIVE_FILE]) {
2336 unsigned long nr_anon, nr_file, percentage;
2337 unsigned long nr_scanned;
2339 for_each_evictable_lru(lru) {
2341 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2342 nr[lru] -= nr_to_scan;
2344 nr_reclaimed += shrink_list(lru, nr_to_scan,
2349 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2353 * For kswapd and memcg, reclaim at least the number of pages
2354 * requested. Ensure that the anon and file LRUs are scanned
2355 * proportionally what was requested by get_scan_count(). We
2356 * stop reclaiming one LRU and reduce the amount scanning
2357 * proportional to the original scan target.
2359 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2360 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2363 * It's just vindictive to attack the larger once the smaller
2364 * has gone to zero. And given the way we stop scanning the
2365 * smaller below, this makes sure that we only make one nudge
2366 * towards proportionality once we've got nr_to_reclaim.
2368 if (!nr_file || !nr_anon)
2371 if (nr_file > nr_anon) {
2372 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2373 targets[LRU_ACTIVE_ANON] + 1;
2375 percentage = nr_anon * 100 / scan_target;
2377 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2378 targets[LRU_ACTIVE_FILE] + 1;
2380 percentage = nr_file * 100 / scan_target;
2383 /* Stop scanning the smaller of the LRU */
2385 nr[lru + LRU_ACTIVE] = 0;
2388 * Recalculate the other LRU scan count based on its original
2389 * scan target and the percentage scanning already complete
2391 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2392 nr_scanned = targets[lru] - nr[lru];
2393 nr[lru] = targets[lru] * (100 - percentage) / 100;
2394 nr[lru] -= min(nr[lru], nr_scanned);
2397 nr_scanned = targets[lru] - nr[lru];
2398 nr[lru] = targets[lru] * (100 - percentage) / 100;
2399 nr[lru] -= min(nr[lru], nr_scanned);
2401 scan_adjusted = true;
2403 blk_finish_plug(&plug);
2404 sc->nr_reclaimed += nr_reclaimed;
2407 * Even if we did not try to evict anon pages at all, we want to
2408 * rebalance the anon lru active/inactive ratio.
2410 if (inactive_list_is_low(lruvec, false, sc))
2411 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2412 sc, LRU_ACTIVE_ANON);
2414 throttle_vm_writeout(sc->gfp_mask);
2417 /* Use reclaim/compaction for costly allocs or under memory pressure */
2418 static bool in_reclaim_compaction(struct scan_control *sc)
2420 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2421 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2422 sc->priority < DEF_PRIORITY - 2))
2429 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2430 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2431 * true if more pages should be reclaimed such that when the page allocator
2432 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2433 * It will give up earlier than that if there is difficulty reclaiming pages.
2435 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2436 unsigned long nr_reclaimed,
2437 unsigned long nr_scanned,
2438 struct scan_control *sc)
2440 unsigned long pages_for_compaction;
2441 unsigned long inactive_lru_pages;
2444 /* If not in reclaim/compaction mode, stop */
2445 if (!in_reclaim_compaction(sc))
2448 /* Consider stopping depending on scan and reclaim activity */
2449 if (sc->gfp_mask & __GFP_REPEAT) {
2451 * For __GFP_REPEAT allocations, stop reclaiming if the
2452 * full LRU list has been scanned and we are still failing
2453 * to reclaim pages. This full LRU scan is potentially
2454 * expensive but a __GFP_REPEAT caller really wants to succeed
2456 if (!nr_reclaimed && !nr_scanned)
2460 * For non-__GFP_REPEAT allocations which can presumably
2461 * fail without consequence, stop if we failed to reclaim
2462 * any pages from the last SWAP_CLUSTER_MAX number of
2463 * pages that were scanned. This will return to the
2464 * caller faster at the risk reclaim/compaction and
2465 * the resulting allocation attempt fails
2472 * If we have not reclaimed enough pages for compaction and the
2473 * inactive lists are large enough, continue reclaiming
2475 pages_for_compaction = (2UL << sc->order);
2476 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2477 if (get_nr_swap_pages() > 0)
2478 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2479 if (sc->nr_reclaimed < pages_for_compaction &&
2480 inactive_lru_pages > pages_for_compaction)
2483 /* If compaction would go ahead or the allocation would succeed, stop */
2484 for (z = 0; z <= sc->reclaim_idx; z++) {
2485 struct zone *zone = &pgdat->node_zones[z];
2486 if (!populated_zone(zone))
2489 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2490 case COMPACT_PARTIAL:
2491 case COMPACT_CONTINUE:
2494 /* check next zone */
2501 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2503 struct reclaim_state *reclaim_state = current->reclaim_state;
2504 unsigned long nr_reclaimed, nr_scanned;
2505 bool reclaimable = false;
2508 struct mem_cgroup *root = sc->target_mem_cgroup;
2509 struct mem_cgroup_reclaim_cookie reclaim = {
2511 .priority = sc->priority,
2513 unsigned long node_lru_pages = 0;
2514 struct mem_cgroup *memcg;
2516 nr_reclaimed = sc->nr_reclaimed;
2517 nr_scanned = sc->nr_scanned;
2519 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2521 unsigned long lru_pages;
2522 unsigned long reclaimed;
2523 unsigned long scanned;
2525 if (mem_cgroup_low(root, memcg)) {
2526 if (!sc->may_thrash)
2528 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2531 reclaimed = sc->nr_reclaimed;
2532 scanned = sc->nr_scanned;
2534 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2535 node_lru_pages += lru_pages;
2537 if (!global_reclaim(sc))
2538 shrink_slab(sc->gfp_mask, pgdat->node_id,
2539 memcg, sc->nr_scanned - scanned,
2542 /* Record the group's reclaim efficiency */
2543 vmpressure(sc->gfp_mask, memcg, false,
2544 sc->nr_scanned - scanned,
2545 sc->nr_reclaimed - reclaimed);
2548 * Direct reclaim and kswapd have to scan all memory
2549 * cgroups to fulfill the overall scan target for the
2552 * Limit reclaim, on the other hand, only cares about
2553 * nr_to_reclaim pages to be reclaimed and it will
2554 * retry with decreasing priority if one round over the
2555 * whole hierarchy is not sufficient.
2557 if (!global_reclaim(sc) &&
2558 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2559 mem_cgroup_iter_break(root, memcg);
2562 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2565 * Shrink the slab caches in the same proportion that
2566 * the eligible LRU pages were scanned.
2568 if (global_reclaim(sc))
2569 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2570 sc->nr_scanned - nr_scanned,
2573 if (reclaim_state) {
2574 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2575 reclaim_state->reclaimed_slab = 0;
2578 /* Record the subtree's reclaim efficiency */
2579 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2580 sc->nr_scanned - nr_scanned,
2581 sc->nr_reclaimed - nr_reclaimed);
2583 if (sc->nr_reclaimed - nr_reclaimed)
2586 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2587 sc->nr_scanned - nr_scanned, sc));
2593 * Returns true if compaction should go ahead for a high-order request, or
2594 * the high-order allocation would succeed without compaction.
2596 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2598 unsigned long watermark;
2602 * Compaction takes time to run and there are potentially other
2603 * callers using the pages just freed. Continue reclaiming until
2604 * there is a buffer of free pages available to give compaction
2605 * a reasonable chance of completing and allocating the page
2607 watermark = high_wmark_pages(zone) + (2UL << sc->order);
2608 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2611 * If compaction is deferred, reclaim up to a point where
2612 * compaction will have a chance of success when re-enabled
2614 if (compaction_deferred(zone, sc->order))
2615 return watermark_ok;
2618 * If compaction is not ready to start and allocation is not likely
2619 * to succeed without it, then keep reclaiming.
2621 if (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx) == COMPACT_SKIPPED)
2624 return watermark_ok;
2628 * This is the direct reclaim path, for page-allocating processes. We only
2629 * try to reclaim pages from zones which will satisfy the caller's allocation
2632 * If a zone is deemed to be full of pinned pages then just give it a light
2633 * scan then give up on it.
2635 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2639 unsigned long nr_soft_reclaimed;
2640 unsigned long nr_soft_scanned;
2642 pg_data_t *last_pgdat = NULL;
2645 * If the number of buffer_heads in the machine exceeds the maximum
2646 * allowed level, force direct reclaim to scan the highmem zone as
2647 * highmem pages could be pinning lowmem pages storing buffer_heads
2649 orig_mask = sc->gfp_mask;
2650 if (buffer_heads_over_limit) {
2651 sc->gfp_mask |= __GFP_HIGHMEM;
2652 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2655 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2656 sc->reclaim_idx, sc->nodemask) {
2658 * Take care memory controller reclaiming has small influence
2661 if (global_reclaim(sc)) {
2662 if (!cpuset_zone_allowed(zone,
2663 GFP_KERNEL | __GFP_HARDWALL))
2666 if (sc->priority != DEF_PRIORITY &&
2667 !pgdat_reclaimable(zone->zone_pgdat))
2668 continue; /* Let kswapd poll it */
2671 * If we already have plenty of memory free for
2672 * compaction in this zone, don't free any more.
2673 * Even though compaction is invoked for any
2674 * non-zero order, only frequent costly order
2675 * reclamation is disruptive enough to become a
2676 * noticeable problem, like transparent huge
2679 if (IS_ENABLED(CONFIG_COMPACTION) &&
2680 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2681 compaction_ready(zone, sc)) {
2682 sc->compaction_ready = true;
2687 * Shrink each node in the zonelist once. If the
2688 * zonelist is ordered by zone (not the default) then a
2689 * node may be shrunk multiple times but in that case
2690 * the user prefers lower zones being preserved.
2692 if (zone->zone_pgdat == last_pgdat)
2696 * This steals pages from memory cgroups over softlimit
2697 * and returns the number of reclaimed pages and
2698 * scanned pages. This works for global memory pressure
2699 * and balancing, not for a memcg's limit.
2701 nr_soft_scanned = 0;
2702 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2703 sc->order, sc->gfp_mask,
2705 sc->nr_reclaimed += nr_soft_reclaimed;
2706 sc->nr_scanned += nr_soft_scanned;
2707 /* need some check for avoid more shrink_zone() */
2710 /* See comment about same check for global reclaim above */
2711 if (zone->zone_pgdat == last_pgdat)
2713 last_pgdat = zone->zone_pgdat;
2714 shrink_node(zone->zone_pgdat, sc);
2718 * Restore to original mask to avoid the impact on the caller if we
2719 * promoted it to __GFP_HIGHMEM.
2721 sc->gfp_mask = orig_mask;
2725 * This is the main entry point to direct page reclaim.
2727 * If a full scan of the inactive list fails to free enough memory then we
2728 * are "out of memory" and something needs to be killed.
2730 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2731 * high - the zone may be full of dirty or under-writeback pages, which this
2732 * caller can't do much about. We kick the writeback threads and take explicit
2733 * naps in the hope that some of these pages can be written. But if the
2734 * allocating task holds filesystem locks which prevent writeout this might not
2735 * work, and the allocation attempt will fail.
2737 * returns: 0, if no pages reclaimed
2738 * else, the number of pages reclaimed
2740 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2741 struct scan_control *sc)
2743 int initial_priority = sc->priority;
2744 unsigned long total_scanned = 0;
2745 unsigned long writeback_threshold;
2747 delayacct_freepages_start();
2749 if (global_reclaim(sc))
2750 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2753 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2756 shrink_zones(zonelist, sc);
2758 total_scanned += sc->nr_scanned;
2759 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2762 if (sc->compaction_ready)
2766 * If we're getting trouble reclaiming, start doing
2767 * writepage even in laptop mode.
2769 if (sc->priority < DEF_PRIORITY - 2)
2770 sc->may_writepage = 1;
2773 * Try to write back as many pages as we just scanned. This
2774 * tends to cause slow streaming writers to write data to the
2775 * disk smoothly, at the dirtying rate, which is nice. But
2776 * that's undesirable in laptop mode, where we *want* lumpy
2777 * writeout. So in laptop mode, write out the whole world.
2779 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2780 if (total_scanned > writeback_threshold) {
2781 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2782 WB_REASON_TRY_TO_FREE_PAGES);
2783 sc->may_writepage = 1;
2785 } while (--sc->priority >= 0);
2787 delayacct_freepages_end();
2789 if (sc->nr_reclaimed)
2790 return sc->nr_reclaimed;
2792 /* Aborted reclaim to try compaction? don't OOM, then */
2793 if (sc->compaction_ready)
2796 /* Untapped cgroup reserves? Don't OOM, retry. */
2797 if (!sc->may_thrash) {
2798 sc->priority = initial_priority;
2806 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2809 unsigned long pfmemalloc_reserve = 0;
2810 unsigned long free_pages = 0;
2814 for (i = 0; i <= ZONE_NORMAL; i++) {
2815 zone = &pgdat->node_zones[i];
2816 if (!populated_zone(zone) ||
2817 pgdat_reclaimable_pages(pgdat) == 0)
2820 pfmemalloc_reserve += min_wmark_pages(zone);
2821 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2824 /* If there are no reserves (unexpected config) then do not throttle */
2825 if (!pfmemalloc_reserve)
2828 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2830 /* kswapd must be awake if processes are being throttled */
2831 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2832 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2833 (enum zone_type)ZONE_NORMAL);
2834 wake_up_interruptible(&pgdat->kswapd_wait);
2841 * Throttle direct reclaimers if backing storage is backed by the network
2842 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2843 * depleted. kswapd will continue to make progress and wake the processes
2844 * when the low watermark is reached.
2846 * Returns true if a fatal signal was delivered during throttling. If this
2847 * happens, the page allocator should not consider triggering the OOM killer.
2849 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2850 nodemask_t *nodemask)
2854 pg_data_t *pgdat = NULL;
2857 * Kernel threads should not be throttled as they may be indirectly
2858 * responsible for cleaning pages necessary for reclaim to make forward
2859 * progress. kjournald for example may enter direct reclaim while
2860 * committing a transaction where throttling it could forcing other
2861 * processes to block on log_wait_commit().
2863 if (current->flags & PF_KTHREAD)
2867 * If a fatal signal is pending, this process should not throttle.
2868 * It should return quickly so it can exit and free its memory
2870 if (fatal_signal_pending(current))
2874 * Check if the pfmemalloc reserves are ok by finding the first node
2875 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2876 * GFP_KERNEL will be required for allocating network buffers when
2877 * swapping over the network so ZONE_HIGHMEM is unusable.
2879 * Throttling is based on the first usable node and throttled processes
2880 * wait on a queue until kswapd makes progress and wakes them. There
2881 * is an affinity then between processes waking up and where reclaim
2882 * progress has been made assuming the process wakes on the same node.
2883 * More importantly, processes running on remote nodes will not compete
2884 * for remote pfmemalloc reserves and processes on different nodes
2885 * should make reasonable progress.
2887 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2888 gfp_zone(gfp_mask), nodemask) {
2889 if (zone_idx(zone) > ZONE_NORMAL)
2892 /* Throttle based on the first usable node */
2893 pgdat = zone->zone_pgdat;
2894 if (pfmemalloc_watermark_ok(pgdat))
2899 /* If no zone was usable by the allocation flags then do not throttle */
2903 /* Account for the throttling */
2904 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2907 * If the caller cannot enter the filesystem, it's possible that it
2908 * is due to the caller holding an FS lock or performing a journal
2909 * transaction in the case of a filesystem like ext[3|4]. In this case,
2910 * it is not safe to block on pfmemalloc_wait as kswapd could be
2911 * blocked waiting on the same lock. Instead, throttle for up to a
2912 * second before continuing.
2914 if (!(gfp_mask & __GFP_FS)) {
2915 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2916 pfmemalloc_watermark_ok(pgdat), HZ);
2921 /* Throttle until kswapd wakes the process */
2922 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2923 pfmemalloc_watermark_ok(pgdat));
2926 if (fatal_signal_pending(current))
2933 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2934 gfp_t gfp_mask, nodemask_t *nodemask)
2936 unsigned long nr_reclaimed;
2937 struct scan_control sc = {
2938 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2939 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2940 .reclaim_idx = gfp_zone(gfp_mask),
2942 .nodemask = nodemask,
2943 .priority = DEF_PRIORITY,
2944 .may_writepage = !laptop_mode,
2950 * Do not enter reclaim if fatal signal was delivered while throttled.
2951 * 1 is returned so that the page allocator does not OOM kill at this
2954 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2957 trace_mm_vmscan_direct_reclaim_begin(order,
2962 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2964 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2966 return nr_reclaimed;
2971 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2972 gfp_t gfp_mask, bool noswap,
2974 unsigned long *nr_scanned)
2976 struct scan_control sc = {
2977 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2978 .target_mem_cgroup = memcg,
2979 .may_writepage = !laptop_mode,
2981 .reclaim_idx = MAX_NR_ZONES - 1,
2982 .may_swap = !noswap,
2984 unsigned long lru_pages;
2986 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2987 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2989 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2995 * NOTE: Although we can get the priority field, using it
2996 * here is not a good idea, since it limits the pages we can scan.
2997 * if we don't reclaim here, the shrink_node from balance_pgdat
2998 * will pick up pages from other mem cgroup's as well. We hack
2999 * the priority and make it zero.
3001 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3003 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3005 *nr_scanned = sc.nr_scanned;
3006 return sc.nr_reclaimed;
3009 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3010 unsigned long nr_pages,
3014 struct zonelist *zonelist;
3015 unsigned long nr_reclaimed;
3017 struct scan_control sc = {
3018 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3019 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3020 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3021 .reclaim_idx = MAX_NR_ZONES - 1,
3022 .target_mem_cgroup = memcg,
3023 .priority = DEF_PRIORITY,
3024 .may_writepage = !laptop_mode,
3026 .may_swap = may_swap,
3030 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3031 * take care of from where we get pages. So the node where we start the
3032 * scan does not need to be the current node.
3034 nid = mem_cgroup_select_victim_node(memcg);
3036 zonelist = NODE_DATA(nid)->node_zonelists;
3038 trace_mm_vmscan_memcg_reclaim_begin(0,
3043 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3045 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3047 return nr_reclaimed;
3051 static void age_active_anon(struct pglist_data *pgdat,
3052 struct scan_control *sc)
3054 struct mem_cgroup *memcg;
3056 if (!total_swap_pages)
3059 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3061 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3063 if (inactive_list_is_low(lruvec, false, sc))
3064 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3065 sc, LRU_ACTIVE_ANON);
3067 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3071 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3073 unsigned long mark = high_wmark_pages(zone);
3075 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3079 * If any eligible zone is balanced then the node is not considered
3080 * to be congested or dirty
3082 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3083 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3089 * Prepare kswapd for sleeping. This verifies that there are no processes
3090 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3092 * Returns true if kswapd is ready to sleep
3094 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3099 * The throttled processes are normally woken up in balance_pgdat() as
3100 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3101 * race between when kswapd checks the watermarks and a process gets
3102 * throttled. There is also a potential race if processes get
3103 * throttled, kswapd wakes, a large process exits thereby balancing the
3104 * zones, which causes kswapd to exit balance_pgdat() before reaching
3105 * the wake up checks. If kswapd is going to sleep, no process should
3106 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3107 * the wake up is premature, processes will wake kswapd and get
3108 * throttled again. The difference from wake ups in balance_pgdat() is
3109 * that here we are under prepare_to_wait().
3111 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3112 wake_up_all(&pgdat->pfmemalloc_wait);
3114 for (i = 0; i <= classzone_idx; i++) {
3115 struct zone *zone = pgdat->node_zones + i;
3117 if (!populated_zone(zone))
3120 if (!zone_balanced(zone, order, classzone_idx))
3128 * kswapd shrinks a node of pages that are at or below the highest usable
3129 * zone that is currently unbalanced.
3131 * Returns true if kswapd scanned at least the requested number of pages to
3132 * reclaim or if the lack of progress was due to pages under writeback.
3133 * This is used to determine if the scanning priority needs to be raised.
3135 static bool kswapd_shrink_node(pg_data_t *pgdat,
3136 struct scan_control *sc)
3141 /* Reclaim a number of pages proportional to the number of zones */
3142 sc->nr_to_reclaim = 0;
3143 for (z = 0; z <= sc->reclaim_idx; z++) {
3144 zone = pgdat->node_zones + z;
3145 if (!populated_zone(zone))
3148 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3152 * Historically care was taken to put equal pressure on all zones but
3153 * now pressure is applied based on node LRU order.
3155 shrink_node(pgdat, sc);
3158 * Fragmentation may mean that the system cannot be rebalanced for
3159 * high-order allocations. If twice the allocation size has been
3160 * reclaimed then recheck watermarks only at order-0 to prevent
3161 * excessive reclaim. Assume that a process requested a high-order
3162 * can direct reclaim/compact.
3164 if (sc->order && sc->nr_reclaimed >= 2UL << sc->order)
3167 return sc->nr_scanned >= sc->nr_to_reclaim;
3171 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3172 * that are eligible for use by the caller until at least one zone is
3175 * Returns the order kswapd finished reclaiming at.
3177 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3178 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3179 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3180 * or lower is eligible for reclaim until at least one usable zone is
3183 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3186 unsigned long nr_soft_reclaimed;
3187 unsigned long nr_soft_scanned;
3189 struct scan_control sc = {
3190 .gfp_mask = GFP_KERNEL,
3192 .priority = DEF_PRIORITY,
3193 .may_writepage = !laptop_mode,
3197 count_vm_event(PAGEOUTRUN);
3200 bool raise_priority = true;
3202 sc.nr_reclaimed = 0;
3203 sc.reclaim_idx = classzone_idx;
3206 * If the number of buffer_heads exceeds the maximum allowed
3207 * then consider reclaiming from all zones. This has a dual
3208 * purpose -- on 64-bit systems it is expected that
3209 * buffer_heads are stripped during active rotation. On 32-bit
3210 * systems, highmem pages can pin lowmem memory and shrinking
3211 * buffers can relieve lowmem pressure. Reclaim may still not
3212 * go ahead if all eligible zones for the original allocation
3213 * request are balanced to avoid excessive reclaim from kswapd.
3215 if (buffer_heads_over_limit) {
3216 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3217 zone = pgdat->node_zones + i;
3218 if (!populated_zone(zone))
3227 * Only reclaim if there are no eligible zones. Check from
3228 * high to low zone as allocations prefer higher zones.
3229 * Scanning from low to high zone would allow congestion to be
3230 * cleared during a very small window when a small low
3231 * zone was balanced even under extreme pressure when the
3232 * overall node may be congested. Note that sc.reclaim_idx
3233 * is not used as buffer_heads_over_limit may have adjusted
3236 for (i = classzone_idx; i >= 0; i--) {
3237 zone = pgdat->node_zones + i;
3238 if (!populated_zone(zone))
3241 if (zone_balanced(zone, sc.order, classzone_idx))
3246 * Do some background aging of the anon list, to give
3247 * pages a chance to be referenced before reclaiming. All
3248 * pages are rotated regardless of classzone as this is
3249 * about consistent aging.
3251 age_active_anon(pgdat, &sc);
3254 * If we're getting trouble reclaiming, start doing writepage
3255 * even in laptop mode.
3257 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3258 sc.may_writepage = 1;
3260 /* Call soft limit reclaim before calling shrink_node. */
3262 nr_soft_scanned = 0;
3263 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3264 sc.gfp_mask, &nr_soft_scanned);
3265 sc.nr_reclaimed += nr_soft_reclaimed;
3268 * There should be no need to raise the scanning priority if
3269 * enough pages are already being scanned that that high
3270 * watermark would be met at 100% efficiency.
3272 if (kswapd_shrink_node(pgdat, &sc))
3273 raise_priority = false;
3276 * If the low watermark is met there is no need for processes
3277 * to be throttled on pfmemalloc_wait as they should not be
3278 * able to safely make forward progress. Wake them
3280 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3281 pfmemalloc_watermark_ok(pgdat))
3282 wake_up_all(&pgdat->pfmemalloc_wait);
3284 /* Check if kswapd should be suspending */
3285 if (try_to_freeze() || kthread_should_stop())
3289 * Raise priority if scanning rate is too low or there was no
3290 * progress in reclaiming pages
3292 if (raise_priority || !sc.nr_reclaimed)
3294 } while (sc.priority >= 1);
3298 * Return the order kswapd stopped reclaiming at as
3299 * prepare_kswapd_sleep() takes it into account. If another caller
3300 * entered the allocator slow path while kswapd was awake, order will
3301 * remain at the higher level.
3306 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3307 unsigned int classzone_idx)
3312 if (freezing(current) || kthread_should_stop())
3315 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3317 /* Try to sleep for a short interval */
3318 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3320 * Compaction records what page blocks it recently failed to
3321 * isolate pages from and skips them in the future scanning.
3322 * When kswapd is going to sleep, it is reasonable to assume
3323 * that pages and compaction may succeed so reset the cache.
3325 reset_isolation_suitable(pgdat);
3328 * We have freed the memory, now we should compact it to make
3329 * allocation of the requested order possible.
3331 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3333 remaining = schedule_timeout(HZ/10);
3336 * If woken prematurely then reset kswapd_classzone_idx and
3337 * order. The values will either be from a wakeup request or
3338 * the previous request that slept prematurely.
3341 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3342 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3345 finish_wait(&pgdat->kswapd_wait, &wait);
3346 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3350 * After a short sleep, check if it was a premature sleep. If not, then
3351 * go fully to sleep until explicitly woken up.
3354 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3355 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3358 * vmstat counters are not perfectly accurate and the estimated
3359 * value for counters such as NR_FREE_PAGES can deviate from the
3360 * true value by nr_online_cpus * threshold. To avoid the zone
3361 * watermarks being breached while under pressure, we reduce the
3362 * per-cpu vmstat threshold while kswapd is awake and restore
3363 * them before going back to sleep.
3365 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3367 if (!kthread_should_stop())
3370 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3373 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3375 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3377 finish_wait(&pgdat->kswapd_wait, &wait);
3381 * The background pageout daemon, started as a kernel thread
3382 * from the init process.
3384 * This basically trickles out pages so that we have _some_
3385 * free memory available even if there is no other activity
3386 * that frees anything up. This is needed for things like routing
3387 * etc, where we otherwise might have all activity going on in
3388 * asynchronous contexts that cannot page things out.
3390 * If there are applications that are active memory-allocators
3391 * (most normal use), this basically shouldn't matter.
3393 static int kswapd(void *p)
3395 unsigned int alloc_order, reclaim_order, classzone_idx;
3396 pg_data_t *pgdat = (pg_data_t*)p;
3397 struct task_struct *tsk = current;
3399 struct reclaim_state reclaim_state = {
3400 .reclaimed_slab = 0,
3402 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3404 lockdep_set_current_reclaim_state(GFP_KERNEL);
3406 if (!cpumask_empty(cpumask))
3407 set_cpus_allowed_ptr(tsk, cpumask);
3408 current->reclaim_state = &reclaim_state;
3411 * Tell the memory management that we're a "memory allocator",
3412 * and that if we need more memory we should get access to it
3413 * regardless (see "__alloc_pages()"). "kswapd" should
3414 * never get caught in the normal page freeing logic.
3416 * (Kswapd normally doesn't need memory anyway, but sometimes
3417 * you need a small amount of memory in order to be able to
3418 * page out something else, and this flag essentially protects
3419 * us from recursively trying to free more memory as we're
3420 * trying to free the first piece of memory in the first place).
3422 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3425 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3426 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3431 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3434 /* Read the new order and classzone_idx */
3435 alloc_order = reclaim_order = pgdat->kswapd_order;
3436 classzone_idx = pgdat->kswapd_classzone_idx;
3437 pgdat->kswapd_order = 0;
3438 pgdat->kswapd_classzone_idx = 0;
3440 ret = try_to_freeze();
3441 if (kthread_should_stop())
3445 * We can speed up thawing tasks if we don't call balance_pgdat
3446 * after returning from the refrigerator
3452 * Reclaim begins at the requested order but if a high-order
3453 * reclaim fails then kswapd falls back to reclaiming for
3454 * order-0. If that happens, kswapd will consider sleeping
3455 * for the order it finished reclaiming at (reclaim_order)
3456 * but kcompactd is woken to compact for the original
3457 * request (alloc_order).
3459 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3461 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3462 if (reclaim_order < alloc_order)
3463 goto kswapd_try_sleep;
3465 alloc_order = reclaim_order = pgdat->kswapd_order;
3466 classzone_idx = pgdat->kswapd_classzone_idx;
3469 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3470 current->reclaim_state = NULL;
3471 lockdep_clear_current_reclaim_state();
3477 * A zone is low on free memory, so wake its kswapd task to service it.
3479 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3484 if (!populated_zone(zone))
3487 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3489 pgdat = zone->zone_pgdat;
3490 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3491 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3492 if (!waitqueue_active(&pgdat->kswapd_wait))
3495 /* Only wake kswapd if all zones are unbalanced */
3496 for (z = 0; z <= classzone_idx; z++) {
3497 zone = pgdat->node_zones + z;
3498 if (!populated_zone(zone))
3501 if (zone_balanced(zone, order, classzone_idx))
3505 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3506 wake_up_interruptible(&pgdat->kswapd_wait);
3509 #ifdef CONFIG_HIBERNATION
3511 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3514 * Rather than trying to age LRUs the aim is to preserve the overall
3515 * LRU order by reclaiming preferentially
3516 * inactive > active > active referenced > active mapped
3518 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3520 struct reclaim_state reclaim_state;
3521 struct scan_control sc = {
3522 .nr_to_reclaim = nr_to_reclaim,
3523 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3524 .reclaim_idx = MAX_NR_ZONES - 1,
3525 .priority = DEF_PRIORITY,
3529 .hibernation_mode = 1,
3531 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3532 struct task_struct *p = current;
3533 unsigned long nr_reclaimed;
3535 p->flags |= PF_MEMALLOC;
3536 lockdep_set_current_reclaim_state(sc.gfp_mask);
3537 reclaim_state.reclaimed_slab = 0;
3538 p->reclaim_state = &reclaim_state;
3540 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3542 p->reclaim_state = NULL;
3543 lockdep_clear_current_reclaim_state();
3544 p->flags &= ~PF_MEMALLOC;
3546 return nr_reclaimed;
3548 #endif /* CONFIG_HIBERNATION */
3550 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3551 not required for correctness. So if the last cpu in a node goes
3552 away, we get changed to run anywhere: as the first one comes back,
3553 restore their cpu bindings. */
3554 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3559 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3560 for_each_node_state(nid, N_MEMORY) {
3561 pg_data_t *pgdat = NODE_DATA(nid);
3562 const struct cpumask *mask;
3564 mask = cpumask_of_node(pgdat->node_id);
3566 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3567 /* One of our CPUs online: restore mask */
3568 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3575 * This kswapd start function will be called by init and node-hot-add.
3576 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3578 int kswapd_run(int nid)
3580 pg_data_t *pgdat = NODE_DATA(nid);
3586 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3587 if (IS_ERR(pgdat->kswapd)) {
3588 /* failure at boot is fatal */
3589 BUG_ON(system_state == SYSTEM_BOOTING);
3590 pr_err("Failed to start kswapd on node %d\n", nid);
3591 ret = PTR_ERR(pgdat->kswapd);
3592 pgdat->kswapd = NULL;
3598 * Called by memory hotplug when all memory in a node is offlined. Caller must
3599 * hold mem_hotplug_begin/end().
3601 void kswapd_stop(int nid)
3603 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3606 kthread_stop(kswapd);
3607 NODE_DATA(nid)->kswapd = NULL;
3611 static int __init kswapd_init(void)
3616 for_each_node_state(nid, N_MEMORY)
3618 hotcpu_notifier(cpu_callback, 0);
3622 module_init(kswapd_init)
3628 * If non-zero call node_reclaim when the number of free pages falls below
3631 int node_reclaim_mode __read_mostly;
3633 #define RECLAIM_OFF 0
3634 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3635 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3636 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3639 * Priority for NODE_RECLAIM. This determines the fraction of pages
3640 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3643 #define NODE_RECLAIM_PRIORITY 4
3646 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3649 int sysctl_min_unmapped_ratio = 1;
3652 * If the number of slab pages in a zone grows beyond this percentage then
3653 * slab reclaim needs to occur.
3655 int sysctl_min_slab_ratio = 5;
3657 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3659 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3660 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3661 node_page_state(pgdat, NR_ACTIVE_FILE);
3664 * It's possible for there to be more file mapped pages than
3665 * accounted for by the pages on the file LRU lists because
3666 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3668 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3671 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3672 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3674 unsigned long nr_pagecache_reclaimable;
3675 unsigned long delta = 0;
3678 * If RECLAIM_UNMAP is set, then all file pages are considered
3679 * potentially reclaimable. Otherwise, we have to worry about
3680 * pages like swapcache and node_unmapped_file_pages() provides
3683 if (node_reclaim_mode & RECLAIM_UNMAP)
3684 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3686 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3688 /* If we can't clean pages, remove dirty pages from consideration */
3689 if (!(node_reclaim_mode & RECLAIM_WRITE))
3690 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3692 /* Watch for any possible underflows due to delta */
3693 if (unlikely(delta > nr_pagecache_reclaimable))
3694 delta = nr_pagecache_reclaimable;
3696 return nr_pagecache_reclaimable - delta;
3700 * Try to free up some pages from this node through reclaim.
3702 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3704 /* Minimum pages needed in order to stay on node */
3705 const unsigned long nr_pages = 1 << order;
3706 struct task_struct *p = current;
3707 struct reclaim_state reclaim_state;
3708 int classzone_idx = gfp_zone(gfp_mask);
3709 struct scan_control sc = {
3710 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3711 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3713 .priority = NODE_RECLAIM_PRIORITY,
3714 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3715 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3717 .reclaim_idx = classzone_idx,
3722 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3723 * and we also need to be able to write out pages for RECLAIM_WRITE
3724 * and RECLAIM_UNMAP.
3726 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3727 lockdep_set_current_reclaim_state(gfp_mask);
3728 reclaim_state.reclaimed_slab = 0;
3729 p->reclaim_state = &reclaim_state;
3731 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3733 * Free memory by calling shrink zone with increasing
3734 * priorities until we have enough memory freed.
3737 shrink_node(pgdat, &sc);
3738 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3741 p->reclaim_state = NULL;
3742 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3743 lockdep_clear_current_reclaim_state();
3744 return sc.nr_reclaimed >= nr_pages;
3747 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3752 * Node reclaim reclaims unmapped file backed pages and
3753 * slab pages if we are over the defined limits.
3755 * A small portion of unmapped file backed pages is needed for
3756 * file I/O otherwise pages read by file I/O will be immediately
3757 * thrown out if the node is overallocated. So we do not reclaim
3758 * if less than a specified percentage of the node is used by
3759 * unmapped file backed pages.
3761 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3762 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3763 return NODE_RECLAIM_FULL;
3765 if (!pgdat_reclaimable(pgdat))
3766 return NODE_RECLAIM_FULL;
3769 * Do not scan if the allocation should not be delayed.
3771 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3772 return NODE_RECLAIM_NOSCAN;
3775 * Only run node reclaim on the local node or on nodes that do not
3776 * have associated processors. This will favor the local processor
3777 * over remote processors and spread off node memory allocations
3778 * as wide as possible.
3780 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3781 return NODE_RECLAIM_NOSCAN;
3783 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3784 return NODE_RECLAIM_NOSCAN;
3786 ret = __node_reclaim(pgdat, gfp_mask, order);
3787 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3790 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3797 * page_evictable - test whether a page is evictable
3798 * @page: the page to test
3800 * Test whether page is evictable--i.e., should be placed on active/inactive
3801 * lists vs unevictable list.
3803 * Reasons page might not be evictable:
3804 * (1) page's mapping marked unevictable
3805 * (2) page is part of an mlocked VMA
3808 int page_evictable(struct page *page)
3810 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3815 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3816 * @pages: array of pages to check
3817 * @nr_pages: number of pages to check
3819 * Checks pages for evictability and moves them to the appropriate lru list.
3821 * This function is only used for SysV IPC SHM_UNLOCK.
3823 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3825 struct lruvec *lruvec;
3826 struct pglist_data *pgdat = NULL;
3831 for (i = 0; i < nr_pages; i++) {
3832 struct page *page = pages[i];
3833 struct pglist_data *pagepgdat = page_pgdat(page);
3836 if (pagepgdat != pgdat) {
3838 spin_unlock_irq(&pgdat->lru_lock);
3840 spin_lock_irq(&pgdat->lru_lock);
3842 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3844 if (!PageLRU(page) || !PageUnevictable(page))
3847 if (page_evictable(page)) {
3848 enum lru_list lru = page_lru_base_type(page);
3850 VM_BUG_ON_PAGE(PageActive(page), page);
3851 ClearPageUnevictable(page);
3852 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3853 add_page_to_lru_list(page, lruvec, lru);
3859 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3860 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3861 spin_unlock_irq(&pgdat->lru_lock);
3864 #endif /* CONFIG_SHMEM */