Merge branch 'pm-opp'
[linux/fpc-iii.git] / mm / vmscan.c
blob532a2a750952daffed4e4242d36a449eb9ba8837
1 /*
2 * linux/mm/vmscan.c
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
16 #include <linux/mm.h>
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>
57 #include "internal.h"
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
62 struct scan_control {
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 /* This context's GFP mask */
67 gfp_t gfp_mask;
69 /* Allocation order */
70 int order;
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
76 nodemask_t *nodemask;
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 */
85 int priority;
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) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
122 } while (0)
123 #else
124 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
127 #ifdef ARCH_HAS_PREFETCHW
128 #define prefetchw_prev_lru_page(_page, _base, _field) \
129 do { \
130 if ((_page)->lru.prev != _base) { \
131 struct page *prev; \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
136 } while (0)
137 #else
138 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139 #endif
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
147 * zones.
149 unsigned long vm_total_pages;
151 static LIST_HEAD(shrinker_list);
152 static DECLARE_RWSEM(shrinker_rwsem);
154 #ifdef CONFIG_MEMCG
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;
177 if (!memcg)
178 return true;
179 #ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
181 return true;
182 #endif
183 return false;
185 #else
186 static bool global_reclaim(struct scan_control *sc)
188 return true;
191 static bool sane_reclaim(struct scan_control *sc)
193 return true;
195 #endif
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)
204 unsigned long nr;
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);
212 return nr;
215 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
217 unsigned long nr;
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);
228 return nr;
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);
245 unsigned long lruvec_zone_lru_size(struct lruvec *lruvec, enum lru_list lru,
246 int zone_idx)
248 if (!mem_cgroup_disabled())
249 return mem_cgroup_get_zone_lru_size(lruvec, lru, zone_idx);
251 return zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zone_idx],
252 NR_ZONE_LRU_BASE + lru);
256 * Add a shrinker callback to be called from the vm.
258 int register_shrinker(struct shrinker *shrinker)
260 size_t size = sizeof(*shrinker->nr_deferred);
262 if (shrinker->flags & SHRINKER_NUMA_AWARE)
263 size *= nr_node_ids;
265 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
266 if (!shrinker->nr_deferred)
267 return -ENOMEM;
269 down_write(&shrinker_rwsem);
270 list_add_tail(&shrinker->list, &shrinker_list);
271 up_write(&shrinker_rwsem);
272 return 0;
274 EXPORT_SYMBOL(register_shrinker);
277 * Remove one
279 void unregister_shrinker(struct shrinker *shrinker)
281 down_write(&shrinker_rwsem);
282 list_del(&shrinker->list);
283 up_write(&shrinker_rwsem);
284 kfree(shrinker->nr_deferred);
286 EXPORT_SYMBOL(unregister_shrinker);
288 #define SHRINK_BATCH 128
290 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
291 struct shrinker *shrinker,
292 unsigned long nr_scanned,
293 unsigned long nr_eligible)
295 unsigned long freed = 0;
296 unsigned long long delta;
297 long total_scan;
298 long freeable;
299 long nr;
300 long new_nr;
301 int nid = shrinkctl->nid;
302 long batch_size = shrinker->batch ? shrinker->batch
303 : SHRINK_BATCH;
304 long scanned = 0, next_deferred;
306 freeable = shrinker->count_objects(shrinker, shrinkctl);
307 if (freeable == 0)
308 return 0;
311 * copy the current shrinker scan count into a local variable
312 * and zero it so that other concurrent shrinker invocations
313 * don't also do this scanning work.
315 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
317 total_scan = nr;
318 delta = (4 * nr_scanned) / shrinker->seeks;
319 delta *= freeable;
320 do_div(delta, nr_eligible + 1);
321 total_scan += delta;
322 if (total_scan < 0) {
323 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
324 shrinker->scan_objects, total_scan);
325 total_scan = freeable;
326 next_deferred = nr;
327 } else
328 next_deferred = total_scan;
331 * We need to avoid excessive windup on filesystem shrinkers
332 * due to large numbers of GFP_NOFS allocations causing the
333 * shrinkers to return -1 all the time. This results in a large
334 * nr being built up so when a shrink that can do some work
335 * comes along it empties the entire cache due to nr >>>
336 * freeable. This is bad for sustaining a working set in
337 * memory.
339 * Hence only allow the shrinker to scan the entire cache when
340 * a large delta change is calculated directly.
342 if (delta < freeable / 4)
343 total_scan = min(total_scan, freeable / 2);
346 * Avoid risking looping forever due to too large nr value:
347 * never try to free more than twice the estimate number of
348 * freeable entries.
350 if (total_scan > freeable * 2)
351 total_scan = freeable * 2;
353 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
354 nr_scanned, nr_eligible,
355 freeable, delta, total_scan);
358 * Normally, we should not scan less than batch_size objects in one
359 * pass to avoid too frequent shrinker calls, but if the slab has less
360 * than batch_size objects in total and we are really tight on memory,
361 * we will try to reclaim all available objects, otherwise we can end
362 * up failing allocations although there are plenty of reclaimable
363 * objects spread over several slabs with usage less than the
364 * batch_size.
366 * We detect the "tight on memory" situations by looking at the total
367 * number of objects we want to scan (total_scan). If it is greater
368 * than the total number of objects on slab (freeable), we must be
369 * scanning at high prio and therefore should try to reclaim as much as
370 * possible.
372 while (total_scan >= batch_size ||
373 total_scan >= freeable) {
374 unsigned long ret;
375 unsigned long nr_to_scan = min(batch_size, total_scan);
377 shrinkctl->nr_to_scan = nr_to_scan;
378 ret = shrinker->scan_objects(shrinker, shrinkctl);
379 if (ret == SHRINK_STOP)
380 break;
381 freed += ret;
383 count_vm_events(SLABS_SCANNED, nr_to_scan);
384 total_scan -= nr_to_scan;
385 scanned += nr_to_scan;
387 cond_resched();
390 if (next_deferred >= scanned)
391 next_deferred -= scanned;
392 else
393 next_deferred = 0;
395 * move the unused scan count back into the shrinker in a
396 * manner that handles concurrent updates. If we exhausted the
397 * scan, there is no need to do an update.
399 if (next_deferred > 0)
400 new_nr = atomic_long_add_return(next_deferred,
401 &shrinker->nr_deferred[nid]);
402 else
403 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
405 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
406 return freed;
410 * shrink_slab - shrink slab caches
411 * @gfp_mask: allocation context
412 * @nid: node whose slab caches to target
413 * @memcg: memory cgroup whose slab caches to target
414 * @nr_scanned: pressure numerator
415 * @nr_eligible: pressure denominator
417 * Call the shrink functions to age shrinkable caches.
419 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
420 * unaware shrinkers will receive a node id of 0 instead.
422 * @memcg specifies the memory cgroup to target. If it is not NULL,
423 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
424 * objects from the memory cgroup specified. Otherwise, only unaware
425 * shrinkers are called.
427 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
428 * the available objects should be scanned. Page reclaim for example
429 * passes the number of pages scanned and the number of pages on the
430 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
431 * when it encountered mapped pages. The ratio is further biased by
432 * the ->seeks setting of the shrink function, which indicates the
433 * cost to recreate an object relative to that of an LRU page.
435 * Returns the number of reclaimed slab objects.
437 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
438 struct mem_cgroup *memcg,
439 unsigned long nr_scanned,
440 unsigned long nr_eligible)
442 struct shrinker *shrinker;
443 unsigned long freed = 0;
445 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
446 return 0;
448 if (nr_scanned == 0)
449 nr_scanned = SWAP_CLUSTER_MAX;
451 if (!down_read_trylock(&shrinker_rwsem)) {
453 * If we would return 0, our callers would understand that we
454 * have nothing else to shrink and give up trying. By returning
455 * 1 we keep it going and assume we'll be able to shrink next
456 * time.
458 freed = 1;
459 goto out;
462 list_for_each_entry(shrinker, &shrinker_list, list) {
463 struct shrink_control sc = {
464 .gfp_mask = gfp_mask,
465 .nid = nid,
466 .memcg = memcg,
470 * If kernel memory accounting is disabled, we ignore
471 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
472 * passing NULL for memcg.
474 if (memcg_kmem_enabled() &&
475 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
476 continue;
478 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
479 sc.nid = 0;
481 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
484 up_read(&shrinker_rwsem);
485 out:
486 cond_resched();
487 return freed;
490 void drop_slab_node(int nid)
492 unsigned long freed;
494 do {
495 struct mem_cgroup *memcg = NULL;
497 freed = 0;
498 do {
499 freed += shrink_slab(GFP_KERNEL, nid, memcg,
500 1000, 1000);
501 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
502 } while (freed > 10);
505 void drop_slab(void)
507 int nid;
509 for_each_online_node(nid)
510 drop_slab_node(nid);
513 static inline int is_page_cache_freeable(struct page *page)
516 * A freeable page cache page is referenced only by the caller
517 * that isolated the page, the page cache radix tree and
518 * optional buffer heads at page->private.
520 return page_count(page) - page_has_private(page) == 2;
523 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
525 if (current->flags & PF_SWAPWRITE)
526 return 1;
527 if (!inode_write_congested(inode))
528 return 1;
529 if (inode_to_bdi(inode) == current->backing_dev_info)
530 return 1;
531 return 0;
535 * We detected a synchronous write error writing a page out. Probably
536 * -ENOSPC. We need to propagate that into the address_space for a subsequent
537 * fsync(), msync() or close().
539 * The tricky part is that after writepage we cannot touch the mapping: nothing
540 * prevents it from being freed up. But we have a ref on the page and once
541 * that page is locked, the mapping is pinned.
543 * We're allowed to run sleeping lock_page() here because we know the caller has
544 * __GFP_FS.
546 static void handle_write_error(struct address_space *mapping,
547 struct page *page, int error)
549 lock_page(page);
550 if (page_mapping(page) == mapping)
551 mapping_set_error(mapping, error);
552 unlock_page(page);
555 /* possible outcome of pageout() */
556 typedef enum {
557 /* failed to write page out, page is locked */
558 PAGE_KEEP,
559 /* move page to the active list, page is locked */
560 PAGE_ACTIVATE,
561 /* page has been sent to the disk successfully, page is unlocked */
562 PAGE_SUCCESS,
563 /* page is clean and locked */
564 PAGE_CLEAN,
565 } pageout_t;
568 * pageout is called by shrink_page_list() for each dirty page.
569 * Calls ->writepage().
571 static pageout_t pageout(struct page *page, struct address_space *mapping,
572 struct scan_control *sc)
575 * If the page is dirty, only perform writeback if that write
576 * will be non-blocking. To prevent this allocation from being
577 * stalled by pagecache activity. But note that there may be
578 * stalls if we need to run get_block(). We could test
579 * PagePrivate for that.
581 * If this process is currently in __generic_file_write_iter() against
582 * this page's queue, we can perform writeback even if that
583 * will block.
585 * If the page is swapcache, write it back even if that would
586 * block, for some throttling. This happens by accident, because
587 * swap_backing_dev_info is bust: it doesn't reflect the
588 * congestion state of the swapdevs. Easy to fix, if needed.
590 if (!is_page_cache_freeable(page))
591 return PAGE_KEEP;
592 if (!mapping) {
594 * Some data journaling orphaned pages can have
595 * page->mapping == NULL while being dirty with clean buffers.
597 if (page_has_private(page)) {
598 if (try_to_free_buffers(page)) {
599 ClearPageDirty(page);
600 pr_info("%s: orphaned page\n", __func__);
601 return PAGE_CLEAN;
604 return PAGE_KEEP;
606 if (mapping->a_ops->writepage == NULL)
607 return PAGE_ACTIVATE;
608 if (!may_write_to_inode(mapping->host, sc))
609 return PAGE_KEEP;
611 if (clear_page_dirty_for_io(page)) {
612 int res;
613 struct writeback_control wbc = {
614 .sync_mode = WB_SYNC_NONE,
615 .nr_to_write = SWAP_CLUSTER_MAX,
616 .range_start = 0,
617 .range_end = LLONG_MAX,
618 .for_reclaim = 1,
621 SetPageReclaim(page);
622 res = mapping->a_ops->writepage(page, &wbc);
623 if (res < 0)
624 handle_write_error(mapping, page, res);
625 if (res == AOP_WRITEPAGE_ACTIVATE) {
626 ClearPageReclaim(page);
627 return PAGE_ACTIVATE;
630 if (!PageWriteback(page)) {
631 /* synchronous write or broken a_ops? */
632 ClearPageReclaim(page);
634 trace_mm_vmscan_writepage(page);
635 inc_node_page_state(page, NR_VMSCAN_WRITE);
636 return PAGE_SUCCESS;
639 return PAGE_CLEAN;
643 * Same as remove_mapping, but if the page is removed from the mapping, it
644 * gets returned with a refcount of 0.
646 static int __remove_mapping(struct address_space *mapping, struct page *page,
647 bool reclaimed)
649 unsigned long flags;
651 BUG_ON(!PageLocked(page));
652 BUG_ON(mapping != page_mapping(page));
654 spin_lock_irqsave(&mapping->tree_lock, flags);
656 * The non racy check for a busy page.
658 * Must be careful with the order of the tests. When someone has
659 * a ref to the page, it may be possible that they dirty it then
660 * drop the reference. So if PageDirty is tested before page_count
661 * here, then the following race may occur:
663 * get_user_pages(&page);
664 * [user mapping goes away]
665 * write_to(page);
666 * !PageDirty(page) [good]
667 * SetPageDirty(page);
668 * put_page(page);
669 * !page_count(page) [good, discard it]
671 * [oops, our write_to data is lost]
673 * Reversing the order of the tests ensures such a situation cannot
674 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
675 * load is not satisfied before that of page->_refcount.
677 * Note that if SetPageDirty is always performed via set_page_dirty,
678 * and thus under tree_lock, then this ordering is not required.
680 if (!page_ref_freeze(page, 2))
681 goto cannot_free;
682 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
683 if (unlikely(PageDirty(page))) {
684 page_ref_unfreeze(page, 2);
685 goto cannot_free;
688 if (PageSwapCache(page)) {
689 swp_entry_t swap = { .val = page_private(page) };
690 mem_cgroup_swapout(page, swap);
691 __delete_from_swap_cache(page);
692 spin_unlock_irqrestore(&mapping->tree_lock, flags);
693 swapcache_free(swap);
694 } else {
695 void (*freepage)(struct page *);
696 void *shadow = NULL;
698 freepage = mapping->a_ops->freepage;
700 * Remember a shadow entry for reclaimed file cache in
701 * order to detect refaults, thus thrashing, later on.
703 * But don't store shadows in an address space that is
704 * already exiting. This is not just an optizimation,
705 * inode reclaim needs to empty out the radix tree or
706 * the nodes are lost. Don't plant shadows behind its
707 * back.
709 * We also don't store shadows for DAX mappings because the
710 * only page cache pages found in these are zero pages
711 * covering holes, and because we don't want to mix DAX
712 * exceptional entries and shadow exceptional entries in the
713 * same page_tree.
715 if (reclaimed && page_is_file_cache(page) &&
716 !mapping_exiting(mapping) && !dax_mapping(mapping))
717 shadow = workingset_eviction(mapping, page);
718 __delete_from_page_cache(page, shadow);
719 spin_unlock_irqrestore(&mapping->tree_lock, flags);
721 if (freepage != NULL)
722 freepage(page);
725 return 1;
727 cannot_free:
728 spin_unlock_irqrestore(&mapping->tree_lock, flags);
729 return 0;
733 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
734 * someone else has a ref on the page, abort and return 0. If it was
735 * successfully detached, return 1. Assumes the caller has a single ref on
736 * this page.
738 int remove_mapping(struct address_space *mapping, struct page *page)
740 if (__remove_mapping(mapping, page, false)) {
742 * Unfreezing the refcount with 1 rather than 2 effectively
743 * drops the pagecache ref for us without requiring another
744 * atomic operation.
746 page_ref_unfreeze(page, 1);
747 return 1;
749 return 0;
753 * putback_lru_page - put previously isolated page onto appropriate LRU list
754 * @page: page to be put back to appropriate lru list
756 * Add previously isolated @page to appropriate LRU list.
757 * Page may still be unevictable for other reasons.
759 * lru_lock must not be held, interrupts must be enabled.
761 void putback_lru_page(struct page *page)
763 bool is_unevictable;
764 int was_unevictable = PageUnevictable(page);
766 VM_BUG_ON_PAGE(PageLRU(page), page);
768 redo:
769 ClearPageUnevictable(page);
771 if (page_evictable(page)) {
773 * For evictable pages, we can use the cache.
774 * In event of a race, worst case is we end up with an
775 * unevictable page on [in]active list.
776 * We know how to handle that.
778 is_unevictable = false;
779 lru_cache_add(page);
780 } else {
782 * Put unevictable pages directly on zone's unevictable
783 * list.
785 is_unevictable = true;
786 add_page_to_unevictable_list(page);
788 * When racing with an mlock or AS_UNEVICTABLE clearing
789 * (page is unlocked) make sure that if the other thread
790 * does not observe our setting of PG_lru and fails
791 * isolation/check_move_unevictable_pages,
792 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
793 * the page back to the evictable list.
795 * The other side is TestClearPageMlocked() or shmem_lock().
797 smp_mb();
801 * page's status can change while we move it among lru. If an evictable
802 * page is on unevictable list, it never be freed. To avoid that,
803 * check after we added it to the list, again.
805 if (is_unevictable && page_evictable(page)) {
806 if (!isolate_lru_page(page)) {
807 put_page(page);
808 goto redo;
810 /* This means someone else dropped this page from LRU
811 * So, it will be freed or putback to LRU again. There is
812 * nothing to do here.
816 if (was_unevictable && !is_unevictable)
817 count_vm_event(UNEVICTABLE_PGRESCUED);
818 else if (!was_unevictable && is_unevictable)
819 count_vm_event(UNEVICTABLE_PGCULLED);
821 put_page(page); /* drop ref from isolate */
824 enum page_references {
825 PAGEREF_RECLAIM,
826 PAGEREF_RECLAIM_CLEAN,
827 PAGEREF_KEEP,
828 PAGEREF_ACTIVATE,
831 static enum page_references page_check_references(struct page *page,
832 struct scan_control *sc)
834 int referenced_ptes, referenced_page;
835 unsigned long vm_flags;
837 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
838 &vm_flags);
839 referenced_page = TestClearPageReferenced(page);
842 * Mlock lost the isolation race with us. Let try_to_unmap()
843 * move the page to the unevictable list.
845 if (vm_flags & VM_LOCKED)
846 return PAGEREF_RECLAIM;
848 if (referenced_ptes) {
849 if (PageSwapBacked(page))
850 return PAGEREF_ACTIVATE;
852 * All mapped pages start out with page table
853 * references from the instantiating fault, so we need
854 * to look twice if a mapped file page is used more
855 * than once.
857 * Mark it and spare it for another trip around the
858 * inactive list. Another page table reference will
859 * lead to its activation.
861 * Note: the mark is set for activated pages as well
862 * so that recently deactivated but used pages are
863 * quickly recovered.
865 SetPageReferenced(page);
867 if (referenced_page || referenced_ptes > 1)
868 return PAGEREF_ACTIVATE;
871 * Activate file-backed executable pages after first usage.
873 if (vm_flags & VM_EXEC)
874 return PAGEREF_ACTIVATE;
876 return PAGEREF_KEEP;
879 /* Reclaim if clean, defer dirty pages to writeback */
880 if (referenced_page && !PageSwapBacked(page))
881 return PAGEREF_RECLAIM_CLEAN;
883 return PAGEREF_RECLAIM;
886 /* Check if a page is dirty or under writeback */
887 static void page_check_dirty_writeback(struct page *page,
888 bool *dirty, bool *writeback)
890 struct address_space *mapping;
893 * Anonymous pages are not handled by flushers and must be written
894 * from reclaim context. Do not stall reclaim based on them
896 if (!page_is_file_cache(page)) {
897 *dirty = false;
898 *writeback = false;
899 return;
902 /* By default assume that the page flags are accurate */
903 *dirty = PageDirty(page);
904 *writeback = PageWriteback(page);
906 /* Verify dirty/writeback state if the filesystem supports it */
907 if (!page_has_private(page))
908 return;
910 mapping = page_mapping(page);
911 if (mapping && mapping->a_ops->is_dirty_writeback)
912 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
916 * shrink_page_list() returns the number of reclaimed pages
918 static unsigned long shrink_page_list(struct list_head *page_list,
919 struct pglist_data *pgdat,
920 struct scan_control *sc,
921 enum ttu_flags ttu_flags,
922 unsigned long *ret_nr_dirty,
923 unsigned long *ret_nr_unqueued_dirty,
924 unsigned long *ret_nr_congested,
925 unsigned long *ret_nr_writeback,
926 unsigned long *ret_nr_immediate,
927 bool force_reclaim)
929 LIST_HEAD(ret_pages);
930 LIST_HEAD(free_pages);
931 int pgactivate = 0;
932 unsigned long nr_unqueued_dirty = 0;
933 unsigned long nr_dirty = 0;
934 unsigned long nr_congested = 0;
935 unsigned long nr_reclaimed = 0;
936 unsigned long nr_writeback = 0;
937 unsigned long nr_immediate = 0;
939 cond_resched();
941 while (!list_empty(page_list)) {
942 struct address_space *mapping;
943 struct page *page;
944 int may_enter_fs;
945 enum page_references references = PAGEREF_RECLAIM_CLEAN;
946 bool dirty, writeback;
947 bool lazyfree = false;
948 int ret = SWAP_SUCCESS;
950 cond_resched();
952 page = lru_to_page(page_list);
953 list_del(&page->lru);
955 if (!trylock_page(page))
956 goto keep;
958 VM_BUG_ON_PAGE(PageActive(page), page);
960 sc->nr_scanned++;
962 if (unlikely(!page_evictable(page)))
963 goto cull_mlocked;
965 if (!sc->may_unmap && page_mapped(page))
966 goto keep_locked;
968 /* Double the slab pressure for mapped and swapcache pages */
969 if (page_mapped(page) || PageSwapCache(page))
970 sc->nr_scanned++;
972 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
973 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
976 * The number of dirty pages determines if a zone is marked
977 * reclaim_congested which affects wait_iff_congested. kswapd
978 * will stall and start writing pages if the tail of the LRU
979 * is all dirty unqueued pages.
981 page_check_dirty_writeback(page, &dirty, &writeback);
982 if (dirty || writeback)
983 nr_dirty++;
985 if (dirty && !writeback)
986 nr_unqueued_dirty++;
989 * Treat this page as congested if the underlying BDI is or if
990 * pages are cycling through the LRU so quickly that the
991 * pages marked for immediate reclaim are making it to the
992 * end of the LRU a second time.
994 mapping = page_mapping(page);
995 if (((dirty || writeback) && mapping &&
996 inode_write_congested(mapping->host)) ||
997 (writeback && PageReclaim(page)))
998 nr_congested++;
1001 * If a page at the tail of the LRU is under writeback, there
1002 * are three cases to consider.
1004 * 1) If reclaim is encountering an excessive number of pages
1005 * under writeback and this page is both under writeback and
1006 * PageReclaim then it indicates that pages are being queued
1007 * for IO but are being recycled through the LRU before the
1008 * IO can complete. Waiting on the page itself risks an
1009 * indefinite stall if it is impossible to writeback the
1010 * page due to IO error or disconnected storage so instead
1011 * note that the LRU is being scanned too quickly and the
1012 * caller can stall after page list has been processed.
1014 * 2) Global or new memcg reclaim encounters a page that is
1015 * not marked for immediate reclaim, or the caller does not
1016 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1017 * not to fs). In this case mark the page for immediate
1018 * reclaim and continue scanning.
1020 * Require may_enter_fs because we would wait on fs, which
1021 * may not have submitted IO yet. And the loop driver might
1022 * enter reclaim, and deadlock if it waits on a page for
1023 * which it is needed to do the write (loop masks off
1024 * __GFP_IO|__GFP_FS for this reason); but more thought
1025 * would probably show more reasons.
1027 * 3) Legacy memcg encounters a page that is already marked
1028 * PageReclaim. memcg does not have any dirty pages
1029 * throttling so we could easily OOM just because too many
1030 * pages are in writeback and there is nothing else to
1031 * reclaim. Wait for the writeback to complete.
1033 if (PageWriteback(page)) {
1034 /* Case 1 above */
1035 if (current_is_kswapd() &&
1036 PageReclaim(page) &&
1037 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1038 nr_immediate++;
1039 goto keep_locked;
1041 /* Case 2 above */
1042 } else if (sane_reclaim(sc) ||
1043 !PageReclaim(page) || !may_enter_fs) {
1045 * This is slightly racy - end_page_writeback()
1046 * might have just cleared PageReclaim, then
1047 * setting PageReclaim here end up interpreted
1048 * as PageReadahead - but that does not matter
1049 * enough to care. What we do want is for this
1050 * page to have PageReclaim set next time memcg
1051 * reclaim reaches the tests above, so it will
1052 * then wait_on_page_writeback() to avoid OOM;
1053 * and it's also appropriate in global reclaim.
1055 SetPageReclaim(page);
1056 nr_writeback++;
1057 goto keep_locked;
1059 /* Case 3 above */
1060 } else {
1061 unlock_page(page);
1062 wait_on_page_writeback(page);
1063 /* then go back and try same page again */
1064 list_add_tail(&page->lru, page_list);
1065 continue;
1069 if (!force_reclaim)
1070 references = page_check_references(page, sc);
1072 switch (references) {
1073 case PAGEREF_ACTIVATE:
1074 goto activate_locked;
1075 case PAGEREF_KEEP:
1076 goto keep_locked;
1077 case PAGEREF_RECLAIM:
1078 case PAGEREF_RECLAIM_CLEAN:
1079 ; /* try to reclaim the page below */
1083 * Anonymous process memory has backing store?
1084 * Try to allocate it some swap space here.
1086 if (PageAnon(page) && !PageSwapCache(page)) {
1087 if (!(sc->gfp_mask & __GFP_IO))
1088 goto keep_locked;
1089 if (!add_to_swap(page, page_list))
1090 goto activate_locked;
1091 lazyfree = true;
1092 may_enter_fs = 1;
1094 /* Adding to swap updated mapping */
1095 mapping = page_mapping(page);
1096 } else if (unlikely(PageTransHuge(page))) {
1097 /* Split file THP */
1098 if (split_huge_page_to_list(page, page_list))
1099 goto keep_locked;
1102 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1105 * The page is mapped into the page tables of one or more
1106 * processes. Try to unmap it here.
1108 if (page_mapped(page) && mapping) {
1109 switch (ret = try_to_unmap(page, lazyfree ?
1110 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1111 (ttu_flags | TTU_BATCH_FLUSH))) {
1112 case SWAP_FAIL:
1113 goto activate_locked;
1114 case SWAP_AGAIN:
1115 goto keep_locked;
1116 case SWAP_MLOCK:
1117 goto cull_mlocked;
1118 case SWAP_LZFREE:
1119 goto lazyfree;
1120 case SWAP_SUCCESS:
1121 ; /* try to free the page below */
1125 if (PageDirty(page)) {
1127 * Only kswapd can writeback filesystem pages to
1128 * avoid risk of stack overflow but only writeback
1129 * if many dirty pages have been encountered.
1131 if (page_is_file_cache(page) &&
1132 (!current_is_kswapd() ||
1133 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1135 * Immediately reclaim when written back.
1136 * Similar in principal to deactivate_page()
1137 * except we already have the page isolated
1138 * and know it's dirty
1140 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1141 SetPageReclaim(page);
1143 goto keep_locked;
1146 if (references == PAGEREF_RECLAIM_CLEAN)
1147 goto keep_locked;
1148 if (!may_enter_fs)
1149 goto keep_locked;
1150 if (!sc->may_writepage)
1151 goto keep_locked;
1154 * Page is dirty. Flush the TLB if a writable entry
1155 * potentially exists to avoid CPU writes after IO
1156 * starts and then write it out here.
1158 try_to_unmap_flush_dirty();
1159 switch (pageout(page, mapping, sc)) {
1160 case PAGE_KEEP:
1161 goto keep_locked;
1162 case PAGE_ACTIVATE:
1163 goto activate_locked;
1164 case PAGE_SUCCESS:
1165 if (PageWriteback(page))
1166 goto keep;
1167 if (PageDirty(page))
1168 goto keep;
1171 * A synchronous write - probably a ramdisk. Go
1172 * ahead and try to reclaim the page.
1174 if (!trylock_page(page))
1175 goto keep;
1176 if (PageDirty(page) || PageWriteback(page))
1177 goto keep_locked;
1178 mapping = page_mapping(page);
1179 case PAGE_CLEAN:
1180 ; /* try to free the page below */
1185 * If the page has buffers, try to free the buffer mappings
1186 * associated with this page. If we succeed we try to free
1187 * the page as well.
1189 * We do this even if the page is PageDirty().
1190 * try_to_release_page() does not perform I/O, but it is
1191 * possible for a page to have PageDirty set, but it is actually
1192 * clean (all its buffers are clean). This happens if the
1193 * buffers were written out directly, with submit_bh(). ext3
1194 * will do this, as well as the blockdev mapping.
1195 * try_to_release_page() will discover that cleanness and will
1196 * drop the buffers and mark the page clean - it can be freed.
1198 * Rarely, pages can have buffers and no ->mapping. These are
1199 * the pages which were not successfully invalidated in
1200 * truncate_complete_page(). We try to drop those buffers here
1201 * and if that worked, and the page is no longer mapped into
1202 * process address space (page_count == 1) it can be freed.
1203 * Otherwise, leave the page on the LRU so it is swappable.
1205 if (page_has_private(page)) {
1206 if (!try_to_release_page(page, sc->gfp_mask))
1207 goto activate_locked;
1208 if (!mapping && page_count(page) == 1) {
1209 unlock_page(page);
1210 if (put_page_testzero(page))
1211 goto free_it;
1212 else {
1214 * rare race with speculative reference.
1215 * the speculative reference will free
1216 * this page shortly, so we may
1217 * increment nr_reclaimed here (and
1218 * leave it off the LRU).
1220 nr_reclaimed++;
1221 continue;
1226 lazyfree:
1227 if (!mapping || !__remove_mapping(mapping, page, true))
1228 goto keep_locked;
1231 * At this point, we have no other references and there is
1232 * no way to pick any more up (removed from LRU, removed
1233 * from pagecache). Can use non-atomic bitops now (and
1234 * we obviously don't have to worry about waking up a process
1235 * waiting on the page lock, because there are no references.
1237 __ClearPageLocked(page);
1238 free_it:
1239 if (ret == SWAP_LZFREE)
1240 count_vm_event(PGLAZYFREED);
1242 nr_reclaimed++;
1245 * Is there need to periodically free_page_list? It would
1246 * appear not as the counts should be low
1248 list_add(&page->lru, &free_pages);
1249 continue;
1251 cull_mlocked:
1252 if (PageSwapCache(page))
1253 try_to_free_swap(page);
1254 unlock_page(page);
1255 list_add(&page->lru, &ret_pages);
1256 continue;
1258 activate_locked:
1259 /* Not a candidate for swapping, so reclaim swap space. */
1260 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1261 try_to_free_swap(page);
1262 VM_BUG_ON_PAGE(PageActive(page), page);
1263 SetPageActive(page);
1264 pgactivate++;
1265 keep_locked:
1266 unlock_page(page);
1267 keep:
1268 list_add(&page->lru, &ret_pages);
1269 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1272 mem_cgroup_uncharge_list(&free_pages);
1273 try_to_unmap_flush();
1274 free_hot_cold_page_list(&free_pages, true);
1276 list_splice(&ret_pages, page_list);
1277 count_vm_events(PGACTIVATE, pgactivate);
1279 *ret_nr_dirty += nr_dirty;
1280 *ret_nr_congested += nr_congested;
1281 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1282 *ret_nr_writeback += nr_writeback;
1283 *ret_nr_immediate += nr_immediate;
1284 return nr_reclaimed;
1287 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1288 struct list_head *page_list)
1290 struct scan_control sc = {
1291 .gfp_mask = GFP_KERNEL,
1292 .priority = DEF_PRIORITY,
1293 .may_unmap = 1,
1295 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1296 struct page *page, *next;
1297 LIST_HEAD(clean_pages);
1299 list_for_each_entry_safe(page, next, page_list, lru) {
1300 if (page_is_file_cache(page) && !PageDirty(page) &&
1301 !__PageMovable(page)) {
1302 ClearPageActive(page);
1303 list_move(&page->lru, &clean_pages);
1307 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1308 TTU_UNMAP|TTU_IGNORE_ACCESS,
1309 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1310 list_splice(&clean_pages, page_list);
1311 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1312 return ret;
1316 * Attempt to remove the specified page from its LRU. Only take this page
1317 * if it is of the appropriate PageActive status. Pages which are being
1318 * freed elsewhere are also ignored.
1320 * page: page to consider
1321 * mode: one of the LRU isolation modes defined above
1323 * returns 0 on success, -ve errno on failure.
1325 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1327 int ret = -EINVAL;
1329 /* Only take pages on the LRU. */
1330 if (!PageLRU(page))
1331 return ret;
1333 /* Compaction should not handle unevictable pages but CMA can do so */
1334 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1335 return ret;
1337 ret = -EBUSY;
1340 * To minimise LRU disruption, the caller can indicate that it only
1341 * wants to isolate pages it will be able to operate on without
1342 * blocking - clean pages for the most part.
1344 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1345 * is used by reclaim when it is cannot write to backing storage
1347 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1348 * that it is possible to migrate without blocking
1350 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1351 /* All the caller can do on PageWriteback is block */
1352 if (PageWriteback(page))
1353 return ret;
1355 if (PageDirty(page)) {
1356 struct address_space *mapping;
1358 /* ISOLATE_CLEAN means only clean pages */
1359 if (mode & ISOLATE_CLEAN)
1360 return ret;
1363 * Only pages without mappings or that have a
1364 * ->migratepage callback are possible to migrate
1365 * without blocking
1367 mapping = page_mapping(page);
1368 if (mapping && !mapping->a_ops->migratepage)
1369 return ret;
1373 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1374 return ret;
1376 if (likely(get_page_unless_zero(page))) {
1378 * Be careful not to clear PageLRU until after we're
1379 * sure the page is not being freed elsewhere -- the
1380 * page release code relies on it.
1382 ClearPageLRU(page);
1383 ret = 0;
1386 return ret;
1391 * Update LRU sizes after isolating pages. The LRU size updates must
1392 * be complete before mem_cgroup_update_lru_size due to a santity check.
1394 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1395 enum lru_list lru, unsigned long *nr_zone_taken)
1397 int zid;
1399 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1400 if (!nr_zone_taken[zid])
1401 continue;
1403 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1404 #ifdef CONFIG_MEMCG
1405 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1406 #endif
1412 * zone_lru_lock is heavily contended. Some of the functions that
1413 * shrink the lists perform better by taking out a batch of pages
1414 * and working on them outside the LRU lock.
1416 * For pagecache intensive workloads, this function is the hottest
1417 * spot in the kernel (apart from copy_*_user functions).
1419 * Appropriate locks must be held before calling this function.
1421 * @nr_to_scan: The number of pages to look through on the list.
1422 * @lruvec: The LRU vector to pull pages from.
1423 * @dst: The temp list to put pages on to.
1424 * @nr_scanned: The number of pages that were scanned.
1425 * @sc: The scan_control struct for this reclaim session
1426 * @mode: One of the LRU isolation modes
1427 * @lru: LRU list id for isolating
1429 * returns how many pages were moved onto *@dst.
1431 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1432 struct lruvec *lruvec, struct list_head *dst,
1433 unsigned long *nr_scanned, struct scan_control *sc,
1434 isolate_mode_t mode, enum lru_list lru)
1436 struct list_head *src = &lruvec->lists[lru];
1437 unsigned long nr_taken = 0;
1438 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1439 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1440 unsigned long scan, nr_pages;
1441 LIST_HEAD(pages_skipped);
1443 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1444 !list_empty(src);) {
1445 struct page *page;
1447 page = lru_to_page(src);
1448 prefetchw_prev_lru_page(page, src, flags);
1450 VM_BUG_ON_PAGE(!PageLRU(page), page);
1452 if (page_zonenum(page) > sc->reclaim_idx) {
1453 list_move(&page->lru, &pages_skipped);
1454 nr_skipped[page_zonenum(page)]++;
1455 continue;
1459 * Account for scanned and skipped separetly to avoid the pgdat
1460 * being prematurely marked unreclaimable by pgdat_reclaimable.
1462 scan++;
1464 switch (__isolate_lru_page(page, mode)) {
1465 case 0:
1466 nr_pages = hpage_nr_pages(page);
1467 nr_taken += nr_pages;
1468 nr_zone_taken[page_zonenum(page)] += nr_pages;
1469 list_move(&page->lru, dst);
1470 break;
1472 case -EBUSY:
1473 /* else it is being freed elsewhere */
1474 list_move(&page->lru, src);
1475 continue;
1477 default:
1478 BUG();
1483 * Splice any skipped pages to the start of the LRU list. Note that
1484 * this disrupts the LRU order when reclaiming for lower zones but
1485 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1486 * scanning would soon rescan the same pages to skip and put the
1487 * system at risk of premature OOM.
1489 if (!list_empty(&pages_skipped)) {
1490 int zid;
1491 unsigned long total_skipped = 0;
1493 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1494 if (!nr_skipped[zid])
1495 continue;
1497 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1498 total_skipped += nr_skipped[zid];
1502 * Account skipped pages as a partial scan as the pgdat may be
1503 * close to unreclaimable. If the LRU list is empty, account
1504 * skipped pages as a full scan.
1506 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1508 list_splice(&pages_skipped, src);
1510 *nr_scanned = scan;
1511 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
1512 nr_taken, mode, is_file_lru(lru));
1513 update_lru_sizes(lruvec, lru, nr_zone_taken);
1514 return nr_taken;
1518 * isolate_lru_page - tries to isolate a page from its LRU list
1519 * @page: page to isolate from its LRU list
1521 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1522 * vmstat statistic corresponding to whatever LRU list the page was on.
1524 * Returns 0 if the page was removed from an LRU list.
1525 * Returns -EBUSY if the page was not on an LRU list.
1527 * The returned page will have PageLRU() cleared. If it was found on
1528 * the active list, it will have PageActive set. If it was found on
1529 * the unevictable list, it will have the PageUnevictable bit set. That flag
1530 * may need to be cleared by the caller before letting the page go.
1532 * The vmstat statistic corresponding to the list on which the page was
1533 * found will be decremented.
1535 * Restrictions:
1536 * (1) Must be called with an elevated refcount on the page. This is a
1537 * fundamentnal difference from isolate_lru_pages (which is called
1538 * without a stable reference).
1539 * (2) the lru_lock must not be held.
1540 * (3) interrupts must be enabled.
1542 int isolate_lru_page(struct page *page)
1544 int ret = -EBUSY;
1546 VM_BUG_ON_PAGE(!page_count(page), page);
1547 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1549 if (PageLRU(page)) {
1550 struct zone *zone = page_zone(page);
1551 struct lruvec *lruvec;
1553 spin_lock_irq(zone_lru_lock(zone));
1554 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1555 if (PageLRU(page)) {
1556 int lru = page_lru(page);
1557 get_page(page);
1558 ClearPageLRU(page);
1559 del_page_from_lru_list(page, lruvec, lru);
1560 ret = 0;
1562 spin_unlock_irq(zone_lru_lock(zone));
1564 return ret;
1568 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1569 * then get resheduled. When there are massive number of tasks doing page
1570 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1571 * the LRU list will go small and be scanned faster than necessary, leading to
1572 * unnecessary swapping, thrashing and OOM.
1574 static int too_many_isolated(struct pglist_data *pgdat, int file,
1575 struct scan_control *sc)
1577 unsigned long inactive, isolated;
1579 if (current_is_kswapd())
1580 return 0;
1582 if (!sane_reclaim(sc))
1583 return 0;
1585 if (file) {
1586 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1587 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1588 } else {
1589 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1590 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1594 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1595 * won't get blocked by normal direct-reclaimers, forming a circular
1596 * deadlock.
1598 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1599 inactive >>= 3;
1601 return isolated > inactive;
1604 static noinline_for_stack void
1605 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1607 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1608 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1609 LIST_HEAD(pages_to_free);
1612 * Put back any unfreeable pages.
1614 while (!list_empty(page_list)) {
1615 struct page *page = lru_to_page(page_list);
1616 int lru;
1618 VM_BUG_ON_PAGE(PageLRU(page), page);
1619 list_del(&page->lru);
1620 if (unlikely(!page_evictable(page))) {
1621 spin_unlock_irq(&pgdat->lru_lock);
1622 putback_lru_page(page);
1623 spin_lock_irq(&pgdat->lru_lock);
1624 continue;
1627 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1629 SetPageLRU(page);
1630 lru = page_lru(page);
1631 add_page_to_lru_list(page, lruvec, lru);
1633 if (is_active_lru(lru)) {
1634 int file = is_file_lru(lru);
1635 int numpages = hpage_nr_pages(page);
1636 reclaim_stat->recent_rotated[file] += numpages;
1638 if (put_page_testzero(page)) {
1639 __ClearPageLRU(page);
1640 __ClearPageActive(page);
1641 del_page_from_lru_list(page, lruvec, lru);
1643 if (unlikely(PageCompound(page))) {
1644 spin_unlock_irq(&pgdat->lru_lock);
1645 mem_cgroup_uncharge(page);
1646 (*get_compound_page_dtor(page))(page);
1647 spin_lock_irq(&pgdat->lru_lock);
1648 } else
1649 list_add(&page->lru, &pages_to_free);
1654 * To save our caller's stack, now use input list for pages to free.
1656 list_splice(&pages_to_free, page_list);
1660 * If a kernel thread (such as nfsd for loop-back mounts) services
1661 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1662 * In that case we should only throttle if the backing device it is
1663 * writing to is congested. In other cases it is safe to throttle.
1665 static int current_may_throttle(void)
1667 return !(current->flags & PF_LESS_THROTTLE) ||
1668 current->backing_dev_info == NULL ||
1669 bdi_write_congested(current->backing_dev_info);
1672 static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1673 struct scan_control *sc, enum lru_list lru)
1675 int zid;
1676 struct zone *zone;
1677 int file = is_file_lru(lru);
1678 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1680 if (!global_reclaim(sc))
1681 return true;
1683 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1684 zone = &pgdat->node_zones[zid];
1685 if (!managed_zone(zone))
1686 continue;
1688 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1689 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1690 return true;
1693 return false;
1697 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1698 * of reclaimed pages
1700 static noinline_for_stack unsigned long
1701 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1702 struct scan_control *sc, enum lru_list lru)
1704 LIST_HEAD(page_list);
1705 unsigned long nr_scanned;
1706 unsigned long nr_reclaimed = 0;
1707 unsigned long nr_taken;
1708 unsigned long nr_dirty = 0;
1709 unsigned long nr_congested = 0;
1710 unsigned long nr_unqueued_dirty = 0;
1711 unsigned long nr_writeback = 0;
1712 unsigned long nr_immediate = 0;
1713 isolate_mode_t isolate_mode = 0;
1714 int file = is_file_lru(lru);
1715 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1716 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1718 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1719 return 0;
1721 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1722 congestion_wait(BLK_RW_ASYNC, HZ/10);
1724 /* We are about to die and free our memory. Return now. */
1725 if (fatal_signal_pending(current))
1726 return SWAP_CLUSTER_MAX;
1729 lru_add_drain();
1731 if (!sc->may_unmap)
1732 isolate_mode |= ISOLATE_UNMAPPED;
1733 if (!sc->may_writepage)
1734 isolate_mode |= ISOLATE_CLEAN;
1736 spin_lock_irq(&pgdat->lru_lock);
1738 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1739 &nr_scanned, sc, isolate_mode, lru);
1741 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1742 reclaim_stat->recent_scanned[file] += nr_taken;
1744 if (global_reclaim(sc)) {
1745 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1746 if (current_is_kswapd())
1747 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1748 else
1749 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1751 spin_unlock_irq(&pgdat->lru_lock);
1753 if (nr_taken == 0)
1754 return 0;
1756 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1757 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1758 &nr_writeback, &nr_immediate,
1759 false);
1761 spin_lock_irq(&pgdat->lru_lock);
1763 if (global_reclaim(sc)) {
1764 if (current_is_kswapd())
1765 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1766 else
1767 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1770 putback_inactive_pages(lruvec, &page_list);
1772 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1774 spin_unlock_irq(&pgdat->lru_lock);
1776 mem_cgroup_uncharge_list(&page_list);
1777 free_hot_cold_page_list(&page_list, true);
1780 * If reclaim is isolating dirty pages under writeback, it implies
1781 * that the long-lived page allocation rate is exceeding the page
1782 * laundering rate. Either the global limits are not being effective
1783 * at throttling processes due to the page distribution throughout
1784 * zones or there is heavy usage of a slow backing device. The
1785 * only option is to throttle from reclaim context which is not ideal
1786 * as there is no guarantee the dirtying process is throttled in the
1787 * same way balance_dirty_pages() manages.
1789 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1790 * of pages under pages flagged for immediate reclaim and stall if any
1791 * are encountered in the nr_immediate check below.
1793 if (nr_writeback && nr_writeback == nr_taken)
1794 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1797 * Legacy memcg will stall in page writeback so avoid forcibly
1798 * stalling here.
1800 if (sane_reclaim(sc)) {
1802 * Tag a zone as congested if all the dirty pages scanned were
1803 * backed by a congested BDI and wait_iff_congested will stall.
1805 if (nr_dirty && nr_dirty == nr_congested)
1806 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1809 * If dirty pages are scanned that are not queued for IO, it
1810 * implies that flushers are not keeping up. In this case, flag
1811 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1812 * reclaim context.
1814 if (nr_unqueued_dirty == nr_taken)
1815 set_bit(PGDAT_DIRTY, &pgdat->flags);
1818 * If kswapd scans pages marked marked for immediate
1819 * reclaim and under writeback (nr_immediate), it implies
1820 * that pages are cycling through the LRU faster than
1821 * they are written so also forcibly stall.
1823 if (nr_immediate && current_may_throttle())
1824 congestion_wait(BLK_RW_ASYNC, HZ/10);
1828 * Stall direct reclaim for IO completions if underlying BDIs or zone
1829 * is congested. Allow kswapd to continue until it starts encountering
1830 * unqueued dirty pages or cycling through the LRU too quickly.
1832 if (!sc->hibernation_mode && !current_is_kswapd() &&
1833 current_may_throttle())
1834 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1836 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1837 nr_scanned, nr_reclaimed,
1838 sc->priority, file);
1839 return nr_reclaimed;
1843 * This moves pages from the active list to the inactive list.
1845 * We move them the other way if the page is referenced by one or more
1846 * processes, from rmap.
1848 * If the pages are mostly unmapped, the processing is fast and it is
1849 * appropriate to hold zone_lru_lock across the whole operation. But if
1850 * the pages are mapped, the processing is slow (page_referenced()) so we
1851 * should drop zone_lru_lock around each page. It's impossible to balance
1852 * this, so instead we remove the pages from the LRU while processing them.
1853 * It is safe to rely on PG_active against the non-LRU pages in here because
1854 * nobody will play with that bit on a non-LRU page.
1856 * The downside is that we have to touch page->_refcount against each page.
1857 * But we had to alter page->flags anyway.
1860 static void move_active_pages_to_lru(struct lruvec *lruvec,
1861 struct list_head *list,
1862 struct list_head *pages_to_free,
1863 enum lru_list lru)
1865 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1866 unsigned long pgmoved = 0;
1867 struct page *page;
1868 int nr_pages;
1870 while (!list_empty(list)) {
1871 page = lru_to_page(list);
1872 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1874 VM_BUG_ON_PAGE(PageLRU(page), page);
1875 SetPageLRU(page);
1877 nr_pages = hpage_nr_pages(page);
1878 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1879 list_move(&page->lru, &lruvec->lists[lru]);
1880 pgmoved += nr_pages;
1882 if (put_page_testzero(page)) {
1883 __ClearPageLRU(page);
1884 __ClearPageActive(page);
1885 del_page_from_lru_list(page, lruvec, lru);
1887 if (unlikely(PageCompound(page))) {
1888 spin_unlock_irq(&pgdat->lru_lock);
1889 mem_cgroup_uncharge(page);
1890 (*get_compound_page_dtor(page))(page);
1891 spin_lock_irq(&pgdat->lru_lock);
1892 } else
1893 list_add(&page->lru, pages_to_free);
1897 if (!is_active_lru(lru))
1898 __count_vm_events(PGDEACTIVATE, pgmoved);
1901 static void shrink_active_list(unsigned long nr_to_scan,
1902 struct lruvec *lruvec,
1903 struct scan_control *sc,
1904 enum lru_list lru)
1906 unsigned long nr_taken;
1907 unsigned long nr_scanned;
1908 unsigned long vm_flags;
1909 LIST_HEAD(l_hold); /* The pages which were snipped off */
1910 LIST_HEAD(l_active);
1911 LIST_HEAD(l_inactive);
1912 struct page *page;
1913 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1914 unsigned long nr_rotated = 0;
1915 isolate_mode_t isolate_mode = 0;
1916 int file = is_file_lru(lru);
1917 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1919 lru_add_drain();
1921 if (!sc->may_unmap)
1922 isolate_mode |= ISOLATE_UNMAPPED;
1923 if (!sc->may_writepage)
1924 isolate_mode |= ISOLATE_CLEAN;
1926 spin_lock_irq(&pgdat->lru_lock);
1928 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1929 &nr_scanned, sc, isolate_mode, lru);
1931 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1932 reclaim_stat->recent_scanned[file] += nr_taken;
1934 if (global_reclaim(sc))
1935 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1936 __count_vm_events(PGREFILL, nr_scanned);
1938 spin_unlock_irq(&pgdat->lru_lock);
1940 while (!list_empty(&l_hold)) {
1941 cond_resched();
1942 page = lru_to_page(&l_hold);
1943 list_del(&page->lru);
1945 if (unlikely(!page_evictable(page))) {
1946 putback_lru_page(page);
1947 continue;
1950 if (unlikely(buffer_heads_over_limit)) {
1951 if (page_has_private(page) && trylock_page(page)) {
1952 if (page_has_private(page))
1953 try_to_release_page(page, 0);
1954 unlock_page(page);
1958 if (page_referenced(page, 0, sc->target_mem_cgroup,
1959 &vm_flags)) {
1960 nr_rotated += hpage_nr_pages(page);
1962 * Identify referenced, file-backed active pages and
1963 * give them one more trip around the active list. So
1964 * that executable code get better chances to stay in
1965 * memory under moderate memory pressure. Anon pages
1966 * are not likely to be evicted by use-once streaming
1967 * IO, plus JVM can create lots of anon VM_EXEC pages,
1968 * so we ignore them here.
1970 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1971 list_add(&page->lru, &l_active);
1972 continue;
1976 ClearPageActive(page); /* we are de-activating */
1977 list_add(&page->lru, &l_inactive);
1981 * Move pages back to the lru list.
1983 spin_lock_irq(&pgdat->lru_lock);
1985 * Count referenced pages from currently used mappings as rotated,
1986 * even though only some of them are actually re-activated. This
1987 * helps balance scan pressure between file and anonymous pages in
1988 * get_scan_count.
1990 reclaim_stat->recent_rotated[file] += nr_rotated;
1992 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1993 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1994 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1995 spin_unlock_irq(&pgdat->lru_lock);
1997 mem_cgroup_uncharge_list(&l_hold);
1998 free_hot_cold_page_list(&l_hold, true);
2002 * The inactive anon list should be small enough that the VM never has
2003 * to do too much work.
2005 * The inactive file list should be small enough to leave most memory
2006 * to the established workingset on the scan-resistant active list,
2007 * but large enough to avoid thrashing the aggregate readahead window.
2009 * Both inactive lists should also be large enough that each inactive
2010 * page has a chance to be referenced again before it is reclaimed.
2012 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2013 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2014 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2016 * total target max
2017 * memory ratio inactive
2018 * -------------------------------------
2019 * 10MB 1 5MB
2020 * 100MB 1 50MB
2021 * 1GB 3 250MB
2022 * 10GB 10 0.9GB
2023 * 100GB 31 3GB
2024 * 1TB 101 10GB
2025 * 10TB 320 32GB
2027 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2028 struct scan_control *sc)
2030 unsigned long inactive_ratio;
2031 unsigned long inactive;
2032 unsigned long active;
2033 unsigned long gb;
2034 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2035 int zid;
2038 * If we don't have swap space, anonymous page deactivation
2039 * is pointless.
2041 if (!file && !total_swap_pages)
2042 return false;
2044 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
2045 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
2048 * For zone-constrained allocations, it is necessary to check if
2049 * deactivations are required for lowmem to be reclaimed. This
2050 * calculates the inactive/active pages available in eligible zones.
2052 for (zid = sc->reclaim_idx + 1; zid < MAX_NR_ZONES; zid++) {
2053 struct zone *zone = &pgdat->node_zones[zid];
2054 unsigned long inactive_zone, active_zone;
2056 if (!managed_zone(zone))
2057 continue;
2059 inactive_zone = lruvec_zone_lru_size(lruvec, file * LRU_FILE, zid);
2060 active_zone = lruvec_zone_lru_size(lruvec, (file * LRU_FILE) + LRU_ACTIVE, zid);
2062 inactive -= min(inactive, inactive_zone);
2063 active -= min(active, active_zone);
2066 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2067 if (gb)
2068 inactive_ratio = int_sqrt(10 * gb);
2069 else
2070 inactive_ratio = 1;
2072 return inactive * inactive_ratio < active;
2075 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2076 struct lruvec *lruvec, struct scan_control *sc)
2078 if (is_active_lru(lru)) {
2079 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
2080 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2081 return 0;
2084 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2087 enum scan_balance {
2088 SCAN_EQUAL,
2089 SCAN_FRACT,
2090 SCAN_ANON,
2091 SCAN_FILE,
2095 * Determine how aggressively the anon and file LRU lists should be
2096 * scanned. The relative value of each set of LRU lists is determined
2097 * by looking at the fraction of the pages scanned we did rotate back
2098 * onto the active list instead of evict.
2100 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2101 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2103 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2104 struct scan_control *sc, unsigned long *nr,
2105 unsigned long *lru_pages)
2107 int swappiness = mem_cgroup_swappiness(memcg);
2108 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2109 u64 fraction[2];
2110 u64 denominator = 0; /* gcc */
2111 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2112 unsigned long anon_prio, file_prio;
2113 enum scan_balance scan_balance;
2114 unsigned long anon, file;
2115 bool force_scan = false;
2116 unsigned long ap, fp;
2117 enum lru_list lru;
2118 bool some_scanned;
2119 int pass;
2122 * If the zone or memcg is small, nr[l] can be 0. This
2123 * results in no scanning on this priority and a potential
2124 * priority drop. Global direct reclaim can go to the next
2125 * zone and tends to have no problems. Global kswapd is for
2126 * zone balancing and it needs to scan a minimum amount. When
2127 * reclaiming for a memcg, a priority drop can cause high
2128 * latencies, so it's better to scan a minimum amount there as
2129 * well.
2131 if (current_is_kswapd()) {
2132 if (!pgdat_reclaimable(pgdat))
2133 force_scan = true;
2134 if (!mem_cgroup_online(memcg))
2135 force_scan = true;
2137 if (!global_reclaim(sc))
2138 force_scan = true;
2140 /* If we have no swap space, do not bother scanning anon pages. */
2141 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2142 scan_balance = SCAN_FILE;
2143 goto out;
2147 * Global reclaim will swap to prevent OOM even with no
2148 * swappiness, but memcg users want to use this knob to
2149 * disable swapping for individual groups completely when
2150 * using the memory controller's swap limit feature would be
2151 * too expensive.
2153 if (!global_reclaim(sc) && !swappiness) {
2154 scan_balance = SCAN_FILE;
2155 goto out;
2159 * Do not apply any pressure balancing cleverness when the
2160 * system is close to OOM, scan both anon and file equally
2161 * (unless the swappiness setting disagrees with swapping).
2163 if (!sc->priority && swappiness) {
2164 scan_balance = SCAN_EQUAL;
2165 goto out;
2169 * Prevent the reclaimer from falling into the cache trap: as
2170 * cache pages start out inactive, every cache fault will tip
2171 * the scan balance towards the file LRU. And as the file LRU
2172 * shrinks, so does the window for rotation from references.
2173 * This means we have a runaway feedback loop where a tiny
2174 * thrashing file LRU becomes infinitely more attractive than
2175 * anon pages. Try to detect this based on file LRU size.
2177 if (global_reclaim(sc)) {
2178 unsigned long pgdatfile;
2179 unsigned long pgdatfree;
2180 int z;
2181 unsigned long total_high_wmark = 0;
2183 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2184 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2185 node_page_state(pgdat, NR_INACTIVE_FILE);
2187 for (z = 0; z < MAX_NR_ZONES; z++) {
2188 struct zone *zone = &pgdat->node_zones[z];
2189 if (!managed_zone(zone))
2190 continue;
2192 total_high_wmark += high_wmark_pages(zone);
2195 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2196 scan_balance = SCAN_ANON;
2197 goto out;
2202 * If there is enough inactive page cache, i.e. if the size of the
2203 * inactive list is greater than that of the active list *and* the
2204 * inactive list actually has some pages to scan on this priority, we
2205 * do not reclaim anything from the anonymous working set right now.
2206 * Without the second condition we could end up never scanning an
2207 * lruvec even if it has plenty of old anonymous pages unless the
2208 * system is under heavy pressure.
2210 if (!inactive_list_is_low(lruvec, true, sc) &&
2211 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2212 scan_balance = SCAN_FILE;
2213 goto out;
2216 scan_balance = SCAN_FRACT;
2219 * With swappiness at 100, anonymous and file have the same priority.
2220 * This scanning priority is essentially the inverse of IO cost.
2222 anon_prio = swappiness;
2223 file_prio = 200 - anon_prio;
2226 * OK, so we have swap space and a fair amount of page cache
2227 * pages. We use the recently rotated / recently scanned
2228 * ratios to determine how valuable each cache is.
2230 * Because workloads change over time (and to avoid overflow)
2231 * we keep these statistics as a floating average, which ends
2232 * up weighing recent references more than old ones.
2234 * anon in [0], file in [1]
2237 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2238 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2239 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2240 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2242 spin_lock_irq(&pgdat->lru_lock);
2243 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2244 reclaim_stat->recent_scanned[0] /= 2;
2245 reclaim_stat->recent_rotated[0] /= 2;
2248 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2249 reclaim_stat->recent_scanned[1] /= 2;
2250 reclaim_stat->recent_rotated[1] /= 2;
2254 * The amount of pressure on anon vs file pages is inversely
2255 * proportional to the fraction of recently scanned pages on
2256 * each list that were recently referenced and in active use.
2258 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2259 ap /= reclaim_stat->recent_rotated[0] + 1;
2261 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2262 fp /= reclaim_stat->recent_rotated[1] + 1;
2263 spin_unlock_irq(&pgdat->lru_lock);
2265 fraction[0] = ap;
2266 fraction[1] = fp;
2267 denominator = ap + fp + 1;
2268 out:
2269 some_scanned = false;
2270 /* Only use force_scan on second pass. */
2271 for (pass = 0; !some_scanned && pass < 2; pass++) {
2272 *lru_pages = 0;
2273 for_each_evictable_lru(lru) {
2274 int file = is_file_lru(lru);
2275 unsigned long size;
2276 unsigned long scan;
2278 size = lruvec_lru_size(lruvec, lru);
2279 scan = size >> sc->priority;
2281 if (!scan && pass && force_scan)
2282 scan = min(size, SWAP_CLUSTER_MAX);
2284 switch (scan_balance) {
2285 case SCAN_EQUAL:
2286 /* Scan lists relative to size */
2287 break;
2288 case SCAN_FRACT:
2290 * Scan types proportional to swappiness and
2291 * their relative recent reclaim efficiency.
2293 scan = div64_u64(scan * fraction[file],
2294 denominator);
2295 break;
2296 case SCAN_FILE:
2297 case SCAN_ANON:
2298 /* Scan one type exclusively */
2299 if ((scan_balance == SCAN_FILE) != file) {
2300 size = 0;
2301 scan = 0;
2303 break;
2304 default:
2305 /* Look ma, no brain */
2306 BUG();
2309 *lru_pages += size;
2310 nr[lru] = scan;
2313 * Skip the second pass and don't force_scan,
2314 * if we found something to scan.
2316 some_scanned |= !!scan;
2322 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2324 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2325 struct scan_control *sc, unsigned long *lru_pages)
2327 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2328 unsigned long nr[NR_LRU_LISTS];
2329 unsigned long targets[NR_LRU_LISTS];
2330 unsigned long nr_to_scan;
2331 enum lru_list lru;
2332 unsigned long nr_reclaimed = 0;
2333 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2334 struct blk_plug plug;
2335 bool scan_adjusted;
2337 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2339 /* Record the original scan target for proportional adjustments later */
2340 memcpy(targets, nr, sizeof(nr));
2343 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2344 * event that can occur when there is little memory pressure e.g.
2345 * multiple streaming readers/writers. Hence, we do not abort scanning
2346 * when the requested number of pages are reclaimed when scanning at
2347 * DEF_PRIORITY on the assumption that the fact we are direct
2348 * reclaiming implies that kswapd is not keeping up and it is best to
2349 * do a batch of work at once. For memcg reclaim one check is made to
2350 * abort proportional reclaim if either the file or anon lru has already
2351 * dropped to zero at the first pass.
2353 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2354 sc->priority == DEF_PRIORITY);
2356 blk_start_plug(&plug);
2357 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2358 nr[LRU_INACTIVE_FILE]) {
2359 unsigned long nr_anon, nr_file, percentage;
2360 unsigned long nr_scanned;
2362 for_each_evictable_lru(lru) {
2363 if (nr[lru]) {
2364 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2365 nr[lru] -= nr_to_scan;
2367 nr_reclaimed += shrink_list(lru, nr_to_scan,
2368 lruvec, sc);
2372 cond_resched();
2374 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2375 continue;
2378 * For kswapd and memcg, reclaim at least the number of pages
2379 * requested. Ensure that the anon and file LRUs are scanned
2380 * proportionally what was requested by get_scan_count(). We
2381 * stop reclaiming one LRU and reduce the amount scanning
2382 * proportional to the original scan target.
2384 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2385 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2388 * It's just vindictive to attack the larger once the smaller
2389 * has gone to zero. And given the way we stop scanning the
2390 * smaller below, this makes sure that we only make one nudge
2391 * towards proportionality once we've got nr_to_reclaim.
2393 if (!nr_file || !nr_anon)
2394 break;
2396 if (nr_file > nr_anon) {
2397 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2398 targets[LRU_ACTIVE_ANON] + 1;
2399 lru = LRU_BASE;
2400 percentage = nr_anon * 100 / scan_target;
2401 } else {
2402 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2403 targets[LRU_ACTIVE_FILE] + 1;
2404 lru = LRU_FILE;
2405 percentage = nr_file * 100 / scan_target;
2408 /* Stop scanning the smaller of the LRU */
2409 nr[lru] = 0;
2410 nr[lru + LRU_ACTIVE] = 0;
2413 * Recalculate the other LRU scan count based on its original
2414 * scan target and the percentage scanning already complete
2416 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2417 nr_scanned = targets[lru] - nr[lru];
2418 nr[lru] = targets[lru] * (100 - percentage) / 100;
2419 nr[lru] -= min(nr[lru], nr_scanned);
2421 lru += LRU_ACTIVE;
2422 nr_scanned = targets[lru] - nr[lru];
2423 nr[lru] = targets[lru] * (100 - percentage) / 100;
2424 nr[lru] -= min(nr[lru], nr_scanned);
2426 scan_adjusted = true;
2428 blk_finish_plug(&plug);
2429 sc->nr_reclaimed += nr_reclaimed;
2432 * Even if we did not try to evict anon pages at all, we want to
2433 * rebalance the anon lru active/inactive ratio.
2435 if (inactive_list_is_low(lruvec, false, sc))
2436 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2437 sc, LRU_ACTIVE_ANON);
2440 /* Use reclaim/compaction for costly allocs or under memory pressure */
2441 static bool in_reclaim_compaction(struct scan_control *sc)
2443 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2444 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2445 sc->priority < DEF_PRIORITY - 2))
2446 return true;
2448 return false;
2452 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2453 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2454 * true if more pages should be reclaimed such that when the page allocator
2455 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2456 * It will give up earlier than that if there is difficulty reclaiming pages.
2458 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2459 unsigned long nr_reclaimed,
2460 unsigned long nr_scanned,
2461 struct scan_control *sc)
2463 unsigned long pages_for_compaction;
2464 unsigned long inactive_lru_pages;
2465 int z;
2467 /* If not in reclaim/compaction mode, stop */
2468 if (!in_reclaim_compaction(sc))
2469 return false;
2471 /* Consider stopping depending on scan and reclaim activity */
2472 if (sc->gfp_mask & __GFP_REPEAT) {
2474 * For __GFP_REPEAT allocations, stop reclaiming if the
2475 * full LRU list has been scanned and we are still failing
2476 * to reclaim pages. This full LRU scan is potentially
2477 * expensive but a __GFP_REPEAT caller really wants to succeed
2479 if (!nr_reclaimed && !nr_scanned)
2480 return false;
2481 } else {
2483 * For non-__GFP_REPEAT allocations which can presumably
2484 * fail without consequence, stop if we failed to reclaim
2485 * any pages from the last SWAP_CLUSTER_MAX number of
2486 * pages that were scanned. This will return to the
2487 * caller faster at the risk reclaim/compaction and
2488 * the resulting allocation attempt fails
2490 if (!nr_reclaimed)
2491 return false;
2495 * If we have not reclaimed enough pages for compaction and the
2496 * inactive lists are large enough, continue reclaiming
2498 pages_for_compaction = compact_gap(sc->order);
2499 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2500 if (get_nr_swap_pages() > 0)
2501 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2502 if (sc->nr_reclaimed < pages_for_compaction &&
2503 inactive_lru_pages > pages_for_compaction)
2504 return true;
2506 /* If compaction would go ahead or the allocation would succeed, stop */
2507 for (z = 0; z <= sc->reclaim_idx; z++) {
2508 struct zone *zone = &pgdat->node_zones[z];
2509 if (!managed_zone(zone))
2510 continue;
2512 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2513 case COMPACT_SUCCESS:
2514 case COMPACT_CONTINUE:
2515 return false;
2516 default:
2517 /* check next zone */
2521 return true;
2524 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2526 struct reclaim_state *reclaim_state = current->reclaim_state;
2527 unsigned long nr_reclaimed, nr_scanned;
2528 bool reclaimable = false;
2530 do {
2531 struct mem_cgroup *root = sc->target_mem_cgroup;
2532 struct mem_cgroup_reclaim_cookie reclaim = {
2533 .pgdat = pgdat,
2534 .priority = sc->priority,
2536 unsigned long node_lru_pages = 0;
2537 struct mem_cgroup *memcg;
2539 nr_reclaimed = sc->nr_reclaimed;
2540 nr_scanned = sc->nr_scanned;
2542 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2543 do {
2544 unsigned long lru_pages;
2545 unsigned long reclaimed;
2546 unsigned long scanned;
2548 if (mem_cgroup_low(root, memcg)) {
2549 if (!sc->may_thrash)
2550 continue;
2551 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2554 reclaimed = sc->nr_reclaimed;
2555 scanned = sc->nr_scanned;
2557 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2558 node_lru_pages += lru_pages;
2560 if (memcg)
2561 shrink_slab(sc->gfp_mask, pgdat->node_id,
2562 memcg, sc->nr_scanned - scanned,
2563 lru_pages);
2565 /* Record the group's reclaim efficiency */
2566 vmpressure(sc->gfp_mask, memcg, false,
2567 sc->nr_scanned - scanned,
2568 sc->nr_reclaimed - reclaimed);
2571 * Direct reclaim and kswapd have to scan all memory
2572 * cgroups to fulfill the overall scan target for the
2573 * node.
2575 * Limit reclaim, on the other hand, only cares about
2576 * nr_to_reclaim pages to be reclaimed and it will
2577 * retry with decreasing priority if one round over the
2578 * whole hierarchy is not sufficient.
2580 if (!global_reclaim(sc) &&
2581 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2582 mem_cgroup_iter_break(root, memcg);
2583 break;
2585 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2588 * Shrink the slab caches in the same proportion that
2589 * the eligible LRU pages were scanned.
2591 if (global_reclaim(sc))
2592 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2593 sc->nr_scanned - nr_scanned,
2594 node_lru_pages);
2596 if (reclaim_state) {
2597 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2598 reclaim_state->reclaimed_slab = 0;
2601 /* Record the subtree's reclaim efficiency */
2602 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2603 sc->nr_scanned - nr_scanned,
2604 sc->nr_reclaimed - nr_reclaimed);
2606 if (sc->nr_reclaimed - nr_reclaimed)
2607 reclaimable = true;
2609 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2610 sc->nr_scanned - nr_scanned, sc));
2612 return reclaimable;
2616 * Returns true if compaction should go ahead for a costly-order request, or
2617 * the allocation would already succeed without compaction. Return false if we
2618 * should reclaim first.
2620 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2622 unsigned long watermark;
2623 enum compact_result suitable;
2625 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2626 if (suitable == COMPACT_SUCCESS)
2627 /* Allocation should succeed already. Don't reclaim. */
2628 return true;
2629 if (suitable == COMPACT_SKIPPED)
2630 /* Compaction cannot yet proceed. Do reclaim. */
2631 return false;
2634 * Compaction is already possible, but it takes time to run and there
2635 * are potentially other callers using the pages just freed. So proceed
2636 * with reclaim to make a buffer of free pages available to give
2637 * compaction a reasonable chance of completing and allocating the page.
2638 * Note that we won't actually reclaim the whole buffer in one attempt
2639 * as the target watermark in should_continue_reclaim() is lower. But if
2640 * we are already above the high+gap watermark, don't reclaim at all.
2642 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2644 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2648 * This is the direct reclaim path, for page-allocating processes. We only
2649 * try to reclaim pages from zones which will satisfy the caller's allocation
2650 * request.
2652 * If a zone is deemed to be full of pinned pages then just give it a light
2653 * scan then give up on it.
2655 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2657 struct zoneref *z;
2658 struct zone *zone;
2659 unsigned long nr_soft_reclaimed;
2660 unsigned long nr_soft_scanned;
2661 gfp_t orig_mask;
2662 pg_data_t *last_pgdat = NULL;
2665 * If the number of buffer_heads in the machine exceeds the maximum
2666 * allowed level, force direct reclaim to scan the highmem zone as
2667 * highmem pages could be pinning lowmem pages storing buffer_heads
2669 orig_mask = sc->gfp_mask;
2670 if (buffer_heads_over_limit) {
2671 sc->gfp_mask |= __GFP_HIGHMEM;
2672 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2675 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2676 sc->reclaim_idx, sc->nodemask) {
2678 * Take care memory controller reclaiming has small influence
2679 * to global LRU.
2681 if (global_reclaim(sc)) {
2682 if (!cpuset_zone_allowed(zone,
2683 GFP_KERNEL | __GFP_HARDWALL))
2684 continue;
2686 if (sc->priority != DEF_PRIORITY &&
2687 !pgdat_reclaimable(zone->zone_pgdat))
2688 continue; /* Let kswapd poll it */
2691 * If we already have plenty of memory free for
2692 * compaction in this zone, don't free any more.
2693 * Even though compaction is invoked for any
2694 * non-zero order, only frequent costly order
2695 * reclamation is disruptive enough to become a
2696 * noticeable problem, like transparent huge
2697 * page allocations.
2699 if (IS_ENABLED(CONFIG_COMPACTION) &&
2700 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2701 compaction_ready(zone, sc)) {
2702 sc->compaction_ready = true;
2703 continue;
2707 * Shrink each node in the zonelist once. If the
2708 * zonelist is ordered by zone (not the default) then a
2709 * node may be shrunk multiple times but in that case
2710 * the user prefers lower zones being preserved.
2712 if (zone->zone_pgdat == last_pgdat)
2713 continue;
2716 * This steals pages from memory cgroups over softlimit
2717 * and returns the number of reclaimed pages and
2718 * scanned pages. This works for global memory pressure
2719 * and balancing, not for a memcg's limit.
2721 nr_soft_scanned = 0;
2722 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2723 sc->order, sc->gfp_mask,
2724 &nr_soft_scanned);
2725 sc->nr_reclaimed += nr_soft_reclaimed;
2726 sc->nr_scanned += nr_soft_scanned;
2727 /* need some check for avoid more shrink_zone() */
2730 /* See comment about same check for global reclaim above */
2731 if (zone->zone_pgdat == last_pgdat)
2732 continue;
2733 last_pgdat = zone->zone_pgdat;
2734 shrink_node(zone->zone_pgdat, sc);
2738 * Restore to original mask to avoid the impact on the caller if we
2739 * promoted it to __GFP_HIGHMEM.
2741 sc->gfp_mask = orig_mask;
2745 * This is the main entry point to direct page reclaim.
2747 * If a full scan of the inactive list fails to free enough memory then we
2748 * are "out of memory" and something needs to be killed.
2750 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2751 * high - the zone may be full of dirty or under-writeback pages, which this
2752 * caller can't do much about. We kick the writeback threads and take explicit
2753 * naps in the hope that some of these pages can be written. But if the
2754 * allocating task holds filesystem locks which prevent writeout this might not
2755 * work, and the allocation attempt will fail.
2757 * returns: 0, if no pages reclaimed
2758 * else, the number of pages reclaimed
2760 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2761 struct scan_control *sc)
2763 int initial_priority = sc->priority;
2764 unsigned long total_scanned = 0;
2765 unsigned long writeback_threshold;
2766 retry:
2767 delayacct_freepages_start();
2769 if (global_reclaim(sc))
2770 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2772 do {
2773 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2774 sc->priority);
2775 sc->nr_scanned = 0;
2776 shrink_zones(zonelist, sc);
2778 total_scanned += sc->nr_scanned;
2779 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2780 break;
2782 if (sc->compaction_ready)
2783 break;
2786 * If we're getting trouble reclaiming, start doing
2787 * writepage even in laptop mode.
2789 if (sc->priority < DEF_PRIORITY - 2)
2790 sc->may_writepage = 1;
2793 * Try to write back as many pages as we just scanned. This
2794 * tends to cause slow streaming writers to write data to the
2795 * disk smoothly, at the dirtying rate, which is nice. But
2796 * that's undesirable in laptop mode, where we *want* lumpy
2797 * writeout. So in laptop mode, write out the whole world.
2799 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2800 if (total_scanned > writeback_threshold) {
2801 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2802 WB_REASON_TRY_TO_FREE_PAGES);
2803 sc->may_writepage = 1;
2805 } while (--sc->priority >= 0);
2807 delayacct_freepages_end();
2809 if (sc->nr_reclaimed)
2810 return sc->nr_reclaimed;
2812 /* Aborted reclaim to try compaction? don't OOM, then */
2813 if (sc->compaction_ready)
2814 return 1;
2816 /* Untapped cgroup reserves? Don't OOM, retry. */
2817 if (!sc->may_thrash) {
2818 sc->priority = initial_priority;
2819 sc->may_thrash = 1;
2820 goto retry;
2823 return 0;
2826 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2828 struct zone *zone;
2829 unsigned long pfmemalloc_reserve = 0;
2830 unsigned long free_pages = 0;
2831 int i;
2832 bool wmark_ok;
2834 for (i = 0; i <= ZONE_NORMAL; i++) {
2835 zone = &pgdat->node_zones[i];
2836 if (!managed_zone(zone) ||
2837 pgdat_reclaimable_pages(pgdat) == 0)
2838 continue;
2840 pfmemalloc_reserve += min_wmark_pages(zone);
2841 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2844 /* If there are no reserves (unexpected config) then do not throttle */
2845 if (!pfmemalloc_reserve)
2846 return true;
2848 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2850 /* kswapd must be awake if processes are being throttled */
2851 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2852 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2853 (enum zone_type)ZONE_NORMAL);
2854 wake_up_interruptible(&pgdat->kswapd_wait);
2857 return wmark_ok;
2861 * Throttle direct reclaimers if backing storage is backed by the network
2862 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2863 * depleted. kswapd will continue to make progress and wake the processes
2864 * when the low watermark is reached.
2866 * Returns true if a fatal signal was delivered during throttling. If this
2867 * happens, the page allocator should not consider triggering the OOM killer.
2869 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2870 nodemask_t *nodemask)
2872 struct zoneref *z;
2873 struct zone *zone;
2874 pg_data_t *pgdat = NULL;
2877 * Kernel threads should not be throttled as they may be indirectly
2878 * responsible for cleaning pages necessary for reclaim to make forward
2879 * progress. kjournald for example may enter direct reclaim while
2880 * committing a transaction where throttling it could forcing other
2881 * processes to block on log_wait_commit().
2883 if (current->flags & PF_KTHREAD)
2884 goto out;
2887 * If a fatal signal is pending, this process should not throttle.
2888 * It should return quickly so it can exit and free its memory
2890 if (fatal_signal_pending(current))
2891 goto out;
2894 * Check if the pfmemalloc reserves are ok by finding the first node
2895 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2896 * GFP_KERNEL will be required for allocating network buffers when
2897 * swapping over the network so ZONE_HIGHMEM is unusable.
2899 * Throttling is based on the first usable node and throttled processes
2900 * wait on a queue until kswapd makes progress and wakes them. There
2901 * is an affinity then between processes waking up and where reclaim
2902 * progress has been made assuming the process wakes on the same node.
2903 * More importantly, processes running on remote nodes will not compete
2904 * for remote pfmemalloc reserves and processes on different nodes
2905 * should make reasonable progress.
2907 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2908 gfp_zone(gfp_mask), nodemask) {
2909 if (zone_idx(zone) > ZONE_NORMAL)
2910 continue;
2912 /* Throttle based on the first usable node */
2913 pgdat = zone->zone_pgdat;
2914 if (pfmemalloc_watermark_ok(pgdat))
2915 goto out;
2916 break;
2919 /* If no zone was usable by the allocation flags then do not throttle */
2920 if (!pgdat)
2921 goto out;
2923 /* Account for the throttling */
2924 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2927 * If the caller cannot enter the filesystem, it's possible that it
2928 * is due to the caller holding an FS lock or performing a journal
2929 * transaction in the case of a filesystem like ext[3|4]. In this case,
2930 * it is not safe to block on pfmemalloc_wait as kswapd could be
2931 * blocked waiting on the same lock. Instead, throttle for up to a
2932 * second before continuing.
2934 if (!(gfp_mask & __GFP_FS)) {
2935 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2936 pfmemalloc_watermark_ok(pgdat), HZ);
2938 goto check_pending;
2941 /* Throttle until kswapd wakes the process */
2942 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2943 pfmemalloc_watermark_ok(pgdat));
2945 check_pending:
2946 if (fatal_signal_pending(current))
2947 return true;
2949 out:
2950 return false;
2953 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2954 gfp_t gfp_mask, nodemask_t *nodemask)
2956 unsigned long nr_reclaimed;
2957 struct scan_control sc = {
2958 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2959 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2960 .reclaim_idx = gfp_zone(gfp_mask),
2961 .order = order,
2962 .nodemask = nodemask,
2963 .priority = DEF_PRIORITY,
2964 .may_writepage = !laptop_mode,
2965 .may_unmap = 1,
2966 .may_swap = 1,
2970 * Do not enter reclaim if fatal signal was delivered while throttled.
2971 * 1 is returned so that the page allocator does not OOM kill at this
2972 * point.
2974 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2975 return 1;
2977 trace_mm_vmscan_direct_reclaim_begin(order,
2978 sc.may_writepage,
2979 gfp_mask,
2980 sc.reclaim_idx);
2982 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2984 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2986 return nr_reclaimed;
2989 #ifdef CONFIG_MEMCG
2991 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2992 gfp_t gfp_mask, bool noswap,
2993 pg_data_t *pgdat,
2994 unsigned long *nr_scanned)
2996 struct scan_control sc = {
2997 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2998 .target_mem_cgroup = memcg,
2999 .may_writepage = !laptop_mode,
3000 .may_unmap = 1,
3001 .reclaim_idx = MAX_NR_ZONES - 1,
3002 .may_swap = !noswap,
3004 unsigned long lru_pages;
3006 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3007 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3009 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3010 sc.may_writepage,
3011 sc.gfp_mask,
3012 sc.reclaim_idx);
3015 * NOTE: Although we can get the priority field, using it
3016 * here is not a good idea, since it limits the pages we can scan.
3017 * if we don't reclaim here, the shrink_node from balance_pgdat
3018 * will pick up pages from other mem cgroup's as well. We hack
3019 * the priority and make it zero.
3021 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3023 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3025 *nr_scanned = sc.nr_scanned;
3026 return sc.nr_reclaimed;
3029 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3030 unsigned long nr_pages,
3031 gfp_t gfp_mask,
3032 bool may_swap)
3034 struct zonelist *zonelist;
3035 unsigned long nr_reclaimed;
3036 int nid;
3037 struct scan_control sc = {
3038 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3039 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3040 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3041 .reclaim_idx = MAX_NR_ZONES - 1,
3042 .target_mem_cgroup = memcg,
3043 .priority = DEF_PRIORITY,
3044 .may_writepage = !laptop_mode,
3045 .may_unmap = 1,
3046 .may_swap = may_swap,
3050 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3051 * take care of from where we get pages. So the node where we start the
3052 * scan does not need to be the current node.
3054 nid = mem_cgroup_select_victim_node(memcg);
3056 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3058 trace_mm_vmscan_memcg_reclaim_begin(0,
3059 sc.may_writepage,
3060 sc.gfp_mask,
3061 sc.reclaim_idx);
3063 current->flags |= PF_MEMALLOC;
3064 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3065 current->flags &= ~PF_MEMALLOC;
3067 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3069 return nr_reclaimed;
3071 #endif
3073 static void age_active_anon(struct pglist_data *pgdat,
3074 struct scan_control *sc)
3076 struct mem_cgroup *memcg;
3078 if (!total_swap_pages)
3079 return;
3081 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3082 do {
3083 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3085 if (inactive_list_is_low(lruvec, false, sc))
3086 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3087 sc, LRU_ACTIVE_ANON);
3089 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3090 } while (memcg);
3093 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3095 unsigned long mark = high_wmark_pages(zone);
3097 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3098 return false;
3101 * If any eligible zone is balanced then the node is not considered
3102 * to be congested or dirty
3104 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3105 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3107 return true;
3111 * Prepare kswapd for sleeping. This verifies that there are no processes
3112 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3114 * Returns true if kswapd is ready to sleep
3116 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3118 int i;
3121 * The throttled processes are normally woken up in balance_pgdat() as
3122 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3123 * race between when kswapd checks the watermarks and a process gets
3124 * throttled. There is also a potential race if processes get
3125 * throttled, kswapd wakes, a large process exits thereby balancing the
3126 * zones, which causes kswapd to exit balance_pgdat() before reaching
3127 * the wake up checks. If kswapd is going to sleep, no process should
3128 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3129 * the wake up is premature, processes will wake kswapd and get
3130 * throttled again. The difference from wake ups in balance_pgdat() is
3131 * that here we are under prepare_to_wait().
3133 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3134 wake_up_all(&pgdat->pfmemalloc_wait);
3136 for (i = 0; i <= classzone_idx; i++) {
3137 struct zone *zone = pgdat->node_zones + i;
3139 if (!managed_zone(zone))
3140 continue;
3142 if (!zone_balanced(zone, order, classzone_idx))
3143 return false;
3146 return true;
3150 * kswapd shrinks a node of pages that are at or below the highest usable
3151 * zone that is currently unbalanced.
3153 * Returns true if kswapd scanned at least the requested number of pages to
3154 * reclaim or if the lack of progress was due to pages under writeback.
3155 * This is used to determine if the scanning priority needs to be raised.
3157 static bool kswapd_shrink_node(pg_data_t *pgdat,
3158 struct scan_control *sc)
3160 struct zone *zone;
3161 int z;
3163 /* Reclaim a number of pages proportional to the number of zones */
3164 sc->nr_to_reclaim = 0;
3165 for (z = 0; z <= sc->reclaim_idx; z++) {
3166 zone = pgdat->node_zones + z;
3167 if (!managed_zone(zone))
3168 continue;
3170 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3174 * Historically care was taken to put equal pressure on all zones but
3175 * now pressure is applied based on node LRU order.
3177 shrink_node(pgdat, sc);
3180 * Fragmentation may mean that the system cannot be rebalanced for
3181 * high-order allocations. If twice the allocation size has been
3182 * reclaimed then recheck watermarks only at order-0 to prevent
3183 * excessive reclaim. Assume that a process requested a high-order
3184 * can direct reclaim/compact.
3186 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3187 sc->order = 0;
3189 return sc->nr_scanned >= sc->nr_to_reclaim;
3193 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3194 * that are eligible for use by the caller until at least one zone is
3195 * balanced.
3197 * Returns the order kswapd finished reclaiming at.
3199 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3200 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3201 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3202 * or lower is eligible for reclaim until at least one usable zone is
3203 * balanced.
3205 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3207 int i;
3208 unsigned long nr_soft_reclaimed;
3209 unsigned long nr_soft_scanned;
3210 struct zone *zone;
3211 struct scan_control sc = {
3212 .gfp_mask = GFP_KERNEL,
3213 .order = order,
3214 .priority = DEF_PRIORITY,
3215 .may_writepage = !laptop_mode,
3216 .may_unmap = 1,
3217 .may_swap = 1,
3219 count_vm_event(PAGEOUTRUN);
3221 do {
3222 bool raise_priority = true;
3224 sc.nr_reclaimed = 0;
3225 sc.reclaim_idx = classzone_idx;
3228 * If the number of buffer_heads exceeds the maximum allowed
3229 * then consider reclaiming from all zones. This has a dual
3230 * purpose -- on 64-bit systems it is expected that
3231 * buffer_heads are stripped during active rotation. On 32-bit
3232 * systems, highmem pages can pin lowmem memory and shrinking
3233 * buffers can relieve lowmem pressure. Reclaim may still not
3234 * go ahead if all eligible zones for the original allocation
3235 * request are balanced to avoid excessive reclaim from kswapd.
3237 if (buffer_heads_over_limit) {
3238 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3239 zone = pgdat->node_zones + i;
3240 if (!managed_zone(zone))
3241 continue;
3243 sc.reclaim_idx = i;
3244 break;
3249 * Only reclaim if there are no eligible zones. Check from
3250 * high to low zone as allocations prefer higher zones.
3251 * Scanning from low to high zone would allow congestion to be
3252 * cleared during a very small window when a small low
3253 * zone was balanced even under extreme pressure when the
3254 * overall node may be congested. Note that sc.reclaim_idx
3255 * is not used as buffer_heads_over_limit may have adjusted
3256 * it.
3258 for (i = classzone_idx; i >= 0; i--) {
3259 zone = pgdat->node_zones + i;
3260 if (!managed_zone(zone))
3261 continue;
3263 if (zone_balanced(zone, sc.order, classzone_idx))
3264 goto out;
3268 * Do some background aging of the anon list, to give
3269 * pages a chance to be referenced before reclaiming. All
3270 * pages are rotated regardless of classzone as this is
3271 * about consistent aging.
3273 age_active_anon(pgdat, &sc);
3276 * If we're getting trouble reclaiming, start doing writepage
3277 * even in laptop mode.
3279 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3280 sc.may_writepage = 1;
3282 /* Call soft limit reclaim before calling shrink_node. */
3283 sc.nr_scanned = 0;
3284 nr_soft_scanned = 0;
3285 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3286 sc.gfp_mask, &nr_soft_scanned);
3287 sc.nr_reclaimed += nr_soft_reclaimed;
3290 * There should be no need to raise the scanning priority if
3291 * enough pages are already being scanned that that high
3292 * watermark would be met at 100% efficiency.
3294 if (kswapd_shrink_node(pgdat, &sc))
3295 raise_priority = false;
3298 * If the low watermark is met there is no need for processes
3299 * to be throttled on pfmemalloc_wait as they should not be
3300 * able to safely make forward progress. Wake them
3302 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3303 pfmemalloc_watermark_ok(pgdat))
3304 wake_up_all(&pgdat->pfmemalloc_wait);
3306 /* Check if kswapd should be suspending */
3307 if (try_to_freeze() || kthread_should_stop())
3308 break;
3311 * Raise priority if scanning rate is too low or there was no
3312 * progress in reclaiming pages
3314 if (raise_priority || !sc.nr_reclaimed)
3315 sc.priority--;
3316 } while (sc.priority >= 1);
3318 out:
3320 * Return the order kswapd stopped reclaiming at as
3321 * prepare_kswapd_sleep() takes it into account. If another caller
3322 * entered the allocator slow path while kswapd was awake, order will
3323 * remain at the higher level.
3325 return sc.order;
3328 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3329 unsigned int classzone_idx)
3331 long remaining = 0;
3332 DEFINE_WAIT(wait);
3334 if (freezing(current) || kthread_should_stop())
3335 return;
3337 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3339 /* Try to sleep for a short interval */
3340 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3342 * Compaction records what page blocks it recently failed to
3343 * isolate pages from and skips them in the future scanning.
3344 * When kswapd is going to sleep, it is reasonable to assume
3345 * that pages and compaction may succeed so reset the cache.
3347 reset_isolation_suitable(pgdat);
3350 * We have freed the memory, now we should compact it to make
3351 * allocation of the requested order possible.
3353 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3355 remaining = schedule_timeout(HZ/10);
3358 * If woken prematurely then reset kswapd_classzone_idx and
3359 * order. The values will either be from a wakeup request or
3360 * the previous request that slept prematurely.
3362 if (remaining) {
3363 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3364 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3367 finish_wait(&pgdat->kswapd_wait, &wait);
3368 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3372 * After a short sleep, check if it was a premature sleep. If not, then
3373 * go fully to sleep until explicitly woken up.
3375 if (!remaining &&
3376 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3377 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3380 * vmstat counters are not perfectly accurate and the estimated
3381 * value for counters such as NR_FREE_PAGES can deviate from the
3382 * true value by nr_online_cpus * threshold. To avoid the zone
3383 * watermarks being breached while under pressure, we reduce the
3384 * per-cpu vmstat threshold while kswapd is awake and restore
3385 * them before going back to sleep.
3387 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3389 if (!kthread_should_stop())
3390 schedule();
3392 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3393 } else {
3394 if (remaining)
3395 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3396 else
3397 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3399 finish_wait(&pgdat->kswapd_wait, &wait);
3403 * The background pageout daemon, started as a kernel thread
3404 * from the init process.
3406 * This basically trickles out pages so that we have _some_
3407 * free memory available even if there is no other activity
3408 * that frees anything up. This is needed for things like routing
3409 * etc, where we otherwise might have all activity going on in
3410 * asynchronous contexts that cannot page things out.
3412 * If there are applications that are active memory-allocators
3413 * (most normal use), this basically shouldn't matter.
3415 static int kswapd(void *p)
3417 unsigned int alloc_order, reclaim_order, classzone_idx;
3418 pg_data_t *pgdat = (pg_data_t*)p;
3419 struct task_struct *tsk = current;
3421 struct reclaim_state reclaim_state = {
3422 .reclaimed_slab = 0,
3424 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3426 lockdep_set_current_reclaim_state(GFP_KERNEL);
3428 if (!cpumask_empty(cpumask))
3429 set_cpus_allowed_ptr(tsk, cpumask);
3430 current->reclaim_state = &reclaim_state;
3433 * Tell the memory management that we're a "memory allocator",
3434 * and that if we need more memory we should get access to it
3435 * regardless (see "__alloc_pages()"). "kswapd" should
3436 * never get caught in the normal page freeing logic.
3438 * (Kswapd normally doesn't need memory anyway, but sometimes
3439 * you need a small amount of memory in order to be able to
3440 * page out something else, and this flag essentially protects
3441 * us from recursively trying to free more memory as we're
3442 * trying to free the first piece of memory in the first place).
3444 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3445 set_freezable();
3447 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3448 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3449 for ( ; ; ) {
3450 bool ret;
3452 kswapd_try_sleep:
3453 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3454 classzone_idx);
3456 /* Read the new order and classzone_idx */
3457 alloc_order = reclaim_order = pgdat->kswapd_order;
3458 classzone_idx = pgdat->kswapd_classzone_idx;
3459 pgdat->kswapd_order = 0;
3460 pgdat->kswapd_classzone_idx = 0;
3462 ret = try_to_freeze();
3463 if (kthread_should_stop())
3464 break;
3467 * We can speed up thawing tasks if we don't call balance_pgdat
3468 * after returning from the refrigerator
3470 if (ret)
3471 continue;
3474 * Reclaim begins at the requested order but if a high-order
3475 * reclaim fails then kswapd falls back to reclaiming for
3476 * order-0. If that happens, kswapd will consider sleeping
3477 * for the order it finished reclaiming at (reclaim_order)
3478 * but kcompactd is woken to compact for the original
3479 * request (alloc_order).
3481 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3482 alloc_order);
3483 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3484 if (reclaim_order < alloc_order)
3485 goto kswapd_try_sleep;
3487 alloc_order = reclaim_order = pgdat->kswapd_order;
3488 classzone_idx = pgdat->kswapd_classzone_idx;
3491 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3492 current->reclaim_state = NULL;
3493 lockdep_clear_current_reclaim_state();
3495 return 0;
3499 * A zone is low on free memory, so wake its kswapd task to service it.
3501 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3503 pg_data_t *pgdat;
3504 int z;
3506 if (!managed_zone(zone))
3507 return;
3509 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3510 return;
3511 pgdat = zone->zone_pgdat;
3512 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3513 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3514 if (!waitqueue_active(&pgdat->kswapd_wait))
3515 return;
3517 /* Only wake kswapd if all zones are unbalanced */
3518 for (z = 0; z <= classzone_idx; z++) {
3519 zone = pgdat->node_zones + z;
3520 if (!managed_zone(zone))
3521 continue;
3523 if (zone_balanced(zone, order, classzone_idx))
3524 return;
3527 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3528 wake_up_interruptible(&pgdat->kswapd_wait);
3531 #ifdef CONFIG_HIBERNATION
3533 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3534 * freed pages.
3536 * Rather than trying to age LRUs the aim is to preserve the overall
3537 * LRU order by reclaiming preferentially
3538 * inactive > active > active referenced > active mapped
3540 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3542 struct reclaim_state reclaim_state;
3543 struct scan_control sc = {
3544 .nr_to_reclaim = nr_to_reclaim,
3545 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3546 .reclaim_idx = MAX_NR_ZONES - 1,
3547 .priority = DEF_PRIORITY,
3548 .may_writepage = 1,
3549 .may_unmap = 1,
3550 .may_swap = 1,
3551 .hibernation_mode = 1,
3553 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3554 struct task_struct *p = current;
3555 unsigned long nr_reclaimed;
3557 p->flags |= PF_MEMALLOC;
3558 lockdep_set_current_reclaim_state(sc.gfp_mask);
3559 reclaim_state.reclaimed_slab = 0;
3560 p->reclaim_state = &reclaim_state;
3562 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3564 p->reclaim_state = NULL;
3565 lockdep_clear_current_reclaim_state();
3566 p->flags &= ~PF_MEMALLOC;
3568 return nr_reclaimed;
3570 #endif /* CONFIG_HIBERNATION */
3572 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3573 not required for correctness. So if the last cpu in a node goes
3574 away, we get changed to run anywhere: as the first one comes back,
3575 restore their cpu bindings. */
3576 static int kswapd_cpu_online(unsigned int cpu)
3578 int nid;
3580 for_each_node_state(nid, N_MEMORY) {
3581 pg_data_t *pgdat = NODE_DATA(nid);
3582 const struct cpumask *mask;
3584 mask = cpumask_of_node(pgdat->node_id);
3586 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3587 /* One of our CPUs online: restore mask */
3588 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3590 return 0;
3594 * This kswapd start function will be called by init and node-hot-add.
3595 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3597 int kswapd_run(int nid)
3599 pg_data_t *pgdat = NODE_DATA(nid);
3600 int ret = 0;
3602 if (pgdat->kswapd)
3603 return 0;
3605 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3606 if (IS_ERR(pgdat->kswapd)) {
3607 /* failure at boot is fatal */
3608 BUG_ON(system_state == SYSTEM_BOOTING);
3609 pr_err("Failed to start kswapd on node %d\n", nid);
3610 ret = PTR_ERR(pgdat->kswapd);
3611 pgdat->kswapd = NULL;
3613 return ret;
3617 * Called by memory hotplug when all memory in a node is offlined. Caller must
3618 * hold mem_hotplug_begin/end().
3620 void kswapd_stop(int nid)
3622 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3624 if (kswapd) {
3625 kthread_stop(kswapd);
3626 NODE_DATA(nid)->kswapd = NULL;
3630 static int __init kswapd_init(void)
3632 int nid, ret;
3634 swap_setup();
3635 for_each_node_state(nid, N_MEMORY)
3636 kswapd_run(nid);
3637 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3638 "mm/vmscan:online", kswapd_cpu_online,
3639 NULL);
3640 WARN_ON(ret < 0);
3641 return 0;
3644 module_init(kswapd_init)
3646 #ifdef CONFIG_NUMA
3648 * Node reclaim mode
3650 * If non-zero call node_reclaim when the number of free pages falls below
3651 * the watermarks.
3653 int node_reclaim_mode __read_mostly;
3655 #define RECLAIM_OFF 0
3656 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3657 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3658 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3661 * Priority for NODE_RECLAIM. This determines the fraction of pages
3662 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3663 * a zone.
3665 #define NODE_RECLAIM_PRIORITY 4
3668 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3669 * occur.
3671 int sysctl_min_unmapped_ratio = 1;
3674 * If the number of slab pages in a zone grows beyond this percentage then
3675 * slab reclaim needs to occur.
3677 int sysctl_min_slab_ratio = 5;
3679 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3681 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3682 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3683 node_page_state(pgdat, NR_ACTIVE_FILE);
3686 * It's possible for there to be more file mapped pages than
3687 * accounted for by the pages on the file LRU lists because
3688 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3690 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3693 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3694 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3696 unsigned long nr_pagecache_reclaimable;
3697 unsigned long delta = 0;
3700 * If RECLAIM_UNMAP is set, then all file pages are considered
3701 * potentially reclaimable. Otherwise, we have to worry about
3702 * pages like swapcache and node_unmapped_file_pages() provides
3703 * a better estimate
3705 if (node_reclaim_mode & RECLAIM_UNMAP)
3706 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3707 else
3708 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3710 /* If we can't clean pages, remove dirty pages from consideration */
3711 if (!(node_reclaim_mode & RECLAIM_WRITE))
3712 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3714 /* Watch for any possible underflows due to delta */
3715 if (unlikely(delta > nr_pagecache_reclaimable))
3716 delta = nr_pagecache_reclaimable;
3718 return nr_pagecache_reclaimable - delta;
3722 * Try to free up some pages from this node through reclaim.
3724 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3726 /* Minimum pages needed in order to stay on node */
3727 const unsigned long nr_pages = 1 << order;
3728 struct task_struct *p = current;
3729 struct reclaim_state reclaim_state;
3730 int classzone_idx = gfp_zone(gfp_mask);
3731 struct scan_control sc = {
3732 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3733 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3734 .order = order,
3735 .priority = NODE_RECLAIM_PRIORITY,
3736 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3737 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3738 .may_swap = 1,
3739 .reclaim_idx = classzone_idx,
3742 cond_resched();
3744 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3745 * and we also need to be able to write out pages for RECLAIM_WRITE
3746 * and RECLAIM_UNMAP.
3748 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3749 lockdep_set_current_reclaim_state(gfp_mask);
3750 reclaim_state.reclaimed_slab = 0;
3751 p->reclaim_state = &reclaim_state;
3753 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3755 * Free memory by calling shrink zone with increasing
3756 * priorities until we have enough memory freed.
3758 do {
3759 shrink_node(pgdat, &sc);
3760 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3763 p->reclaim_state = NULL;
3764 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3765 lockdep_clear_current_reclaim_state();
3766 return sc.nr_reclaimed >= nr_pages;
3769 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3771 int ret;
3774 * Node reclaim reclaims unmapped file backed pages and
3775 * slab pages if we are over the defined limits.
3777 * A small portion of unmapped file backed pages is needed for
3778 * file I/O otherwise pages read by file I/O will be immediately
3779 * thrown out if the node is overallocated. So we do not reclaim
3780 * if less than a specified percentage of the node is used by
3781 * unmapped file backed pages.
3783 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3784 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3785 return NODE_RECLAIM_FULL;
3787 if (!pgdat_reclaimable(pgdat))
3788 return NODE_RECLAIM_FULL;
3791 * Do not scan if the allocation should not be delayed.
3793 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3794 return NODE_RECLAIM_NOSCAN;
3797 * Only run node reclaim on the local node or on nodes that do not
3798 * have associated processors. This will favor the local processor
3799 * over remote processors and spread off node memory allocations
3800 * as wide as possible.
3802 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3803 return NODE_RECLAIM_NOSCAN;
3805 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3806 return NODE_RECLAIM_NOSCAN;
3808 ret = __node_reclaim(pgdat, gfp_mask, order);
3809 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3811 if (!ret)
3812 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3814 return ret;
3816 #endif
3819 * page_evictable - test whether a page is evictable
3820 * @page: the page to test
3822 * Test whether page is evictable--i.e., should be placed on active/inactive
3823 * lists vs unevictable list.
3825 * Reasons page might not be evictable:
3826 * (1) page's mapping marked unevictable
3827 * (2) page is part of an mlocked VMA
3830 int page_evictable(struct page *page)
3832 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3835 #ifdef CONFIG_SHMEM
3837 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3838 * @pages: array of pages to check
3839 * @nr_pages: number of pages to check
3841 * Checks pages for evictability and moves them to the appropriate lru list.
3843 * This function is only used for SysV IPC SHM_UNLOCK.
3845 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3847 struct lruvec *lruvec;
3848 struct pglist_data *pgdat = NULL;
3849 int pgscanned = 0;
3850 int pgrescued = 0;
3851 int i;
3853 for (i = 0; i < nr_pages; i++) {
3854 struct page *page = pages[i];
3855 struct pglist_data *pagepgdat = page_pgdat(page);
3857 pgscanned++;
3858 if (pagepgdat != pgdat) {
3859 if (pgdat)
3860 spin_unlock_irq(&pgdat->lru_lock);
3861 pgdat = pagepgdat;
3862 spin_lock_irq(&pgdat->lru_lock);
3864 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3866 if (!PageLRU(page) || !PageUnevictable(page))
3867 continue;
3869 if (page_evictable(page)) {
3870 enum lru_list lru = page_lru_base_type(page);
3872 VM_BUG_ON_PAGE(PageActive(page), page);
3873 ClearPageUnevictable(page);
3874 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3875 add_page_to_lru_list(page, lruvec, lru);
3876 pgrescued++;
3880 if (pgdat) {
3881 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3882 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3883 spin_unlock_irq(&pgdat->lru_lock);
3886 #endif /* CONFIG_SHMEM */