Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
15#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
17#include <linux/mm.h>
18#include <linux/sched/mm.h>
19#include <linux/module.h>
20#include <linux/gfp.h>
21#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
26#include <linux/vmpressure.h>
27#include <linux/vmstat.h>
28#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
34#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
39#include <linux/compaction.h>
40#include <linux/notifier.h>
41#include <linux/rwsem.h>
42#include <linux/delay.h>
43#include <linux/kthread.h>
44#include <linux/freezer.h>
45#include <linux/memcontrol.h>
46#include <linux/delayacct.h>
47#include <linux/sysctl.h>
48#include <linux/oom.h>
49#include <linux/prefetch.h>
50#include <linux/printk.h>
51#include <linux/dax.h>
52
53#include <asm/tlbflush.h>
54#include <asm/div64.h>
55
56#include <linux/swapops.h>
57#include <linux/balloon_compaction.h>
58
59#include "internal.h"
60
61#define CREATE_TRACE_POINTS
62#include <trace/events/vmscan.h>
63
64struct scan_control {
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
68 /* This context's GFP mask */
69 gfp_t gfp_mask;
70
71 /* Allocation order */
72 int order;
73
74 /*
75 * Nodemask of nodes allowed by the caller. If NULL, all nodes
76 * are scanned.
77 */
78 nodemask_t *nodemask;
79
80 /*
81 * The memory cgroup that hit its limit and as a result is the
82 * primary target of this reclaim invocation.
83 */
84 struct mem_cgroup *target_mem_cgroup;
85
86 /* Scan (total_size >> priority) pages at once */
87 int priority;
88
89 /* The highest zone to isolate pages for reclaim from */
90 enum zone_type reclaim_idx;
91
92 /* Writepage batching in laptop mode; RECLAIM_WRITE */
93 unsigned int may_writepage:1;
94
95 /* Can mapped pages be reclaimed? */
96 unsigned int may_unmap:1;
97
98 /* Can pages be swapped as part of reclaim? */
99 unsigned int may_swap:1;
100
101 /*
102 * Cgroups are not reclaimed below their configured memory.low,
103 * unless we threaten to OOM. If any cgroups are skipped due to
104 * memory.low and nothing was reclaimed, go back for memory.low.
105 */
106 unsigned int memcg_low_reclaim:1;
107 unsigned int memcg_low_skipped:1;
108
109 unsigned int hibernation_mode:1;
110
111 /* One of the zones is ready for compaction */
112 unsigned int compaction_ready:1;
113
114 /* Incremented by the number of inactive pages that were scanned */
115 unsigned long nr_scanned;
116
117 /* Number of pages freed so far during a call to shrink_zones() */
118 unsigned long nr_reclaimed;
119
120 struct {
121 unsigned int dirty;
122 unsigned int unqueued_dirty;
123 unsigned int congested;
124 unsigned int writeback;
125 unsigned int immediate;
126 unsigned int file_taken;
127 unsigned int taken;
128 } nr;
129};
130
131#ifdef ARCH_HAS_PREFETCH
132#define prefetch_prev_lru_page(_page, _base, _field) \
133 do { \
134 if ((_page)->lru.prev != _base) { \
135 struct page *prev; \
136 \
137 prev = lru_to_page(&(_page->lru)); \
138 prefetch(&prev->_field); \
139 } \
140 } while (0)
141#else
142#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
143#endif
144
145#ifdef ARCH_HAS_PREFETCHW
146#define prefetchw_prev_lru_page(_page, _base, _field) \
147 do { \
148 if ((_page)->lru.prev != _base) { \
149 struct page *prev; \
150 \
151 prev = lru_to_page(&(_page->lru)); \
152 prefetchw(&prev->_field); \
153 } \
154 } while (0)
155#else
156#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
157#endif
158
159/*
160 * From 0 .. 100. Higher means more swappy.
161 */
162int vm_swappiness = 60;
163/*
164 * The total number of pages which are beyond the high watermark within all
165 * zones.
166 */
167unsigned long vm_total_pages;
168
169static LIST_HEAD(shrinker_list);
170static DECLARE_RWSEM(shrinker_rwsem);
171
172#ifdef CONFIG_MEMCG
173static bool global_reclaim(struct scan_control *sc)
174{
175 return !sc->target_mem_cgroup;
176}
177
178/**
179 * sane_reclaim - is the usual dirty throttling mechanism operational?
180 * @sc: scan_control in question
181 *
182 * The normal page dirty throttling mechanism in balance_dirty_pages() is
183 * completely broken with the legacy memcg and direct stalling in
184 * shrink_page_list() is used for throttling instead, which lacks all the
185 * niceties such as fairness, adaptive pausing, bandwidth proportional
186 * allocation and configurability.
187 *
188 * This function tests whether the vmscan currently in progress can assume
189 * that the normal dirty throttling mechanism is operational.
190 */
191static bool sane_reclaim(struct scan_control *sc)
192{
193 struct mem_cgroup *memcg = sc->target_mem_cgroup;
194
195 if (!memcg)
196 return true;
197#ifdef CONFIG_CGROUP_WRITEBACK
198 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
199 return true;
200#endif
201 return false;
202}
203
204static void set_memcg_congestion(pg_data_t *pgdat,
205 struct mem_cgroup *memcg,
206 bool congested)
207{
208 struct mem_cgroup_per_node *mn;
209
210 if (!memcg)
211 return;
212
213 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
214 WRITE_ONCE(mn->congested, congested);
215}
216
217static bool memcg_congested(pg_data_t *pgdat,
218 struct mem_cgroup *memcg)
219{
220 struct mem_cgroup_per_node *mn;
221
222 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
223 return READ_ONCE(mn->congested);
224
225}
226#else
227static bool global_reclaim(struct scan_control *sc)
228{
229 return true;
230}
231
232static bool sane_reclaim(struct scan_control *sc)
233{
234 return true;
235}
236
237static inline void set_memcg_congestion(struct pglist_data *pgdat,
238 struct mem_cgroup *memcg, bool congested)
239{
240}
241
242static inline bool memcg_congested(struct pglist_data *pgdat,
243 struct mem_cgroup *memcg)
244{
245 return false;
246
247}
248#endif
249
250/*
251 * This misses isolated pages which are not accounted for to save counters.
252 * As the data only determines if reclaim or compaction continues, it is
253 * not expected that isolated pages will be a dominating factor.
254 */
255unsigned long zone_reclaimable_pages(struct zone *zone)
256{
257 unsigned long nr;
258
259 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
260 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
261 if (get_nr_swap_pages() > 0)
262 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
263 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
264
265 return nr;
266}
267
268/**
269 * lruvec_lru_size - Returns the number of pages on the given LRU list.
270 * @lruvec: lru vector
271 * @lru: lru to use
272 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
273 */
274unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
275{
276 unsigned long lru_size;
277 int zid;
278
279 if (!mem_cgroup_disabled())
280 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
281 else
282 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
283
284 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
285 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
286 unsigned long size;
287
288 if (!managed_zone(zone))
289 continue;
290
291 if (!mem_cgroup_disabled())
292 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
293 else
294 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
295 NR_ZONE_LRU_BASE + lru);
296 lru_size -= min(size, lru_size);
297 }
298
299 return lru_size;
300
301}
302
303/*
304 * Add a shrinker callback to be called from the vm.
305 */
306int prealloc_shrinker(struct shrinker *shrinker)
307{
308 size_t size = sizeof(*shrinker->nr_deferred);
309
310 if (shrinker->flags & SHRINKER_NUMA_AWARE)
311 size *= nr_node_ids;
312
313 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
314 if (!shrinker->nr_deferred)
315 return -ENOMEM;
316 return 0;
317}
318
319void free_prealloced_shrinker(struct shrinker *shrinker)
320{
321 kfree(shrinker->nr_deferred);
322 shrinker->nr_deferred = NULL;
323}
324
325void register_shrinker_prepared(struct shrinker *shrinker)
326{
327 down_write(&shrinker_rwsem);
328 list_add_tail(&shrinker->list, &shrinker_list);
329 up_write(&shrinker_rwsem);
330}
331
332int register_shrinker(struct shrinker *shrinker)
333{
334 int err = prealloc_shrinker(shrinker);
335
336 if (err)
337 return err;
338 register_shrinker_prepared(shrinker);
339 return 0;
340}
341EXPORT_SYMBOL(register_shrinker);
342
343/*
344 * Remove one
345 */
346void unregister_shrinker(struct shrinker *shrinker)
347{
348 if (!shrinker->nr_deferred)
349 return;
350 down_write(&shrinker_rwsem);
351 list_del(&shrinker->list);
352 up_write(&shrinker_rwsem);
353 kfree(shrinker->nr_deferred);
354 shrinker->nr_deferred = NULL;
355}
356EXPORT_SYMBOL(unregister_shrinker);
357
358#define SHRINK_BATCH 128
359
360static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
361 struct shrinker *shrinker, int priority)
362{
363 unsigned long freed = 0;
364 unsigned long long delta;
365 long total_scan;
366 long freeable;
367 long nr;
368 long new_nr;
369 int nid = shrinkctl->nid;
370 long batch_size = shrinker->batch ? shrinker->batch
371 : SHRINK_BATCH;
372 long scanned = 0, next_deferred;
373
374 freeable = shrinker->count_objects(shrinker, shrinkctl);
375 if (freeable == 0)
376 return 0;
377
378 /*
379 * copy the current shrinker scan count into a local variable
380 * and zero it so that other concurrent shrinker invocations
381 * don't also do this scanning work.
382 */
383 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
384
385 total_scan = nr;
386 delta = freeable >> priority;
387 delta *= 4;
388 do_div(delta, shrinker->seeks);
389 total_scan += delta;
390 if (total_scan < 0) {
391 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
392 shrinker->scan_objects, total_scan);
393 total_scan = freeable;
394 next_deferred = nr;
395 } else
396 next_deferred = total_scan;
397
398 /*
399 * We need to avoid excessive windup on filesystem shrinkers
400 * due to large numbers of GFP_NOFS allocations causing the
401 * shrinkers to return -1 all the time. This results in a large
402 * nr being built up so when a shrink that can do some work
403 * comes along it empties the entire cache due to nr >>>
404 * freeable. This is bad for sustaining a working set in
405 * memory.
406 *
407 * Hence only allow the shrinker to scan the entire cache when
408 * a large delta change is calculated directly.
409 */
410 if (delta < freeable / 4)
411 total_scan = min(total_scan, freeable / 2);
412
413 /*
414 * Avoid risking looping forever due to too large nr value:
415 * never try to free more than twice the estimate number of
416 * freeable entries.
417 */
418 if (total_scan > freeable * 2)
419 total_scan = freeable * 2;
420
421 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
422 freeable, delta, total_scan, priority);
423
424 /*
425 * Normally, we should not scan less than batch_size objects in one
426 * pass to avoid too frequent shrinker calls, but if the slab has less
427 * than batch_size objects in total and we are really tight on memory,
428 * we will try to reclaim all available objects, otherwise we can end
429 * up failing allocations although there are plenty of reclaimable
430 * objects spread over several slabs with usage less than the
431 * batch_size.
432 *
433 * We detect the "tight on memory" situations by looking at the total
434 * number of objects we want to scan (total_scan). If it is greater
435 * than the total number of objects on slab (freeable), we must be
436 * scanning at high prio and therefore should try to reclaim as much as
437 * possible.
438 */
439 while (total_scan >= batch_size ||
440 total_scan >= freeable) {
441 unsigned long ret;
442 unsigned long nr_to_scan = min(batch_size, total_scan);
443
444 shrinkctl->nr_to_scan = nr_to_scan;
445 shrinkctl->nr_scanned = nr_to_scan;
446 ret = shrinker->scan_objects(shrinker, shrinkctl);
447 if (ret == SHRINK_STOP)
448 break;
449 freed += ret;
450
451 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
452 total_scan -= shrinkctl->nr_scanned;
453 scanned += shrinkctl->nr_scanned;
454
455 cond_resched();
456 }
457
458 if (next_deferred >= scanned)
459 next_deferred -= scanned;
460 else
461 next_deferred = 0;
462 /*
463 * move the unused scan count back into the shrinker in a
464 * manner that handles concurrent updates. If we exhausted the
465 * scan, there is no need to do an update.
466 */
467 if (next_deferred > 0)
468 new_nr = atomic_long_add_return(next_deferred,
469 &shrinker->nr_deferred[nid]);
470 else
471 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
472
473 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
474 return freed;
475}
476
477/**
478 * shrink_slab - shrink slab caches
479 * @gfp_mask: allocation context
480 * @nid: node whose slab caches to target
481 * @memcg: memory cgroup whose slab caches to target
482 * @priority: the reclaim priority
483 *
484 * Call the shrink functions to age shrinkable caches.
485 *
486 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
487 * unaware shrinkers will receive a node id of 0 instead.
488 *
489 * @memcg specifies the memory cgroup to target. If it is not NULL,
490 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
491 * objects from the memory cgroup specified. Otherwise, only unaware
492 * shrinkers are called.
493 *
494 * @priority is sc->priority, we take the number of objects and >> by priority
495 * in order to get the scan target.
496 *
497 * Returns the number of reclaimed slab objects.
498 */
499static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
500 struct mem_cgroup *memcg,
501 int priority)
502{
503 struct shrinker *shrinker;
504 unsigned long freed = 0;
505
506 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
507 return 0;
508
509 if (!down_read_trylock(&shrinker_rwsem))
510 goto out;
511
512 list_for_each_entry(shrinker, &shrinker_list, list) {
513 struct shrink_control sc = {
514 .gfp_mask = gfp_mask,
515 .nid = nid,
516 .memcg = memcg,
517 };
518
519 /*
520 * If kernel memory accounting is disabled, we ignore
521 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
522 * passing NULL for memcg.
523 */
524 if (memcg_kmem_enabled() &&
525 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
526 continue;
527
528 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
529 sc.nid = 0;
530
531 freed += do_shrink_slab(&sc, shrinker, priority);
532 /*
533 * Bail out if someone want to register a new shrinker to
534 * prevent the regsitration from being stalled for long periods
535 * by parallel ongoing shrinking.
536 */
537 if (rwsem_is_contended(&shrinker_rwsem)) {
538 freed = freed ? : 1;
539 break;
540 }
541 }
542
543 up_read(&shrinker_rwsem);
544out:
545 cond_resched();
546 return freed;
547}
548
549void drop_slab_node(int nid)
550{
551 unsigned long freed;
552
553 do {
554 struct mem_cgroup *memcg = NULL;
555
556 freed = 0;
557 do {
558 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
559 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
560 } while (freed > 10);
561}
562
563void drop_slab(void)
564{
565 int nid;
566
567 for_each_online_node(nid)
568 drop_slab_node(nid);
569}
570
571static inline int is_page_cache_freeable(struct page *page)
572{
573 /*
574 * A freeable page cache page is referenced only by the caller
575 * that isolated the page, the page cache radix tree and
576 * optional buffer heads at page->private.
577 */
578 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
579 HPAGE_PMD_NR : 1;
580 return page_count(page) - page_has_private(page) == 1 + radix_pins;
581}
582
583static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
584{
585 if (current->flags & PF_SWAPWRITE)
586 return 1;
587 if (!inode_write_congested(inode))
588 return 1;
589 if (inode_to_bdi(inode) == current->backing_dev_info)
590 return 1;
591 return 0;
592}
593
594/*
595 * We detected a synchronous write error writing a page out. Probably
596 * -ENOSPC. We need to propagate that into the address_space for a subsequent
597 * fsync(), msync() or close().
598 *
599 * The tricky part is that after writepage we cannot touch the mapping: nothing
600 * prevents it from being freed up. But we have a ref on the page and once
601 * that page is locked, the mapping is pinned.
602 *
603 * We're allowed to run sleeping lock_page() here because we know the caller has
604 * __GFP_FS.
605 */
606static void handle_write_error(struct address_space *mapping,
607 struct page *page, int error)
608{
609 lock_page(page);
610 if (page_mapping(page) == mapping)
611 mapping_set_error(mapping, error);
612 unlock_page(page);
613}
614
615/* possible outcome of pageout() */
616typedef enum {
617 /* failed to write page out, page is locked */
618 PAGE_KEEP,
619 /* move page to the active list, page is locked */
620 PAGE_ACTIVATE,
621 /* page has been sent to the disk successfully, page is unlocked */
622 PAGE_SUCCESS,
623 /* page is clean and locked */
624 PAGE_CLEAN,
625} pageout_t;
626
627/*
628 * pageout is called by shrink_page_list() for each dirty page.
629 * Calls ->writepage().
630 */
631static pageout_t pageout(struct page *page, struct address_space *mapping,
632 struct scan_control *sc)
633{
634 /*
635 * If the page is dirty, only perform writeback if that write
636 * will be non-blocking. To prevent this allocation from being
637 * stalled by pagecache activity. But note that there may be
638 * stalls if we need to run get_block(). We could test
639 * PagePrivate for that.
640 *
641 * If this process is currently in __generic_file_write_iter() against
642 * this page's queue, we can perform writeback even if that
643 * will block.
644 *
645 * If the page is swapcache, write it back even if that would
646 * block, for some throttling. This happens by accident, because
647 * swap_backing_dev_info is bust: it doesn't reflect the
648 * congestion state of the swapdevs. Easy to fix, if needed.
649 */
650 if (!is_page_cache_freeable(page))
651 return PAGE_KEEP;
652 if (!mapping) {
653 /*
654 * Some data journaling orphaned pages can have
655 * page->mapping == NULL while being dirty with clean buffers.
656 */
657 if (page_has_private(page)) {
658 if (try_to_free_buffers(page)) {
659 ClearPageDirty(page);
660 pr_info("%s: orphaned page\n", __func__);
661 return PAGE_CLEAN;
662 }
663 }
664 return PAGE_KEEP;
665 }
666 if (mapping->a_ops->writepage == NULL)
667 return PAGE_ACTIVATE;
668 if (!may_write_to_inode(mapping->host, sc))
669 return PAGE_KEEP;
670
671 if (clear_page_dirty_for_io(page)) {
672 int res;
673 struct writeback_control wbc = {
674 .sync_mode = WB_SYNC_NONE,
675 .nr_to_write = SWAP_CLUSTER_MAX,
676 .range_start = 0,
677 .range_end = LLONG_MAX,
678 .for_reclaim = 1,
679 };
680
681 SetPageReclaim(page);
682 res = mapping->a_ops->writepage(page, &wbc);
683 if (res < 0)
684 handle_write_error(mapping, page, res);
685 if (res == AOP_WRITEPAGE_ACTIVATE) {
686 ClearPageReclaim(page);
687 return PAGE_ACTIVATE;
688 }
689
690 if (!PageWriteback(page)) {
691 /* synchronous write or broken a_ops? */
692 ClearPageReclaim(page);
693 }
694 trace_mm_vmscan_writepage(page);
695 inc_node_page_state(page, NR_VMSCAN_WRITE);
696 return PAGE_SUCCESS;
697 }
698
699 return PAGE_CLEAN;
700}
701
702/*
703 * Same as remove_mapping, but if the page is removed from the mapping, it
704 * gets returned with a refcount of 0.
705 */
706static int __remove_mapping(struct address_space *mapping, struct page *page,
707 bool reclaimed)
708{
709 unsigned long flags;
710 int refcount;
711
712 BUG_ON(!PageLocked(page));
713 BUG_ON(mapping != page_mapping(page));
714
715 xa_lock_irqsave(&mapping->i_pages, flags);
716 /*
717 * The non racy check for a busy page.
718 *
719 * Must be careful with the order of the tests. When someone has
720 * a ref to the page, it may be possible that they dirty it then
721 * drop the reference. So if PageDirty is tested before page_count
722 * here, then the following race may occur:
723 *
724 * get_user_pages(&page);
725 * [user mapping goes away]
726 * write_to(page);
727 * !PageDirty(page) [good]
728 * SetPageDirty(page);
729 * put_page(page);
730 * !page_count(page) [good, discard it]
731 *
732 * [oops, our write_to data is lost]
733 *
734 * Reversing the order of the tests ensures such a situation cannot
735 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
736 * load is not satisfied before that of page->_refcount.
737 *
738 * Note that if SetPageDirty is always performed via set_page_dirty,
739 * and thus under the i_pages lock, then this ordering is not required.
740 */
741 if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
742 refcount = 1 + HPAGE_PMD_NR;
743 else
744 refcount = 2;
745 if (!page_ref_freeze(page, refcount))
746 goto cannot_free;
747 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
748 if (unlikely(PageDirty(page))) {
749 page_ref_unfreeze(page, refcount);
750 goto cannot_free;
751 }
752
753 if (PageSwapCache(page)) {
754 swp_entry_t swap = { .val = page_private(page) };
755 mem_cgroup_swapout(page, swap);
756 __delete_from_swap_cache(page);
757 xa_unlock_irqrestore(&mapping->i_pages, flags);
758 put_swap_page(page, swap);
759 } else {
760 void (*freepage)(struct page *);
761 void *shadow = NULL;
762
763 freepage = mapping->a_ops->freepage;
764 /*
765 * Remember a shadow entry for reclaimed file cache in
766 * order to detect refaults, thus thrashing, later on.
767 *
768 * But don't store shadows in an address space that is
769 * already exiting. This is not just an optizimation,
770 * inode reclaim needs to empty out the radix tree or
771 * the nodes are lost. Don't plant shadows behind its
772 * back.
773 *
774 * We also don't store shadows for DAX mappings because the
775 * only page cache pages found in these are zero pages
776 * covering holes, and because we don't want to mix DAX
777 * exceptional entries and shadow exceptional entries in the
778 * same address_space.
779 */
780 if (reclaimed && page_is_file_cache(page) &&
781 !mapping_exiting(mapping) && !dax_mapping(mapping))
782 shadow = workingset_eviction(mapping, page);
783 __delete_from_page_cache(page, shadow);
784 xa_unlock_irqrestore(&mapping->i_pages, flags);
785
786 if (freepage != NULL)
787 freepage(page);
788 }
789
790 return 1;
791
792cannot_free:
793 xa_unlock_irqrestore(&mapping->i_pages, flags);
794 return 0;
795}
796
797/*
798 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
799 * someone else has a ref on the page, abort and return 0. If it was
800 * successfully detached, return 1. Assumes the caller has a single ref on
801 * this page.
802 */
803int remove_mapping(struct address_space *mapping, struct page *page)
804{
805 if (__remove_mapping(mapping, page, false)) {
806 /*
807 * Unfreezing the refcount with 1 rather than 2 effectively
808 * drops the pagecache ref for us without requiring another
809 * atomic operation.
810 */
811 page_ref_unfreeze(page, 1);
812 return 1;
813 }
814 return 0;
815}
816
817/**
818 * putback_lru_page - put previously isolated page onto appropriate LRU list
819 * @page: page to be put back to appropriate lru list
820 *
821 * Add previously isolated @page to appropriate LRU list.
822 * Page may still be unevictable for other reasons.
823 *
824 * lru_lock must not be held, interrupts must be enabled.
825 */
826void putback_lru_page(struct page *page)
827{
828 lru_cache_add(page);
829 put_page(page); /* drop ref from isolate */
830}
831
832enum page_references {
833 PAGEREF_RECLAIM,
834 PAGEREF_RECLAIM_CLEAN,
835 PAGEREF_KEEP,
836 PAGEREF_ACTIVATE,
837};
838
839static enum page_references page_check_references(struct page *page,
840 struct scan_control *sc)
841{
842 int referenced_ptes, referenced_page;
843 unsigned long vm_flags;
844
845 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
846 &vm_flags);
847 referenced_page = TestClearPageReferenced(page);
848
849 /*
850 * Mlock lost the isolation race with us. Let try_to_unmap()
851 * move the page to the unevictable list.
852 */
853 if (vm_flags & VM_LOCKED)
854 return PAGEREF_RECLAIM;
855
856 if (referenced_ptes) {
857 if (PageSwapBacked(page))
858 return PAGEREF_ACTIVATE;
859 /*
860 * All mapped pages start out with page table
861 * references from the instantiating fault, so we need
862 * to look twice if a mapped file page is used more
863 * than once.
864 *
865 * Mark it and spare it for another trip around the
866 * inactive list. Another page table reference will
867 * lead to its activation.
868 *
869 * Note: the mark is set for activated pages as well
870 * so that recently deactivated but used pages are
871 * quickly recovered.
872 */
873 SetPageReferenced(page);
874
875 if (referenced_page || referenced_ptes > 1)
876 return PAGEREF_ACTIVATE;
877
878 /*
879 * Activate file-backed executable pages after first usage.
880 */
881 if (vm_flags & VM_EXEC)
882 return PAGEREF_ACTIVATE;
883
884 return PAGEREF_KEEP;
885 }
886
887 /* Reclaim if clean, defer dirty pages to writeback */
888 if (referenced_page && !PageSwapBacked(page))
889 return PAGEREF_RECLAIM_CLEAN;
890
891 return PAGEREF_RECLAIM;
892}
893
894/* Check if a page is dirty or under writeback */
895static void page_check_dirty_writeback(struct page *page,
896 bool *dirty, bool *writeback)
897{
898 struct address_space *mapping;
899
900 /*
901 * Anonymous pages are not handled by flushers and must be written
902 * from reclaim context. Do not stall reclaim based on them
903 */
904 if (!page_is_file_cache(page) ||
905 (PageAnon(page) && !PageSwapBacked(page))) {
906 *dirty = false;
907 *writeback = false;
908 return;
909 }
910
911 /* By default assume that the page flags are accurate */
912 *dirty = PageDirty(page);
913 *writeback = PageWriteback(page);
914
915 /* Verify dirty/writeback state if the filesystem supports it */
916 if (!page_has_private(page))
917 return;
918
919 mapping = page_mapping(page);
920 if (mapping && mapping->a_ops->is_dirty_writeback)
921 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
922}
923
924/*
925 * shrink_page_list() returns the number of reclaimed pages
926 */
927static unsigned long shrink_page_list(struct list_head *page_list,
928 struct pglist_data *pgdat,
929 struct scan_control *sc,
930 enum ttu_flags ttu_flags,
931 struct reclaim_stat *stat,
932 bool force_reclaim)
933{
934 LIST_HEAD(ret_pages);
935 LIST_HEAD(free_pages);
936 int pgactivate = 0;
937 unsigned nr_unqueued_dirty = 0;
938 unsigned nr_dirty = 0;
939 unsigned nr_congested = 0;
940 unsigned nr_reclaimed = 0;
941 unsigned nr_writeback = 0;
942 unsigned nr_immediate = 0;
943 unsigned nr_ref_keep = 0;
944 unsigned nr_unmap_fail = 0;
945
946 cond_resched();
947
948 while (!list_empty(page_list)) {
949 struct address_space *mapping;
950 struct page *page;
951 int may_enter_fs;
952 enum page_references references = PAGEREF_RECLAIM_CLEAN;
953 bool dirty, writeback;
954
955 cond_resched();
956
957 page = lru_to_page(page_list);
958 list_del(&page->lru);
959
960 if (!trylock_page(page))
961 goto keep;
962
963 VM_BUG_ON_PAGE(PageActive(page), page);
964
965 sc->nr_scanned++;
966
967 if (unlikely(!page_evictable(page)))
968 goto activate_locked;
969
970 if (!sc->may_unmap && page_mapped(page))
971 goto keep_locked;
972
973 /* Double the slab pressure for mapped and swapcache pages */
974 if ((page_mapped(page) || PageSwapCache(page)) &&
975 !(PageAnon(page) && !PageSwapBacked(page)))
976 sc->nr_scanned++;
977
978 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
979 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
980
981 /*
982 * The number of dirty pages determines if a node is marked
983 * reclaim_congested which affects wait_iff_congested. kswapd
984 * will stall and start writing pages if the tail of the LRU
985 * is all dirty unqueued pages.
986 */
987 page_check_dirty_writeback(page, &dirty, &writeback);
988 if (dirty || writeback)
989 nr_dirty++;
990
991 if (dirty && !writeback)
992 nr_unqueued_dirty++;
993
994 /*
995 * Treat this page as congested if the underlying BDI is or if
996 * pages are cycling through the LRU so quickly that the
997 * pages marked for immediate reclaim are making it to the
998 * end of the LRU a second time.
999 */
1000 mapping = page_mapping(page);
1001 if (((dirty || writeback) && mapping &&
1002 inode_write_congested(mapping->host)) ||
1003 (writeback && PageReclaim(page)))
1004 nr_congested++;
1005
1006 /*
1007 * If a page at the tail of the LRU is under writeback, there
1008 * are three cases to consider.
1009 *
1010 * 1) If reclaim is encountering an excessive number of pages
1011 * under writeback and this page is both under writeback and
1012 * PageReclaim then it indicates that pages are being queued
1013 * for IO but are being recycled through the LRU before the
1014 * IO can complete. Waiting on the page itself risks an
1015 * indefinite stall if it is impossible to writeback the
1016 * page due to IO error or disconnected storage so instead
1017 * note that the LRU is being scanned too quickly and the
1018 * caller can stall after page list has been processed.
1019 *
1020 * 2) Global or new memcg reclaim encounters a page that is
1021 * not marked for immediate reclaim, or the caller does not
1022 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1023 * not to fs). In this case mark the page for immediate
1024 * reclaim and continue scanning.
1025 *
1026 * Require may_enter_fs because we would wait on fs, which
1027 * may not have submitted IO yet. And the loop driver might
1028 * enter reclaim, and deadlock if it waits on a page for
1029 * which it is needed to do the write (loop masks off
1030 * __GFP_IO|__GFP_FS for this reason); but more thought
1031 * would probably show more reasons.
1032 *
1033 * 3) Legacy memcg encounters a page that is already marked
1034 * PageReclaim. memcg does not have any dirty pages
1035 * throttling so we could easily OOM just because too many
1036 * pages are in writeback and there is nothing else to
1037 * reclaim. Wait for the writeback to complete.
1038 *
1039 * In cases 1) and 2) we activate the pages to get them out of
1040 * the way while we continue scanning for clean pages on the
1041 * inactive list and refilling from the active list. The
1042 * observation here is that waiting for disk writes is more
1043 * expensive than potentially causing reloads down the line.
1044 * Since they're marked for immediate reclaim, they won't put
1045 * memory pressure on the cache working set any longer than it
1046 * takes to write them to disk.
1047 */
1048 if (PageWriteback(page)) {
1049 /* Case 1 above */
1050 if (current_is_kswapd() &&
1051 PageReclaim(page) &&
1052 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1053 nr_immediate++;
1054 goto activate_locked;
1055
1056 /* Case 2 above */
1057 } else if (sane_reclaim(sc) ||
1058 !PageReclaim(page) || !may_enter_fs) {
1059 /*
1060 * This is slightly racy - end_page_writeback()
1061 * might have just cleared PageReclaim, then
1062 * setting PageReclaim here end up interpreted
1063 * as PageReadahead - but that does not matter
1064 * enough to care. What we do want is for this
1065 * page to have PageReclaim set next time memcg
1066 * reclaim reaches the tests above, so it will
1067 * then wait_on_page_writeback() to avoid OOM;
1068 * and it's also appropriate in global reclaim.
1069 */
1070 SetPageReclaim(page);
1071 nr_writeback++;
1072 goto activate_locked;
1073
1074 /* Case 3 above */
1075 } else {
1076 unlock_page(page);
1077 wait_on_page_writeback(page);
1078 /* then go back and try same page again */
1079 list_add_tail(&page->lru, page_list);
1080 continue;
1081 }
1082 }
1083
1084 if (!force_reclaim)
1085 references = page_check_references(page, sc);
1086
1087 switch (references) {
1088 case PAGEREF_ACTIVATE:
1089 goto activate_locked;
1090 case PAGEREF_KEEP:
1091 nr_ref_keep++;
1092 goto keep_locked;
1093 case PAGEREF_RECLAIM:
1094 case PAGEREF_RECLAIM_CLEAN:
1095 ; /* try to reclaim the page below */
1096 }
1097
1098 /*
1099 * Anonymous process memory has backing store?
1100 * Try to allocate it some swap space here.
1101 * Lazyfree page could be freed directly
1102 */
1103 if (PageAnon(page) && PageSwapBacked(page)) {
1104 if (!PageSwapCache(page)) {
1105 if (!(sc->gfp_mask & __GFP_IO))
1106 goto keep_locked;
1107 if (PageTransHuge(page)) {
1108 /* cannot split THP, skip it */
1109 if (!can_split_huge_page(page, NULL))
1110 goto activate_locked;
1111 /*
1112 * Split pages without a PMD map right
1113 * away. Chances are some or all of the
1114 * tail pages can be freed without IO.
1115 */
1116 if (!compound_mapcount(page) &&
1117 split_huge_page_to_list(page,
1118 page_list))
1119 goto activate_locked;
1120 }
1121 if (!add_to_swap(page)) {
1122 if (!PageTransHuge(page))
1123 goto activate_locked;
1124 /* Fallback to swap normal pages */
1125 if (split_huge_page_to_list(page,
1126 page_list))
1127 goto activate_locked;
1128#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1129 count_vm_event(THP_SWPOUT_FALLBACK);
1130#endif
1131 if (!add_to_swap(page))
1132 goto activate_locked;
1133 }
1134
1135 may_enter_fs = 1;
1136
1137 /* Adding to swap updated mapping */
1138 mapping = page_mapping(page);
1139 }
1140 } else if (unlikely(PageTransHuge(page))) {
1141 /* Split file THP */
1142 if (split_huge_page_to_list(page, page_list))
1143 goto keep_locked;
1144 }
1145
1146 /*
1147 * The page is mapped into the page tables of one or more
1148 * processes. Try to unmap it here.
1149 */
1150 if (page_mapped(page)) {
1151 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1152
1153 if (unlikely(PageTransHuge(page)))
1154 flags |= TTU_SPLIT_HUGE_PMD;
1155 if (!try_to_unmap(page, flags)) {
1156 nr_unmap_fail++;
1157 goto activate_locked;
1158 }
1159 }
1160
1161 if (PageDirty(page)) {
1162 /*
1163 * Only kswapd can writeback filesystem pages
1164 * to avoid risk of stack overflow. But avoid
1165 * injecting inefficient single-page IO into
1166 * flusher writeback as much as possible: only
1167 * write pages when we've encountered many
1168 * dirty pages, and when we've already scanned
1169 * the rest of the LRU for clean pages and see
1170 * the same dirty pages again (PageReclaim).
1171 */
1172 if (page_is_file_cache(page) &&
1173 (!current_is_kswapd() || !PageReclaim(page) ||
1174 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1175 /*
1176 * Immediately reclaim when written back.
1177 * Similar in principal to deactivate_page()
1178 * except we already have the page isolated
1179 * and know it's dirty
1180 */
1181 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1182 SetPageReclaim(page);
1183
1184 goto activate_locked;
1185 }
1186
1187 if (references == PAGEREF_RECLAIM_CLEAN)
1188 goto keep_locked;
1189 if (!may_enter_fs)
1190 goto keep_locked;
1191 if (!sc->may_writepage)
1192 goto keep_locked;
1193
1194 /*
1195 * Page is dirty. Flush the TLB if a writable entry
1196 * potentially exists to avoid CPU writes after IO
1197 * starts and then write it out here.
1198 */
1199 try_to_unmap_flush_dirty();
1200 switch (pageout(page, mapping, sc)) {
1201 case PAGE_KEEP:
1202 goto keep_locked;
1203 case PAGE_ACTIVATE:
1204 goto activate_locked;
1205 case PAGE_SUCCESS:
1206 if (PageWriteback(page))
1207 goto keep;
1208 if (PageDirty(page))
1209 goto keep;
1210
1211 /*
1212 * A synchronous write - probably a ramdisk. Go
1213 * ahead and try to reclaim the page.
1214 */
1215 if (!trylock_page(page))
1216 goto keep;
1217 if (PageDirty(page) || PageWriteback(page))
1218 goto keep_locked;
1219 mapping = page_mapping(page);
1220 case PAGE_CLEAN:
1221 ; /* try to free the page below */
1222 }
1223 }
1224
1225 /*
1226 * If the page has buffers, try to free the buffer mappings
1227 * associated with this page. If we succeed we try to free
1228 * the page as well.
1229 *
1230 * We do this even if the page is PageDirty().
1231 * try_to_release_page() does not perform I/O, but it is
1232 * possible for a page to have PageDirty set, but it is actually
1233 * clean (all its buffers are clean). This happens if the
1234 * buffers were written out directly, with submit_bh(). ext3
1235 * will do this, as well as the blockdev mapping.
1236 * try_to_release_page() will discover that cleanness and will
1237 * drop the buffers and mark the page clean - it can be freed.
1238 *
1239 * Rarely, pages can have buffers and no ->mapping. These are
1240 * the pages which were not successfully invalidated in
1241 * truncate_complete_page(). We try to drop those buffers here
1242 * and if that worked, and the page is no longer mapped into
1243 * process address space (page_count == 1) it can be freed.
1244 * Otherwise, leave the page on the LRU so it is swappable.
1245 */
1246 if (page_has_private(page)) {
1247 if (!try_to_release_page(page, sc->gfp_mask))
1248 goto activate_locked;
1249 if (!mapping && page_count(page) == 1) {
1250 unlock_page(page);
1251 if (put_page_testzero(page))
1252 goto free_it;
1253 else {
1254 /*
1255 * rare race with speculative reference.
1256 * the speculative reference will free
1257 * this page shortly, so we may
1258 * increment nr_reclaimed here (and
1259 * leave it off the LRU).
1260 */
1261 nr_reclaimed++;
1262 continue;
1263 }
1264 }
1265 }
1266
1267 if (PageAnon(page) && !PageSwapBacked(page)) {
1268 /* follow __remove_mapping for reference */
1269 if (!page_ref_freeze(page, 1))
1270 goto keep_locked;
1271 if (PageDirty(page)) {
1272 page_ref_unfreeze(page, 1);
1273 goto keep_locked;
1274 }
1275
1276 count_vm_event(PGLAZYFREED);
1277 count_memcg_page_event(page, PGLAZYFREED);
1278 } else if (!mapping || !__remove_mapping(mapping, page, true))
1279 goto keep_locked;
1280 /*
1281 * At this point, we have no other references and there is
1282 * no way to pick any more up (removed from LRU, removed
1283 * from pagecache). Can use non-atomic bitops now (and
1284 * we obviously don't have to worry about waking up a process
1285 * waiting on the page lock, because there are no references.
1286 */
1287 __ClearPageLocked(page);
1288free_it:
1289 nr_reclaimed++;
1290
1291 /*
1292 * Is there need to periodically free_page_list? It would
1293 * appear not as the counts should be low
1294 */
1295 if (unlikely(PageTransHuge(page))) {
1296 mem_cgroup_uncharge(page);
1297 (*get_compound_page_dtor(page))(page);
1298 } else
1299 list_add(&page->lru, &free_pages);
1300 continue;
1301
1302activate_locked:
1303 /* Not a candidate for swapping, so reclaim swap space. */
1304 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1305 PageMlocked(page)))
1306 try_to_free_swap(page);
1307 VM_BUG_ON_PAGE(PageActive(page), page);
1308 if (!PageMlocked(page)) {
1309 SetPageActive(page);
1310 pgactivate++;
1311 count_memcg_page_event(page, PGACTIVATE);
1312 }
1313keep_locked:
1314 unlock_page(page);
1315keep:
1316 list_add(&page->lru, &ret_pages);
1317 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1318 }
1319
1320 mem_cgroup_uncharge_list(&free_pages);
1321 try_to_unmap_flush();
1322 free_unref_page_list(&free_pages);
1323
1324 list_splice(&ret_pages, page_list);
1325 count_vm_events(PGACTIVATE, pgactivate);
1326
1327 if (stat) {
1328 stat->nr_dirty = nr_dirty;
1329 stat->nr_congested = nr_congested;
1330 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1331 stat->nr_writeback = nr_writeback;
1332 stat->nr_immediate = nr_immediate;
1333 stat->nr_activate = pgactivate;
1334 stat->nr_ref_keep = nr_ref_keep;
1335 stat->nr_unmap_fail = nr_unmap_fail;
1336 }
1337 return nr_reclaimed;
1338}
1339
1340unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1341 struct list_head *page_list)
1342{
1343 struct scan_control sc = {
1344 .gfp_mask = GFP_KERNEL,
1345 .priority = DEF_PRIORITY,
1346 .may_unmap = 1,
1347 };
1348 unsigned long ret;
1349 struct page *page, *next;
1350 LIST_HEAD(clean_pages);
1351
1352 list_for_each_entry_safe(page, next, page_list, lru) {
1353 if (page_is_file_cache(page) && !PageDirty(page) &&
1354 !__PageMovable(page)) {
1355 ClearPageActive(page);
1356 list_move(&page->lru, &clean_pages);
1357 }
1358 }
1359
1360 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1361 TTU_IGNORE_ACCESS, NULL, true);
1362 list_splice(&clean_pages, page_list);
1363 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1364 return ret;
1365}
1366
1367/*
1368 * Attempt to remove the specified page from its LRU. Only take this page
1369 * if it is of the appropriate PageActive status. Pages which are being
1370 * freed elsewhere are also ignored.
1371 *
1372 * page: page to consider
1373 * mode: one of the LRU isolation modes defined above
1374 *
1375 * returns 0 on success, -ve errno on failure.
1376 */
1377int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1378{
1379 int ret = -EINVAL;
1380
1381 /* Only take pages on the LRU. */
1382 if (!PageLRU(page))
1383 return ret;
1384
1385 /* Compaction should not handle unevictable pages but CMA can do so */
1386 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1387 return ret;
1388
1389 ret = -EBUSY;
1390
1391 /*
1392 * To minimise LRU disruption, the caller can indicate that it only
1393 * wants to isolate pages it will be able to operate on without
1394 * blocking - clean pages for the most part.
1395 *
1396 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1397 * that it is possible to migrate without blocking
1398 */
1399 if (mode & ISOLATE_ASYNC_MIGRATE) {
1400 /* All the caller can do on PageWriteback is block */
1401 if (PageWriteback(page))
1402 return ret;
1403
1404 if (PageDirty(page)) {
1405 struct address_space *mapping;
1406 bool migrate_dirty;
1407
1408 /*
1409 * Only pages without mappings or that have a
1410 * ->migratepage callback are possible to migrate
1411 * without blocking. However, we can be racing with
1412 * truncation so it's necessary to lock the page
1413 * to stabilise the mapping as truncation holds
1414 * the page lock until after the page is removed
1415 * from the page cache.
1416 */
1417 if (!trylock_page(page))
1418 return ret;
1419
1420 mapping = page_mapping(page);
1421 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1422 unlock_page(page);
1423 if (!migrate_dirty)
1424 return ret;
1425 }
1426 }
1427
1428 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1429 return ret;
1430
1431 if (likely(get_page_unless_zero(page))) {
1432 /*
1433 * Be careful not to clear PageLRU until after we're
1434 * sure the page is not being freed elsewhere -- the
1435 * page release code relies on it.
1436 */
1437 ClearPageLRU(page);
1438 ret = 0;
1439 }
1440
1441 return ret;
1442}
1443
1444
1445/*
1446 * Update LRU sizes after isolating pages. The LRU size updates must
1447 * be complete before mem_cgroup_update_lru_size due to a santity check.
1448 */
1449static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1450 enum lru_list lru, unsigned long *nr_zone_taken)
1451{
1452 int zid;
1453
1454 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1455 if (!nr_zone_taken[zid])
1456 continue;
1457
1458 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1459#ifdef CONFIG_MEMCG
1460 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1461#endif
1462 }
1463
1464}
1465
1466/*
1467 * zone_lru_lock is heavily contended. Some of the functions that
1468 * shrink the lists perform better by taking out a batch of pages
1469 * and working on them outside the LRU lock.
1470 *
1471 * For pagecache intensive workloads, this function is the hottest
1472 * spot in the kernel (apart from copy_*_user functions).
1473 *
1474 * Appropriate locks must be held before calling this function.
1475 *
1476 * @nr_to_scan: The number of eligible pages to look through on the list.
1477 * @lruvec: The LRU vector to pull pages from.
1478 * @dst: The temp list to put pages on to.
1479 * @nr_scanned: The number of pages that were scanned.
1480 * @sc: The scan_control struct for this reclaim session
1481 * @mode: One of the LRU isolation modes
1482 * @lru: LRU list id for isolating
1483 *
1484 * returns how many pages were moved onto *@dst.
1485 */
1486static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1487 struct lruvec *lruvec, struct list_head *dst,
1488 unsigned long *nr_scanned, struct scan_control *sc,
1489 isolate_mode_t mode, enum lru_list lru)
1490{
1491 struct list_head *src = &lruvec->lists[lru];
1492 unsigned long nr_taken = 0;
1493 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1494 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1495 unsigned long skipped = 0;
1496 unsigned long scan, total_scan, nr_pages;
1497 LIST_HEAD(pages_skipped);
1498
1499 scan = 0;
1500 for (total_scan = 0;
1501 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1502 total_scan++) {
1503 struct page *page;
1504
1505 page = lru_to_page(src);
1506 prefetchw_prev_lru_page(page, src, flags);
1507
1508 VM_BUG_ON_PAGE(!PageLRU(page), page);
1509
1510 if (page_zonenum(page) > sc->reclaim_idx) {
1511 list_move(&page->lru, &pages_skipped);
1512 nr_skipped[page_zonenum(page)]++;
1513 continue;
1514 }
1515
1516 /*
1517 * Do not count skipped pages because that makes the function
1518 * return with no isolated pages if the LRU mostly contains
1519 * ineligible pages. This causes the VM to not reclaim any
1520 * pages, triggering a premature OOM.
1521 */
1522 scan++;
1523 switch (__isolate_lru_page(page, mode)) {
1524 case 0:
1525 nr_pages = hpage_nr_pages(page);
1526 nr_taken += nr_pages;
1527 nr_zone_taken[page_zonenum(page)] += nr_pages;
1528 list_move(&page->lru, dst);
1529 break;
1530
1531 case -EBUSY:
1532 /* else it is being freed elsewhere */
1533 list_move(&page->lru, src);
1534 continue;
1535
1536 default:
1537 BUG();
1538 }
1539 }
1540
1541 /*
1542 * Splice any skipped pages to the start of the LRU list. Note that
1543 * this disrupts the LRU order when reclaiming for lower zones but
1544 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1545 * scanning would soon rescan the same pages to skip and put the
1546 * system at risk of premature OOM.
1547 */
1548 if (!list_empty(&pages_skipped)) {
1549 int zid;
1550
1551 list_splice(&pages_skipped, src);
1552 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1553 if (!nr_skipped[zid])
1554 continue;
1555
1556 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1557 skipped += nr_skipped[zid];
1558 }
1559 }
1560 *nr_scanned = total_scan;
1561 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1562 total_scan, skipped, nr_taken, mode, lru);
1563 update_lru_sizes(lruvec, lru, nr_zone_taken);
1564 return nr_taken;
1565}
1566
1567/**
1568 * isolate_lru_page - tries to isolate a page from its LRU list
1569 * @page: page to isolate from its LRU list
1570 *
1571 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1572 * vmstat statistic corresponding to whatever LRU list the page was on.
1573 *
1574 * Returns 0 if the page was removed from an LRU list.
1575 * Returns -EBUSY if the page was not on an LRU list.
1576 *
1577 * The returned page will have PageLRU() cleared. If it was found on
1578 * the active list, it will have PageActive set. If it was found on
1579 * the unevictable list, it will have the PageUnevictable bit set. That flag
1580 * may need to be cleared by the caller before letting the page go.
1581 *
1582 * The vmstat statistic corresponding to the list on which the page was
1583 * found will be decremented.
1584 *
1585 * Restrictions:
1586 *
1587 * (1) Must be called with an elevated refcount on the page. This is a
1588 * fundamentnal difference from isolate_lru_pages (which is called
1589 * without a stable reference).
1590 * (2) the lru_lock must not be held.
1591 * (3) interrupts must be enabled.
1592 */
1593int isolate_lru_page(struct page *page)
1594{
1595 int ret = -EBUSY;
1596
1597 VM_BUG_ON_PAGE(!page_count(page), page);
1598 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1599
1600 if (PageLRU(page)) {
1601 struct zone *zone = page_zone(page);
1602 struct lruvec *lruvec;
1603
1604 spin_lock_irq(zone_lru_lock(zone));
1605 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1606 if (PageLRU(page)) {
1607 int lru = page_lru(page);
1608 get_page(page);
1609 ClearPageLRU(page);
1610 del_page_from_lru_list(page, lruvec, lru);
1611 ret = 0;
1612 }
1613 spin_unlock_irq(zone_lru_lock(zone));
1614 }
1615 return ret;
1616}
1617
1618/*
1619 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1620 * then get resheduled. When there are massive number of tasks doing page
1621 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1622 * the LRU list will go small and be scanned faster than necessary, leading to
1623 * unnecessary swapping, thrashing and OOM.
1624 */
1625static int too_many_isolated(struct pglist_data *pgdat, int file,
1626 struct scan_control *sc)
1627{
1628 unsigned long inactive, isolated;
1629
1630 if (current_is_kswapd())
1631 return 0;
1632
1633 if (!sane_reclaim(sc))
1634 return 0;
1635
1636 if (file) {
1637 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1638 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1639 } else {
1640 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1641 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1642 }
1643
1644 /*
1645 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1646 * won't get blocked by normal direct-reclaimers, forming a circular
1647 * deadlock.
1648 */
1649 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1650 inactive >>= 3;
1651
1652 return isolated > inactive;
1653}
1654
1655static noinline_for_stack void
1656putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1657{
1658 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1659 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1660 LIST_HEAD(pages_to_free);
1661
1662 /*
1663 * Put back any unfreeable pages.
1664 */
1665 while (!list_empty(page_list)) {
1666 struct page *page = lru_to_page(page_list);
1667 int lru;
1668
1669 VM_BUG_ON_PAGE(PageLRU(page), page);
1670 list_del(&page->lru);
1671 if (unlikely(!page_evictable(page))) {
1672 spin_unlock_irq(&pgdat->lru_lock);
1673 putback_lru_page(page);
1674 spin_lock_irq(&pgdat->lru_lock);
1675 continue;
1676 }
1677
1678 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1679
1680 SetPageLRU(page);
1681 lru = page_lru(page);
1682 add_page_to_lru_list(page, lruvec, lru);
1683
1684 if (is_active_lru(lru)) {
1685 int file = is_file_lru(lru);
1686 int numpages = hpage_nr_pages(page);
1687 reclaim_stat->recent_rotated[file] += numpages;
1688 }
1689 if (put_page_testzero(page)) {
1690 __ClearPageLRU(page);
1691 __ClearPageActive(page);
1692 del_page_from_lru_list(page, lruvec, lru);
1693
1694 if (unlikely(PageCompound(page))) {
1695 spin_unlock_irq(&pgdat->lru_lock);
1696 mem_cgroup_uncharge(page);
1697 (*get_compound_page_dtor(page))(page);
1698 spin_lock_irq(&pgdat->lru_lock);
1699 } else
1700 list_add(&page->lru, &pages_to_free);
1701 }
1702 }
1703
1704 /*
1705 * To save our caller's stack, now use input list for pages to free.
1706 */
1707 list_splice(&pages_to_free, page_list);
1708}
1709
1710/*
1711 * If a kernel thread (such as nfsd for loop-back mounts) services
1712 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1713 * In that case we should only throttle if the backing device it is
1714 * writing to is congested. In other cases it is safe to throttle.
1715 */
1716static int current_may_throttle(void)
1717{
1718 return !(current->flags & PF_LESS_THROTTLE) ||
1719 current->backing_dev_info == NULL ||
1720 bdi_write_congested(current->backing_dev_info);
1721}
1722
1723/*
1724 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1725 * of reclaimed pages
1726 */
1727static noinline_for_stack unsigned long
1728shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1729 struct scan_control *sc, enum lru_list lru)
1730{
1731 LIST_HEAD(page_list);
1732 unsigned long nr_scanned;
1733 unsigned long nr_reclaimed = 0;
1734 unsigned long nr_taken;
1735 struct reclaim_stat stat = {};
1736 isolate_mode_t isolate_mode = 0;
1737 int file = is_file_lru(lru);
1738 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1739 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1740 bool stalled = false;
1741
1742 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1743 if (stalled)
1744 return 0;
1745
1746 /* wait a bit for the reclaimer. */
1747 msleep(100);
1748 stalled = true;
1749
1750 /* We are about to die and free our memory. Return now. */
1751 if (fatal_signal_pending(current))
1752 return SWAP_CLUSTER_MAX;
1753 }
1754
1755 lru_add_drain();
1756
1757 if (!sc->may_unmap)
1758 isolate_mode |= ISOLATE_UNMAPPED;
1759
1760 spin_lock_irq(&pgdat->lru_lock);
1761
1762 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1763 &nr_scanned, sc, isolate_mode, lru);
1764
1765 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1766 reclaim_stat->recent_scanned[file] += nr_taken;
1767
1768 if (current_is_kswapd()) {
1769 if (global_reclaim(sc))
1770 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1771 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1772 nr_scanned);
1773 } else {
1774 if (global_reclaim(sc))
1775 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1776 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1777 nr_scanned);
1778 }
1779 spin_unlock_irq(&pgdat->lru_lock);
1780
1781 if (nr_taken == 0)
1782 return 0;
1783
1784 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1785 &stat, false);
1786
1787 spin_lock_irq(&pgdat->lru_lock);
1788
1789 if (current_is_kswapd()) {
1790 if (global_reclaim(sc))
1791 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1792 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1793 nr_reclaimed);
1794 } else {
1795 if (global_reclaim(sc))
1796 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1797 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1798 nr_reclaimed);
1799 }
1800
1801 putback_inactive_pages(lruvec, &page_list);
1802
1803 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1804
1805 spin_unlock_irq(&pgdat->lru_lock);
1806
1807 mem_cgroup_uncharge_list(&page_list);
1808 free_unref_page_list(&page_list);
1809
1810 /*
1811 * If dirty pages are scanned that are not queued for IO, it
1812 * implies that flushers are not doing their job. This can
1813 * happen when memory pressure pushes dirty pages to the end of
1814 * the LRU before the dirty limits are breached and the dirty
1815 * data has expired. It can also happen when the proportion of
1816 * dirty pages grows not through writes but through memory
1817 * pressure reclaiming all the clean cache. And in some cases,
1818 * the flushers simply cannot keep up with the allocation
1819 * rate. Nudge the flusher threads in case they are asleep.
1820 */
1821 if (stat.nr_unqueued_dirty == nr_taken)
1822 wakeup_flusher_threads(WB_REASON_VMSCAN);
1823
1824 sc->nr.dirty += stat.nr_dirty;
1825 sc->nr.congested += stat.nr_congested;
1826 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1827 sc->nr.writeback += stat.nr_writeback;
1828 sc->nr.immediate += stat.nr_immediate;
1829 sc->nr.taken += nr_taken;
1830 if (file)
1831 sc->nr.file_taken += nr_taken;
1832
1833 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1834 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1835 return nr_reclaimed;
1836}
1837
1838/*
1839 * This moves pages from the active list to the inactive list.
1840 *
1841 * We move them the other way if the page is referenced by one or more
1842 * processes, from rmap.
1843 *
1844 * If the pages are mostly unmapped, the processing is fast and it is
1845 * appropriate to hold zone_lru_lock across the whole operation. But if
1846 * the pages are mapped, the processing is slow (page_referenced()) so we
1847 * should drop zone_lru_lock around each page. It's impossible to balance
1848 * this, so instead we remove the pages from the LRU while processing them.
1849 * It is safe to rely on PG_active against the non-LRU pages in here because
1850 * nobody will play with that bit on a non-LRU page.
1851 *
1852 * The downside is that we have to touch page->_refcount against each page.
1853 * But we had to alter page->flags anyway.
1854 *
1855 * Returns the number of pages moved to the given lru.
1856 */
1857
1858static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1859 struct list_head *list,
1860 struct list_head *pages_to_free,
1861 enum lru_list lru)
1862{
1863 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1864 struct page *page;
1865 int nr_pages;
1866 int nr_moved = 0;
1867
1868 while (!list_empty(list)) {
1869 page = lru_to_page(list);
1870 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1871
1872 VM_BUG_ON_PAGE(PageLRU(page), page);
1873 SetPageLRU(page);
1874
1875 nr_pages = hpage_nr_pages(page);
1876 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1877 list_move(&page->lru, &lruvec->lists[lru]);
1878
1879 if (put_page_testzero(page)) {
1880 __ClearPageLRU(page);
1881 __ClearPageActive(page);
1882 del_page_from_lru_list(page, lruvec, lru);
1883
1884 if (unlikely(PageCompound(page))) {
1885 spin_unlock_irq(&pgdat->lru_lock);
1886 mem_cgroup_uncharge(page);
1887 (*get_compound_page_dtor(page))(page);
1888 spin_lock_irq(&pgdat->lru_lock);
1889 } else
1890 list_add(&page->lru, pages_to_free);
1891 } else {
1892 nr_moved += nr_pages;
1893 }
1894 }
1895
1896 if (!is_active_lru(lru)) {
1897 __count_vm_events(PGDEACTIVATE, nr_moved);
1898 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1899 nr_moved);
1900 }
1901
1902 return nr_moved;
1903}
1904
1905static void shrink_active_list(unsigned long nr_to_scan,
1906 struct lruvec *lruvec,
1907 struct scan_control *sc,
1908 enum lru_list lru)
1909{
1910 unsigned long nr_taken;
1911 unsigned long nr_scanned;
1912 unsigned long vm_flags;
1913 LIST_HEAD(l_hold); /* The pages which were snipped off */
1914 LIST_HEAD(l_active);
1915 LIST_HEAD(l_inactive);
1916 struct page *page;
1917 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1918 unsigned nr_deactivate, nr_activate;
1919 unsigned nr_rotated = 0;
1920 isolate_mode_t isolate_mode = 0;
1921 int file = is_file_lru(lru);
1922 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1923
1924 lru_add_drain();
1925
1926 if (!sc->may_unmap)
1927 isolate_mode |= ISOLATE_UNMAPPED;
1928
1929 spin_lock_irq(&pgdat->lru_lock);
1930
1931 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1932 &nr_scanned, sc, isolate_mode, lru);
1933
1934 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1935 reclaim_stat->recent_scanned[file] += nr_taken;
1936
1937 __count_vm_events(PGREFILL, nr_scanned);
1938 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
1939
1940 spin_unlock_irq(&pgdat->lru_lock);
1941
1942 while (!list_empty(&l_hold)) {
1943 cond_resched();
1944 page = lru_to_page(&l_hold);
1945 list_del(&page->lru);
1946
1947 if (unlikely(!page_evictable(page))) {
1948 putback_lru_page(page);
1949 continue;
1950 }
1951
1952 if (unlikely(buffer_heads_over_limit)) {
1953 if (page_has_private(page) && trylock_page(page)) {
1954 if (page_has_private(page))
1955 try_to_release_page(page, 0);
1956 unlock_page(page);
1957 }
1958 }
1959
1960 if (page_referenced(page, 0, sc->target_mem_cgroup,
1961 &vm_flags)) {
1962 nr_rotated += hpage_nr_pages(page);
1963 /*
1964 * Identify referenced, file-backed active pages and
1965 * give them one more trip around the active list. So
1966 * that executable code get better chances to stay in
1967 * memory under moderate memory pressure. Anon pages
1968 * are not likely to be evicted by use-once streaming
1969 * IO, plus JVM can create lots of anon VM_EXEC pages,
1970 * so we ignore them here.
1971 */
1972 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1973 list_add(&page->lru, &l_active);
1974 continue;
1975 }
1976 }
1977
1978 ClearPageActive(page); /* we are de-activating */
1979 list_add(&page->lru, &l_inactive);
1980 }
1981
1982 /*
1983 * Move pages back to the lru list.
1984 */
1985 spin_lock_irq(&pgdat->lru_lock);
1986 /*
1987 * Count referenced pages from currently used mappings as rotated,
1988 * even though only some of them are actually re-activated. This
1989 * helps balance scan pressure between file and anonymous pages in
1990 * get_scan_count.
1991 */
1992 reclaim_stat->recent_rotated[file] += nr_rotated;
1993
1994 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1995 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1996 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1997 spin_unlock_irq(&pgdat->lru_lock);
1998
1999 mem_cgroup_uncharge_list(&l_hold);
2000 free_unref_page_list(&l_hold);
2001 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2002 nr_deactivate, nr_rotated, sc->priority, file);
2003}
2004
2005/*
2006 * The inactive anon list should be small enough that the VM never has
2007 * to do too much work.
2008 *
2009 * The inactive file list should be small enough to leave most memory
2010 * to the established workingset on the scan-resistant active list,
2011 * but large enough to avoid thrashing the aggregate readahead window.
2012 *
2013 * Both inactive lists should also be large enough that each inactive
2014 * page has a chance to be referenced again before it is reclaimed.
2015 *
2016 * If that fails and refaulting is observed, the inactive list grows.
2017 *
2018 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2019 * on this LRU, maintained by the pageout code. An inactive_ratio
2020 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2021 *
2022 * total target max
2023 * memory ratio inactive
2024 * -------------------------------------
2025 * 10MB 1 5MB
2026 * 100MB 1 50MB
2027 * 1GB 3 250MB
2028 * 10GB 10 0.9GB
2029 * 100GB 31 3GB
2030 * 1TB 101 10GB
2031 * 10TB 320 32GB
2032 */
2033static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2034 struct mem_cgroup *memcg,
2035 struct scan_control *sc, bool actual_reclaim)
2036{
2037 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2038 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2039 enum lru_list inactive_lru = file * LRU_FILE;
2040 unsigned long inactive, active;
2041 unsigned long inactive_ratio;
2042 unsigned long refaults;
2043 unsigned long gb;
2044
2045 /*
2046 * If we don't have swap space, anonymous page deactivation
2047 * is pointless.
2048 */
2049 if (!file && !total_swap_pages)
2050 return false;
2051
2052 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2053 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2054
2055 if (memcg)
2056 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2057 else
2058 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2059
2060 /*
2061 * When refaults are being observed, it means a new workingset
2062 * is being established. Disable active list protection to get
2063 * rid of the stale workingset quickly.
2064 */
2065 if (file && actual_reclaim && lruvec->refaults != refaults) {
2066 inactive_ratio = 0;
2067 } else {
2068 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2069 if (gb)
2070 inactive_ratio = int_sqrt(10 * gb);
2071 else
2072 inactive_ratio = 1;
2073 }
2074
2075 if (actual_reclaim)
2076 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2077 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2078 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2079 inactive_ratio, file);
2080
2081 return inactive * inactive_ratio < active;
2082}
2083
2084static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2085 struct lruvec *lruvec, struct mem_cgroup *memcg,
2086 struct scan_control *sc)
2087{
2088 if (is_active_lru(lru)) {
2089 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2090 memcg, sc, true))
2091 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2092 return 0;
2093 }
2094
2095 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2096}
2097
2098enum scan_balance {
2099 SCAN_EQUAL,
2100 SCAN_FRACT,
2101 SCAN_ANON,
2102 SCAN_FILE,
2103};
2104
2105/*
2106 * Determine how aggressively the anon and file LRU lists should be
2107 * scanned. The relative value of each set of LRU lists is determined
2108 * by looking at the fraction of the pages scanned we did rotate back
2109 * onto the active list instead of evict.
2110 *
2111 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2112 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2113 */
2114static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2115 struct scan_control *sc, unsigned long *nr,
2116 unsigned long *lru_pages)
2117{
2118 int swappiness = mem_cgroup_swappiness(memcg);
2119 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2120 u64 fraction[2];
2121 u64 denominator = 0; /* gcc */
2122 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2123 unsigned long anon_prio, file_prio;
2124 enum scan_balance scan_balance;
2125 unsigned long anon, file;
2126 unsigned long ap, fp;
2127 enum lru_list lru;
2128
2129 /* If we have no swap space, do not bother scanning anon pages. */
2130 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2131 scan_balance = SCAN_FILE;
2132 goto out;
2133 }
2134
2135 /*
2136 * Global reclaim will swap to prevent OOM even with no
2137 * swappiness, but memcg users want to use this knob to
2138 * disable swapping for individual groups completely when
2139 * using the memory controller's swap limit feature would be
2140 * too expensive.
2141 */
2142 if (!global_reclaim(sc) && !swappiness) {
2143 scan_balance = SCAN_FILE;
2144 goto out;
2145 }
2146
2147 /*
2148 * Do not apply any pressure balancing cleverness when the
2149 * system is close to OOM, scan both anon and file equally
2150 * (unless the swappiness setting disagrees with swapping).
2151 */
2152 if (!sc->priority && swappiness) {
2153 scan_balance = SCAN_EQUAL;
2154 goto out;
2155 }
2156
2157 /*
2158 * Prevent the reclaimer from falling into the cache trap: as
2159 * cache pages start out inactive, every cache fault will tip
2160 * the scan balance towards the file LRU. And as the file LRU
2161 * shrinks, so does the window for rotation from references.
2162 * This means we have a runaway feedback loop where a tiny
2163 * thrashing file LRU becomes infinitely more attractive than
2164 * anon pages. Try to detect this based on file LRU size.
2165 */
2166 if (global_reclaim(sc)) {
2167 unsigned long pgdatfile;
2168 unsigned long pgdatfree;
2169 int z;
2170 unsigned long total_high_wmark = 0;
2171
2172 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2173 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2174 node_page_state(pgdat, NR_INACTIVE_FILE);
2175
2176 for (z = 0; z < MAX_NR_ZONES; z++) {
2177 struct zone *zone = &pgdat->node_zones[z];
2178 if (!managed_zone(zone))
2179 continue;
2180
2181 total_high_wmark += high_wmark_pages(zone);
2182 }
2183
2184 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2185 /*
2186 * Force SCAN_ANON if there are enough inactive
2187 * anonymous pages on the LRU in eligible zones.
2188 * Otherwise, the small LRU gets thrashed.
2189 */
2190 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2191 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2192 >> sc->priority) {
2193 scan_balance = SCAN_ANON;
2194 goto out;
2195 }
2196 }
2197 }
2198
2199 /*
2200 * If there is enough inactive page cache, i.e. if the size of the
2201 * inactive list is greater than that of the active list *and* the
2202 * inactive list actually has some pages to scan on this priority, we
2203 * do not reclaim anything from the anonymous working set right now.
2204 * Without the second condition we could end up never scanning an
2205 * lruvec even if it has plenty of old anonymous pages unless the
2206 * system is under heavy pressure.
2207 */
2208 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2209 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2210 scan_balance = SCAN_FILE;
2211 goto out;
2212 }
2213
2214 scan_balance = SCAN_FRACT;
2215
2216 /*
2217 * With swappiness at 100, anonymous and file have the same priority.
2218 * This scanning priority is essentially the inverse of IO cost.
2219 */
2220 anon_prio = swappiness;
2221 file_prio = 200 - anon_prio;
2222
2223 /*
2224 * OK, so we have swap space and a fair amount of page cache
2225 * pages. We use the recently rotated / recently scanned
2226 * ratios to determine how valuable each cache is.
2227 *
2228 * Because workloads change over time (and to avoid overflow)
2229 * we keep these statistics as a floating average, which ends
2230 * up weighing recent references more than old ones.
2231 *
2232 * anon in [0], file in [1]
2233 */
2234
2235 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2236 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2237 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2238 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2239
2240 spin_lock_irq(&pgdat->lru_lock);
2241 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2242 reclaim_stat->recent_scanned[0] /= 2;
2243 reclaim_stat->recent_rotated[0] /= 2;
2244 }
2245
2246 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2247 reclaim_stat->recent_scanned[1] /= 2;
2248 reclaim_stat->recent_rotated[1] /= 2;
2249 }
2250
2251 /*
2252 * The amount of pressure on anon vs file pages is inversely
2253 * proportional to the fraction of recently scanned pages on
2254 * each list that were recently referenced and in active use.
2255 */
2256 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2257 ap /= reclaim_stat->recent_rotated[0] + 1;
2258
2259 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2260 fp /= reclaim_stat->recent_rotated[1] + 1;
2261 spin_unlock_irq(&pgdat->lru_lock);
2262
2263 fraction[0] = ap;
2264 fraction[1] = fp;
2265 denominator = ap + fp + 1;
2266out:
2267 *lru_pages = 0;
2268 for_each_evictable_lru(lru) {
2269 int file = is_file_lru(lru);
2270 unsigned long size;
2271 unsigned long scan;
2272
2273 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2274 scan = size >> sc->priority;
2275 /*
2276 * If the cgroup's already been deleted, make sure to
2277 * scrape out the remaining cache.
2278 */
2279 if (!scan && !mem_cgroup_online(memcg))
2280 scan = min(size, SWAP_CLUSTER_MAX);
2281
2282 switch (scan_balance) {
2283 case SCAN_EQUAL:
2284 /* Scan lists relative to size */
2285 break;
2286 case SCAN_FRACT:
2287 /*
2288 * Scan types proportional to swappiness and
2289 * their relative recent reclaim efficiency.
2290 */
2291 scan = div64_u64(scan * fraction[file],
2292 denominator);
2293 break;
2294 case SCAN_FILE:
2295 case SCAN_ANON:
2296 /* Scan one type exclusively */
2297 if ((scan_balance == SCAN_FILE) != file) {
2298 size = 0;
2299 scan = 0;
2300 }
2301 break;
2302 default:
2303 /* Look ma, no brain */
2304 BUG();
2305 }
2306
2307 *lru_pages += size;
2308 nr[lru] = scan;
2309 }
2310}
2311
2312/*
2313 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2314 */
2315static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2316 struct scan_control *sc, unsigned long *lru_pages)
2317{
2318 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2319 unsigned long nr[NR_LRU_LISTS];
2320 unsigned long targets[NR_LRU_LISTS];
2321 unsigned long nr_to_scan;
2322 enum lru_list lru;
2323 unsigned long nr_reclaimed = 0;
2324 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2325 struct blk_plug plug;
2326 bool scan_adjusted;
2327
2328 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2329
2330 /* Record the original scan target for proportional adjustments later */
2331 memcpy(targets, nr, sizeof(nr));
2332
2333 /*
2334 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2335 * event that can occur when there is little memory pressure e.g.
2336 * multiple streaming readers/writers. Hence, we do not abort scanning
2337 * when the requested number of pages are reclaimed when scanning at
2338 * DEF_PRIORITY on the assumption that the fact we are direct
2339 * reclaiming implies that kswapd is not keeping up and it is best to
2340 * do a batch of work at once. For memcg reclaim one check is made to
2341 * abort proportional reclaim if either the file or anon lru has already
2342 * dropped to zero at the first pass.
2343 */
2344 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2345 sc->priority == DEF_PRIORITY);
2346
2347 blk_start_plug(&plug);
2348 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2349 nr[LRU_INACTIVE_FILE]) {
2350 unsigned long nr_anon, nr_file, percentage;
2351 unsigned long nr_scanned;
2352
2353 for_each_evictable_lru(lru) {
2354 if (nr[lru]) {
2355 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2356 nr[lru] -= nr_to_scan;
2357
2358 nr_reclaimed += shrink_list(lru, nr_to_scan,
2359 lruvec, memcg, sc);
2360 }
2361 }
2362
2363 cond_resched();
2364
2365 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2366 continue;
2367
2368 /*
2369 * For kswapd and memcg, reclaim at least the number of pages
2370 * requested. Ensure that the anon and file LRUs are scanned
2371 * proportionally what was requested by get_scan_count(). We
2372 * stop reclaiming one LRU and reduce the amount scanning
2373 * proportional to the original scan target.
2374 */
2375 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2376 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2377
2378 /*
2379 * It's just vindictive to attack the larger once the smaller
2380 * has gone to zero. And given the way we stop scanning the
2381 * smaller below, this makes sure that we only make one nudge
2382 * towards proportionality once we've got nr_to_reclaim.
2383 */
2384 if (!nr_file || !nr_anon)
2385 break;
2386
2387 if (nr_file > nr_anon) {
2388 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2389 targets[LRU_ACTIVE_ANON] + 1;
2390 lru = LRU_BASE;
2391 percentage = nr_anon * 100 / scan_target;
2392 } else {
2393 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2394 targets[LRU_ACTIVE_FILE] + 1;
2395 lru = LRU_FILE;
2396 percentage = nr_file * 100 / scan_target;
2397 }
2398
2399 /* Stop scanning the smaller of the LRU */
2400 nr[lru] = 0;
2401 nr[lru + LRU_ACTIVE] = 0;
2402
2403 /*
2404 * Recalculate the other LRU scan count based on its original
2405 * scan target and the percentage scanning already complete
2406 */
2407 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2408 nr_scanned = targets[lru] - nr[lru];
2409 nr[lru] = targets[lru] * (100 - percentage) / 100;
2410 nr[lru] -= min(nr[lru], nr_scanned);
2411
2412 lru += LRU_ACTIVE;
2413 nr_scanned = targets[lru] - nr[lru];
2414 nr[lru] = targets[lru] * (100 - percentage) / 100;
2415 nr[lru] -= min(nr[lru], nr_scanned);
2416
2417 scan_adjusted = true;
2418 }
2419 blk_finish_plug(&plug);
2420 sc->nr_reclaimed += nr_reclaimed;
2421
2422 /*
2423 * Even if we did not try to evict anon pages at all, we want to
2424 * rebalance the anon lru active/inactive ratio.
2425 */
2426 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2427 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2428 sc, LRU_ACTIVE_ANON);
2429}
2430
2431/* Use reclaim/compaction for costly allocs or under memory pressure */
2432static bool in_reclaim_compaction(struct scan_control *sc)
2433{
2434 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2435 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2436 sc->priority < DEF_PRIORITY - 2))
2437 return true;
2438
2439 return false;
2440}
2441
2442/*
2443 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2444 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2445 * true if more pages should be reclaimed such that when the page allocator
2446 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2447 * It will give up earlier than that if there is difficulty reclaiming pages.
2448 */
2449static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2450 unsigned long nr_reclaimed,
2451 unsigned long nr_scanned,
2452 struct scan_control *sc)
2453{
2454 unsigned long pages_for_compaction;
2455 unsigned long inactive_lru_pages;
2456 int z;
2457
2458 /* If not in reclaim/compaction mode, stop */
2459 if (!in_reclaim_compaction(sc))
2460 return false;
2461
2462 /* Consider stopping depending on scan and reclaim activity */
2463 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2464 /*
2465 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2466 * full LRU list has been scanned and we are still failing
2467 * to reclaim pages. This full LRU scan is potentially
2468 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2469 */
2470 if (!nr_reclaimed && !nr_scanned)
2471 return false;
2472 } else {
2473 /*
2474 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2475 * fail without consequence, stop if we failed to reclaim
2476 * any pages from the last SWAP_CLUSTER_MAX number of
2477 * pages that were scanned. This will return to the
2478 * caller faster at the risk reclaim/compaction and
2479 * the resulting allocation attempt fails
2480 */
2481 if (!nr_reclaimed)
2482 return false;
2483 }
2484
2485 /*
2486 * If we have not reclaimed enough pages for compaction and the
2487 * inactive lists are large enough, continue reclaiming
2488 */
2489 pages_for_compaction = compact_gap(sc->order);
2490 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2491 if (get_nr_swap_pages() > 0)
2492 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2493 if (sc->nr_reclaimed < pages_for_compaction &&
2494 inactive_lru_pages > pages_for_compaction)
2495 return true;
2496
2497 /* If compaction would go ahead or the allocation would succeed, stop */
2498 for (z = 0; z <= sc->reclaim_idx; z++) {
2499 struct zone *zone = &pgdat->node_zones[z];
2500 if (!managed_zone(zone))
2501 continue;
2502
2503 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2504 case COMPACT_SUCCESS:
2505 case COMPACT_CONTINUE:
2506 return false;
2507 default:
2508 /* check next zone */
2509 ;
2510 }
2511 }
2512 return true;
2513}
2514
2515static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2516{
2517 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2518 (memcg && memcg_congested(pgdat, memcg));
2519}
2520
2521static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2522{
2523 struct reclaim_state *reclaim_state = current->reclaim_state;
2524 unsigned long nr_reclaimed, nr_scanned;
2525 bool reclaimable = false;
2526
2527 do {
2528 struct mem_cgroup *root = sc->target_mem_cgroup;
2529 struct mem_cgroup_reclaim_cookie reclaim = {
2530 .pgdat = pgdat,
2531 .priority = sc->priority,
2532 };
2533 unsigned long node_lru_pages = 0;
2534 struct mem_cgroup *memcg;
2535
2536 memset(&sc->nr, 0, sizeof(sc->nr));
2537
2538 nr_reclaimed = sc->nr_reclaimed;
2539 nr_scanned = sc->nr_scanned;
2540
2541 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2542 do {
2543 unsigned long lru_pages;
2544 unsigned long reclaimed;
2545 unsigned long scanned;
2546
2547 if (mem_cgroup_low(root, memcg)) {
2548 if (!sc->memcg_low_reclaim) {
2549 sc->memcg_low_skipped = 1;
2550 continue;
2551 }
2552 memcg_memory_event(memcg, MEMCG_LOW);
2553 }
2554
2555 reclaimed = sc->nr_reclaimed;
2556 scanned = sc->nr_scanned;
2557 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2558 node_lru_pages += lru_pages;
2559
2560 if (memcg)
2561 shrink_slab(sc->gfp_mask, pgdat->node_id,
2562 memcg, sc->priority);
2563
2564 /* Record the group's reclaim efficiency */
2565 vmpressure(sc->gfp_mask, memcg, false,
2566 sc->nr_scanned - scanned,
2567 sc->nr_reclaimed - reclaimed);
2568
2569 /*
2570 * Direct reclaim and kswapd have to scan all memory
2571 * cgroups to fulfill the overall scan target for the
2572 * node.
2573 *
2574 * Limit reclaim, on the other hand, only cares about
2575 * nr_to_reclaim pages to be reclaimed and it will
2576 * retry with decreasing priority if one round over the
2577 * whole hierarchy is not sufficient.
2578 */
2579 if (!global_reclaim(sc) &&
2580 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2581 mem_cgroup_iter_break(root, memcg);
2582 break;
2583 }
2584 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2585
2586 if (global_reclaim(sc))
2587 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2588 sc->priority);
2589
2590 if (reclaim_state) {
2591 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2592 reclaim_state->reclaimed_slab = 0;
2593 }
2594
2595 /* Record the subtree's reclaim efficiency */
2596 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2597 sc->nr_scanned - nr_scanned,
2598 sc->nr_reclaimed - nr_reclaimed);
2599
2600 if (sc->nr_reclaimed - nr_reclaimed)
2601 reclaimable = true;
2602
2603 if (current_is_kswapd()) {
2604 /*
2605 * If reclaim is isolating dirty pages under writeback,
2606 * it implies that the long-lived page allocation rate
2607 * is exceeding the page laundering rate. Either the
2608 * global limits are not being effective at throttling
2609 * processes due to the page distribution throughout
2610 * zones or there is heavy usage of a slow backing
2611 * device. The only option is to throttle from reclaim
2612 * context which is not ideal as there is no guarantee
2613 * the dirtying process is throttled in the same way
2614 * balance_dirty_pages() manages.
2615 *
2616 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2617 * count the number of pages under pages flagged for
2618 * immediate reclaim and stall if any are encountered
2619 * in the nr_immediate check below.
2620 */
2621 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2622 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2623
2624 /*
2625 * Tag a node as congested if all the dirty pages
2626 * scanned were backed by a congested BDI and
2627 * wait_iff_congested will stall.
2628 */
2629 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2630 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2631
2632 /* Allow kswapd to start writing pages during reclaim.*/
2633 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2634 set_bit(PGDAT_DIRTY, &pgdat->flags);
2635
2636 /*
2637 * If kswapd scans pages marked marked for immediate
2638 * reclaim and under writeback (nr_immediate), it
2639 * implies that pages are cycling through the LRU
2640 * faster than they are written so also forcibly stall.
2641 */
2642 if (sc->nr.immediate)
2643 congestion_wait(BLK_RW_ASYNC, HZ/10);
2644 }
2645
2646 /*
2647 * Legacy memcg will stall in page writeback so avoid forcibly
2648 * stalling in wait_iff_congested().
2649 */
2650 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2651 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2652 set_memcg_congestion(pgdat, root, true);
2653
2654 /*
2655 * Stall direct reclaim for IO completions if underlying BDIs
2656 * and node is congested. Allow kswapd to continue until it
2657 * starts encountering unqueued dirty pages or cycling through
2658 * the LRU too quickly.
2659 */
2660 if (!sc->hibernation_mode && !current_is_kswapd() &&
2661 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2662 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2663
2664 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2665 sc->nr_scanned - nr_scanned, sc));
2666
2667 /*
2668 * Kswapd gives up on balancing particular nodes after too
2669 * many failures to reclaim anything from them and goes to
2670 * sleep. On reclaim progress, reset the failure counter. A
2671 * successful direct reclaim run will revive a dormant kswapd.
2672 */
2673 if (reclaimable)
2674 pgdat->kswapd_failures = 0;
2675
2676 return reclaimable;
2677}
2678
2679/*
2680 * Returns true if compaction should go ahead for a costly-order request, or
2681 * the allocation would already succeed without compaction. Return false if we
2682 * should reclaim first.
2683 */
2684static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2685{
2686 unsigned long watermark;
2687 enum compact_result suitable;
2688
2689 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2690 if (suitable == COMPACT_SUCCESS)
2691 /* Allocation should succeed already. Don't reclaim. */
2692 return true;
2693 if (suitable == COMPACT_SKIPPED)
2694 /* Compaction cannot yet proceed. Do reclaim. */
2695 return false;
2696
2697 /*
2698 * Compaction is already possible, but it takes time to run and there
2699 * are potentially other callers using the pages just freed. So proceed
2700 * with reclaim to make a buffer of free pages available to give
2701 * compaction a reasonable chance of completing and allocating the page.
2702 * Note that we won't actually reclaim the whole buffer in one attempt
2703 * as the target watermark in should_continue_reclaim() is lower. But if
2704 * we are already above the high+gap watermark, don't reclaim at all.
2705 */
2706 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2707
2708 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2709}
2710
2711/*
2712 * This is the direct reclaim path, for page-allocating processes. We only
2713 * try to reclaim pages from zones which will satisfy the caller's allocation
2714 * request.
2715 *
2716 * If a zone is deemed to be full of pinned pages then just give it a light
2717 * scan then give up on it.
2718 */
2719static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2720{
2721 struct zoneref *z;
2722 struct zone *zone;
2723 unsigned long nr_soft_reclaimed;
2724 unsigned long nr_soft_scanned;
2725 gfp_t orig_mask;
2726 pg_data_t *last_pgdat = NULL;
2727
2728 /*
2729 * If the number of buffer_heads in the machine exceeds the maximum
2730 * allowed level, force direct reclaim to scan the highmem zone as
2731 * highmem pages could be pinning lowmem pages storing buffer_heads
2732 */
2733 orig_mask = sc->gfp_mask;
2734 if (buffer_heads_over_limit) {
2735 sc->gfp_mask |= __GFP_HIGHMEM;
2736 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2737 }
2738
2739 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2740 sc->reclaim_idx, sc->nodemask) {
2741 /*
2742 * Take care memory controller reclaiming has small influence
2743 * to global LRU.
2744 */
2745 if (global_reclaim(sc)) {
2746 if (!cpuset_zone_allowed(zone,
2747 GFP_KERNEL | __GFP_HARDWALL))
2748 continue;
2749
2750 /*
2751 * If we already have plenty of memory free for
2752 * compaction in this zone, don't free any more.
2753 * Even though compaction is invoked for any
2754 * non-zero order, only frequent costly order
2755 * reclamation is disruptive enough to become a
2756 * noticeable problem, like transparent huge
2757 * page allocations.
2758 */
2759 if (IS_ENABLED(CONFIG_COMPACTION) &&
2760 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2761 compaction_ready(zone, sc)) {
2762 sc->compaction_ready = true;
2763 continue;
2764 }
2765
2766 /*
2767 * Shrink each node in the zonelist once. If the
2768 * zonelist is ordered by zone (not the default) then a
2769 * node may be shrunk multiple times but in that case
2770 * the user prefers lower zones being preserved.
2771 */
2772 if (zone->zone_pgdat == last_pgdat)
2773 continue;
2774
2775 /*
2776 * This steals pages from memory cgroups over softlimit
2777 * and returns the number of reclaimed pages and
2778 * scanned pages. This works for global memory pressure
2779 * and balancing, not for a memcg's limit.
2780 */
2781 nr_soft_scanned = 0;
2782 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2783 sc->order, sc->gfp_mask,
2784 &nr_soft_scanned);
2785 sc->nr_reclaimed += nr_soft_reclaimed;
2786 sc->nr_scanned += nr_soft_scanned;
2787 /* need some check for avoid more shrink_zone() */
2788 }
2789
2790 /* See comment about same check for global reclaim above */
2791 if (zone->zone_pgdat == last_pgdat)
2792 continue;
2793 last_pgdat = zone->zone_pgdat;
2794 shrink_node(zone->zone_pgdat, sc);
2795 }
2796
2797 /*
2798 * Restore to original mask to avoid the impact on the caller if we
2799 * promoted it to __GFP_HIGHMEM.
2800 */
2801 sc->gfp_mask = orig_mask;
2802}
2803
2804static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2805{
2806 struct mem_cgroup *memcg;
2807
2808 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2809 do {
2810 unsigned long refaults;
2811 struct lruvec *lruvec;
2812
2813 if (memcg)
2814 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2815 else
2816 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2817
2818 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2819 lruvec->refaults = refaults;
2820 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2821}
2822
2823/*
2824 * This is the main entry point to direct page reclaim.
2825 *
2826 * If a full scan of the inactive list fails to free enough memory then we
2827 * are "out of memory" and something needs to be killed.
2828 *
2829 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2830 * high - the zone may be full of dirty or under-writeback pages, which this
2831 * caller can't do much about. We kick the writeback threads and take explicit
2832 * naps in the hope that some of these pages can be written. But if the
2833 * allocating task holds filesystem locks which prevent writeout this might not
2834 * work, and the allocation attempt will fail.
2835 *
2836 * returns: 0, if no pages reclaimed
2837 * else, the number of pages reclaimed
2838 */
2839static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2840 struct scan_control *sc)
2841{
2842 int initial_priority = sc->priority;
2843 pg_data_t *last_pgdat;
2844 struct zoneref *z;
2845 struct zone *zone;
2846retry:
2847 delayacct_freepages_start();
2848
2849 if (global_reclaim(sc))
2850 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2851
2852 do {
2853 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2854 sc->priority);
2855 sc->nr_scanned = 0;
2856 shrink_zones(zonelist, sc);
2857
2858 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2859 break;
2860
2861 if (sc->compaction_ready)
2862 break;
2863
2864 /*
2865 * If we're getting trouble reclaiming, start doing
2866 * writepage even in laptop mode.
2867 */
2868 if (sc->priority < DEF_PRIORITY - 2)
2869 sc->may_writepage = 1;
2870 } while (--sc->priority >= 0);
2871
2872 last_pgdat = NULL;
2873 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2874 sc->nodemask) {
2875 if (zone->zone_pgdat == last_pgdat)
2876 continue;
2877 last_pgdat = zone->zone_pgdat;
2878 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2879 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
2880 }
2881
2882 delayacct_freepages_end();
2883
2884 if (sc->nr_reclaimed)
2885 return sc->nr_reclaimed;
2886
2887 /* Aborted reclaim to try compaction? don't OOM, then */
2888 if (sc->compaction_ready)
2889 return 1;
2890
2891 /* Untapped cgroup reserves? Don't OOM, retry. */
2892 if (sc->memcg_low_skipped) {
2893 sc->priority = initial_priority;
2894 sc->memcg_low_reclaim = 1;
2895 sc->memcg_low_skipped = 0;
2896 goto retry;
2897 }
2898
2899 return 0;
2900}
2901
2902static bool allow_direct_reclaim(pg_data_t *pgdat)
2903{
2904 struct zone *zone;
2905 unsigned long pfmemalloc_reserve = 0;
2906 unsigned long free_pages = 0;
2907 int i;
2908 bool wmark_ok;
2909
2910 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2911 return true;
2912
2913 for (i = 0; i <= ZONE_NORMAL; i++) {
2914 zone = &pgdat->node_zones[i];
2915 if (!managed_zone(zone))
2916 continue;
2917
2918 if (!zone_reclaimable_pages(zone))
2919 continue;
2920
2921 pfmemalloc_reserve += min_wmark_pages(zone);
2922 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2923 }
2924
2925 /* If there are no reserves (unexpected config) then do not throttle */
2926 if (!pfmemalloc_reserve)
2927 return true;
2928
2929 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2930
2931 /* kswapd must be awake if processes are being throttled */
2932 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2933 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2934 (enum zone_type)ZONE_NORMAL);
2935 wake_up_interruptible(&pgdat->kswapd_wait);
2936 }
2937
2938 return wmark_ok;
2939}
2940
2941/*
2942 * Throttle direct reclaimers if backing storage is backed by the network
2943 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2944 * depleted. kswapd will continue to make progress and wake the processes
2945 * when the low watermark is reached.
2946 *
2947 * Returns true if a fatal signal was delivered during throttling. If this
2948 * happens, the page allocator should not consider triggering the OOM killer.
2949 */
2950static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2951 nodemask_t *nodemask)
2952{
2953 struct zoneref *z;
2954 struct zone *zone;
2955 pg_data_t *pgdat = NULL;
2956
2957 /*
2958 * Kernel threads should not be throttled as they may be indirectly
2959 * responsible for cleaning pages necessary for reclaim to make forward
2960 * progress. kjournald for example may enter direct reclaim while
2961 * committing a transaction where throttling it could forcing other
2962 * processes to block on log_wait_commit().
2963 */
2964 if (current->flags & PF_KTHREAD)
2965 goto out;
2966
2967 /*
2968 * If a fatal signal is pending, this process should not throttle.
2969 * It should return quickly so it can exit and free its memory
2970 */
2971 if (fatal_signal_pending(current))
2972 goto out;
2973
2974 /*
2975 * Check if the pfmemalloc reserves are ok by finding the first node
2976 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2977 * GFP_KERNEL will be required for allocating network buffers when
2978 * swapping over the network so ZONE_HIGHMEM is unusable.
2979 *
2980 * Throttling is based on the first usable node and throttled processes
2981 * wait on a queue until kswapd makes progress and wakes them. There
2982 * is an affinity then between processes waking up and where reclaim
2983 * progress has been made assuming the process wakes on the same node.
2984 * More importantly, processes running on remote nodes will not compete
2985 * for remote pfmemalloc reserves and processes on different nodes
2986 * should make reasonable progress.
2987 */
2988 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2989 gfp_zone(gfp_mask), nodemask) {
2990 if (zone_idx(zone) > ZONE_NORMAL)
2991 continue;
2992
2993 /* Throttle based on the first usable node */
2994 pgdat = zone->zone_pgdat;
2995 if (allow_direct_reclaim(pgdat))
2996 goto out;
2997 break;
2998 }
2999
3000 /* If no zone was usable by the allocation flags then do not throttle */
3001 if (!pgdat)
3002 goto out;
3003
3004 /* Account for the throttling */
3005 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3006
3007 /*
3008 * If the caller cannot enter the filesystem, it's possible that it
3009 * is due to the caller holding an FS lock or performing a journal
3010 * transaction in the case of a filesystem like ext[3|4]. In this case,
3011 * it is not safe to block on pfmemalloc_wait as kswapd could be
3012 * blocked waiting on the same lock. Instead, throttle for up to a
3013 * second before continuing.
3014 */
3015 if (!(gfp_mask & __GFP_FS)) {
3016 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3017 allow_direct_reclaim(pgdat), HZ);
3018
3019 goto check_pending;
3020 }
3021
3022 /* Throttle until kswapd wakes the process */
3023 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3024 allow_direct_reclaim(pgdat));
3025
3026check_pending:
3027 if (fatal_signal_pending(current))
3028 return true;
3029
3030out:
3031 return false;
3032}
3033
3034unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3035 gfp_t gfp_mask, nodemask_t *nodemask)
3036{
3037 unsigned long nr_reclaimed;
3038 struct scan_control sc = {
3039 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3040 .gfp_mask = current_gfp_context(gfp_mask),
3041 .reclaim_idx = gfp_zone(gfp_mask),
3042 .order = order,
3043 .nodemask = nodemask,
3044 .priority = DEF_PRIORITY,
3045 .may_writepage = !laptop_mode,
3046 .may_unmap = 1,
3047 .may_swap = 1,
3048 };
3049
3050 /*
3051 * Do not enter reclaim if fatal signal was delivered while throttled.
3052 * 1 is returned so that the page allocator does not OOM kill at this
3053 * point.
3054 */
3055 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3056 return 1;
3057
3058 trace_mm_vmscan_direct_reclaim_begin(order,
3059 sc.may_writepage,
3060 sc.gfp_mask,
3061 sc.reclaim_idx);
3062
3063 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3064
3065 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3066
3067 return nr_reclaimed;
3068}
3069
3070#ifdef CONFIG_MEMCG
3071
3072unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3073 gfp_t gfp_mask, bool noswap,
3074 pg_data_t *pgdat,
3075 unsigned long *nr_scanned)
3076{
3077 struct scan_control sc = {
3078 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3079 .target_mem_cgroup = memcg,
3080 .may_writepage = !laptop_mode,
3081 .may_unmap = 1,
3082 .reclaim_idx = MAX_NR_ZONES - 1,
3083 .may_swap = !noswap,
3084 };
3085 unsigned long lru_pages;
3086
3087 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3088 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3089
3090 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3091 sc.may_writepage,
3092 sc.gfp_mask,
3093 sc.reclaim_idx);
3094
3095 /*
3096 * NOTE: Although we can get the priority field, using it
3097 * here is not a good idea, since it limits the pages we can scan.
3098 * if we don't reclaim here, the shrink_node from balance_pgdat
3099 * will pick up pages from other mem cgroup's as well. We hack
3100 * the priority and make it zero.
3101 */
3102 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3103
3104 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3105
3106 *nr_scanned = sc.nr_scanned;
3107 return sc.nr_reclaimed;
3108}
3109
3110unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3111 unsigned long nr_pages,
3112 gfp_t gfp_mask,
3113 bool may_swap)
3114{
3115 struct zonelist *zonelist;
3116 unsigned long nr_reclaimed;
3117 int nid;
3118 unsigned int noreclaim_flag;
3119 struct scan_control sc = {
3120 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3121 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3122 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3123 .reclaim_idx = MAX_NR_ZONES - 1,
3124 .target_mem_cgroup = memcg,
3125 .priority = DEF_PRIORITY,
3126 .may_writepage = !laptop_mode,
3127 .may_unmap = 1,
3128 .may_swap = may_swap,
3129 };
3130
3131 /*
3132 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3133 * take care of from where we get pages. So the node where we start the
3134 * scan does not need to be the current node.
3135 */
3136 nid = mem_cgroup_select_victim_node(memcg);
3137
3138 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3139
3140 trace_mm_vmscan_memcg_reclaim_begin(0,
3141 sc.may_writepage,
3142 sc.gfp_mask,
3143 sc.reclaim_idx);
3144
3145 noreclaim_flag = memalloc_noreclaim_save();
3146 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3147 memalloc_noreclaim_restore(noreclaim_flag);
3148
3149 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3150
3151 return nr_reclaimed;
3152}
3153#endif
3154
3155static void age_active_anon(struct pglist_data *pgdat,
3156 struct scan_control *sc)
3157{
3158 struct mem_cgroup *memcg;
3159
3160 if (!total_swap_pages)
3161 return;
3162
3163 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3164 do {
3165 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3166
3167 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3168 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3169 sc, LRU_ACTIVE_ANON);
3170
3171 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3172 } while (memcg);
3173}
3174
3175/*
3176 * Returns true if there is an eligible zone balanced for the request order
3177 * and classzone_idx
3178 */
3179static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3180{
3181 int i;
3182 unsigned long mark = -1;
3183 struct zone *zone;
3184
3185 for (i = 0; i <= classzone_idx; i++) {
3186 zone = pgdat->node_zones + i;
3187
3188 if (!managed_zone(zone))
3189 continue;
3190
3191 mark = high_wmark_pages(zone);
3192 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3193 return true;
3194 }
3195
3196 /*
3197 * If a node has no populated zone within classzone_idx, it does not
3198 * need balancing by definition. This can happen if a zone-restricted
3199 * allocation tries to wake a remote kswapd.
3200 */
3201 if (mark == -1)
3202 return true;
3203
3204 return false;
3205}
3206
3207/* Clear pgdat state for congested, dirty or under writeback. */
3208static void clear_pgdat_congested(pg_data_t *pgdat)
3209{
3210 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3211 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3212 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3213}
3214
3215/*
3216 * Prepare kswapd for sleeping. This verifies that there are no processes
3217 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3218 *
3219 * Returns true if kswapd is ready to sleep
3220 */
3221static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3222{
3223 /*
3224 * The throttled processes are normally woken up in balance_pgdat() as
3225 * soon as allow_direct_reclaim() is true. But there is a potential
3226 * race between when kswapd checks the watermarks and a process gets
3227 * throttled. There is also a potential race if processes get
3228 * throttled, kswapd wakes, a large process exits thereby balancing the
3229 * zones, which causes kswapd to exit balance_pgdat() before reaching
3230 * the wake up checks. If kswapd is going to sleep, no process should
3231 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3232 * the wake up is premature, processes will wake kswapd and get
3233 * throttled again. The difference from wake ups in balance_pgdat() is
3234 * that here we are under prepare_to_wait().
3235 */
3236 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3237 wake_up_all(&pgdat->pfmemalloc_wait);
3238
3239 /* Hopeless node, leave it to direct reclaim */
3240 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3241 return true;
3242
3243 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3244 clear_pgdat_congested(pgdat);
3245 return true;
3246 }
3247
3248 return false;
3249}
3250
3251/*
3252 * kswapd shrinks a node of pages that are at or below the highest usable
3253 * zone that is currently unbalanced.
3254 *
3255 * Returns true if kswapd scanned at least the requested number of pages to
3256 * reclaim or if the lack of progress was due to pages under writeback.
3257 * This is used to determine if the scanning priority needs to be raised.
3258 */
3259static bool kswapd_shrink_node(pg_data_t *pgdat,
3260 struct scan_control *sc)
3261{
3262 struct zone *zone;
3263 int z;
3264
3265 /* Reclaim a number of pages proportional to the number of zones */
3266 sc->nr_to_reclaim = 0;
3267 for (z = 0; z <= sc->reclaim_idx; z++) {
3268 zone = pgdat->node_zones + z;
3269 if (!managed_zone(zone))
3270 continue;
3271
3272 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3273 }
3274
3275 /*
3276 * Historically care was taken to put equal pressure on all zones but
3277 * now pressure is applied based on node LRU order.
3278 */
3279 shrink_node(pgdat, sc);
3280
3281 /*
3282 * Fragmentation may mean that the system cannot be rebalanced for
3283 * high-order allocations. If twice the allocation size has been
3284 * reclaimed then recheck watermarks only at order-0 to prevent
3285 * excessive reclaim. Assume that a process requested a high-order
3286 * can direct reclaim/compact.
3287 */
3288 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3289 sc->order = 0;
3290
3291 return sc->nr_scanned >= sc->nr_to_reclaim;
3292}
3293
3294/*
3295 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3296 * that are eligible for use by the caller until at least one zone is
3297 * balanced.
3298 *
3299 * Returns the order kswapd finished reclaiming at.
3300 *
3301 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3302 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3303 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3304 * or lower is eligible for reclaim until at least one usable zone is
3305 * balanced.
3306 */
3307static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3308{
3309 int i;
3310 unsigned long nr_soft_reclaimed;
3311 unsigned long nr_soft_scanned;
3312 struct zone *zone;
3313 struct scan_control sc = {
3314 .gfp_mask = GFP_KERNEL,
3315 .order = order,
3316 .priority = DEF_PRIORITY,
3317 .may_writepage = !laptop_mode,
3318 .may_unmap = 1,
3319 .may_swap = 1,
3320 };
3321 count_vm_event(PAGEOUTRUN);
3322
3323 do {
3324 unsigned long nr_reclaimed = sc.nr_reclaimed;
3325 bool raise_priority = true;
3326
3327 sc.reclaim_idx = classzone_idx;
3328
3329 /*
3330 * If the number of buffer_heads exceeds the maximum allowed
3331 * then consider reclaiming from all zones. This has a dual
3332 * purpose -- on 64-bit systems it is expected that
3333 * buffer_heads are stripped during active rotation. On 32-bit
3334 * systems, highmem pages can pin lowmem memory and shrinking
3335 * buffers can relieve lowmem pressure. Reclaim may still not
3336 * go ahead if all eligible zones for the original allocation
3337 * request are balanced to avoid excessive reclaim from kswapd.
3338 */
3339 if (buffer_heads_over_limit) {
3340 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3341 zone = pgdat->node_zones + i;
3342 if (!managed_zone(zone))
3343 continue;
3344
3345 sc.reclaim_idx = i;
3346 break;
3347 }
3348 }
3349
3350 /*
3351 * Only reclaim if there are no eligible zones. Note that
3352 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3353 * have adjusted it.
3354 */
3355 if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3356 goto out;
3357
3358 /*
3359 * Do some background aging of the anon list, to give
3360 * pages a chance to be referenced before reclaiming. All
3361 * pages are rotated regardless of classzone as this is
3362 * about consistent aging.
3363 */
3364 age_active_anon(pgdat, &sc);
3365
3366 /*
3367 * If we're getting trouble reclaiming, start doing writepage
3368 * even in laptop mode.
3369 */
3370 if (sc.priority < DEF_PRIORITY - 2)
3371 sc.may_writepage = 1;
3372
3373 /* Call soft limit reclaim before calling shrink_node. */
3374 sc.nr_scanned = 0;
3375 nr_soft_scanned = 0;
3376 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3377 sc.gfp_mask, &nr_soft_scanned);
3378 sc.nr_reclaimed += nr_soft_reclaimed;
3379
3380 /*
3381 * There should be no need to raise the scanning priority if
3382 * enough pages are already being scanned that that high
3383 * watermark would be met at 100% efficiency.
3384 */
3385 if (kswapd_shrink_node(pgdat, &sc))
3386 raise_priority = false;
3387
3388 /*
3389 * If the low watermark is met there is no need for processes
3390 * to be throttled on pfmemalloc_wait as they should not be
3391 * able to safely make forward progress. Wake them
3392 */
3393 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3394 allow_direct_reclaim(pgdat))
3395 wake_up_all(&pgdat->pfmemalloc_wait);
3396
3397 /* Check if kswapd should be suspending */
3398 if (try_to_freeze() || kthread_should_stop())
3399 break;
3400
3401 /*
3402 * Raise priority if scanning rate is too low or there was no
3403 * progress in reclaiming pages
3404 */
3405 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3406 if (raise_priority || !nr_reclaimed)
3407 sc.priority--;
3408 } while (sc.priority >= 1);
3409
3410 if (!sc.nr_reclaimed)
3411 pgdat->kswapd_failures++;
3412
3413out:
3414 snapshot_refaults(NULL, pgdat);
3415 /*
3416 * Return the order kswapd stopped reclaiming at as
3417 * prepare_kswapd_sleep() takes it into account. If another caller
3418 * entered the allocator slow path while kswapd was awake, order will
3419 * remain at the higher level.
3420 */
3421 return sc.order;
3422}
3423
3424/*
3425 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3426 * allocation request woke kswapd for. When kswapd has not woken recently,
3427 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3428 * given classzone and returns it or the highest classzone index kswapd
3429 * was recently woke for.
3430 */
3431static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3432 enum zone_type classzone_idx)
3433{
3434 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3435 return classzone_idx;
3436
3437 return max(pgdat->kswapd_classzone_idx, classzone_idx);
3438}
3439
3440static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3441 unsigned int classzone_idx)
3442{
3443 long remaining = 0;
3444 DEFINE_WAIT(wait);
3445
3446 if (freezing(current) || kthread_should_stop())
3447 return;
3448
3449 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3450
3451 /*
3452 * Try to sleep for a short interval. Note that kcompactd will only be
3453 * woken if it is possible to sleep for a short interval. This is
3454 * deliberate on the assumption that if reclaim cannot keep an
3455 * eligible zone balanced that it's also unlikely that compaction will
3456 * succeed.
3457 */
3458 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3459 /*
3460 * Compaction records what page blocks it recently failed to
3461 * isolate pages from and skips them in the future scanning.
3462 * When kswapd is going to sleep, it is reasonable to assume
3463 * that pages and compaction may succeed so reset the cache.
3464 */
3465 reset_isolation_suitable(pgdat);
3466
3467 /*
3468 * We have freed the memory, now we should compact it to make
3469 * allocation of the requested order possible.
3470 */
3471 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3472
3473 remaining = schedule_timeout(HZ/10);
3474
3475 /*
3476 * If woken prematurely then reset kswapd_classzone_idx and
3477 * order. The values will either be from a wakeup request or
3478 * the previous request that slept prematurely.
3479 */
3480 if (remaining) {
3481 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3482 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3483 }
3484
3485 finish_wait(&pgdat->kswapd_wait, &wait);
3486 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3487 }
3488
3489 /*
3490 * After a short sleep, check if it was a premature sleep. If not, then
3491 * go fully to sleep until explicitly woken up.
3492 */
3493 if (!remaining &&
3494 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3495 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3496
3497 /*
3498 * vmstat counters are not perfectly accurate and the estimated
3499 * value for counters such as NR_FREE_PAGES can deviate from the
3500 * true value by nr_online_cpus * threshold. To avoid the zone
3501 * watermarks being breached while under pressure, we reduce the
3502 * per-cpu vmstat threshold while kswapd is awake and restore
3503 * them before going back to sleep.
3504 */
3505 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3506
3507 if (!kthread_should_stop())
3508 schedule();
3509
3510 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3511 } else {
3512 if (remaining)
3513 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3514 else
3515 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3516 }
3517 finish_wait(&pgdat->kswapd_wait, &wait);
3518}
3519
3520/*
3521 * The background pageout daemon, started as a kernel thread
3522 * from the init process.
3523 *
3524 * This basically trickles out pages so that we have _some_
3525 * free memory available even if there is no other activity
3526 * that frees anything up. This is needed for things like routing
3527 * etc, where we otherwise might have all activity going on in
3528 * asynchronous contexts that cannot page things out.
3529 *
3530 * If there are applications that are active memory-allocators
3531 * (most normal use), this basically shouldn't matter.
3532 */
3533static int kswapd(void *p)
3534{
3535 unsigned int alloc_order, reclaim_order;
3536 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3537 pg_data_t *pgdat = (pg_data_t*)p;
3538 struct task_struct *tsk = current;
3539
3540 struct reclaim_state reclaim_state = {
3541 .reclaimed_slab = 0,
3542 };
3543 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3544
3545 if (!cpumask_empty(cpumask))
3546 set_cpus_allowed_ptr(tsk, cpumask);
3547 current->reclaim_state = &reclaim_state;
3548
3549 /*
3550 * Tell the memory management that we're a "memory allocator",
3551 * and that if we need more memory we should get access to it
3552 * regardless (see "__alloc_pages()"). "kswapd" should
3553 * never get caught in the normal page freeing logic.
3554 *
3555 * (Kswapd normally doesn't need memory anyway, but sometimes
3556 * you need a small amount of memory in order to be able to
3557 * page out something else, and this flag essentially protects
3558 * us from recursively trying to free more memory as we're
3559 * trying to free the first piece of memory in the first place).
3560 */
3561 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3562 set_freezable();
3563
3564 pgdat->kswapd_order = 0;
3565 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3566 for ( ; ; ) {
3567 bool ret;
3568
3569 alloc_order = reclaim_order = pgdat->kswapd_order;
3570 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3571
3572kswapd_try_sleep:
3573 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3574 classzone_idx);
3575
3576 /* Read the new order and classzone_idx */
3577 alloc_order = reclaim_order = pgdat->kswapd_order;
3578 classzone_idx = kswapd_classzone_idx(pgdat, 0);
3579 pgdat->kswapd_order = 0;
3580 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3581
3582 ret = try_to_freeze();
3583 if (kthread_should_stop())
3584 break;
3585
3586 /*
3587 * We can speed up thawing tasks if we don't call balance_pgdat
3588 * after returning from the refrigerator
3589 */
3590 if (ret)
3591 continue;
3592
3593 /*
3594 * Reclaim begins at the requested order but if a high-order
3595 * reclaim fails then kswapd falls back to reclaiming for
3596 * order-0. If that happens, kswapd will consider sleeping
3597 * for the order it finished reclaiming at (reclaim_order)
3598 * but kcompactd is woken to compact for the original
3599 * request (alloc_order).
3600 */
3601 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3602 alloc_order);
3603 fs_reclaim_acquire(GFP_KERNEL);
3604 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3605 fs_reclaim_release(GFP_KERNEL);
3606 if (reclaim_order < alloc_order)
3607 goto kswapd_try_sleep;
3608 }
3609
3610 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3611 current->reclaim_state = NULL;
3612
3613 return 0;
3614}
3615
3616/*
3617 * A zone is low on free memory or too fragmented for high-order memory. If
3618 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3619 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3620 * has failed or is not needed, still wake up kcompactd if only compaction is
3621 * needed.
3622 */
3623void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3624 enum zone_type classzone_idx)
3625{
3626 pg_data_t *pgdat;
3627
3628 if (!managed_zone(zone))
3629 return;
3630
3631 if (!cpuset_zone_allowed(zone, gfp_flags))
3632 return;
3633 pgdat = zone->zone_pgdat;
3634 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3635 classzone_idx);
3636 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3637 if (!waitqueue_active(&pgdat->kswapd_wait))
3638 return;
3639
3640 /* Hopeless node, leave it to direct reclaim if possible */
3641 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3642 pgdat_balanced(pgdat, order, classzone_idx)) {
3643 /*
3644 * There may be plenty of free memory available, but it's too
3645 * fragmented for high-order allocations. Wake up kcompactd
3646 * and rely on compaction_suitable() to determine if it's
3647 * needed. If it fails, it will defer subsequent attempts to
3648 * ratelimit its work.
3649 */
3650 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3651 wakeup_kcompactd(pgdat, order, classzone_idx);
3652 return;
3653 }
3654
3655 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3656 gfp_flags);
3657 wake_up_interruptible(&pgdat->kswapd_wait);
3658}
3659
3660#ifdef CONFIG_HIBERNATION
3661/*
3662 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3663 * freed pages.
3664 *
3665 * Rather than trying to age LRUs the aim is to preserve the overall
3666 * LRU order by reclaiming preferentially
3667 * inactive > active > active referenced > active mapped
3668 */
3669unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3670{
3671 struct reclaim_state reclaim_state;
3672 struct scan_control sc = {
3673 .nr_to_reclaim = nr_to_reclaim,
3674 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3675 .reclaim_idx = MAX_NR_ZONES - 1,
3676 .priority = DEF_PRIORITY,
3677 .may_writepage = 1,
3678 .may_unmap = 1,
3679 .may_swap = 1,
3680 .hibernation_mode = 1,
3681 };
3682 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3683 struct task_struct *p = current;
3684 unsigned long nr_reclaimed;
3685 unsigned int noreclaim_flag;
3686
3687 noreclaim_flag = memalloc_noreclaim_save();
3688 fs_reclaim_acquire(sc.gfp_mask);
3689 reclaim_state.reclaimed_slab = 0;
3690 p->reclaim_state = &reclaim_state;
3691
3692 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3693
3694 p->reclaim_state = NULL;
3695 fs_reclaim_release(sc.gfp_mask);
3696 memalloc_noreclaim_restore(noreclaim_flag);
3697
3698 return nr_reclaimed;
3699}
3700#endif /* CONFIG_HIBERNATION */
3701
3702/* It's optimal to keep kswapds on the same CPUs as their memory, but
3703 not required for correctness. So if the last cpu in a node goes
3704 away, we get changed to run anywhere: as the first one comes back,
3705 restore their cpu bindings. */
3706static int kswapd_cpu_online(unsigned int cpu)
3707{
3708 int nid;
3709
3710 for_each_node_state(nid, N_MEMORY) {
3711 pg_data_t *pgdat = NODE_DATA(nid);
3712 const struct cpumask *mask;
3713
3714 mask = cpumask_of_node(pgdat->node_id);
3715
3716 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3717 /* One of our CPUs online: restore mask */
3718 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3719 }
3720 return 0;
3721}
3722
3723/*
3724 * This kswapd start function will be called by init and node-hot-add.
3725 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3726 */
3727int kswapd_run(int nid)
3728{
3729 pg_data_t *pgdat = NODE_DATA(nid);
3730 int ret = 0;
3731
3732 if (pgdat->kswapd)
3733 return 0;
3734
3735 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3736 if (IS_ERR(pgdat->kswapd)) {
3737 /* failure at boot is fatal */
3738 BUG_ON(system_state < SYSTEM_RUNNING);
3739 pr_err("Failed to start kswapd on node %d\n", nid);
3740 ret = PTR_ERR(pgdat->kswapd);
3741 pgdat->kswapd = NULL;
3742 }
3743 return ret;
3744}
3745
3746/*
3747 * Called by memory hotplug when all memory in a node is offlined. Caller must
3748 * hold mem_hotplug_begin/end().
3749 */
3750void kswapd_stop(int nid)
3751{
3752 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3753
3754 if (kswapd) {
3755 kthread_stop(kswapd);
3756 NODE_DATA(nid)->kswapd = NULL;
3757 }
3758}
3759
3760static int __init kswapd_init(void)
3761{
3762 int nid, ret;
3763
3764 swap_setup();
3765 for_each_node_state(nid, N_MEMORY)
3766 kswapd_run(nid);
3767 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3768 "mm/vmscan:online", kswapd_cpu_online,
3769 NULL);
3770 WARN_ON(ret < 0);
3771 return 0;
3772}
3773
3774module_init(kswapd_init)
3775
3776#ifdef CONFIG_NUMA
3777/*
3778 * Node reclaim mode
3779 *
3780 * If non-zero call node_reclaim when the number of free pages falls below
3781 * the watermarks.
3782 */
3783int node_reclaim_mode __read_mostly;
3784
3785#define RECLAIM_OFF 0
3786#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3787#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3788#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3789
3790/*
3791 * Priority for NODE_RECLAIM. This determines the fraction of pages
3792 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3793 * a zone.
3794 */
3795#define NODE_RECLAIM_PRIORITY 4
3796
3797/*
3798 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3799 * occur.
3800 */
3801int sysctl_min_unmapped_ratio = 1;
3802
3803/*
3804 * If the number of slab pages in a zone grows beyond this percentage then
3805 * slab reclaim needs to occur.
3806 */
3807int sysctl_min_slab_ratio = 5;
3808
3809static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3810{
3811 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3812 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3813 node_page_state(pgdat, NR_ACTIVE_FILE);
3814
3815 /*
3816 * It's possible for there to be more file mapped pages than
3817 * accounted for by the pages on the file LRU lists because
3818 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3819 */
3820 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3821}
3822
3823/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3824static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3825{
3826 unsigned long nr_pagecache_reclaimable;
3827 unsigned long delta = 0;
3828
3829 /*
3830 * If RECLAIM_UNMAP is set, then all file pages are considered
3831 * potentially reclaimable. Otherwise, we have to worry about
3832 * pages like swapcache and node_unmapped_file_pages() provides
3833 * a better estimate
3834 */
3835 if (node_reclaim_mode & RECLAIM_UNMAP)
3836 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3837 else
3838 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3839
3840 /* If we can't clean pages, remove dirty pages from consideration */
3841 if (!(node_reclaim_mode & RECLAIM_WRITE))
3842 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3843
3844 /* Watch for any possible underflows due to delta */
3845 if (unlikely(delta > nr_pagecache_reclaimable))
3846 delta = nr_pagecache_reclaimable;
3847
3848 return nr_pagecache_reclaimable - delta;
3849}
3850
3851/*
3852 * Try to free up some pages from this node through reclaim.
3853 */
3854static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3855{
3856 /* Minimum pages needed in order to stay on node */
3857 const unsigned long nr_pages = 1 << order;
3858 struct task_struct *p = current;
3859 struct reclaim_state reclaim_state;
3860 unsigned int noreclaim_flag;
3861 struct scan_control sc = {
3862 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3863 .gfp_mask = current_gfp_context(gfp_mask),
3864 .order = order,
3865 .priority = NODE_RECLAIM_PRIORITY,
3866 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3867 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3868 .may_swap = 1,
3869 .reclaim_idx = gfp_zone(gfp_mask),
3870 };
3871
3872 cond_resched();
3873 /*
3874 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3875 * and we also need to be able to write out pages for RECLAIM_WRITE
3876 * and RECLAIM_UNMAP.
3877 */
3878 noreclaim_flag = memalloc_noreclaim_save();
3879 p->flags |= PF_SWAPWRITE;
3880 fs_reclaim_acquire(sc.gfp_mask);
3881 reclaim_state.reclaimed_slab = 0;
3882 p->reclaim_state = &reclaim_state;
3883
3884 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3885 /*
3886 * Free memory by calling shrink node with increasing
3887 * priorities until we have enough memory freed.
3888 */
3889 do {
3890 shrink_node(pgdat, &sc);
3891 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3892 }
3893
3894 p->reclaim_state = NULL;
3895 fs_reclaim_release(gfp_mask);
3896 current->flags &= ~PF_SWAPWRITE;
3897 memalloc_noreclaim_restore(noreclaim_flag);
3898 return sc.nr_reclaimed >= nr_pages;
3899}
3900
3901int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3902{
3903 int ret;
3904
3905 /*
3906 * Node reclaim reclaims unmapped file backed pages and
3907 * slab pages if we are over the defined limits.
3908 *
3909 * A small portion of unmapped file backed pages is needed for
3910 * file I/O otherwise pages read by file I/O will be immediately
3911 * thrown out if the node is overallocated. So we do not reclaim
3912 * if less than a specified percentage of the node is used by
3913 * unmapped file backed pages.
3914 */
3915 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3916 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3917 return NODE_RECLAIM_FULL;
3918
3919 /*
3920 * Do not scan if the allocation should not be delayed.
3921 */
3922 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3923 return NODE_RECLAIM_NOSCAN;
3924
3925 /*
3926 * Only run node reclaim on the local node or on nodes that do not
3927 * have associated processors. This will favor the local processor
3928 * over remote processors and spread off node memory allocations
3929 * as wide as possible.
3930 */
3931 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3932 return NODE_RECLAIM_NOSCAN;
3933
3934 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3935 return NODE_RECLAIM_NOSCAN;
3936
3937 ret = __node_reclaim(pgdat, gfp_mask, order);
3938 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3939
3940 if (!ret)
3941 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3942
3943 return ret;
3944}
3945#endif
3946
3947/*
3948 * page_evictable - test whether a page is evictable
3949 * @page: the page to test
3950 *
3951 * Test whether page is evictable--i.e., should be placed on active/inactive
3952 * lists vs unevictable list.
3953 *
3954 * Reasons page might not be evictable:
3955 * (1) page's mapping marked unevictable
3956 * (2) page is part of an mlocked VMA
3957 *
3958 */
3959int page_evictable(struct page *page)
3960{
3961 int ret;
3962
3963 /* Prevent address_space of inode and swap cache from being freed */
3964 rcu_read_lock();
3965 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3966 rcu_read_unlock();
3967 return ret;
3968}
3969
3970#ifdef CONFIG_SHMEM
3971/**
3972 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3973 * @pages: array of pages to check
3974 * @nr_pages: number of pages to check
3975 *
3976 * Checks pages for evictability and moves them to the appropriate lru list.
3977 *
3978 * This function is only used for SysV IPC SHM_UNLOCK.
3979 */
3980void check_move_unevictable_pages(struct page **pages, int nr_pages)
3981{
3982 struct lruvec *lruvec;
3983 struct pglist_data *pgdat = NULL;
3984 int pgscanned = 0;
3985 int pgrescued = 0;
3986 int i;
3987
3988 for (i = 0; i < nr_pages; i++) {
3989 struct page *page = pages[i];
3990 struct pglist_data *pagepgdat = page_pgdat(page);
3991
3992 pgscanned++;
3993 if (pagepgdat != pgdat) {
3994 if (pgdat)
3995 spin_unlock_irq(&pgdat->lru_lock);
3996 pgdat = pagepgdat;
3997 spin_lock_irq(&pgdat->lru_lock);
3998 }
3999 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4000
4001 if (!PageLRU(page) || !PageUnevictable(page))
4002 continue;
4003
4004 if (page_evictable(page)) {
4005 enum lru_list lru = page_lru_base_type(page);
4006
4007 VM_BUG_ON_PAGE(PageActive(page), page);
4008 ClearPageUnevictable(page);
4009 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4010 add_page_to_lru_list(page, lruvec, lru);
4011 pgrescued++;
4012 }
4013 }
4014
4015 if (pgdat) {
4016 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4017 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4018 spin_unlock_irq(&pgdat->lru_lock);
4019 }
4020}
4021#endif /* CONFIG_SHMEM */
1/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
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.
12 */
13
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
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>
50
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
55#include <linux/balloon_compaction.h>
56
57#include "internal.h"
58
59#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
62struct scan_control {
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
66 /* This context's GFP mask */
67 gfp_t gfp_mask;
68
69 /* Allocation order */
70 int order;
71
72 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
77
78 /*
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
83
84 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
87 unsigned int may_writepage:1;
88
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
91
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
94
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
97
98 unsigned int hibernation_mode:1;
99
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
102
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
105
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
108};
109
110#ifdef ARCH_HAS_PREFETCH
111#define prefetch_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124#ifdef ARCH_HAS_PREFETCHW
125#define prefetchw_prev_lru_page(_page, _base, _field) \
126 do { \
127 if ((_page)->lru.prev != _base) { \
128 struct page *prev; \
129 \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
132 } \
133 } while (0)
134#else
135#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136#endif
137
138/*
139 * From 0 .. 100. Higher means more swappy.
140 */
141int vm_swappiness = 60;
142/*
143 * The total number of pages which are beyond the high watermark within all
144 * zones.
145 */
146unsigned long vm_total_pages;
147
148static LIST_HEAD(shrinker_list);
149static DECLARE_RWSEM(shrinker_rwsem);
150
151#ifdef CONFIG_MEMCG
152static bool global_reclaim(struct scan_control *sc)
153{
154 return !sc->target_mem_cgroup;
155}
156
157/**
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
160 *
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
166 *
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
169 */
170static bool sane_reclaim(struct scan_control *sc)
171{
172 struct mem_cgroup *memcg = sc->target_mem_cgroup;
173
174 if (!memcg)
175 return true;
176#ifdef CONFIG_CGROUP_WRITEBACK
177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
178 return true;
179#endif
180 return false;
181}
182#else
183static bool global_reclaim(struct scan_control *sc)
184{
185 return true;
186}
187
188static bool sane_reclaim(struct scan_control *sc)
189{
190 return true;
191}
192#endif
193
194static unsigned long zone_reclaimable_pages(struct zone *zone)
195{
196 unsigned long nr;
197
198 nr = zone_page_state_snapshot(zone, NR_ACTIVE_FILE) +
199 zone_page_state_snapshot(zone, NR_INACTIVE_FILE) +
200 zone_page_state_snapshot(zone, NR_ISOLATED_FILE);
201
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state_snapshot(zone, NR_ACTIVE_ANON) +
204 zone_page_state_snapshot(zone, NR_INACTIVE_ANON) +
205 zone_page_state_snapshot(zone, NR_ISOLATED_ANON);
206
207 return nr;
208}
209
210bool zone_reclaimable(struct zone *zone)
211{
212 return zone_page_state_snapshot(zone, NR_PAGES_SCANNED) <
213 zone_reclaimable_pages(zone) * 6;
214}
215
216unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
217{
218 if (!mem_cgroup_disabled())
219 return mem_cgroup_get_lru_size(lruvec, lru);
220
221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
222}
223
224/*
225 * Add a shrinker callback to be called from the vm.
226 */
227int register_shrinker(struct shrinker *shrinker)
228{
229 size_t size = sizeof(*shrinker->nr_deferred);
230
231 if (shrinker->flags & SHRINKER_NUMA_AWARE)
232 size *= nr_node_ids;
233
234 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
235 if (!shrinker->nr_deferred)
236 return -ENOMEM;
237
238 down_write(&shrinker_rwsem);
239 list_add_tail(&shrinker->list, &shrinker_list);
240 up_write(&shrinker_rwsem);
241 return 0;
242}
243EXPORT_SYMBOL(register_shrinker);
244
245/*
246 * Remove one
247 */
248void unregister_shrinker(struct shrinker *shrinker)
249{
250 down_write(&shrinker_rwsem);
251 list_del(&shrinker->list);
252 up_write(&shrinker_rwsem);
253 kfree(shrinker->nr_deferred);
254}
255EXPORT_SYMBOL(unregister_shrinker);
256
257#define SHRINK_BATCH 128
258
259static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
260 struct shrinker *shrinker,
261 unsigned long nr_scanned,
262 unsigned long nr_eligible)
263{
264 unsigned long freed = 0;
265 unsigned long long delta;
266 long total_scan;
267 long freeable;
268 long nr;
269 long new_nr;
270 int nid = shrinkctl->nid;
271 long batch_size = shrinker->batch ? shrinker->batch
272 : SHRINK_BATCH;
273
274 freeable = shrinker->count_objects(shrinker, shrinkctl);
275 if (freeable == 0)
276 return 0;
277
278 /*
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
282 */
283 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
284
285 total_scan = nr;
286 delta = (4 * nr_scanned) / shrinker->seeks;
287 delta *= freeable;
288 do_div(delta, nr_eligible + 1);
289 total_scan += delta;
290 if (total_scan < 0) {
291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
292 shrinker->scan_objects, total_scan);
293 total_scan = freeable;
294 }
295
296 /*
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
302 * freeable. This is bad for sustaining a working set in
303 * memory.
304 *
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
307 */
308 if (delta < freeable / 4)
309 total_scan = min(total_scan, freeable / 2);
310
311 /*
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
314 * freeable entries.
315 */
316 if (total_scan > freeable * 2)
317 total_scan = freeable * 2;
318
319 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
320 nr_scanned, nr_eligible,
321 freeable, delta, total_scan);
322
323 /*
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
330 * batch_size.
331 *
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
334 * than the total number of objects on slab (freeable), we must be
335 * scanning at high prio and therefore should try to reclaim as much as
336 * possible.
337 */
338 while (total_scan >= batch_size ||
339 total_scan >= freeable) {
340 unsigned long ret;
341 unsigned long nr_to_scan = min(batch_size, total_scan);
342
343 shrinkctl->nr_to_scan = nr_to_scan;
344 ret = shrinker->scan_objects(shrinker, shrinkctl);
345 if (ret == SHRINK_STOP)
346 break;
347 freed += ret;
348
349 count_vm_events(SLABS_SCANNED, nr_to_scan);
350 total_scan -= nr_to_scan;
351
352 cond_resched();
353 }
354
355 /*
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
359 */
360 if (total_scan > 0)
361 new_nr = atomic_long_add_return(total_scan,
362 &shrinker->nr_deferred[nid]);
363 else
364 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
365
366 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
367 return freed;
368}
369
370/**
371 * shrink_slab - shrink slab caches
372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
374 * @memcg: memory cgroup whose slab caches to target
375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
377 *
378 * Call the shrink functions to age shrinkable caches.
379 *
380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
382 *
383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
385 * objects from the memory cgroup specified. Otherwise, only unaware
386 * shrinkers are called.
387 *
388 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
389 * the available objects should be scanned. Page reclaim for example
390 * passes the number of pages scanned and the number of pages on the
391 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
392 * when it encountered mapped pages. The ratio is further biased by
393 * the ->seeks setting of the shrink function, which indicates the
394 * cost to recreate an object relative to that of an LRU page.
395 *
396 * Returns the number of reclaimed slab objects.
397 */
398static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
399 struct mem_cgroup *memcg,
400 unsigned long nr_scanned,
401 unsigned long nr_eligible)
402{
403 struct shrinker *shrinker;
404 unsigned long freed = 0;
405
406 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
407 return 0;
408
409 if (nr_scanned == 0)
410 nr_scanned = SWAP_CLUSTER_MAX;
411
412 if (!down_read_trylock(&shrinker_rwsem)) {
413 /*
414 * If we would return 0, our callers would understand that we
415 * have nothing else to shrink and give up trying. By returning
416 * 1 we keep it going and assume we'll be able to shrink next
417 * time.
418 */
419 freed = 1;
420 goto out;
421 }
422
423 list_for_each_entry(shrinker, &shrinker_list, list) {
424 struct shrink_control sc = {
425 .gfp_mask = gfp_mask,
426 .nid = nid,
427 .memcg = memcg,
428 };
429
430 /*
431 * If kernel memory accounting is disabled, we ignore
432 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
433 * passing NULL for memcg.
434 */
435 if (memcg_kmem_enabled() &&
436 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
437 continue;
438
439 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
440 sc.nid = 0;
441
442 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
443 }
444
445 up_read(&shrinker_rwsem);
446out:
447 cond_resched();
448 return freed;
449}
450
451void drop_slab_node(int nid)
452{
453 unsigned long freed;
454
455 do {
456 struct mem_cgroup *memcg = NULL;
457
458 freed = 0;
459 do {
460 freed += shrink_slab(GFP_KERNEL, nid, memcg,
461 1000, 1000);
462 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
463 } while (freed > 10);
464}
465
466void drop_slab(void)
467{
468 int nid;
469
470 for_each_online_node(nid)
471 drop_slab_node(nid);
472}
473
474static inline int is_page_cache_freeable(struct page *page)
475{
476 /*
477 * A freeable page cache page is referenced only by the caller
478 * that isolated the page, the page cache radix tree and
479 * optional buffer heads at page->private.
480 */
481 return page_count(page) - page_has_private(page) == 2;
482}
483
484static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
485{
486 if (current->flags & PF_SWAPWRITE)
487 return 1;
488 if (!inode_write_congested(inode))
489 return 1;
490 if (inode_to_bdi(inode) == current->backing_dev_info)
491 return 1;
492 return 0;
493}
494
495/*
496 * We detected a synchronous write error writing a page out. Probably
497 * -ENOSPC. We need to propagate that into the address_space for a subsequent
498 * fsync(), msync() or close().
499 *
500 * The tricky part is that after writepage we cannot touch the mapping: nothing
501 * prevents it from being freed up. But we have a ref on the page and once
502 * that page is locked, the mapping is pinned.
503 *
504 * We're allowed to run sleeping lock_page() here because we know the caller has
505 * __GFP_FS.
506 */
507static void handle_write_error(struct address_space *mapping,
508 struct page *page, int error)
509{
510 lock_page(page);
511 if (page_mapping(page) == mapping)
512 mapping_set_error(mapping, error);
513 unlock_page(page);
514}
515
516/* possible outcome of pageout() */
517typedef enum {
518 /* failed to write page out, page is locked */
519 PAGE_KEEP,
520 /* move page to the active list, page is locked */
521 PAGE_ACTIVATE,
522 /* page has been sent to the disk successfully, page is unlocked */
523 PAGE_SUCCESS,
524 /* page is clean and locked */
525 PAGE_CLEAN,
526} pageout_t;
527
528/*
529 * pageout is called by shrink_page_list() for each dirty page.
530 * Calls ->writepage().
531 */
532static pageout_t pageout(struct page *page, struct address_space *mapping,
533 struct scan_control *sc)
534{
535 /*
536 * If the page is dirty, only perform writeback if that write
537 * will be non-blocking. To prevent this allocation from being
538 * stalled by pagecache activity. But note that there may be
539 * stalls if we need to run get_block(). We could test
540 * PagePrivate for that.
541 *
542 * If this process is currently in __generic_file_write_iter() against
543 * this page's queue, we can perform writeback even if that
544 * will block.
545 *
546 * If the page is swapcache, write it back even if that would
547 * block, for some throttling. This happens by accident, because
548 * swap_backing_dev_info is bust: it doesn't reflect the
549 * congestion state of the swapdevs. Easy to fix, if needed.
550 */
551 if (!is_page_cache_freeable(page))
552 return PAGE_KEEP;
553 if (!mapping) {
554 /*
555 * Some data journaling orphaned pages can have
556 * page->mapping == NULL while being dirty with clean buffers.
557 */
558 if (page_has_private(page)) {
559 if (try_to_free_buffers(page)) {
560 ClearPageDirty(page);
561 pr_info("%s: orphaned page\n", __func__);
562 return PAGE_CLEAN;
563 }
564 }
565 return PAGE_KEEP;
566 }
567 if (mapping->a_ops->writepage == NULL)
568 return PAGE_ACTIVATE;
569 if (!may_write_to_inode(mapping->host, sc))
570 return PAGE_KEEP;
571
572 if (clear_page_dirty_for_io(page)) {
573 int res;
574 struct writeback_control wbc = {
575 .sync_mode = WB_SYNC_NONE,
576 .nr_to_write = SWAP_CLUSTER_MAX,
577 .range_start = 0,
578 .range_end = LLONG_MAX,
579 .for_reclaim = 1,
580 };
581
582 SetPageReclaim(page);
583 res = mapping->a_ops->writepage(page, &wbc);
584 if (res < 0)
585 handle_write_error(mapping, page, res);
586 if (res == AOP_WRITEPAGE_ACTIVATE) {
587 ClearPageReclaim(page);
588 return PAGE_ACTIVATE;
589 }
590
591 if (!PageWriteback(page)) {
592 /* synchronous write or broken a_ops? */
593 ClearPageReclaim(page);
594 }
595 trace_mm_vmscan_writepage(page);
596 inc_zone_page_state(page, NR_VMSCAN_WRITE);
597 return PAGE_SUCCESS;
598 }
599
600 return PAGE_CLEAN;
601}
602
603/*
604 * Same as remove_mapping, but if the page is removed from the mapping, it
605 * gets returned with a refcount of 0.
606 */
607static int __remove_mapping(struct address_space *mapping, struct page *page,
608 bool reclaimed)
609{
610 unsigned long flags;
611
612 BUG_ON(!PageLocked(page));
613 BUG_ON(mapping != page_mapping(page));
614
615 spin_lock_irqsave(&mapping->tree_lock, flags);
616 /*
617 * The non racy check for a busy page.
618 *
619 * Must be careful with the order of the tests. When someone has
620 * a ref to the page, it may be possible that they dirty it then
621 * drop the reference. So if PageDirty is tested before page_count
622 * here, then the following race may occur:
623 *
624 * get_user_pages(&page);
625 * [user mapping goes away]
626 * write_to(page);
627 * !PageDirty(page) [good]
628 * SetPageDirty(page);
629 * put_page(page);
630 * !page_count(page) [good, discard it]
631 *
632 * [oops, our write_to data is lost]
633 *
634 * Reversing the order of the tests ensures such a situation cannot
635 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
636 * load is not satisfied before that of page->_count.
637 *
638 * Note that if SetPageDirty is always performed via set_page_dirty,
639 * and thus under tree_lock, then this ordering is not required.
640 */
641 if (!page_ref_freeze(page, 2))
642 goto cannot_free;
643 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
644 if (unlikely(PageDirty(page))) {
645 page_ref_unfreeze(page, 2);
646 goto cannot_free;
647 }
648
649 if (PageSwapCache(page)) {
650 swp_entry_t swap = { .val = page_private(page) };
651 mem_cgroup_swapout(page, swap);
652 __delete_from_swap_cache(page);
653 spin_unlock_irqrestore(&mapping->tree_lock, flags);
654 swapcache_free(swap);
655 } else {
656 void (*freepage)(struct page *);
657 void *shadow = NULL;
658
659 freepage = mapping->a_ops->freepage;
660 /*
661 * Remember a shadow entry for reclaimed file cache in
662 * order to detect refaults, thus thrashing, later on.
663 *
664 * But don't store shadows in an address space that is
665 * already exiting. This is not just an optizimation,
666 * inode reclaim needs to empty out the radix tree or
667 * the nodes are lost. Don't plant shadows behind its
668 * back.
669 *
670 * We also don't store shadows for DAX mappings because the
671 * only page cache pages found in these are zero pages
672 * covering holes, and because we don't want to mix DAX
673 * exceptional entries and shadow exceptional entries in the
674 * same page_tree.
675 */
676 if (reclaimed && page_is_file_cache(page) &&
677 !mapping_exiting(mapping) && !dax_mapping(mapping))
678 shadow = workingset_eviction(mapping, page);
679 __delete_from_page_cache(page, shadow);
680 spin_unlock_irqrestore(&mapping->tree_lock, flags);
681
682 if (freepage != NULL)
683 freepage(page);
684 }
685
686 return 1;
687
688cannot_free:
689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
690 return 0;
691}
692
693/*
694 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
695 * someone else has a ref on the page, abort and return 0. If it was
696 * successfully detached, return 1. Assumes the caller has a single ref on
697 * this page.
698 */
699int remove_mapping(struct address_space *mapping, struct page *page)
700{
701 if (__remove_mapping(mapping, page, false)) {
702 /*
703 * Unfreezing the refcount with 1 rather than 2 effectively
704 * drops the pagecache ref for us without requiring another
705 * atomic operation.
706 */
707 page_ref_unfreeze(page, 1);
708 return 1;
709 }
710 return 0;
711}
712
713/**
714 * putback_lru_page - put previously isolated page onto appropriate LRU list
715 * @page: page to be put back to appropriate lru list
716 *
717 * Add previously isolated @page to appropriate LRU list.
718 * Page may still be unevictable for other reasons.
719 *
720 * lru_lock must not be held, interrupts must be enabled.
721 */
722void putback_lru_page(struct page *page)
723{
724 bool is_unevictable;
725 int was_unevictable = PageUnevictable(page);
726
727 VM_BUG_ON_PAGE(PageLRU(page), page);
728
729redo:
730 ClearPageUnevictable(page);
731
732 if (page_evictable(page)) {
733 /*
734 * For evictable pages, we can use the cache.
735 * In event of a race, worst case is we end up with an
736 * unevictable page on [in]active list.
737 * We know how to handle that.
738 */
739 is_unevictable = false;
740 lru_cache_add(page);
741 } else {
742 /*
743 * Put unevictable pages directly on zone's unevictable
744 * list.
745 */
746 is_unevictable = true;
747 add_page_to_unevictable_list(page);
748 /*
749 * When racing with an mlock or AS_UNEVICTABLE clearing
750 * (page is unlocked) make sure that if the other thread
751 * does not observe our setting of PG_lru and fails
752 * isolation/check_move_unevictable_pages,
753 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
754 * the page back to the evictable list.
755 *
756 * The other side is TestClearPageMlocked() or shmem_lock().
757 */
758 smp_mb();
759 }
760
761 /*
762 * page's status can change while we move it among lru. If an evictable
763 * page is on unevictable list, it never be freed. To avoid that,
764 * check after we added it to the list, again.
765 */
766 if (is_unevictable && page_evictable(page)) {
767 if (!isolate_lru_page(page)) {
768 put_page(page);
769 goto redo;
770 }
771 /* This means someone else dropped this page from LRU
772 * So, it will be freed or putback to LRU again. There is
773 * nothing to do here.
774 */
775 }
776
777 if (was_unevictable && !is_unevictable)
778 count_vm_event(UNEVICTABLE_PGRESCUED);
779 else if (!was_unevictable && is_unevictable)
780 count_vm_event(UNEVICTABLE_PGCULLED);
781
782 put_page(page); /* drop ref from isolate */
783}
784
785enum page_references {
786 PAGEREF_RECLAIM,
787 PAGEREF_RECLAIM_CLEAN,
788 PAGEREF_KEEP,
789 PAGEREF_ACTIVATE,
790};
791
792static enum page_references page_check_references(struct page *page,
793 struct scan_control *sc)
794{
795 int referenced_ptes, referenced_page;
796 unsigned long vm_flags;
797
798 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
799 &vm_flags);
800 referenced_page = TestClearPageReferenced(page);
801
802 /*
803 * Mlock lost the isolation race with us. Let try_to_unmap()
804 * move the page to the unevictable list.
805 */
806 if (vm_flags & VM_LOCKED)
807 return PAGEREF_RECLAIM;
808
809 if (referenced_ptes) {
810 if (PageSwapBacked(page))
811 return PAGEREF_ACTIVATE;
812 /*
813 * All mapped pages start out with page table
814 * references from the instantiating fault, so we need
815 * to look twice if a mapped file page is used more
816 * than once.
817 *
818 * Mark it and spare it for another trip around the
819 * inactive list. Another page table reference will
820 * lead to its activation.
821 *
822 * Note: the mark is set for activated pages as well
823 * so that recently deactivated but used pages are
824 * quickly recovered.
825 */
826 SetPageReferenced(page);
827
828 if (referenced_page || referenced_ptes > 1)
829 return PAGEREF_ACTIVATE;
830
831 /*
832 * Activate file-backed executable pages after first usage.
833 */
834 if (vm_flags & VM_EXEC)
835 return PAGEREF_ACTIVATE;
836
837 return PAGEREF_KEEP;
838 }
839
840 /* Reclaim if clean, defer dirty pages to writeback */
841 if (referenced_page && !PageSwapBacked(page))
842 return PAGEREF_RECLAIM_CLEAN;
843
844 return PAGEREF_RECLAIM;
845}
846
847/* Check if a page is dirty or under writeback */
848static void page_check_dirty_writeback(struct page *page,
849 bool *dirty, bool *writeback)
850{
851 struct address_space *mapping;
852
853 /*
854 * Anonymous pages are not handled by flushers and must be written
855 * from reclaim context. Do not stall reclaim based on them
856 */
857 if (!page_is_file_cache(page)) {
858 *dirty = false;
859 *writeback = false;
860 return;
861 }
862
863 /* By default assume that the page flags are accurate */
864 *dirty = PageDirty(page);
865 *writeback = PageWriteback(page);
866
867 /* Verify dirty/writeback state if the filesystem supports it */
868 if (!page_has_private(page))
869 return;
870
871 mapping = page_mapping(page);
872 if (mapping && mapping->a_ops->is_dirty_writeback)
873 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
874}
875
876/*
877 * shrink_page_list() returns the number of reclaimed pages
878 */
879static unsigned long shrink_page_list(struct list_head *page_list,
880 struct zone *zone,
881 struct scan_control *sc,
882 enum ttu_flags ttu_flags,
883 unsigned long *ret_nr_dirty,
884 unsigned long *ret_nr_unqueued_dirty,
885 unsigned long *ret_nr_congested,
886 unsigned long *ret_nr_writeback,
887 unsigned long *ret_nr_immediate,
888 bool force_reclaim)
889{
890 LIST_HEAD(ret_pages);
891 LIST_HEAD(free_pages);
892 int pgactivate = 0;
893 unsigned long nr_unqueued_dirty = 0;
894 unsigned long nr_dirty = 0;
895 unsigned long nr_congested = 0;
896 unsigned long nr_reclaimed = 0;
897 unsigned long nr_writeback = 0;
898 unsigned long nr_immediate = 0;
899
900 cond_resched();
901
902 while (!list_empty(page_list)) {
903 struct address_space *mapping;
904 struct page *page;
905 int may_enter_fs;
906 enum page_references references = PAGEREF_RECLAIM_CLEAN;
907 bool dirty, writeback;
908 bool lazyfree = false;
909 int ret = SWAP_SUCCESS;
910
911 cond_resched();
912
913 page = lru_to_page(page_list);
914 list_del(&page->lru);
915
916 if (!trylock_page(page))
917 goto keep;
918
919 VM_BUG_ON_PAGE(PageActive(page), page);
920 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
921
922 sc->nr_scanned++;
923
924 if (unlikely(!page_evictable(page)))
925 goto cull_mlocked;
926
927 if (!sc->may_unmap && page_mapped(page))
928 goto keep_locked;
929
930 /* Double the slab pressure for mapped and swapcache pages */
931 if (page_mapped(page) || PageSwapCache(page))
932 sc->nr_scanned++;
933
934 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
935 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
936
937 /*
938 * The number of dirty pages determines if a zone is marked
939 * reclaim_congested which affects wait_iff_congested. kswapd
940 * will stall and start writing pages if the tail of the LRU
941 * is all dirty unqueued pages.
942 */
943 page_check_dirty_writeback(page, &dirty, &writeback);
944 if (dirty || writeback)
945 nr_dirty++;
946
947 if (dirty && !writeback)
948 nr_unqueued_dirty++;
949
950 /*
951 * Treat this page as congested if the underlying BDI is or if
952 * pages are cycling through the LRU so quickly that the
953 * pages marked for immediate reclaim are making it to the
954 * end of the LRU a second time.
955 */
956 mapping = page_mapping(page);
957 if (((dirty || writeback) && mapping &&
958 inode_write_congested(mapping->host)) ||
959 (writeback && PageReclaim(page)))
960 nr_congested++;
961
962 /*
963 * If a page at the tail of the LRU is under writeback, there
964 * are three cases to consider.
965 *
966 * 1) If reclaim is encountering an excessive number of pages
967 * under writeback and this page is both under writeback and
968 * PageReclaim then it indicates that pages are being queued
969 * for IO but are being recycled through the LRU before the
970 * IO can complete. Waiting on the page itself risks an
971 * indefinite stall if it is impossible to writeback the
972 * page due to IO error or disconnected storage so instead
973 * note that the LRU is being scanned too quickly and the
974 * caller can stall after page list has been processed.
975 *
976 * 2) Global or new memcg reclaim encounters a page that is
977 * not marked for immediate reclaim, or the caller does not
978 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
979 * not to fs). In this case mark the page for immediate
980 * reclaim and continue scanning.
981 *
982 * Require may_enter_fs because we would wait on fs, which
983 * may not have submitted IO yet. And the loop driver might
984 * enter reclaim, and deadlock if it waits on a page for
985 * which it is needed to do the write (loop masks off
986 * __GFP_IO|__GFP_FS for this reason); but more thought
987 * would probably show more reasons.
988 *
989 * 3) Legacy memcg encounters a page that is already marked
990 * PageReclaim. memcg does not have any dirty pages
991 * throttling so we could easily OOM just because too many
992 * pages are in writeback and there is nothing else to
993 * reclaim. Wait for the writeback to complete.
994 */
995 if (PageWriteback(page)) {
996 /* Case 1 above */
997 if (current_is_kswapd() &&
998 PageReclaim(page) &&
999 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1000 nr_immediate++;
1001 goto keep_locked;
1002
1003 /* Case 2 above */
1004 } else if (sane_reclaim(sc) ||
1005 !PageReclaim(page) || !may_enter_fs) {
1006 /*
1007 * This is slightly racy - end_page_writeback()
1008 * might have just cleared PageReclaim, then
1009 * setting PageReclaim here end up interpreted
1010 * as PageReadahead - but that does not matter
1011 * enough to care. What we do want is for this
1012 * page to have PageReclaim set next time memcg
1013 * reclaim reaches the tests above, so it will
1014 * then wait_on_page_writeback() to avoid OOM;
1015 * and it's also appropriate in global reclaim.
1016 */
1017 SetPageReclaim(page);
1018 nr_writeback++;
1019 goto keep_locked;
1020
1021 /* Case 3 above */
1022 } else {
1023 unlock_page(page);
1024 wait_on_page_writeback(page);
1025 /* then go back and try same page again */
1026 list_add_tail(&page->lru, page_list);
1027 continue;
1028 }
1029 }
1030
1031 if (!force_reclaim)
1032 references = page_check_references(page, sc);
1033
1034 switch (references) {
1035 case PAGEREF_ACTIVATE:
1036 goto activate_locked;
1037 case PAGEREF_KEEP:
1038 goto keep_locked;
1039 case PAGEREF_RECLAIM:
1040 case PAGEREF_RECLAIM_CLEAN:
1041 ; /* try to reclaim the page below */
1042 }
1043
1044 /*
1045 * Anonymous process memory has backing store?
1046 * Try to allocate it some swap space here.
1047 */
1048 if (PageAnon(page) && !PageSwapCache(page)) {
1049 if (!(sc->gfp_mask & __GFP_IO))
1050 goto keep_locked;
1051 if (!add_to_swap(page, page_list))
1052 goto activate_locked;
1053 lazyfree = true;
1054 may_enter_fs = 1;
1055
1056 /* Adding to swap updated mapping */
1057 mapping = page_mapping(page);
1058 }
1059
1060 /*
1061 * The page is mapped into the page tables of one or more
1062 * processes. Try to unmap it here.
1063 */
1064 if (page_mapped(page) && mapping) {
1065 switch (ret = try_to_unmap(page, lazyfree ?
1066 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1067 (ttu_flags | TTU_BATCH_FLUSH))) {
1068 case SWAP_FAIL:
1069 goto activate_locked;
1070 case SWAP_AGAIN:
1071 goto keep_locked;
1072 case SWAP_MLOCK:
1073 goto cull_mlocked;
1074 case SWAP_LZFREE:
1075 goto lazyfree;
1076 case SWAP_SUCCESS:
1077 ; /* try to free the page below */
1078 }
1079 }
1080
1081 if (PageDirty(page)) {
1082 /*
1083 * Only kswapd can writeback filesystem pages to
1084 * avoid risk of stack overflow but only writeback
1085 * if many dirty pages have been encountered.
1086 */
1087 if (page_is_file_cache(page) &&
1088 (!current_is_kswapd() ||
1089 !test_bit(ZONE_DIRTY, &zone->flags))) {
1090 /*
1091 * Immediately reclaim when written back.
1092 * Similar in principal to deactivate_page()
1093 * except we already have the page isolated
1094 * and know it's dirty
1095 */
1096 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1097 SetPageReclaim(page);
1098
1099 goto keep_locked;
1100 }
1101
1102 if (references == PAGEREF_RECLAIM_CLEAN)
1103 goto keep_locked;
1104 if (!may_enter_fs)
1105 goto keep_locked;
1106 if (!sc->may_writepage)
1107 goto keep_locked;
1108
1109 /*
1110 * Page is dirty. Flush the TLB if a writable entry
1111 * potentially exists to avoid CPU writes after IO
1112 * starts and then write it out here.
1113 */
1114 try_to_unmap_flush_dirty();
1115 switch (pageout(page, mapping, sc)) {
1116 case PAGE_KEEP:
1117 goto keep_locked;
1118 case PAGE_ACTIVATE:
1119 goto activate_locked;
1120 case PAGE_SUCCESS:
1121 if (PageWriteback(page))
1122 goto keep;
1123 if (PageDirty(page))
1124 goto keep;
1125
1126 /*
1127 * A synchronous write - probably a ramdisk. Go
1128 * ahead and try to reclaim the page.
1129 */
1130 if (!trylock_page(page))
1131 goto keep;
1132 if (PageDirty(page) || PageWriteback(page))
1133 goto keep_locked;
1134 mapping = page_mapping(page);
1135 case PAGE_CLEAN:
1136 ; /* try to free the page below */
1137 }
1138 }
1139
1140 /*
1141 * If the page has buffers, try to free the buffer mappings
1142 * associated with this page. If we succeed we try to free
1143 * the page as well.
1144 *
1145 * We do this even if the page is PageDirty().
1146 * try_to_release_page() does not perform I/O, but it is
1147 * possible for a page to have PageDirty set, but it is actually
1148 * clean (all its buffers are clean). This happens if the
1149 * buffers were written out directly, with submit_bh(). ext3
1150 * will do this, as well as the blockdev mapping.
1151 * try_to_release_page() will discover that cleanness and will
1152 * drop the buffers and mark the page clean - it can be freed.
1153 *
1154 * Rarely, pages can have buffers and no ->mapping. These are
1155 * the pages which were not successfully invalidated in
1156 * truncate_complete_page(). We try to drop those buffers here
1157 * and if that worked, and the page is no longer mapped into
1158 * process address space (page_count == 1) it can be freed.
1159 * Otherwise, leave the page on the LRU so it is swappable.
1160 */
1161 if (page_has_private(page)) {
1162 if (!try_to_release_page(page, sc->gfp_mask))
1163 goto activate_locked;
1164 if (!mapping && page_count(page) == 1) {
1165 unlock_page(page);
1166 if (put_page_testzero(page))
1167 goto free_it;
1168 else {
1169 /*
1170 * rare race with speculative reference.
1171 * the speculative reference will free
1172 * this page shortly, so we may
1173 * increment nr_reclaimed here (and
1174 * leave it off the LRU).
1175 */
1176 nr_reclaimed++;
1177 continue;
1178 }
1179 }
1180 }
1181
1182lazyfree:
1183 if (!mapping || !__remove_mapping(mapping, page, true))
1184 goto keep_locked;
1185
1186 /*
1187 * At this point, we have no other references and there is
1188 * no way to pick any more up (removed from LRU, removed
1189 * from pagecache). Can use non-atomic bitops now (and
1190 * we obviously don't have to worry about waking up a process
1191 * waiting on the page lock, because there are no references.
1192 */
1193 __ClearPageLocked(page);
1194free_it:
1195 if (ret == SWAP_LZFREE)
1196 count_vm_event(PGLAZYFREED);
1197
1198 nr_reclaimed++;
1199
1200 /*
1201 * Is there need to periodically free_page_list? It would
1202 * appear not as the counts should be low
1203 */
1204 list_add(&page->lru, &free_pages);
1205 continue;
1206
1207cull_mlocked:
1208 if (PageSwapCache(page))
1209 try_to_free_swap(page);
1210 unlock_page(page);
1211 list_add(&page->lru, &ret_pages);
1212 continue;
1213
1214activate_locked:
1215 /* Not a candidate for swapping, so reclaim swap space. */
1216 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1217 try_to_free_swap(page);
1218 VM_BUG_ON_PAGE(PageActive(page), page);
1219 SetPageActive(page);
1220 pgactivate++;
1221keep_locked:
1222 unlock_page(page);
1223keep:
1224 list_add(&page->lru, &ret_pages);
1225 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1226 }
1227
1228 mem_cgroup_uncharge_list(&free_pages);
1229 try_to_unmap_flush();
1230 free_hot_cold_page_list(&free_pages, true);
1231
1232 list_splice(&ret_pages, page_list);
1233 count_vm_events(PGACTIVATE, pgactivate);
1234
1235 *ret_nr_dirty += nr_dirty;
1236 *ret_nr_congested += nr_congested;
1237 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1238 *ret_nr_writeback += nr_writeback;
1239 *ret_nr_immediate += nr_immediate;
1240 return nr_reclaimed;
1241}
1242
1243unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1244 struct list_head *page_list)
1245{
1246 struct scan_control sc = {
1247 .gfp_mask = GFP_KERNEL,
1248 .priority = DEF_PRIORITY,
1249 .may_unmap = 1,
1250 };
1251 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1252 struct page *page, *next;
1253 LIST_HEAD(clean_pages);
1254
1255 list_for_each_entry_safe(page, next, page_list, lru) {
1256 if (page_is_file_cache(page) && !PageDirty(page) &&
1257 !isolated_balloon_page(page)) {
1258 ClearPageActive(page);
1259 list_move(&page->lru, &clean_pages);
1260 }
1261 }
1262
1263 ret = shrink_page_list(&clean_pages, zone, &sc,
1264 TTU_UNMAP|TTU_IGNORE_ACCESS,
1265 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1266 list_splice(&clean_pages, page_list);
1267 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1268 return ret;
1269}
1270
1271/*
1272 * Attempt to remove the specified page from its LRU. Only take this page
1273 * if it is of the appropriate PageActive status. Pages which are being
1274 * freed elsewhere are also ignored.
1275 *
1276 * page: page to consider
1277 * mode: one of the LRU isolation modes defined above
1278 *
1279 * returns 0 on success, -ve errno on failure.
1280 */
1281int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1282{
1283 int ret = -EINVAL;
1284
1285 /* Only take pages on the LRU. */
1286 if (!PageLRU(page))
1287 return ret;
1288
1289 /* Compaction should not handle unevictable pages but CMA can do so */
1290 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1291 return ret;
1292
1293 ret = -EBUSY;
1294
1295 /*
1296 * To minimise LRU disruption, the caller can indicate that it only
1297 * wants to isolate pages it will be able to operate on without
1298 * blocking - clean pages for the most part.
1299 *
1300 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1301 * is used by reclaim when it is cannot write to backing storage
1302 *
1303 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1304 * that it is possible to migrate without blocking
1305 */
1306 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1307 /* All the caller can do on PageWriteback is block */
1308 if (PageWriteback(page))
1309 return ret;
1310
1311 if (PageDirty(page)) {
1312 struct address_space *mapping;
1313
1314 /* ISOLATE_CLEAN means only clean pages */
1315 if (mode & ISOLATE_CLEAN)
1316 return ret;
1317
1318 /*
1319 * Only pages without mappings or that have a
1320 * ->migratepage callback are possible to migrate
1321 * without blocking
1322 */
1323 mapping = page_mapping(page);
1324 if (mapping && !mapping->a_ops->migratepage)
1325 return ret;
1326 }
1327 }
1328
1329 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1330 return ret;
1331
1332 if (likely(get_page_unless_zero(page))) {
1333 /*
1334 * Be careful not to clear PageLRU until after we're
1335 * sure the page is not being freed elsewhere -- the
1336 * page release code relies on it.
1337 */
1338 ClearPageLRU(page);
1339 ret = 0;
1340 }
1341
1342 return ret;
1343}
1344
1345/*
1346 * zone->lru_lock is heavily contended. Some of the functions that
1347 * shrink the lists perform better by taking out a batch of pages
1348 * and working on them outside the LRU lock.
1349 *
1350 * For pagecache intensive workloads, this function is the hottest
1351 * spot in the kernel (apart from copy_*_user functions).
1352 *
1353 * Appropriate locks must be held before calling this function.
1354 *
1355 * @nr_to_scan: The number of pages to look through on the list.
1356 * @lruvec: The LRU vector to pull pages from.
1357 * @dst: The temp list to put pages on to.
1358 * @nr_scanned: The number of pages that were scanned.
1359 * @sc: The scan_control struct for this reclaim session
1360 * @mode: One of the LRU isolation modes
1361 * @lru: LRU list id for isolating
1362 *
1363 * returns how many pages were moved onto *@dst.
1364 */
1365static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1366 struct lruvec *lruvec, struct list_head *dst,
1367 unsigned long *nr_scanned, struct scan_control *sc,
1368 isolate_mode_t mode, enum lru_list lru)
1369{
1370 struct list_head *src = &lruvec->lists[lru];
1371 unsigned long nr_taken = 0;
1372 unsigned long scan;
1373
1374 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1375 !list_empty(src); scan++) {
1376 struct page *page;
1377 int nr_pages;
1378
1379 page = lru_to_page(src);
1380 prefetchw_prev_lru_page(page, src, flags);
1381
1382 VM_BUG_ON_PAGE(!PageLRU(page), page);
1383
1384 switch (__isolate_lru_page(page, mode)) {
1385 case 0:
1386 nr_pages = hpage_nr_pages(page);
1387 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1388 list_move(&page->lru, dst);
1389 nr_taken += nr_pages;
1390 break;
1391
1392 case -EBUSY:
1393 /* else it is being freed elsewhere */
1394 list_move(&page->lru, src);
1395 continue;
1396
1397 default:
1398 BUG();
1399 }
1400 }
1401
1402 *nr_scanned = scan;
1403 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1404 nr_taken, mode, is_file_lru(lru));
1405 return nr_taken;
1406}
1407
1408/**
1409 * isolate_lru_page - tries to isolate a page from its LRU list
1410 * @page: page to isolate from its LRU list
1411 *
1412 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1413 * vmstat statistic corresponding to whatever LRU list the page was on.
1414 *
1415 * Returns 0 if the page was removed from an LRU list.
1416 * Returns -EBUSY if the page was not on an LRU list.
1417 *
1418 * The returned page will have PageLRU() cleared. If it was found on
1419 * the active list, it will have PageActive set. If it was found on
1420 * the unevictable list, it will have the PageUnevictable bit set. That flag
1421 * may need to be cleared by the caller before letting the page go.
1422 *
1423 * The vmstat statistic corresponding to the list on which the page was
1424 * found will be decremented.
1425 *
1426 * Restrictions:
1427 * (1) Must be called with an elevated refcount on the page. This is a
1428 * fundamentnal difference from isolate_lru_pages (which is called
1429 * without a stable reference).
1430 * (2) the lru_lock must not be held.
1431 * (3) interrupts must be enabled.
1432 */
1433int isolate_lru_page(struct page *page)
1434{
1435 int ret = -EBUSY;
1436
1437 VM_BUG_ON_PAGE(!page_count(page), page);
1438 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1439
1440 if (PageLRU(page)) {
1441 struct zone *zone = page_zone(page);
1442 struct lruvec *lruvec;
1443
1444 spin_lock_irq(&zone->lru_lock);
1445 lruvec = mem_cgroup_page_lruvec(page, zone);
1446 if (PageLRU(page)) {
1447 int lru = page_lru(page);
1448 get_page(page);
1449 ClearPageLRU(page);
1450 del_page_from_lru_list(page, lruvec, lru);
1451 ret = 0;
1452 }
1453 spin_unlock_irq(&zone->lru_lock);
1454 }
1455 return ret;
1456}
1457
1458/*
1459 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1460 * then get resheduled. When there are massive number of tasks doing page
1461 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1462 * the LRU list will go small and be scanned faster than necessary, leading to
1463 * unnecessary swapping, thrashing and OOM.
1464 */
1465static int too_many_isolated(struct zone *zone, int file,
1466 struct scan_control *sc)
1467{
1468 unsigned long inactive, isolated;
1469
1470 if (current_is_kswapd())
1471 return 0;
1472
1473 if (!sane_reclaim(sc))
1474 return 0;
1475
1476 if (file) {
1477 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1478 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1479 } else {
1480 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1481 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1482 }
1483
1484 /*
1485 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1486 * won't get blocked by normal direct-reclaimers, forming a circular
1487 * deadlock.
1488 */
1489 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1490 inactive >>= 3;
1491
1492 return isolated > inactive;
1493}
1494
1495static noinline_for_stack void
1496putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1497{
1498 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1499 struct zone *zone = lruvec_zone(lruvec);
1500 LIST_HEAD(pages_to_free);
1501
1502 /*
1503 * Put back any unfreeable pages.
1504 */
1505 while (!list_empty(page_list)) {
1506 struct page *page = lru_to_page(page_list);
1507 int lru;
1508
1509 VM_BUG_ON_PAGE(PageLRU(page), page);
1510 list_del(&page->lru);
1511 if (unlikely(!page_evictable(page))) {
1512 spin_unlock_irq(&zone->lru_lock);
1513 putback_lru_page(page);
1514 spin_lock_irq(&zone->lru_lock);
1515 continue;
1516 }
1517
1518 lruvec = mem_cgroup_page_lruvec(page, zone);
1519
1520 SetPageLRU(page);
1521 lru = page_lru(page);
1522 add_page_to_lru_list(page, lruvec, lru);
1523
1524 if (is_active_lru(lru)) {
1525 int file = is_file_lru(lru);
1526 int numpages = hpage_nr_pages(page);
1527 reclaim_stat->recent_rotated[file] += numpages;
1528 }
1529 if (put_page_testzero(page)) {
1530 __ClearPageLRU(page);
1531 __ClearPageActive(page);
1532 del_page_from_lru_list(page, lruvec, lru);
1533
1534 if (unlikely(PageCompound(page))) {
1535 spin_unlock_irq(&zone->lru_lock);
1536 mem_cgroup_uncharge(page);
1537 (*get_compound_page_dtor(page))(page);
1538 spin_lock_irq(&zone->lru_lock);
1539 } else
1540 list_add(&page->lru, &pages_to_free);
1541 }
1542 }
1543
1544 /*
1545 * To save our caller's stack, now use input list for pages to free.
1546 */
1547 list_splice(&pages_to_free, page_list);
1548}
1549
1550/*
1551 * If a kernel thread (such as nfsd for loop-back mounts) services
1552 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1553 * In that case we should only throttle if the backing device it is
1554 * writing to is congested. In other cases it is safe to throttle.
1555 */
1556static int current_may_throttle(void)
1557{
1558 return !(current->flags & PF_LESS_THROTTLE) ||
1559 current->backing_dev_info == NULL ||
1560 bdi_write_congested(current->backing_dev_info);
1561}
1562
1563/*
1564 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1565 * of reclaimed pages
1566 */
1567static noinline_for_stack unsigned long
1568shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1569 struct scan_control *sc, enum lru_list lru)
1570{
1571 LIST_HEAD(page_list);
1572 unsigned long nr_scanned;
1573 unsigned long nr_reclaimed = 0;
1574 unsigned long nr_taken;
1575 unsigned long nr_dirty = 0;
1576 unsigned long nr_congested = 0;
1577 unsigned long nr_unqueued_dirty = 0;
1578 unsigned long nr_writeback = 0;
1579 unsigned long nr_immediate = 0;
1580 isolate_mode_t isolate_mode = 0;
1581 int file = is_file_lru(lru);
1582 struct zone *zone = lruvec_zone(lruvec);
1583 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1584
1585 while (unlikely(too_many_isolated(zone, file, sc))) {
1586 congestion_wait(BLK_RW_ASYNC, HZ/10);
1587
1588 /* We are about to die and free our memory. Return now. */
1589 if (fatal_signal_pending(current))
1590 return SWAP_CLUSTER_MAX;
1591 }
1592
1593 lru_add_drain();
1594
1595 if (!sc->may_unmap)
1596 isolate_mode |= ISOLATE_UNMAPPED;
1597 if (!sc->may_writepage)
1598 isolate_mode |= ISOLATE_CLEAN;
1599
1600 spin_lock_irq(&zone->lru_lock);
1601
1602 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1603 &nr_scanned, sc, isolate_mode, lru);
1604
1605 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1606 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1607
1608 if (global_reclaim(sc)) {
1609 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1610 if (current_is_kswapd())
1611 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1612 else
1613 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1614 }
1615 spin_unlock_irq(&zone->lru_lock);
1616
1617 if (nr_taken == 0)
1618 return 0;
1619
1620 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1621 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1622 &nr_writeback, &nr_immediate,
1623 false);
1624
1625 spin_lock_irq(&zone->lru_lock);
1626
1627 reclaim_stat->recent_scanned[file] += nr_taken;
1628
1629 if (global_reclaim(sc)) {
1630 if (current_is_kswapd())
1631 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1632 nr_reclaimed);
1633 else
1634 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1635 nr_reclaimed);
1636 }
1637
1638 putback_inactive_pages(lruvec, &page_list);
1639
1640 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1641
1642 spin_unlock_irq(&zone->lru_lock);
1643
1644 mem_cgroup_uncharge_list(&page_list);
1645 free_hot_cold_page_list(&page_list, true);
1646
1647 /*
1648 * If reclaim is isolating dirty pages under writeback, it implies
1649 * that the long-lived page allocation rate is exceeding the page
1650 * laundering rate. Either the global limits are not being effective
1651 * at throttling processes due to the page distribution throughout
1652 * zones or there is heavy usage of a slow backing device. The
1653 * only option is to throttle from reclaim context which is not ideal
1654 * as there is no guarantee the dirtying process is throttled in the
1655 * same way balance_dirty_pages() manages.
1656 *
1657 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1658 * of pages under pages flagged for immediate reclaim and stall if any
1659 * are encountered in the nr_immediate check below.
1660 */
1661 if (nr_writeback && nr_writeback == nr_taken)
1662 set_bit(ZONE_WRITEBACK, &zone->flags);
1663
1664 /*
1665 * Legacy memcg will stall in page writeback so avoid forcibly
1666 * stalling here.
1667 */
1668 if (sane_reclaim(sc)) {
1669 /*
1670 * Tag a zone as congested if all the dirty pages scanned were
1671 * backed by a congested BDI and wait_iff_congested will stall.
1672 */
1673 if (nr_dirty && nr_dirty == nr_congested)
1674 set_bit(ZONE_CONGESTED, &zone->flags);
1675
1676 /*
1677 * If dirty pages are scanned that are not queued for IO, it
1678 * implies that flushers are not keeping up. In this case, flag
1679 * the zone ZONE_DIRTY and kswapd will start writing pages from
1680 * reclaim context.
1681 */
1682 if (nr_unqueued_dirty == nr_taken)
1683 set_bit(ZONE_DIRTY, &zone->flags);
1684
1685 /*
1686 * If kswapd scans pages marked marked for immediate
1687 * reclaim and under writeback (nr_immediate), it implies
1688 * that pages are cycling through the LRU faster than
1689 * they are written so also forcibly stall.
1690 */
1691 if (nr_immediate && current_may_throttle())
1692 congestion_wait(BLK_RW_ASYNC, HZ/10);
1693 }
1694
1695 /*
1696 * Stall direct reclaim for IO completions if underlying BDIs or zone
1697 * is congested. Allow kswapd to continue until it starts encountering
1698 * unqueued dirty pages or cycling through the LRU too quickly.
1699 */
1700 if (!sc->hibernation_mode && !current_is_kswapd() &&
1701 current_may_throttle())
1702 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1703
1704 trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed,
1705 sc->priority, file);
1706 return nr_reclaimed;
1707}
1708
1709/*
1710 * This moves pages from the active list to the inactive list.
1711 *
1712 * We move them the other way if the page is referenced by one or more
1713 * processes, from rmap.
1714 *
1715 * If the pages are mostly unmapped, the processing is fast and it is
1716 * appropriate to hold zone->lru_lock across the whole operation. But if
1717 * the pages are mapped, the processing is slow (page_referenced()) so we
1718 * should drop zone->lru_lock around each page. It's impossible to balance
1719 * this, so instead we remove the pages from the LRU while processing them.
1720 * It is safe to rely on PG_active against the non-LRU pages in here because
1721 * nobody will play with that bit on a non-LRU page.
1722 *
1723 * The downside is that we have to touch page->_count against each page.
1724 * But we had to alter page->flags anyway.
1725 */
1726
1727static void move_active_pages_to_lru(struct lruvec *lruvec,
1728 struct list_head *list,
1729 struct list_head *pages_to_free,
1730 enum lru_list lru)
1731{
1732 struct zone *zone = lruvec_zone(lruvec);
1733 unsigned long pgmoved = 0;
1734 struct page *page;
1735 int nr_pages;
1736
1737 while (!list_empty(list)) {
1738 page = lru_to_page(list);
1739 lruvec = mem_cgroup_page_lruvec(page, zone);
1740
1741 VM_BUG_ON_PAGE(PageLRU(page), page);
1742 SetPageLRU(page);
1743
1744 nr_pages = hpage_nr_pages(page);
1745 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1746 list_move(&page->lru, &lruvec->lists[lru]);
1747 pgmoved += nr_pages;
1748
1749 if (put_page_testzero(page)) {
1750 __ClearPageLRU(page);
1751 __ClearPageActive(page);
1752 del_page_from_lru_list(page, lruvec, lru);
1753
1754 if (unlikely(PageCompound(page))) {
1755 spin_unlock_irq(&zone->lru_lock);
1756 mem_cgroup_uncharge(page);
1757 (*get_compound_page_dtor(page))(page);
1758 spin_lock_irq(&zone->lru_lock);
1759 } else
1760 list_add(&page->lru, pages_to_free);
1761 }
1762 }
1763 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1764 if (!is_active_lru(lru))
1765 __count_vm_events(PGDEACTIVATE, pgmoved);
1766}
1767
1768static void shrink_active_list(unsigned long nr_to_scan,
1769 struct lruvec *lruvec,
1770 struct scan_control *sc,
1771 enum lru_list lru)
1772{
1773 unsigned long nr_taken;
1774 unsigned long nr_scanned;
1775 unsigned long vm_flags;
1776 LIST_HEAD(l_hold); /* The pages which were snipped off */
1777 LIST_HEAD(l_active);
1778 LIST_HEAD(l_inactive);
1779 struct page *page;
1780 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1781 unsigned long nr_rotated = 0;
1782 isolate_mode_t isolate_mode = 0;
1783 int file = is_file_lru(lru);
1784 struct zone *zone = lruvec_zone(lruvec);
1785
1786 lru_add_drain();
1787
1788 if (!sc->may_unmap)
1789 isolate_mode |= ISOLATE_UNMAPPED;
1790 if (!sc->may_writepage)
1791 isolate_mode |= ISOLATE_CLEAN;
1792
1793 spin_lock_irq(&zone->lru_lock);
1794
1795 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1796 &nr_scanned, sc, isolate_mode, lru);
1797 if (global_reclaim(sc))
1798 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1799
1800 reclaim_stat->recent_scanned[file] += nr_taken;
1801
1802 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1803 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1804 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1805 spin_unlock_irq(&zone->lru_lock);
1806
1807 while (!list_empty(&l_hold)) {
1808 cond_resched();
1809 page = lru_to_page(&l_hold);
1810 list_del(&page->lru);
1811
1812 if (unlikely(!page_evictable(page))) {
1813 putback_lru_page(page);
1814 continue;
1815 }
1816
1817 if (unlikely(buffer_heads_over_limit)) {
1818 if (page_has_private(page) && trylock_page(page)) {
1819 if (page_has_private(page))
1820 try_to_release_page(page, 0);
1821 unlock_page(page);
1822 }
1823 }
1824
1825 if (page_referenced(page, 0, sc->target_mem_cgroup,
1826 &vm_flags)) {
1827 nr_rotated += hpage_nr_pages(page);
1828 /*
1829 * Identify referenced, file-backed active pages and
1830 * give them one more trip around the active list. So
1831 * that executable code get better chances to stay in
1832 * memory under moderate memory pressure. Anon pages
1833 * are not likely to be evicted by use-once streaming
1834 * IO, plus JVM can create lots of anon VM_EXEC pages,
1835 * so we ignore them here.
1836 */
1837 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1838 list_add(&page->lru, &l_active);
1839 continue;
1840 }
1841 }
1842
1843 ClearPageActive(page); /* we are de-activating */
1844 list_add(&page->lru, &l_inactive);
1845 }
1846
1847 /*
1848 * Move pages back to the lru list.
1849 */
1850 spin_lock_irq(&zone->lru_lock);
1851 /*
1852 * Count referenced pages from currently used mappings as rotated,
1853 * even though only some of them are actually re-activated. This
1854 * helps balance scan pressure between file and anonymous pages in
1855 * get_scan_count.
1856 */
1857 reclaim_stat->recent_rotated[file] += nr_rotated;
1858
1859 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1860 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1861 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1862 spin_unlock_irq(&zone->lru_lock);
1863
1864 mem_cgroup_uncharge_list(&l_hold);
1865 free_hot_cold_page_list(&l_hold, true);
1866}
1867
1868#ifdef CONFIG_SWAP
1869static bool inactive_anon_is_low_global(struct zone *zone)
1870{
1871 unsigned long active, inactive;
1872
1873 active = zone_page_state(zone, NR_ACTIVE_ANON);
1874 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1875
1876 return inactive * zone->inactive_ratio < active;
1877}
1878
1879/**
1880 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1881 * @lruvec: LRU vector to check
1882 *
1883 * Returns true if the zone does not have enough inactive anon pages,
1884 * meaning some active anon pages need to be deactivated.
1885 */
1886static bool inactive_anon_is_low(struct lruvec *lruvec)
1887{
1888 /*
1889 * If we don't have swap space, anonymous page deactivation
1890 * is pointless.
1891 */
1892 if (!total_swap_pages)
1893 return false;
1894
1895 if (!mem_cgroup_disabled())
1896 return mem_cgroup_inactive_anon_is_low(lruvec);
1897
1898 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1899}
1900#else
1901static inline bool inactive_anon_is_low(struct lruvec *lruvec)
1902{
1903 return false;
1904}
1905#endif
1906
1907/**
1908 * inactive_file_is_low - check if file pages need to be deactivated
1909 * @lruvec: LRU vector to check
1910 *
1911 * When the system is doing streaming IO, memory pressure here
1912 * ensures that active file pages get deactivated, until more
1913 * than half of the file pages are on the inactive list.
1914 *
1915 * Once we get to that situation, protect the system's working
1916 * set from being evicted by disabling active file page aging.
1917 *
1918 * This uses a different ratio than the anonymous pages, because
1919 * the page cache uses a use-once replacement algorithm.
1920 */
1921static bool inactive_file_is_low(struct lruvec *lruvec)
1922{
1923 unsigned long inactive;
1924 unsigned long active;
1925
1926 inactive = lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
1927 active = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE);
1928
1929 return active > inactive;
1930}
1931
1932static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1933{
1934 if (is_file_lru(lru))
1935 return inactive_file_is_low(lruvec);
1936 else
1937 return inactive_anon_is_low(lruvec);
1938}
1939
1940static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1941 struct lruvec *lruvec, struct scan_control *sc)
1942{
1943 if (is_active_lru(lru)) {
1944 if (inactive_list_is_low(lruvec, lru))
1945 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1946 return 0;
1947 }
1948
1949 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1950}
1951
1952enum scan_balance {
1953 SCAN_EQUAL,
1954 SCAN_FRACT,
1955 SCAN_ANON,
1956 SCAN_FILE,
1957};
1958
1959/*
1960 * Determine how aggressively the anon and file LRU lists should be
1961 * scanned. The relative value of each set of LRU lists is determined
1962 * by looking at the fraction of the pages scanned we did rotate back
1963 * onto the active list instead of evict.
1964 *
1965 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1966 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1967 */
1968static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
1969 struct scan_control *sc, unsigned long *nr,
1970 unsigned long *lru_pages)
1971{
1972 int swappiness = mem_cgroup_swappiness(memcg);
1973 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1974 u64 fraction[2];
1975 u64 denominator = 0; /* gcc */
1976 struct zone *zone = lruvec_zone(lruvec);
1977 unsigned long anon_prio, file_prio;
1978 enum scan_balance scan_balance;
1979 unsigned long anon, file;
1980 bool force_scan = false;
1981 unsigned long ap, fp;
1982 enum lru_list lru;
1983 bool some_scanned;
1984 int pass;
1985
1986 /*
1987 * If the zone or memcg is small, nr[l] can be 0. This
1988 * results in no scanning on this priority and a potential
1989 * priority drop. Global direct reclaim can go to the next
1990 * zone and tends to have no problems. Global kswapd is for
1991 * zone balancing and it needs to scan a minimum amount. When
1992 * reclaiming for a memcg, a priority drop can cause high
1993 * latencies, so it's better to scan a minimum amount there as
1994 * well.
1995 */
1996 if (current_is_kswapd()) {
1997 if (!zone_reclaimable(zone))
1998 force_scan = true;
1999 if (!mem_cgroup_online(memcg))
2000 force_scan = true;
2001 }
2002 if (!global_reclaim(sc))
2003 force_scan = true;
2004
2005 /* If we have no swap space, do not bother scanning anon pages. */
2006 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2007 scan_balance = SCAN_FILE;
2008 goto out;
2009 }
2010
2011 /*
2012 * Global reclaim will swap to prevent OOM even with no
2013 * swappiness, but memcg users want to use this knob to
2014 * disable swapping for individual groups completely when
2015 * using the memory controller's swap limit feature would be
2016 * too expensive.
2017 */
2018 if (!global_reclaim(sc) && !swappiness) {
2019 scan_balance = SCAN_FILE;
2020 goto out;
2021 }
2022
2023 /*
2024 * Do not apply any pressure balancing cleverness when the
2025 * system is close to OOM, scan both anon and file equally
2026 * (unless the swappiness setting disagrees with swapping).
2027 */
2028 if (!sc->priority && swappiness) {
2029 scan_balance = SCAN_EQUAL;
2030 goto out;
2031 }
2032
2033 /*
2034 * Prevent the reclaimer from falling into the cache trap: as
2035 * cache pages start out inactive, every cache fault will tip
2036 * the scan balance towards the file LRU. And as the file LRU
2037 * shrinks, so does the window for rotation from references.
2038 * This means we have a runaway feedback loop where a tiny
2039 * thrashing file LRU becomes infinitely more attractive than
2040 * anon pages. Try to detect this based on file LRU size.
2041 */
2042 if (global_reclaim(sc)) {
2043 unsigned long zonefile;
2044 unsigned long zonefree;
2045
2046 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2047 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2048 zone_page_state(zone, NR_INACTIVE_FILE);
2049
2050 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2051 scan_balance = SCAN_ANON;
2052 goto out;
2053 }
2054 }
2055
2056 /*
2057 * If there is enough inactive page cache, i.e. if the size of the
2058 * inactive list is greater than that of the active list *and* the
2059 * inactive list actually has some pages to scan on this priority, we
2060 * do not reclaim anything from the anonymous working set right now.
2061 * Without the second condition we could end up never scanning an
2062 * lruvec even if it has plenty of old anonymous pages unless the
2063 * system is under heavy pressure.
2064 */
2065 if (!inactive_file_is_low(lruvec) &&
2066 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2067 scan_balance = SCAN_FILE;
2068 goto out;
2069 }
2070
2071 scan_balance = SCAN_FRACT;
2072
2073 /*
2074 * With swappiness at 100, anonymous and file have the same priority.
2075 * This scanning priority is essentially the inverse of IO cost.
2076 */
2077 anon_prio = swappiness;
2078 file_prio = 200 - anon_prio;
2079
2080 /*
2081 * OK, so we have swap space and a fair amount of page cache
2082 * pages. We use the recently rotated / recently scanned
2083 * ratios to determine how valuable each cache is.
2084 *
2085 * Because workloads change over time (and to avoid overflow)
2086 * we keep these statistics as a floating average, which ends
2087 * up weighing recent references more than old ones.
2088 *
2089 * anon in [0], file in [1]
2090 */
2091
2092 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2093 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2094 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2095 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2096
2097 spin_lock_irq(&zone->lru_lock);
2098 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2099 reclaim_stat->recent_scanned[0] /= 2;
2100 reclaim_stat->recent_rotated[0] /= 2;
2101 }
2102
2103 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2104 reclaim_stat->recent_scanned[1] /= 2;
2105 reclaim_stat->recent_rotated[1] /= 2;
2106 }
2107
2108 /*
2109 * The amount of pressure on anon vs file pages is inversely
2110 * proportional to the fraction of recently scanned pages on
2111 * each list that were recently referenced and in active use.
2112 */
2113 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2114 ap /= reclaim_stat->recent_rotated[0] + 1;
2115
2116 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2117 fp /= reclaim_stat->recent_rotated[1] + 1;
2118 spin_unlock_irq(&zone->lru_lock);
2119
2120 fraction[0] = ap;
2121 fraction[1] = fp;
2122 denominator = ap + fp + 1;
2123out:
2124 some_scanned = false;
2125 /* Only use force_scan on second pass. */
2126 for (pass = 0; !some_scanned && pass < 2; pass++) {
2127 *lru_pages = 0;
2128 for_each_evictable_lru(lru) {
2129 int file = is_file_lru(lru);
2130 unsigned long size;
2131 unsigned long scan;
2132
2133 size = lruvec_lru_size(lruvec, lru);
2134 scan = size >> sc->priority;
2135
2136 if (!scan && pass && force_scan)
2137 scan = min(size, SWAP_CLUSTER_MAX);
2138
2139 switch (scan_balance) {
2140 case SCAN_EQUAL:
2141 /* Scan lists relative to size */
2142 break;
2143 case SCAN_FRACT:
2144 /*
2145 * Scan types proportional to swappiness and
2146 * their relative recent reclaim efficiency.
2147 */
2148 scan = div64_u64(scan * fraction[file],
2149 denominator);
2150 break;
2151 case SCAN_FILE:
2152 case SCAN_ANON:
2153 /* Scan one type exclusively */
2154 if ((scan_balance == SCAN_FILE) != file) {
2155 size = 0;
2156 scan = 0;
2157 }
2158 break;
2159 default:
2160 /* Look ma, no brain */
2161 BUG();
2162 }
2163
2164 *lru_pages += size;
2165 nr[lru] = scan;
2166
2167 /*
2168 * Skip the second pass and don't force_scan,
2169 * if we found something to scan.
2170 */
2171 some_scanned |= !!scan;
2172 }
2173 }
2174}
2175
2176#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2177static void init_tlb_ubc(void)
2178{
2179 /*
2180 * This deliberately does not clear the cpumask as it's expensive
2181 * and unnecessary. If there happens to be data in there then the
2182 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2183 * then will be cleared.
2184 */
2185 current->tlb_ubc.flush_required = false;
2186}
2187#else
2188static inline void init_tlb_ubc(void)
2189{
2190}
2191#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2192
2193/*
2194 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2195 */
2196static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2197 struct scan_control *sc, unsigned long *lru_pages)
2198{
2199 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2200 unsigned long nr[NR_LRU_LISTS];
2201 unsigned long targets[NR_LRU_LISTS];
2202 unsigned long nr_to_scan;
2203 enum lru_list lru;
2204 unsigned long nr_reclaimed = 0;
2205 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2206 struct blk_plug plug;
2207 bool scan_adjusted;
2208
2209 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2210
2211 /* Record the original scan target for proportional adjustments later */
2212 memcpy(targets, nr, sizeof(nr));
2213
2214 /*
2215 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2216 * event that can occur when there is little memory pressure e.g.
2217 * multiple streaming readers/writers. Hence, we do not abort scanning
2218 * when the requested number of pages are reclaimed when scanning at
2219 * DEF_PRIORITY on the assumption that the fact we are direct
2220 * reclaiming implies that kswapd is not keeping up and it is best to
2221 * do a batch of work at once. For memcg reclaim one check is made to
2222 * abort proportional reclaim if either the file or anon lru has already
2223 * dropped to zero at the first pass.
2224 */
2225 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2226 sc->priority == DEF_PRIORITY);
2227
2228 init_tlb_ubc();
2229
2230 blk_start_plug(&plug);
2231 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2232 nr[LRU_INACTIVE_FILE]) {
2233 unsigned long nr_anon, nr_file, percentage;
2234 unsigned long nr_scanned;
2235
2236 for_each_evictable_lru(lru) {
2237 if (nr[lru]) {
2238 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2239 nr[lru] -= nr_to_scan;
2240
2241 nr_reclaimed += shrink_list(lru, nr_to_scan,
2242 lruvec, sc);
2243 }
2244 }
2245
2246 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2247 continue;
2248
2249 /*
2250 * For kswapd and memcg, reclaim at least the number of pages
2251 * requested. Ensure that the anon and file LRUs are scanned
2252 * proportionally what was requested by get_scan_count(). We
2253 * stop reclaiming one LRU and reduce the amount scanning
2254 * proportional to the original scan target.
2255 */
2256 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2257 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2258
2259 /*
2260 * It's just vindictive to attack the larger once the smaller
2261 * has gone to zero. And given the way we stop scanning the
2262 * smaller below, this makes sure that we only make one nudge
2263 * towards proportionality once we've got nr_to_reclaim.
2264 */
2265 if (!nr_file || !nr_anon)
2266 break;
2267
2268 if (nr_file > nr_anon) {
2269 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2270 targets[LRU_ACTIVE_ANON] + 1;
2271 lru = LRU_BASE;
2272 percentage = nr_anon * 100 / scan_target;
2273 } else {
2274 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2275 targets[LRU_ACTIVE_FILE] + 1;
2276 lru = LRU_FILE;
2277 percentage = nr_file * 100 / scan_target;
2278 }
2279
2280 /* Stop scanning the smaller of the LRU */
2281 nr[lru] = 0;
2282 nr[lru + LRU_ACTIVE] = 0;
2283
2284 /*
2285 * Recalculate the other LRU scan count based on its original
2286 * scan target and the percentage scanning already complete
2287 */
2288 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2289 nr_scanned = targets[lru] - nr[lru];
2290 nr[lru] = targets[lru] * (100 - percentage) / 100;
2291 nr[lru] -= min(nr[lru], nr_scanned);
2292
2293 lru += LRU_ACTIVE;
2294 nr_scanned = targets[lru] - nr[lru];
2295 nr[lru] = targets[lru] * (100 - percentage) / 100;
2296 nr[lru] -= min(nr[lru], nr_scanned);
2297
2298 scan_adjusted = true;
2299 }
2300 blk_finish_plug(&plug);
2301 sc->nr_reclaimed += nr_reclaimed;
2302
2303 /*
2304 * Even if we did not try to evict anon pages at all, we want to
2305 * rebalance the anon lru active/inactive ratio.
2306 */
2307 if (inactive_anon_is_low(lruvec))
2308 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2309 sc, LRU_ACTIVE_ANON);
2310
2311 throttle_vm_writeout(sc->gfp_mask);
2312}
2313
2314/* Use reclaim/compaction for costly allocs or under memory pressure */
2315static bool in_reclaim_compaction(struct scan_control *sc)
2316{
2317 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2318 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2319 sc->priority < DEF_PRIORITY - 2))
2320 return true;
2321
2322 return false;
2323}
2324
2325/*
2326 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2327 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2328 * true if more pages should be reclaimed such that when the page allocator
2329 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2330 * It will give up earlier than that if there is difficulty reclaiming pages.
2331 */
2332static inline bool should_continue_reclaim(struct zone *zone,
2333 unsigned long nr_reclaimed,
2334 unsigned long nr_scanned,
2335 struct scan_control *sc)
2336{
2337 unsigned long pages_for_compaction;
2338 unsigned long inactive_lru_pages;
2339
2340 /* If not in reclaim/compaction mode, stop */
2341 if (!in_reclaim_compaction(sc))
2342 return false;
2343
2344 /* Consider stopping depending on scan and reclaim activity */
2345 if (sc->gfp_mask & __GFP_REPEAT) {
2346 /*
2347 * For __GFP_REPEAT allocations, stop reclaiming if the
2348 * full LRU list has been scanned and we are still failing
2349 * to reclaim pages. This full LRU scan is potentially
2350 * expensive but a __GFP_REPEAT caller really wants to succeed
2351 */
2352 if (!nr_reclaimed && !nr_scanned)
2353 return false;
2354 } else {
2355 /*
2356 * For non-__GFP_REPEAT allocations which can presumably
2357 * fail without consequence, stop if we failed to reclaim
2358 * any pages from the last SWAP_CLUSTER_MAX number of
2359 * pages that were scanned. This will return to the
2360 * caller faster at the risk reclaim/compaction and
2361 * the resulting allocation attempt fails
2362 */
2363 if (!nr_reclaimed)
2364 return false;
2365 }
2366
2367 /*
2368 * If we have not reclaimed enough pages for compaction and the
2369 * inactive lists are large enough, continue reclaiming
2370 */
2371 pages_for_compaction = (2UL << sc->order);
2372 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2373 if (get_nr_swap_pages() > 0)
2374 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2375 if (sc->nr_reclaimed < pages_for_compaction &&
2376 inactive_lru_pages > pages_for_compaction)
2377 return true;
2378
2379 /* If compaction would go ahead or the allocation would succeed, stop */
2380 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2381 case COMPACT_PARTIAL:
2382 case COMPACT_CONTINUE:
2383 return false;
2384 default:
2385 return true;
2386 }
2387}
2388
2389static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2390 bool is_classzone)
2391{
2392 struct reclaim_state *reclaim_state = current->reclaim_state;
2393 unsigned long nr_reclaimed, nr_scanned;
2394 bool reclaimable = false;
2395
2396 do {
2397 struct mem_cgroup *root = sc->target_mem_cgroup;
2398 struct mem_cgroup_reclaim_cookie reclaim = {
2399 .zone = zone,
2400 .priority = sc->priority,
2401 };
2402 unsigned long zone_lru_pages = 0;
2403 struct mem_cgroup *memcg;
2404
2405 nr_reclaimed = sc->nr_reclaimed;
2406 nr_scanned = sc->nr_scanned;
2407
2408 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2409 do {
2410 unsigned long lru_pages;
2411 unsigned long reclaimed;
2412 unsigned long scanned;
2413
2414 if (mem_cgroup_low(root, memcg)) {
2415 if (!sc->may_thrash)
2416 continue;
2417 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2418 }
2419
2420 reclaimed = sc->nr_reclaimed;
2421 scanned = sc->nr_scanned;
2422
2423 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
2424 zone_lru_pages += lru_pages;
2425
2426 if (memcg && is_classzone)
2427 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2428 memcg, sc->nr_scanned - scanned,
2429 lru_pages);
2430
2431 /* Record the group's reclaim efficiency */
2432 vmpressure(sc->gfp_mask, memcg, false,
2433 sc->nr_scanned - scanned,
2434 sc->nr_reclaimed - reclaimed);
2435
2436 /*
2437 * Direct reclaim and kswapd have to scan all memory
2438 * cgroups to fulfill the overall scan target for the
2439 * zone.
2440 *
2441 * Limit reclaim, on the other hand, only cares about
2442 * nr_to_reclaim pages to be reclaimed and it will
2443 * retry with decreasing priority if one round over the
2444 * whole hierarchy is not sufficient.
2445 */
2446 if (!global_reclaim(sc) &&
2447 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2448 mem_cgroup_iter_break(root, memcg);
2449 break;
2450 }
2451 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2452
2453 /*
2454 * Shrink the slab caches in the same proportion that
2455 * the eligible LRU pages were scanned.
2456 */
2457 if (global_reclaim(sc) && is_classzone)
2458 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2459 sc->nr_scanned - nr_scanned,
2460 zone_lru_pages);
2461
2462 if (reclaim_state) {
2463 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2464 reclaim_state->reclaimed_slab = 0;
2465 }
2466
2467 /* Record the subtree's reclaim efficiency */
2468 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2469 sc->nr_scanned - nr_scanned,
2470 sc->nr_reclaimed - nr_reclaimed);
2471
2472 if (sc->nr_reclaimed - nr_reclaimed)
2473 reclaimable = true;
2474
2475 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2476 sc->nr_scanned - nr_scanned, sc));
2477
2478 return reclaimable;
2479}
2480
2481/*
2482 * Returns true if compaction should go ahead for a high-order request, or
2483 * the high-order allocation would succeed without compaction.
2484 */
2485static inline bool compaction_ready(struct zone *zone, int order)
2486{
2487 unsigned long balance_gap, watermark;
2488 bool watermark_ok;
2489
2490 /*
2491 * Compaction takes time to run and there are potentially other
2492 * callers using the pages just freed. Continue reclaiming until
2493 * there is a buffer of free pages available to give compaction
2494 * a reasonable chance of completing and allocating the page
2495 */
2496 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2497 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2498 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2499 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2500
2501 /*
2502 * If compaction is deferred, reclaim up to a point where
2503 * compaction will have a chance of success when re-enabled
2504 */
2505 if (compaction_deferred(zone, order))
2506 return watermark_ok;
2507
2508 /*
2509 * If compaction is not ready to start and allocation is not likely
2510 * to succeed without it, then keep reclaiming.
2511 */
2512 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2513 return false;
2514
2515 return watermark_ok;
2516}
2517
2518/*
2519 * This is the direct reclaim path, for page-allocating processes. We only
2520 * try to reclaim pages from zones which will satisfy the caller's allocation
2521 * request.
2522 *
2523 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2524 * Because:
2525 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2526 * allocation or
2527 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2528 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2529 * zone defense algorithm.
2530 *
2531 * If a zone is deemed to be full of pinned pages then just give it a light
2532 * scan then give up on it.
2533 *
2534 * Returns true if a zone was reclaimable.
2535 */
2536static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2537{
2538 struct zoneref *z;
2539 struct zone *zone;
2540 unsigned long nr_soft_reclaimed;
2541 unsigned long nr_soft_scanned;
2542 gfp_t orig_mask;
2543 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2544 bool reclaimable = false;
2545
2546 /*
2547 * If the number of buffer_heads in the machine exceeds the maximum
2548 * allowed level, force direct reclaim to scan the highmem zone as
2549 * highmem pages could be pinning lowmem pages storing buffer_heads
2550 */
2551 orig_mask = sc->gfp_mask;
2552 if (buffer_heads_over_limit)
2553 sc->gfp_mask |= __GFP_HIGHMEM;
2554
2555 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2556 gfp_zone(sc->gfp_mask), sc->nodemask) {
2557 enum zone_type classzone_idx;
2558
2559 if (!populated_zone(zone))
2560 continue;
2561
2562 classzone_idx = requested_highidx;
2563 while (!populated_zone(zone->zone_pgdat->node_zones +
2564 classzone_idx))
2565 classzone_idx--;
2566
2567 /*
2568 * Take care memory controller reclaiming has small influence
2569 * to global LRU.
2570 */
2571 if (global_reclaim(sc)) {
2572 if (!cpuset_zone_allowed(zone,
2573 GFP_KERNEL | __GFP_HARDWALL))
2574 continue;
2575
2576 if (sc->priority != DEF_PRIORITY &&
2577 !zone_reclaimable(zone))
2578 continue; /* Let kswapd poll it */
2579
2580 /*
2581 * If we already have plenty of memory free for
2582 * compaction in this zone, don't free any more.
2583 * Even though compaction is invoked for any
2584 * non-zero order, only frequent costly order
2585 * reclamation is disruptive enough to become a
2586 * noticeable problem, like transparent huge
2587 * page allocations.
2588 */
2589 if (IS_ENABLED(CONFIG_COMPACTION) &&
2590 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2591 zonelist_zone_idx(z) <= requested_highidx &&
2592 compaction_ready(zone, sc->order)) {
2593 sc->compaction_ready = true;
2594 continue;
2595 }
2596
2597 /*
2598 * This steals pages from memory cgroups over softlimit
2599 * and returns the number of reclaimed pages and
2600 * scanned pages. This works for global memory pressure
2601 * and balancing, not for a memcg's limit.
2602 */
2603 nr_soft_scanned = 0;
2604 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2605 sc->order, sc->gfp_mask,
2606 &nr_soft_scanned);
2607 sc->nr_reclaimed += nr_soft_reclaimed;
2608 sc->nr_scanned += nr_soft_scanned;
2609 if (nr_soft_reclaimed)
2610 reclaimable = true;
2611 /* need some check for avoid more shrink_zone() */
2612 }
2613
2614 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2615 reclaimable = true;
2616
2617 if (global_reclaim(sc) &&
2618 !reclaimable && zone_reclaimable(zone))
2619 reclaimable = true;
2620 }
2621
2622 /*
2623 * Restore to original mask to avoid the impact on the caller if we
2624 * promoted it to __GFP_HIGHMEM.
2625 */
2626 sc->gfp_mask = orig_mask;
2627
2628 return reclaimable;
2629}
2630
2631/*
2632 * This is the main entry point to direct page reclaim.
2633 *
2634 * If a full scan of the inactive list fails to free enough memory then we
2635 * are "out of memory" and something needs to be killed.
2636 *
2637 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2638 * high - the zone may be full of dirty or under-writeback pages, which this
2639 * caller can't do much about. We kick the writeback threads and take explicit
2640 * naps in the hope that some of these pages can be written. But if the
2641 * allocating task holds filesystem locks which prevent writeout this might not
2642 * work, and the allocation attempt will fail.
2643 *
2644 * returns: 0, if no pages reclaimed
2645 * else, the number of pages reclaimed
2646 */
2647static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2648 struct scan_control *sc)
2649{
2650 int initial_priority = sc->priority;
2651 unsigned long total_scanned = 0;
2652 unsigned long writeback_threshold;
2653 bool zones_reclaimable;
2654retry:
2655 delayacct_freepages_start();
2656
2657 if (global_reclaim(sc))
2658 count_vm_event(ALLOCSTALL);
2659
2660 do {
2661 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2662 sc->priority);
2663 sc->nr_scanned = 0;
2664 zones_reclaimable = shrink_zones(zonelist, sc);
2665
2666 total_scanned += sc->nr_scanned;
2667 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2668 break;
2669
2670 if (sc->compaction_ready)
2671 break;
2672
2673 /*
2674 * If we're getting trouble reclaiming, start doing
2675 * writepage even in laptop mode.
2676 */
2677 if (sc->priority < DEF_PRIORITY - 2)
2678 sc->may_writepage = 1;
2679
2680 /*
2681 * Try to write back as many pages as we just scanned. This
2682 * tends to cause slow streaming writers to write data to the
2683 * disk smoothly, at the dirtying rate, which is nice. But
2684 * that's undesirable in laptop mode, where we *want* lumpy
2685 * writeout. So in laptop mode, write out the whole world.
2686 */
2687 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2688 if (total_scanned > writeback_threshold) {
2689 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2690 WB_REASON_TRY_TO_FREE_PAGES);
2691 sc->may_writepage = 1;
2692 }
2693 } while (--sc->priority >= 0);
2694
2695 delayacct_freepages_end();
2696
2697 if (sc->nr_reclaimed)
2698 return sc->nr_reclaimed;
2699
2700 /* Aborted reclaim to try compaction? don't OOM, then */
2701 if (sc->compaction_ready)
2702 return 1;
2703
2704 /* Untapped cgroup reserves? Don't OOM, retry. */
2705 if (!sc->may_thrash) {
2706 sc->priority = initial_priority;
2707 sc->may_thrash = 1;
2708 goto retry;
2709 }
2710
2711 /* Any of the zones still reclaimable? Don't OOM. */
2712 if (zones_reclaimable)
2713 return 1;
2714
2715 return 0;
2716}
2717
2718static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2719{
2720 struct zone *zone;
2721 unsigned long pfmemalloc_reserve = 0;
2722 unsigned long free_pages = 0;
2723 int i;
2724 bool wmark_ok;
2725
2726 for (i = 0; i <= ZONE_NORMAL; i++) {
2727 zone = &pgdat->node_zones[i];
2728 if (!populated_zone(zone) ||
2729 zone_reclaimable_pages(zone) == 0)
2730 continue;
2731
2732 pfmemalloc_reserve += min_wmark_pages(zone);
2733 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2734 }
2735
2736 /* If there are no reserves (unexpected config) then do not throttle */
2737 if (!pfmemalloc_reserve)
2738 return true;
2739
2740 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2741
2742 /* kswapd must be awake if processes are being throttled */
2743 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2744 pgdat->classzone_idx = min(pgdat->classzone_idx,
2745 (enum zone_type)ZONE_NORMAL);
2746 wake_up_interruptible(&pgdat->kswapd_wait);
2747 }
2748
2749 return wmark_ok;
2750}
2751
2752/*
2753 * Throttle direct reclaimers if backing storage is backed by the network
2754 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2755 * depleted. kswapd will continue to make progress and wake the processes
2756 * when the low watermark is reached.
2757 *
2758 * Returns true if a fatal signal was delivered during throttling. If this
2759 * happens, the page allocator should not consider triggering the OOM killer.
2760 */
2761static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2762 nodemask_t *nodemask)
2763{
2764 struct zoneref *z;
2765 struct zone *zone;
2766 pg_data_t *pgdat = NULL;
2767
2768 /*
2769 * Kernel threads should not be throttled as they may be indirectly
2770 * responsible for cleaning pages necessary for reclaim to make forward
2771 * progress. kjournald for example may enter direct reclaim while
2772 * committing a transaction where throttling it could forcing other
2773 * processes to block on log_wait_commit().
2774 */
2775 if (current->flags & PF_KTHREAD)
2776 goto out;
2777
2778 /*
2779 * If a fatal signal is pending, this process should not throttle.
2780 * It should return quickly so it can exit and free its memory
2781 */
2782 if (fatal_signal_pending(current))
2783 goto out;
2784
2785 /*
2786 * Check if the pfmemalloc reserves are ok by finding the first node
2787 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2788 * GFP_KERNEL will be required for allocating network buffers when
2789 * swapping over the network so ZONE_HIGHMEM is unusable.
2790 *
2791 * Throttling is based on the first usable node and throttled processes
2792 * wait on a queue until kswapd makes progress and wakes them. There
2793 * is an affinity then between processes waking up and where reclaim
2794 * progress has been made assuming the process wakes on the same node.
2795 * More importantly, processes running on remote nodes will not compete
2796 * for remote pfmemalloc reserves and processes on different nodes
2797 * should make reasonable progress.
2798 */
2799 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2800 gfp_zone(gfp_mask), nodemask) {
2801 if (zone_idx(zone) > ZONE_NORMAL)
2802 continue;
2803
2804 /* Throttle based on the first usable node */
2805 pgdat = zone->zone_pgdat;
2806 if (pfmemalloc_watermark_ok(pgdat))
2807 goto out;
2808 break;
2809 }
2810
2811 /* If no zone was usable by the allocation flags then do not throttle */
2812 if (!pgdat)
2813 goto out;
2814
2815 /* Account for the throttling */
2816 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2817
2818 /*
2819 * If the caller cannot enter the filesystem, it's possible that it
2820 * is due to the caller holding an FS lock or performing a journal
2821 * transaction in the case of a filesystem like ext[3|4]. In this case,
2822 * it is not safe to block on pfmemalloc_wait as kswapd could be
2823 * blocked waiting on the same lock. Instead, throttle for up to a
2824 * second before continuing.
2825 */
2826 if (!(gfp_mask & __GFP_FS)) {
2827 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2828 pfmemalloc_watermark_ok(pgdat), HZ);
2829
2830 goto check_pending;
2831 }
2832
2833 /* Throttle until kswapd wakes the process */
2834 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2835 pfmemalloc_watermark_ok(pgdat));
2836
2837check_pending:
2838 if (fatal_signal_pending(current))
2839 return true;
2840
2841out:
2842 return false;
2843}
2844
2845unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2846 gfp_t gfp_mask, nodemask_t *nodemask)
2847{
2848 unsigned long nr_reclaimed;
2849 struct scan_control sc = {
2850 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2851 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2852 .order = order,
2853 .nodemask = nodemask,
2854 .priority = DEF_PRIORITY,
2855 .may_writepage = !laptop_mode,
2856 .may_unmap = 1,
2857 .may_swap = 1,
2858 };
2859
2860 /*
2861 * Do not enter reclaim if fatal signal was delivered while throttled.
2862 * 1 is returned so that the page allocator does not OOM kill at this
2863 * point.
2864 */
2865 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2866 return 1;
2867
2868 trace_mm_vmscan_direct_reclaim_begin(order,
2869 sc.may_writepage,
2870 gfp_mask);
2871
2872 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2873
2874 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2875
2876 return nr_reclaimed;
2877}
2878
2879#ifdef CONFIG_MEMCG
2880
2881unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2882 gfp_t gfp_mask, bool noswap,
2883 struct zone *zone,
2884 unsigned long *nr_scanned)
2885{
2886 struct scan_control sc = {
2887 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2888 .target_mem_cgroup = memcg,
2889 .may_writepage = !laptop_mode,
2890 .may_unmap = 1,
2891 .may_swap = !noswap,
2892 };
2893 unsigned long lru_pages;
2894
2895 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2896 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2897
2898 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2899 sc.may_writepage,
2900 sc.gfp_mask);
2901
2902 /*
2903 * NOTE: Although we can get the priority field, using it
2904 * here is not a good idea, since it limits the pages we can scan.
2905 * if we don't reclaim here, the shrink_zone from balance_pgdat
2906 * will pick up pages from other mem cgroup's as well. We hack
2907 * the priority and make it zero.
2908 */
2909 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
2910
2911 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2912
2913 *nr_scanned = sc.nr_scanned;
2914 return sc.nr_reclaimed;
2915}
2916
2917unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2918 unsigned long nr_pages,
2919 gfp_t gfp_mask,
2920 bool may_swap)
2921{
2922 struct zonelist *zonelist;
2923 unsigned long nr_reclaimed;
2924 int nid;
2925 struct scan_control sc = {
2926 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2927 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2928 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2929 .target_mem_cgroup = memcg,
2930 .priority = DEF_PRIORITY,
2931 .may_writepage = !laptop_mode,
2932 .may_unmap = 1,
2933 .may_swap = may_swap,
2934 };
2935
2936 /*
2937 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2938 * take care of from where we get pages. So the node where we start the
2939 * scan does not need to be the current node.
2940 */
2941 nid = mem_cgroup_select_victim_node(memcg);
2942
2943 zonelist = NODE_DATA(nid)->node_zonelists;
2944
2945 trace_mm_vmscan_memcg_reclaim_begin(0,
2946 sc.may_writepage,
2947 sc.gfp_mask);
2948
2949 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2950
2951 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2952
2953 return nr_reclaimed;
2954}
2955#endif
2956
2957static void age_active_anon(struct zone *zone, struct scan_control *sc)
2958{
2959 struct mem_cgroup *memcg;
2960
2961 if (!total_swap_pages)
2962 return;
2963
2964 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2965 do {
2966 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2967
2968 if (inactive_anon_is_low(lruvec))
2969 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2970 sc, LRU_ACTIVE_ANON);
2971
2972 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2973 } while (memcg);
2974}
2975
2976static bool zone_balanced(struct zone *zone, int order, bool highorder,
2977 unsigned long balance_gap, int classzone_idx)
2978{
2979 unsigned long mark = high_wmark_pages(zone) + balance_gap;
2980
2981 /*
2982 * When checking from pgdat_balanced(), kswapd should stop and sleep
2983 * when it reaches the high order-0 watermark and let kcompactd take
2984 * over. Other callers such as wakeup_kswapd() want to determine the
2985 * true high-order watermark.
2986 */
2987 if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) {
2988 mark += (1UL << order);
2989 order = 0;
2990 }
2991
2992 return zone_watermark_ok_safe(zone, order, mark, classzone_idx);
2993}
2994
2995/*
2996 * pgdat_balanced() is used when checking if a node is balanced.
2997 *
2998 * For order-0, all zones must be balanced!
2999 *
3000 * For high-order allocations only zones that meet watermarks and are in a
3001 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3002 * total of balanced pages must be at least 25% of the zones allowed by
3003 * classzone_idx for the node to be considered balanced. Forcing all zones to
3004 * be balanced for high orders can cause excessive reclaim when there are
3005 * imbalanced zones.
3006 * The choice of 25% is due to
3007 * o a 16M DMA zone that is balanced will not balance a zone on any
3008 * reasonable sized machine
3009 * o On all other machines, the top zone must be at least a reasonable
3010 * percentage of the middle zones. For example, on 32-bit x86, highmem
3011 * would need to be at least 256M for it to be balance a whole node.
3012 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3013 * to balance a node on its own. These seemed like reasonable ratios.
3014 */
3015static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3016{
3017 unsigned long managed_pages = 0;
3018 unsigned long balanced_pages = 0;
3019 int i;
3020
3021 /* Check the watermark levels */
3022 for (i = 0; i <= classzone_idx; i++) {
3023 struct zone *zone = pgdat->node_zones + i;
3024
3025 if (!populated_zone(zone))
3026 continue;
3027
3028 managed_pages += zone->managed_pages;
3029
3030 /*
3031 * A special case here:
3032 *
3033 * balance_pgdat() skips over all_unreclaimable after
3034 * DEF_PRIORITY. Effectively, it considers them balanced so
3035 * they must be considered balanced here as well!
3036 */
3037 if (!zone_reclaimable(zone)) {
3038 balanced_pages += zone->managed_pages;
3039 continue;
3040 }
3041
3042 if (zone_balanced(zone, order, false, 0, i))
3043 balanced_pages += zone->managed_pages;
3044 else if (!order)
3045 return false;
3046 }
3047
3048 if (order)
3049 return balanced_pages >= (managed_pages >> 2);
3050 else
3051 return true;
3052}
3053
3054/*
3055 * Prepare kswapd for sleeping. This verifies that there are no processes
3056 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3057 *
3058 * Returns true if kswapd is ready to sleep
3059 */
3060static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3061 int classzone_idx)
3062{
3063 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3064 if (remaining)
3065 return false;
3066
3067 /*
3068 * The throttled processes are normally woken up in balance_pgdat() as
3069 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3070 * race between when kswapd checks the watermarks and a process gets
3071 * throttled. There is also a potential race if processes get
3072 * throttled, kswapd wakes, a large process exits thereby balancing the
3073 * zones, which causes kswapd to exit balance_pgdat() before reaching
3074 * the wake up checks. If kswapd is going to sleep, no process should
3075 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3076 * the wake up is premature, processes will wake kswapd and get
3077 * throttled again. The difference from wake ups in balance_pgdat() is
3078 * that here we are under prepare_to_wait().
3079 */
3080 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3081 wake_up_all(&pgdat->pfmemalloc_wait);
3082
3083 return pgdat_balanced(pgdat, order, classzone_idx);
3084}
3085
3086/*
3087 * kswapd shrinks the zone by the number of pages required to reach
3088 * the high watermark.
3089 *
3090 * Returns true if kswapd scanned at least the requested number of pages to
3091 * reclaim or if the lack of progress was due to pages under writeback.
3092 * This is used to determine if the scanning priority needs to be raised.
3093 */
3094static bool kswapd_shrink_zone(struct zone *zone,
3095 int classzone_idx,
3096 struct scan_control *sc)
3097{
3098 unsigned long balance_gap;
3099 bool lowmem_pressure;
3100
3101 /* Reclaim above the high watermark. */
3102 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3103
3104 /*
3105 * We put equal pressure on every zone, unless one zone has way too
3106 * many pages free already. The "too many pages" is defined as the
3107 * high wmark plus a "gap" where the gap is either the low
3108 * watermark or 1% of the zone, whichever is smaller.
3109 */
3110 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3111 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3112
3113 /*
3114 * If there is no low memory pressure or the zone is balanced then no
3115 * reclaim is necessary
3116 */
3117 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3118 if (!lowmem_pressure && zone_balanced(zone, sc->order, false,
3119 balance_gap, classzone_idx))
3120 return true;
3121
3122 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3123
3124 clear_bit(ZONE_WRITEBACK, &zone->flags);
3125
3126 /*
3127 * If a zone reaches its high watermark, consider it to be no longer
3128 * congested. It's possible there are dirty pages backed by congested
3129 * BDIs but as pressure is relieved, speculatively avoid congestion
3130 * waits.
3131 */
3132 if (zone_reclaimable(zone) &&
3133 zone_balanced(zone, sc->order, false, 0, classzone_idx)) {
3134 clear_bit(ZONE_CONGESTED, &zone->flags);
3135 clear_bit(ZONE_DIRTY, &zone->flags);
3136 }
3137
3138 return sc->nr_scanned >= sc->nr_to_reclaim;
3139}
3140
3141/*
3142 * For kswapd, balance_pgdat() will work across all this node's zones until
3143 * they are all at high_wmark_pages(zone).
3144 *
3145 * Returns the highest zone idx kswapd was reclaiming at
3146 *
3147 * There is special handling here for zones which are full of pinned pages.
3148 * This can happen if the pages are all mlocked, or if they are all used by
3149 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3150 * What we do is to detect the case where all pages in the zone have been
3151 * scanned twice and there has been zero successful reclaim. Mark the zone as
3152 * dead and from now on, only perform a short scan. Basically we're polling
3153 * the zone for when the problem goes away.
3154 *
3155 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3156 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3157 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3158 * lower zones regardless of the number of free pages in the lower zones. This
3159 * interoperates with the page allocator fallback scheme to ensure that aging
3160 * of pages is balanced across the zones.
3161 */
3162static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3163{
3164 int i;
3165 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3166 unsigned long nr_soft_reclaimed;
3167 unsigned long nr_soft_scanned;
3168 struct scan_control sc = {
3169 .gfp_mask = GFP_KERNEL,
3170 .order = order,
3171 .priority = DEF_PRIORITY,
3172 .may_writepage = !laptop_mode,
3173 .may_unmap = 1,
3174 .may_swap = 1,
3175 };
3176 count_vm_event(PAGEOUTRUN);
3177
3178 do {
3179 bool raise_priority = true;
3180
3181 sc.nr_reclaimed = 0;
3182
3183 /*
3184 * Scan in the highmem->dma direction for the highest
3185 * zone which needs scanning
3186 */
3187 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3188 struct zone *zone = pgdat->node_zones + i;
3189
3190 if (!populated_zone(zone))
3191 continue;
3192
3193 if (sc.priority != DEF_PRIORITY &&
3194 !zone_reclaimable(zone))
3195 continue;
3196
3197 /*
3198 * Do some background aging of the anon list, to give
3199 * pages a chance to be referenced before reclaiming.
3200 */
3201 age_active_anon(zone, &sc);
3202
3203 /*
3204 * If the number of buffer_heads in the machine
3205 * exceeds the maximum allowed level and this node
3206 * has a highmem zone, force kswapd to reclaim from
3207 * it to relieve lowmem pressure.
3208 */
3209 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3210 end_zone = i;
3211 break;
3212 }
3213
3214 if (!zone_balanced(zone, order, false, 0, 0)) {
3215 end_zone = i;
3216 break;
3217 } else {
3218 /*
3219 * If balanced, clear the dirty and congested
3220 * flags
3221 */
3222 clear_bit(ZONE_CONGESTED, &zone->flags);
3223 clear_bit(ZONE_DIRTY, &zone->flags);
3224 }
3225 }
3226
3227 if (i < 0)
3228 goto out;
3229
3230 /*
3231 * If we're getting trouble reclaiming, start doing writepage
3232 * even in laptop mode.
3233 */
3234 if (sc.priority < DEF_PRIORITY - 2)
3235 sc.may_writepage = 1;
3236
3237 /*
3238 * Now scan the zone in the dma->highmem direction, stopping
3239 * at the last zone which needs scanning.
3240 *
3241 * We do this because the page allocator works in the opposite
3242 * direction. This prevents the page allocator from allocating
3243 * pages behind kswapd's direction of progress, which would
3244 * cause too much scanning of the lower zones.
3245 */
3246 for (i = 0; i <= end_zone; i++) {
3247 struct zone *zone = pgdat->node_zones + i;
3248
3249 if (!populated_zone(zone))
3250 continue;
3251
3252 if (sc.priority != DEF_PRIORITY &&
3253 !zone_reclaimable(zone))
3254 continue;
3255
3256 sc.nr_scanned = 0;
3257
3258 nr_soft_scanned = 0;
3259 /*
3260 * Call soft limit reclaim before calling shrink_zone.
3261 */
3262 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3263 order, sc.gfp_mask,
3264 &nr_soft_scanned);
3265 sc.nr_reclaimed += nr_soft_reclaimed;
3266
3267 /*
3268 * There should be no need to raise the scanning
3269 * priority if enough pages are already being scanned
3270 * that that high watermark would be met at 100%
3271 * efficiency.
3272 */
3273 if (kswapd_shrink_zone(zone, end_zone, &sc))
3274 raise_priority = false;
3275 }
3276
3277 /*
3278 * If the low watermark is met there is no need for processes
3279 * to be throttled on pfmemalloc_wait as they should not be
3280 * able to safely make forward progress. Wake them
3281 */
3282 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3283 pfmemalloc_watermark_ok(pgdat))
3284 wake_up_all(&pgdat->pfmemalloc_wait);
3285
3286 /* Check if kswapd should be suspending */
3287 if (try_to_freeze() || kthread_should_stop())
3288 break;
3289
3290 /*
3291 * Raise priority if scanning rate is too low or there was no
3292 * progress in reclaiming pages
3293 */
3294 if (raise_priority || !sc.nr_reclaimed)
3295 sc.priority--;
3296 } while (sc.priority >= 1 &&
3297 !pgdat_balanced(pgdat, order, classzone_idx));
3298
3299out:
3300 /*
3301 * Return the highest zone idx we were reclaiming at so
3302 * prepare_kswapd_sleep() makes the same decisions as here.
3303 */
3304 return end_zone;
3305}
3306
3307static void kswapd_try_to_sleep(pg_data_t *pgdat, int order,
3308 int classzone_idx, int balanced_classzone_idx)
3309{
3310 long remaining = 0;
3311 DEFINE_WAIT(wait);
3312
3313 if (freezing(current) || kthread_should_stop())
3314 return;
3315
3316 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3317
3318 /* Try to sleep for a short interval */
3319 if (prepare_kswapd_sleep(pgdat, order, remaining,
3320 balanced_classzone_idx)) {
3321 /*
3322 * Compaction records what page blocks it recently failed to
3323 * isolate pages from and skips them in the future scanning.
3324 * When kswapd is going to sleep, it is reasonable to assume
3325 * that pages and compaction may succeed so reset the cache.
3326 */
3327 reset_isolation_suitable(pgdat);
3328
3329 /*
3330 * We have freed the memory, now we should compact it to make
3331 * allocation of the requested order possible.
3332 */
3333 wakeup_kcompactd(pgdat, order, classzone_idx);
3334
3335 remaining = schedule_timeout(HZ/10);
3336 finish_wait(&pgdat->kswapd_wait, &wait);
3337 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3338 }
3339
3340 /*
3341 * After a short sleep, check if it was a premature sleep. If not, then
3342 * go fully to sleep until explicitly woken up.
3343 */
3344 if (prepare_kswapd_sleep(pgdat, order, remaining,
3345 balanced_classzone_idx)) {
3346 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3347
3348 /*
3349 * vmstat counters are not perfectly accurate and the estimated
3350 * value for counters such as NR_FREE_PAGES can deviate from the
3351 * true value by nr_online_cpus * threshold. To avoid the zone
3352 * watermarks being breached while under pressure, we reduce the
3353 * per-cpu vmstat threshold while kswapd is awake and restore
3354 * them before going back to sleep.
3355 */
3356 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3357
3358 if (!kthread_should_stop())
3359 schedule();
3360
3361 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3362 } else {
3363 if (remaining)
3364 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3365 else
3366 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3367 }
3368 finish_wait(&pgdat->kswapd_wait, &wait);
3369}
3370
3371/*
3372 * The background pageout daemon, started as a kernel thread
3373 * from the init process.
3374 *
3375 * This basically trickles out pages so that we have _some_
3376 * free memory available even if there is no other activity
3377 * that frees anything up. This is needed for things like routing
3378 * etc, where we otherwise might have all activity going on in
3379 * asynchronous contexts that cannot page things out.
3380 *
3381 * If there are applications that are active memory-allocators
3382 * (most normal use), this basically shouldn't matter.
3383 */
3384static int kswapd(void *p)
3385{
3386 unsigned long order, new_order;
3387 int classzone_idx, new_classzone_idx;
3388 int balanced_classzone_idx;
3389 pg_data_t *pgdat = (pg_data_t*)p;
3390 struct task_struct *tsk = current;
3391
3392 struct reclaim_state reclaim_state = {
3393 .reclaimed_slab = 0,
3394 };
3395 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3396
3397 lockdep_set_current_reclaim_state(GFP_KERNEL);
3398
3399 if (!cpumask_empty(cpumask))
3400 set_cpus_allowed_ptr(tsk, cpumask);
3401 current->reclaim_state = &reclaim_state;
3402
3403 /*
3404 * Tell the memory management that we're a "memory allocator",
3405 * and that if we need more memory we should get access to it
3406 * regardless (see "__alloc_pages()"). "kswapd" should
3407 * never get caught in the normal page freeing logic.
3408 *
3409 * (Kswapd normally doesn't need memory anyway, but sometimes
3410 * you need a small amount of memory in order to be able to
3411 * page out something else, and this flag essentially protects
3412 * us from recursively trying to free more memory as we're
3413 * trying to free the first piece of memory in the first place).
3414 */
3415 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3416 set_freezable();
3417
3418 order = new_order = 0;
3419 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3420 balanced_classzone_idx = classzone_idx;
3421 for ( ; ; ) {
3422 bool ret;
3423
3424 /*
3425 * While we were reclaiming, there might have been another
3426 * wakeup, so check the values.
3427 */
3428 new_order = pgdat->kswapd_max_order;
3429 new_classzone_idx = pgdat->classzone_idx;
3430 pgdat->kswapd_max_order = 0;
3431 pgdat->classzone_idx = pgdat->nr_zones - 1;
3432
3433 if (order < new_order || classzone_idx > new_classzone_idx) {
3434 /*
3435 * Don't sleep if someone wants a larger 'order'
3436 * allocation or has tigher zone constraints
3437 */
3438 order = new_order;
3439 classzone_idx = new_classzone_idx;
3440 } else {
3441 kswapd_try_to_sleep(pgdat, order, classzone_idx,
3442 balanced_classzone_idx);
3443 order = pgdat->kswapd_max_order;
3444 classzone_idx = pgdat->classzone_idx;
3445 new_order = order;
3446 new_classzone_idx = classzone_idx;
3447 pgdat->kswapd_max_order = 0;
3448 pgdat->classzone_idx = pgdat->nr_zones - 1;
3449 }
3450
3451 ret = try_to_freeze();
3452 if (kthread_should_stop())
3453 break;
3454
3455 /*
3456 * We can speed up thawing tasks if we don't call balance_pgdat
3457 * after returning from the refrigerator
3458 */
3459 if (!ret) {
3460 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3461 balanced_classzone_idx = balance_pgdat(pgdat, order,
3462 classzone_idx);
3463 }
3464 }
3465
3466 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3467 current->reclaim_state = NULL;
3468 lockdep_clear_current_reclaim_state();
3469
3470 return 0;
3471}
3472
3473/*
3474 * A zone is low on free memory, so wake its kswapd task to service it.
3475 */
3476void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3477{
3478 pg_data_t *pgdat;
3479
3480 if (!populated_zone(zone))
3481 return;
3482
3483 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3484 return;
3485 pgdat = zone->zone_pgdat;
3486 if (pgdat->kswapd_max_order < order) {
3487 pgdat->kswapd_max_order = order;
3488 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3489 }
3490 if (!waitqueue_active(&pgdat->kswapd_wait))
3491 return;
3492 if (zone_balanced(zone, order, true, 0, 0))
3493 return;
3494
3495 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3496 wake_up_interruptible(&pgdat->kswapd_wait);
3497}
3498
3499#ifdef CONFIG_HIBERNATION
3500/*
3501 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3502 * freed pages.
3503 *
3504 * Rather than trying to age LRUs the aim is to preserve the overall
3505 * LRU order by reclaiming preferentially
3506 * inactive > active > active referenced > active mapped
3507 */
3508unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3509{
3510 struct reclaim_state reclaim_state;
3511 struct scan_control sc = {
3512 .nr_to_reclaim = nr_to_reclaim,
3513 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3514 .priority = DEF_PRIORITY,
3515 .may_writepage = 1,
3516 .may_unmap = 1,
3517 .may_swap = 1,
3518 .hibernation_mode = 1,
3519 };
3520 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3521 struct task_struct *p = current;
3522 unsigned long nr_reclaimed;
3523
3524 p->flags |= PF_MEMALLOC;
3525 lockdep_set_current_reclaim_state(sc.gfp_mask);
3526 reclaim_state.reclaimed_slab = 0;
3527 p->reclaim_state = &reclaim_state;
3528
3529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3530
3531 p->reclaim_state = NULL;
3532 lockdep_clear_current_reclaim_state();
3533 p->flags &= ~PF_MEMALLOC;
3534
3535 return nr_reclaimed;
3536}
3537#endif /* CONFIG_HIBERNATION */
3538
3539/* It's optimal to keep kswapds on the same CPUs as their memory, but
3540 not required for correctness. So if the last cpu in a node goes
3541 away, we get changed to run anywhere: as the first one comes back,
3542 restore their cpu bindings. */
3543static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3544 void *hcpu)
3545{
3546 int nid;
3547
3548 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3549 for_each_node_state(nid, N_MEMORY) {
3550 pg_data_t *pgdat = NODE_DATA(nid);
3551 const struct cpumask *mask;
3552
3553 mask = cpumask_of_node(pgdat->node_id);
3554
3555 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3556 /* One of our CPUs online: restore mask */
3557 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3558 }
3559 }
3560 return NOTIFY_OK;
3561}
3562
3563/*
3564 * This kswapd start function will be called by init and node-hot-add.
3565 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3566 */
3567int kswapd_run(int nid)
3568{
3569 pg_data_t *pgdat = NODE_DATA(nid);
3570 int ret = 0;
3571
3572 if (pgdat->kswapd)
3573 return 0;
3574
3575 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3576 if (IS_ERR(pgdat->kswapd)) {
3577 /* failure at boot is fatal */
3578 BUG_ON(system_state == SYSTEM_BOOTING);
3579 pr_err("Failed to start kswapd on node %d\n", nid);
3580 ret = PTR_ERR(pgdat->kswapd);
3581 pgdat->kswapd = NULL;
3582 }
3583 return ret;
3584}
3585
3586/*
3587 * Called by memory hotplug when all memory in a node is offlined. Caller must
3588 * hold mem_hotplug_begin/end().
3589 */
3590void kswapd_stop(int nid)
3591{
3592 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3593
3594 if (kswapd) {
3595 kthread_stop(kswapd);
3596 NODE_DATA(nid)->kswapd = NULL;
3597 }
3598}
3599
3600static int __init kswapd_init(void)
3601{
3602 int nid;
3603
3604 swap_setup();
3605 for_each_node_state(nid, N_MEMORY)
3606 kswapd_run(nid);
3607 hotcpu_notifier(cpu_callback, 0);
3608 return 0;
3609}
3610
3611module_init(kswapd_init)
3612
3613#ifdef CONFIG_NUMA
3614/*
3615 * Zone reclaim mode
3616 *
3617 * If non-zero call zone_reclaim when the number of free pages falls below
3618 * the watermarks.
3619 */
3620int zone_reclaim_mode __read_mostly;
3621
3622#define RECLAIM_OFF 0
3623#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3624#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3625#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3626
3627/*
3628 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3629 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3630 * a zone.
3631 */
3632#define ZONE_RECLAIM_PRIORITY 4
3633
3634/*
3635 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3636 * occur.
3637 */
3638int sysctl_min_unmapped_ratio = 1;
3639
3640/*
3641 * If the number of slab pages in a zone grows beyond this percentage then
3642 * slab reclaim needs to occur.
3643 */
3644int sysctl_min_slab_ratio = 5;
3645
3646static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3647{
3648 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3649 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3650 zone_page_state(zone, NR_ACTIVE_FILE);
3651
3652 /*
3653 * It's possible for there to be more file mapped pages than
3654 * accounted for by the pages on the file LRU lists because
3655 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3656 */
3657 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3658}
3659
3660/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3661static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3662{
3663 unsigned long nr_pagecache_reclaimable;
3664 unsigned long delta = 0;
3665
3666 /*
3667 * If RECLAIM_UNMAP is set, then all file pages are considered
3668 * potentially reclaimable. Otherwise, we have to worry about
3669 * pages like swapcache and zone_unmapped_file_pages() provides
3670 * a better estimate
3671 */
3672 if (zone_reclaim_mode & RECLAIM_UNMAP)
3673 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3674 else
3675 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3676
3677 /* If we can't clean pages, remove dirty pages from consideration */
3678 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3679 delta += zone_page_state(zone, NR_FILE_DIRTY);
3680
3681 /* Watch for any possible underflows due to delta */
3682 if (unlikely(delta > nr_pagecache_reclaimable))
3683 delta = nr_pagecache_reclaimable;
3684
3685 return nr_pagecache_reclaimable - delta;
3686}
3687
3688/*
3689 * Try to free up some pages from this zone through reclaim.
3690 */
3691static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3692{
3693 /* Minimum pages needed in order to stay on node */
3694 const unsigned long nr_pages = 1 << order;
3695 struct task_struct *p = current;
3696 struct reclaim_state reclaim_state;
3697 struct scan_control sc = {
3698 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3699 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3700 .order = order,
3701 .priority = ZONE_RECLAIM_PRIORITY,
3702 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3703 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3704 .may_swap = 1,
3705 };
3706
3707 cond_resched();
3708 /*
3709 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3710 * and we also need to be able to write out pages for RECLAIM_WRITE
3711 * and RECLAIM_UNMAP.
3712 */
3713 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3714 lockdep_set_current_reclaim_state(gfp_mask);
3715 reclaim_state.reclaimed_slab = 0;
3716 p->reclaim_state = &reclaim_state;
3717
3718 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3719 /*
3720 * Free memory by calling shrink zone with increasing
3721 * priorities until we have enough memory freed.
3722 */
3723 do {
3724 shrink_zone(zone, &sc, true);
3725 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3726 }
3727
3728 p->reclaim_state = NULL;
3729 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3730 lockdep_clear_current_reclaim_state();
3731 return sc.nr_reclaimed >= nr_pages;
3732}
3733
3734int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3735{
3736 int node_id;
3737 int ret;
3738
3739 /*
3740 * Zone reclaim reclaims unmapped file backed pages and
3741 * slab pages if we are over the defined limits.
3742 *
3743 * A small portion of unmapped file backed pages is needed for
3744 * file I/O otherwise pages read by file I/O will be immediately
3745 * thrown out if the zone is overallocated. So we do not reclaim
3746 * if less than a specified percentage of the zone is used by
3747 * unmapped file backed pages.
3748 */
3749 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3750 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3751 return ZONE_RECLAIM_FULL;
3752
3753 if (!zone_reclaimable(zone))
3754 return ZONE_RECLAIM_FULL;
3755
3756 /*
3757 * Do not scan if the allocation should not be delayed.
3758 */
3759 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3760 return ZONE_RECLAIM_NOSCAN;
3761
3762 /*
3763 * Only run zone reclaim on the local zone or on zones that do not
3764 * have associated processors. This will favor the local processor
3765 * over remote processors and spread off node memory allocations
3766 * as wide as possible.
3767 */
3768 node_id = zone_to_nid(zone);
3769 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3770 return ZONE_RECLAIM_NOSCAN;
3771
3772 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3773 return ZONE_RECLAIM_NOSCAN;
3774
3775 ret = __zone_reclaim(zone, gfp_mask, order);
3776 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3777
3778 if (!ret)
3779 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3780
3781 return ret;
3782}
3783#endif
3784
3785/*
3786 * page_evictable - test whether a page is evictable
3787 * @page: the page to test
3788 *
3789 * Test whether page is evictable--i.e., should be placed on active/inactive
3790 * lists vs unevictable list.
3791 *
3792 * Reasons page might not be evictable:
3793 * (1) page's mapping marked unevictable
3794 * (2) page is part of an mlocked VMA
3795 *
3796 */
3797int page_evictable(struct page *page)
3798{
3799 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3800}
3801
3802#ifdef CONFIG_SHMEM
3803/**
3804 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3805 * @pages: array of pages to check
3806 * @nr_pages: number of pages to check
3807 *
3808 * Checks pages for evictability and moves them to the appropriate lru list.
3809 *
3810 * This function is only used for SysV IPC SHM_UNLOCK.
3811 */
3812void check_move_unevictable_pages(struct page **pages, int nr_pages)
3813{
3814 struct lruvec *lruvec;
3815 struct zone *zone = NULL;
3816 int pgscanned = 0;
3817 int pgrescued = 0;
3818 int i;
3819
3820 for (i = 0; i < nr_pages; i++) {
3821 struct page *page = pages[i];
3822 struct zone *pagezone;
3823
3824 pgscanned++;
3825 pagezone = page_zone(page);
3826 if (pagezone != zone) {
3827 if (zone)
3828 spin_unlock_irq(&zone->lru_lock);
3829 zone = pagezone;
3830 spin_lock_irq(&zone->lru_lock);
3831 }
3832 lruvec = mem_cgroup_page_lruvec(page, zone);
3833
3834 if (!PageLRU(page) || !PageUnevictable(page))
3835 continue;
3836
3837 if (page_evictable(page)) {
3838 enum lru_list lru = page_lru_base_type(page);
3839
3840 VM_BUG_ON_PAGE(PageActive(page), page);
3841 ClearPageUnevictable(page);
3842 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3843 add_page_to_lru_list(page, lruvec, lru);
3844 pgrescued++;
3845 }
3846 }
3847
3848 if (zone) {
3849 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3850 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3851 spin_unlock_irq(&zone->lru_lock);
3852 }
3853}
3854#endif /* CONFIG_SHMEM */