Loading...
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 */
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/pagevec.h>
50#include <linux/prefetch.h>
51#include <linux/printk.h>
52#include <linux/dax.h>
53#include <linux/psi.h>
54
55#include <asm/tlbflush.h>
56#include <asm/div64.h>
57
58#include <linux/swapops.h>
59#include <linux/balloon_compaction.h>
60
61#include "internal.h"
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/vmscan.h>
65
66struct scan_control {
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
70 /*
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 * are scanned.
73 */
74 nodemask_t *nodemask;
75
76 /*
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
79 */
80 struct mem_cgroup *target_mem_cgroup;
81
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
84
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
87
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
90
91 /*
92 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
95 */
96 unsigned int memcg_low_reclaim:1;
97 unsigned int memcg_low_skipped:1;
98
99 unsigned int hibernation_mode:1;
100
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready:1;
103
104 /* Allocation order */
105 s8 order;
106
107 /* Scan (total_size >> priority) pages at once */
108 s8 priority;
109
110 /* The highest zone to isolate pages for reclaim from */
111 s8 reclaim_idx;
112
113 /* This context's GFP mask */
114 gfp_t gfp_mask;
115
116 /* Incremented by the number of inactive pages that were scanned */
117 unsigned long nr_scanned;
118
119 /* Number of pages freed so far during a call to shrink_zones() */
120 unsigned long nr_reclaimed;
121
122 struct {
123 unsigned int dirty;
124 unsigned int unqueued_dirty;
125 unsigned int congested;
126 unsigned int writeback;
127 unsigned int immediate;
128 unsigned int file_taken;
129 unsigned int taken;
130 } nr;
131
132 /* for recording the reclaimed slab by now */
133 struct reclaim_state reclaim_state;
134};
135
136#ifdef ARCH_HAS_PREFETCH
137#define prefetch_prev_lru_page(_page, _base, _field) \
138 do { \
139 if ((_page)->lru.prev != _base) { \
140 struct page *prev; \
141 \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
144 } \
145 } while (0)
146#else
147#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
148#endif
149
150#ifdef ARCH_HAS_PREFETCHW
151#define prefetchw_prev_lru_page(_page, _base, _field) \
152 do { \
153 if ((_page)->lru.prev != _base) { \
154 struct page *prev; \
155 \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
158 } \
159 } while (0)
160#else
161#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
162#endif
163
164/*
165 * From 0 .. 100. Higher means more swappy.
166 */
167int vm_swappiness = 60;
168/*
169 * The total number of pages which are beyond the high watermark within all
170 * zones.
171 */
172unsigned long vm_total_pages;
173
174static void set_task_reclaim_state(struct task_struct *task,
175 struct reclaim_state *rs)
176{
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs && task->reclaim_state);
179
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs && !task->reclaim_state);
182
183 task->reclaim_state = rs;
184}
185
186static LIST_HEAD(shrinker_list);
187static DECLARE_RWSEM(shrinker_rwsem);
188
189#ifdef CONFIG_MEMCG
190/*
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
197 *
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
200 */
201#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202
203static DEFINE_IDR(shrinker_idr);
204static int shrinker_nr_max;
205
206static int prealloc_memcg_shrinker(struct shrinker *shrinker)
207{
208 int id, ret = -ENOMEM;
209
210 down_write(&shrinker_rwsem);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
213 if (id < 0)
214 goto unlock;
215
216 if (id >= shrinker_nr_max) {
217 if (memcg_expand_shrinker_maps(id)) {
218 idr_remove(&shrinker_idr, id);
219 goto unlock;
220 }
221
222 shrinker_nr_max = id + 1;
223 }
224 shrinker->id = id;
225 ret = 0;
226unlock:
227 up_write(&shrinker_rwsem);
228 return ret;
229}
230
231static void unregister_memcg_shrinker(struct shrinker *shrinker)
232{
233 int id = shrinker->id;
234
235 BUG_ON(id < 0);
236
237 down_write(&shrinker_rwsem);
238 idr_remove(&shrinker_idr, id);
239 up_write(&shrinker_rwsem);
240}
241
242static bool global_reclaim(struct scan_control *sc)
243{
244 return !sc->target_mem_cgroup;
245}
246
247/**
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
250 *
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
256 *
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
259 */
260static bool sane_reclaim(struct scan_control *sc)
261{
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
263
264 if (!memcg)
265 return true;
266#ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
268 return true;
269#endif
270 return false;
271}
272
273static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
275 bool congested)
276{
277 struct mem_cgroup_per_node *mn;
278
279 if (!memcg)
280 return;
281
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
284}
285
286static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
288{
289 struct mem_cgroup_per_node *mn;
290
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
293
294}
295#else
296static int prealloc_memcg_shrinker(struct shrinker *shrinker)
297{
298 return 0;
299}
300
301static void unregister_memcg_shrinker(struct shrinker *shrinker)
302{
303}
304
305static bool global_reclaim(struct scan_control *sc)
306{
307 return true;
308}
309
310static bool sane_reclaim(struct scan_control *sc)
311{
312 return true;
313}
314
315static inline void set_memcg_congestion(struct pglist_data *pgdat,
316 struct mem_cgroup *memcg, bool congested)
317{
318}
319
320static inline bool memcg_congested(struct pglist_data *pgdat,
321 struct mem_cgroup *memcg)
322{
323 return false;
324
325}
326#endif
327
328/*
329 * This misses isolated pages which are not accounted for to save counters.
330 * As the data only determines if reclaim or compaction continues, it is
331 * not expected that isolated pages will be a dominating factor.
332 */
333unsigned long zone_reclaimable_pages(struct zone *zone)
334{
335 unsigned long nr;
336
337 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
338 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
339 if (get_nr_swap_pages() > 0)
340 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
341 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
342
343 return nr;
344}
345
346/**
347 * lruvec_lru_size - Returns the number of pages on the given LRU list.
348 * @lruvec: lru vector
349 * @lru: lru to use
350 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
351 */
352unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
353{
354 unsigned long lru_size = 0;
355 int zid;
356
357 if (!mem_cgroup_disabled()) {
358 for (zid = 0; zid < MAX_NR_ZONES; zid++)
359 lru_size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
360 } else
361 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
362
363 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
364 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
365 unsigned long size;
366
367 if (!managed_zone(zone))
368 continue;
369
370 if (!mem_cgroup_disabled())
371 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
372 else
373 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
374 NR_ZONE_LRU_BASE + lru);
375 lru_size -= min(size, lru_size);
376 }
377
378 return lru_size;
379
380}
381
382/*
383 * Add a shrinker callback to be called from the vm.
384 */
385int prealloc_shrinker(struct shrinker *shrinker)
386{
387 unsigned int size = sizeof(*shrinker->nr_deferred);
388
389 if (shrinker->flags & SHRINKER_NUMA_AWARE)
390 size *= nr_node_ids;
391
392 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
393 if (!shrinker->nr_deferred)
394 return -ENOMEM;
395
396 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
397 if (prealloc_memcg_shrinker(shrinker))
398 goto free_deferred;
399 }
400
401 return 0;
402
403free_deferred:
404 kfree(shrinker->nr_deferred);
405 shrinker->nr_deferred = NULL;
406 return -ENOMEM;
407}
408
409void free_prealloced_shrinker(struct shrinker *shrinker)
410{
411 if (!shrinker->nr_deferred)
412 return;
413
414 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
415 unregister_memcg_shrinker(shrinker);
416
417 kfree(shrinker->nr_deferred);
418 shrinker->nr_deferred = NULL;
419}
420
421void register_shrinker_prepared(struct shrinker *shrinker)
422{
423 down_write(&shrinker_rwsem);
424 list_add_tail(&shrinker->list, &shrinker_list);
425#ifdef CONFIG_MEMCG_KMEM
426 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
427 idr_replace(&shrinker_idr, shrinker, shrinker->id);
428#endif
429 up_write(&shrinker_rwsem);
430}
431
432int register_shrinker(struct shrinker *shrinker)
433{
434 int err = prealloc_shrinker(shrinker);
435
436 if (err)
437 return err;
438 register_shrinker_prepared(shrinker);
439 return 0;
440}
441EXPORT_SYMBOL(register_shrinker);
442
443/*
444 * Remove one
445 */
446void unregister_shrinker(struct shrinker *shrinker)
447{
448 if (!shrinker->nr_deferred)
449 return;
450 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
451 unregister_memcg_shrinker(shrinker);
452 down_write(&shrinker_rwsem);
453 list_del(&shrinker->list);
454 up_write(&shrinker_rwsem);
455 kfree(shrinker->nr_deferred);
456 shrinker->nr_deferred = NULL;
457}
458EXPORT_SYMBOL(unregister_shrinker);
459
460#define SHRINK_BATCH 128
461
462static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
463 struct shrinker *shrinker, int priority)
464{
465 unsigned long freed = 0;
466 unsigned long long delta;
467 long total_scan;
468 long freeable;
469 long nr;
470 long new_nr;
471 int nid = shrinkctl->nid;
472 long batch_size = shrinker->batch ? shrinker->batch
473 : SHRINK_BATCH;
474 long scanned = 0, next_deferred;
475
476 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
477 nid = 0;
478
479 freeable = shrinker->count_objects(shrinker, shrinkctl);
480 if (freeable == 0 || freeable == SHRINK_EMPTY)
481 return freeable;
482
483 /*
484 * copy the current shrinker scan count into a local variable
485 * and zero it so that other concurrent shrinker invocations
486 * don't also do this scanning work.
487 */
488 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
489
490 total_scan = nr;
491 if (shrinker->seeks) {
492 delta = freeable >> priority;
493 delta *= 4;
494 do_div(delta, shrinker->seeks);
495 } else {
496 /*
497 * These objects don't require any IO to create. Trim
498 * them aggressively under memory pressure to keep
499 * them from causing refetches in the IO caches.
500 */
501 delta = freeable / 2;
502 }
503
504 total_scan += delta;
505 if (total_scan < 0) {
506 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
507 shrinker->scan_objects, total_scan);
508 total_scan = freeable;
509 next_deferred = nr;
510 } else
511 next_deferred = total_scan;
512
513 /*
514 * We need to avoid excessive windup on filesystem shrinkers
515 * due to large numbers of GFP_NOFS allocations causing the
516 * shrinkers to return -1 all the time. This results in a large
517 * nr being built up so when a shrink that can do some work
518 * comes along it empties the entire cache due to nr >>>
519 * freeable. This is bad for sustaining a working set in
520 * memory.
521 *
522 * Hence only allow the shrinker to scan the entire cache when
523 * a large delta change is calculated directly.
524 */
525 if (delta < freeable / 4)
526 total_scan = min(total_scan, freeable / 2);
527
528 /*
529 * Avoid risking looping forever due to too large nr value:
530 * never try to free more than twice the estimate number of
531 * freeable entries.
532 */
533 if (total_scan > freeable * 2)
534 total_scan = freeable * 2;
535
536 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
537 freeable, delta, total_scan, priority);
538
539 /*
540 * Normally, we should not scan less than batch_size objects in one
541 * pass to avoid too frequent shrinker calls, but if the slab has less
542 * than batch_size objects in total and we are really tight on memory,
543 * we will try to reclaim all available objects, otherwise we can end
544 * up failing allocations although there are plenty of reclaimable
545 * objects spread over several slabs with usage less than the
546 * batch_size.
547 *
548 * We detect the "tight on memory" situations by looking at the total
549 * number of objects we want to scan (total_scan). If it is greater
550 * than the total number of objects on slab (freeable), we must be
551 * scanning at high prio and therefore should try to reclaim as much as
552 * possible.
553 */
554 while (total_scan >= batch_size ||
555 total_scan >= freeable) {
556 unsigned long ret;
557 unsigned long nr_to_scan = min(batch_size, total_scan);
558
559 shrinkctl->nr_to_scan = nr_to_scan;
560 shrinkctl->nr_scanned = nr_to_scan;
561 ret = shrinker->scan_objects(shrinker, shrinkctl);
562 if (ret == SHRINK_STOP)
563 break;
564 freed += ret;
565
566 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
567 total_scan -= shrinkctl->nr_scanned;
568 scanned += shrinkctl->nr_scanned;
569
570 cond_resched();
571 }
572
573 if (next_deferred >= scanned)
574 next_deferred -= scanned;
575 else
576 next_deferred = 0;
577 /*
578 * move the unused scan count back into the shrinker in a
579 * manner that handles concurrent updates. If we exhausted the
580 * scan, there is no need to do an update.
581 */
582 if (next_deferred > 0)
583 new_nr = atomic_long_add_return(next_deferred,
584 &shrinker->nr_deferred[nid]);
585 else
586 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
587
588 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
589 return freed;
590}
591
592#ifdef CONFIG_MEMCG
593static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
594 struct mem_cgroup *memcg, int priority)
595{
596 struct memcg_shrinker_map *map;
597 unsigned long ret, freed = 0;
598 int i;
599
600 if (!mem_cgroup_online(memcg))
601 return 0;
602
603 if (!down_read_trylock(&shrinker_rwsem))
604 return 0;
605
606 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
607 true);
608 if (unlikely(!map))
609 goto unlock;
610
611 for_each_set_bit(i, map->map, shrinker_nr_max) {
612 struct shrink_control sc = {
613 .gfp_mask = gfp_mask,
614 .nid = nid,
615 .memcg = memcg,
616 };
617 struct shrinker *shrinker;
618
619 shrinker = idr_find(&shrinker_idr, i);
620 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
621 if (!shrinker)
622 clear_bit(i, map->map);
623 continue;
624 }
625
626 /* Call non-slab shrinkers even though kmem is disabled */
627 if (!memcg_kmem_enabled() &&
628 !(shrinker->flags & SHRINKER_NONSLAB))
629 continue;
630
631 ret = do_shrink_slab(&sc, shrinker, priority);
632 if (ret == SHRINK_EMPTY) {
633 clear_bit(i, map->map);
634 /*
635 * After the shrinker reported that it had no objects to
636 * free, but before we cleared the corresponding bit in
637 * the memcg shrinker map, a new object might have been
638 * added. To make sure, we have the bit set in this
639 * case, we invoke the shrinker one more time and reset
640 * the bit if it reports that it is not empty anymore.
641 * The memory barrier here pairs with the barrier in
642 * memcg_set_shrinker_bit():
643 *
644 * list_lru_add() shrink_slab_memcg()
645 * list_add_tail() clear_bit()
646 * <MB> <MB>
647 * set_bit() do_shrink_slab()
648 */
649 smp_mb__after_atomic();
650 ret = do_shrink_slab(&sc, shrinker, priority);
651 if (ret == SHRINK_EMPTY)
652 ret = 0;
653 else
654 memcg_set_shrinker_bit(memcg, nid, i);
655 }
656 freed += ret;
657
658 if (rwsem_is_contended(&shrinker_rwsem)) {
659 freed = freed ? : 1;
660 break;
661 }
662 }
663unlock:
664 up_read(&shrinker_rwsem);
665 return freed;
666}
667#else /* CONFIG_MEMCG */
668static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
669 struct mem_cgroup *memcg, int priority)
670{
671 return 0;
672}
673#endif /* CONFIG_MEMCG */
674
675/**
676 * shrink_slab - shrink slab caches
677 * @gfp_mask: allocation context
678 * @nid: node whose slab caches to target
679 * @memcg: memory cgroup whose slab caches to target
680 * @priority: the reclaim priority
681 *
682 * Call the shrink functions to age shrinkable caches.
683 *
684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 * unaware shrinkers will receive a node id of 0 instead.
686 *
687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 * are called only if it is the root cgroup.
689 *
690 * @priority is sc->priority, we take the number of objects and >> by priority
691 * in order to get the scan target.
692 *
693 * Returns the number of reclaimed slab objects.
694 */
695static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
696 struct mem_cgroup *memcg,
697 int priority)
698{
699 unsigned long ret, freed = 0;
700 struct shrinker *shrinker;
701
702 /*
703 * The root memcg might be allocated even though memcg is disabled
704 * via "cgroup_disable=memory" boot parameter. This could make
705 * mem_cgroup_is_root() return false, then just run memcg slab
706 * shrink, but skip global shrink. This may result in premature
707 * oom.
708 */
709 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
710 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
711
712 if (!down_read_trylock(&shrinker_rwsem))
713 goto out;
714
715 list_for_each_entry(shrinker, &shrinker_list, list) {
716 struct shrink_control sc = {
717 .gfp_mask = gfp_mask,
718 .nid = nid,
719 .memcg = memcg,
720 };
721
722 ret = do_shrink_slab(&sc, shrinker, priority);
723 if (ret == SHRINK_EMPTY)
724 ret = 0;
725 freed += ret;
726 /*
727 * Bail out if someone want to register a new shrinker to
728 * prevent the regsitration from being stalled for long periods
729 * by parallel ongoing shrinking.
730 */
731 if (rwsem_is_contended(&shrinker_rwsem)) {
732 freed = freed ? : 1;
733 break;
734 }
735 }
736
737 up_read(&shrinker_rwsem);
738out:
739 cond_resched();
740 return freed;
741}
742
743void drop_slab_node(int nid)
744{
745 unsigned long freed;
746
747 do {
748 struct mem_cgroup *memcg = NULL;
749
750 freed = 0;
751 memcg = mem_cgroup_iter(NULL, NULL, NULL);
752 do {
753 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
754 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
755 } while (freed > 10);
756}
757
758void drop_slab(void)
759{
760 int nid;
761
762 for_each_online_node(nid)
763 drop_slab_node(nid);
764}
765
766static inline int is_page_cache_freeable(struct page *page)
767{
768 /*
769 * A freeable page cache page is referenced only by the caller
770 * that isolated the page, the page cache and optional buffer
771 * heads at page->private.
772 */
773 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
774 HPAGE_PMD_NR : 1;
775 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
776}
777
778static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
779{
780 if (current->flags & PF_SWAPWRITE)
781 return 1;
782 if (!inode_write_congested(inode))
783 return 1;
784 if (inode_to_bdi(inode) == current->backing_dev_info)
785 return 1;
786 return 0;
787}
788
789/*
790 * We detected a synchronous write error writing a page out. Probably
791 * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 * fsync(), msync() or close().
793 *
794 * The tricky part is that after writepage we cannot touch the mapping: nothing
795 * prevents it from being freed up. But we have a ref on the page and once
796 * that page is locked, the mapping is pinned.
797 *
798 * We're allowed to run sleeping lock_page() here because we know the caller has
799 * __GFP_FS.
800 */
801static void handle_write_error(struct address_space *mapping,
802 struct page *page, int error)
803{
804 lock_page(page);
805 if (page_mapping(page) == mapping)
806 mapping_set_error(mapping, error);
807 unlock_page(page);
808}
809
810/* possible outcome of pageout() */
811typedef enum {
812 /* failed to write page out, page is locked */
813 PAGE_KEEP,
814 /* move page to the active list, page is locked */
815 PAGE_ACTIVATE,
816 /* page has been sent to the disk successfully, page is unlocked */
817 PAGE_SUCCESS,
818 /* page is clean and locked */
819 PAGE_CLEAN,
820} pageout_t;
821
822/*
823 * pageout is called by shrink_page_list() for each dirty page.
824 * Calls ->writepage().
825 */
826static pageout_t pageout(struct page *page, struct address_space *mapping,
827 struct scan_control *sc)
828{
829 /*
830 * If the page is dirty, only perform writeback if that write
831 * will be non-blocking. To prevent this allocation from being
832 * stalled by pagecache activity. But note that there may be
833 * stalls if we need to run get_block(). We could test
834 * PagePrivate for that.
835 *
836 * If this process is currently in __generic_file_write_iter() against
837 * this page's queue, we can perform writeback even if that
838 * will block.
839 *
840 * If the page is swapcache, write it back even if that would
841 * block, for some throttling. This happens by accident, because
842 * swap_backing_dev_info is bust: it doesn't reflect the
843 * congestion state of the swapdevs. Easy to fix, if needed.
844 */
845 if (!is_page_cache_freeable(page))
846 return PAGE_KEEP;
847 if (!mapping) {
848 /*
849 * Some data journaling orphaned pages can have
850 * page->mapping == NULL while being dirty with clean buffers.
851 */
852 if (page_has_private(page)) {
853 if (try_to_free_buffers(page)) {
854 ClearPageDirty(page);
855 pr_info("%s: orphaned page\n", __func__);
856 return PAGE_CLEAN;
857 }
858 }
859 return PAGE_KEEP;
860 }
861 if (mapping->a_ops->writepage == NULL)
862 return PAGE_ACTIVATE;
863 if (!may_write_to_inode(mapping->host, sc))
864 return PAGE_KEEP;
865
866 if (clear_page_dirty_for_io(page)) {
867 int res;
868 struct writeback_control wbc = {
869 .sync_mode = WB_SYNC_NONE,
870 .nr_to_write = SWAP_CLUSTER_MAX,
871 .range_start = 0,
872 .range_end = LLONG_MAX,
873 .for_reclaim = 1,
874 };
875
876 SetPageReclaim(page);
877 res = mapping->a_ops->writepage(page, &wbc);
878 if (res < 0)
879 handle_write_error(mapping, page, res);
880 if (res == AOP_WRITEPAGE_ACTIVATE) {
881 ClearPageReclaim(page);
882 return PAGE_ACTIVATE;
883 }
884
885 if (!PageWriteback(page)) {
886 /* synchronous write or broken a_ops? */
887 ClearPageReclaim(page);
888 }
889 trace_mm_vmscan_writepage(page);
890 inc_node_page_state(page, NR_VMSCAN_WRITE);
891 return PAGE_SUCCESS;
892 }
893
894 return PAGE_CLEAN;
895}
896
897/*
898 * Same as remove_mapping, but if the page is removed from the mapping, it
899 * gets returned with a refcount of 0.
900 */
901static int __remove_mapping(struct address_space *mapping, struct page *page,
902 bool reclaimed)
903{
904 unsigned long flags;
905 int refcount;
906
907 BUG_ON(!PageLocked(page));
908 BUG_ON(mapping != page_mapping(page));
909
910 xa_lock_irqsave(&mapping->i_pages, flags);
911 /*
912 * The non racy check for a busy page.
913 *
914 * Must be careful with the order of the tests. When someone has
915 * a ref to the page, it may be possible that they dirty it then
916 * drop the reference. So if PageDirty is tested before page_count
917 * here, then the following race may occur:
918 *
919 * get_user_pages(&page);
920 * [user mapping goes away]
921 * write_to(page);
922 * !PageDirty(page) [good]
923 * SetPageDirty(page);
924 * put_page(page);
925 * !page_count(page) [good, discard it]
926 *
927 * [oops, our write_to data is lost]
928 *
929 * Reversing the order of the tests ensures such a situation cannot
930 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
931 * load is not satisfied before that of page->_refcount.
932 *
933 * Note that if SetPageDirty is always performed via set_page_dirty,
934 * and thus under the i_pages lock, then this ordering is not required.
935 */
936 refcount = 1 + compound_nr(page);
937 if (!page_ref_freeze(page, refcount))
938 goto cannot_free;
939 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
940 if (unlikely(PageDirty(page))) {
941 page_ref_unfreeze(page, refcount);
942 goto cannot_free;
943 }
944
945 if (PageSwapCache(page)) {
946 swp_entry_t swap = { .val = page_private(page) };
947 mem_cgroup_swapout(page, swap);
948 __delete_from_swap_cache(page, swap);
949 xa_unlock_irqrestore(&mapping->i_pages, flags);
950 put_swap_page(page, swap);
951 } else {
952 void (*freepage)(struct page *);
953 void *shadow = NULL;
954
955 freepage = mapping->a_ops->freepage;
956 /*
957 * Remember a shadow entry for reclaimed file cache in
958 * order to detect refaults, thus thrashing, later on.
959 *
960 * But don't store shadows in an address space that is
961 * already exiting. This is not just an optizimation,
962 * inode reclaim needs to empty out the radix tree or
963 * the nodes are lost. Don't plant shadows behind its
964 * back.
965 *
966 * We also don't store shadows for DAX mappings because the
967 * only page cache pages found in these are zero pages
968 * covering holes, and because we don't want to mix DAX
969 * exceptional entries and shadow exceptional entries in the
970 * same address_space.
971 */
972 if (reclaimed && page_is_file_cache(page) &&
973 !mapping_exiting(mapping) && !dax_mapping(mapping))
974 shadow = workingset_eviction(page);
975 __delete_from_page_cache(page, shadow);
976 xa_unlock_irqrestore(&mapping->i_pages, flags);
977
978 if (freepage != NULL)
979 freepage(page);
980 }
981
982 return 1;
983
984cannot_free:
985 xa_unlock_irqrestore(&mapping->i_pages, flags);
986 return 0;
987}
988
989/*
990 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
991 * someone else has a ref on the page, abort and return 0. If it was
992 * successfully detached, return 1. Assumes the caller has a single ref on
993 * this page.
994 */
995int remove_mapping(struct address_space *mapping, struct page *page)
996{
997 if (__remove_mapping(mapping, page, false)) {
998 /*
999 * Unfreezing the refcount with 1 rather than 2 effectively
1000 * drops the pagecache ref for us without requiring another
1001 * atomic operation.
1002 */
1003 page_ref_unfreeze(page, 1);
1004 return 1;
1005 }
1006 return 0;
1007}
1008
1009/**
1010 * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 * @page: page to be put back to appropriate lru list
1012 *
1013 * Add previously isolated @page to appropriate LRU list.
1014 * Page may still be unevictable for other reasons.
1015 *
1016 * lru_lock must not be held, interrupts must be enabled.
1017 */
1018void putback_lru_page(struct page *page)
1019{
1020 lru_cache_add(page);
1021 put_page(page); /* drop ref from isolate */
1022}
1023
1024enum page_references {
1025 PAGEREF_RECLAIM,
1026 PAGEREF_RECLAIM_CLEAN,
1027 PAGEREF_KEEP,
1028 PAGEREF_ACTIVATE,
1029};
1030
1031static enum page_references page_check_references(struct page *page,
1032 struct scan_control *sc)
1033{
1034 int referenced_ptes, referenced_page;
1035 unsigned long vm_flags;
1036
1037 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1038 &vm_flags);
1039 referenced_page = TestClearPageReferenced(page);
1040
1041 /*
1042 * Mlock lost the isolation race with us. Let try_to_unmap()
1043 * move the page to the unevictable list.
1044 */
1045 if (vm_flags & VM_LOCKED)
1046 return PAGEREF_RECLAIM;
1047
1048 if (referenced_ptes) {
1049 if (PageSwapBacked(page))
1050 return PAGEREF_ACTIVATE;
1051 /*
1052 * All mapped pages start out with page table
1053 * references from the instantiating fault, so we need
1054 * to look twice if a mapped file page is used more
1055 * than once.
1056 *
1057 * Mark it and spare it for another trip around the
1058 * inactive list. Another page table reference will
1059 * lead to its activation.
1060 *
1061 * Note: the mark is set for activated pages as well
1062 * so that recently deactivated but used pages are
1063 * quickly recovered.
1064 */
1065 SetPageReferenced(page);
1066
1067 if (referenced_page || referenced_ptes > 1)
1068 return PAGEREF_ACTIVATE;
1069
1070 /*
1071 * Activate file-backed executable pages after first usage.
1072 */
1073 if (vm_flags & VM_EXEC)
1074 return PAGEREF_ACTIVATE;
1075
1076 return PAGEREF_KEEP;
1077 }
1078
1079 /* Reclaim if clean, defer dirty pages to writeback */
1080 if (referenced_page && !PageSwapBacked(page))
1081 return PAGEREF_RECLAIM_CLEAN;
1082
1083 return PAGEREF_RECLAIM;
1084}
1085
1086/* Check if a page is dirty or under writeback */
1087static void page_check_dirty_writeback(struct page *page,
1088 bool *dirty, bool *writeback)
1089{
1090 struct address_space *mapping;
1091
1092 /*
1093 * Anonymous pages are not handled by flushers and must be written
1094 * from reclaim context. Do not stall reclaim based on them
1095 */
1096 if (!page_is_file_cache(page) ||
1097 (PageAnon(page) && !PageSwapBacked(page))) {
1098 *dirty = false;
1099 *writeback = false;
1100 return;
1101 }
1102
1103 /* By default assume that the page flags are accurate */
1104 *dirty = PageDirty(page);
1105 *writeback = PageWriteback(page);
1106
1107 /* Verify dirty/writeback state if the filesystem supports it */
1108 if (!page_has_private(page))
1109 return;
1110
1111 mapping = page_mapping(page);
1112 if (mapping && mapping->a_ops->is_dirty_writeback)
1113 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1114}
1115
1116/*
1117 * shrink_page_list() returns the number of reclaimed pages
1118 */
1119static unsigned long shrink_page_list(struct list_head *page_list,
1120 struct pglist_data *pgdat,
1121 struct scan_control *sc,
1122 enum ttu_flags ttu_flags,
1123 struct reclaim_stat *stat,
1124 bool ignore_references)
1125{
1126 LIST_HEAD(ret_pages);
1127 LIST_HEAD(free_pages);
1128 unsigned nr_reclaimed = 0;
1129 unsigned pgactivate = 0;
1130
1131 memset(stat, 0, sizeof(*stat));
1132 cond_resched();
1133
1134 while (!list_empty(page_list)) {
1135 struct address_space *mapping;
1136 struct page *page;
1137 int may_enter_fs;
1138 enum page_references references = PAGEREF_RECLAIM;
1139 bool dirty, writeback;
1140 unsigned int nr_pages;
1141
1142 cond_resched();
1143
1144 page = lru_to_page(page_list);
1145 list_del(&page->lru);
1146
1147 if (!trylock_page(page))
1148 goto keep;
1149
1150 VM_BUG_ON_PAGE(PageActive(page), page);
1151
1152 nr_pages = compound_nr(page);
1153
1154 /* Account the number of base pages even though THP */
1155 sc->nr_scanned += nr_pages;
1156
1157 if (unlikely(!page_evictable(page)))
1158 goto activate_locked;
1159
1160 if (!sc->may_unmap && page_mapped(page))
1161 goto keep_locked;
1162
1163 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1164 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1165
1166 /*
1167 * The number of dirty pages determines if a node is marked
1168 * reclaim_congested which affects wait_iff_congested. kswapd
1169 * will stall and start writing pages if the tail of the LRU
1170 * is all dirty unqueued pages.
1171 */
1172 page_check_dirty_writeback(page, &dirty, &writeback);
1173 if (dirty || writeback)
1174 stat->nr_dirty++;
1175
1176 if (dirty && !writeback)
1177 stat->nr_unqueued_dirty++;
1178
1179 /*
1180 * Treat this page as congested if the underlying BDI is or if
1181 * pages are cycling through the LRU so quickly that the
1182 * pages marked for immediate reclaim are making it to the
1183 * end of the LRU a second time.
1184 */
1185 mapping = page_mapping(page);
1186 if (((dirty || writeback) && mapping &&
1187 inode_write_congested(mapping->host)) ||
1188 (writeback && PageReclaim(page)))
1189 stat->nr_congested++;
1190
1191 /*
1192 * If a page at the tail of the LRU is under writeback, there
1193 * are three cases to consider.
1194 *
1195 * 1) If reclaim is encountering an excessive number of pages
1196 * under writeback and this page is both under writeback and
1197 * PageReclaim then it indicates that pages are being queued
1198 * for IO but are being recycled through the LRU before the
1199 * IO can complete. Waiting on the page itself risks an
1200 * indefinite stall if it is impossible to writeback the
1201 * page due to IO error or disconnected storage so instead
1202 * note that the LRU is being scanned too quickly and the
1203 * caller can stall after page list has been processed.
1204 *
1205 * 2) Global or new memcg reclaim encounters a page that is
1206 * not marked for immediate reclaim, or the caller does not
1207 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208 * not to fs). In this case mark the page for immediate
1209 * reclaim and continue scanning.
1210 *
1211 * Require may_enter_fs because we would wait on fs, which
1212 * may not have submitted IO yet. And the loop driver might
1213 * enter reclaim, and deadlock if it waits on a page for
1214 * which it is needed to do the write (loop masks off
1215 * __GFP_IO|__GFP_FS for this reason); but more thought
1216 * would probably show more reasons.
1217 *
1218 * 3) Legacy memcg encounters a page that is already marked
1219 * PageReclaim. memcg does not have any dirty pages
1220 * throttling so we could easily OOM just because too many
1221 * pages are in writeback and there is nothing else to
1222 * reclaim. Wait for the writeback to complete.
1223 *
1224 * In cases 1) and 2) we activate the pages to get them out of
1225 * the way while we continue scanning for clean pages on the
1226 * inactive list and refilling from the active list. The
1227 * observation here is that waiting for disk writes is more
1228 * expensive than potentially causing reloads down the line.
1229 * Since they're marked for immediate reclaim, they won't put
1230 * memory pressure on the cache working set any longer than it
1231 * takes to write them to disk.
1232 */
1233 if (PageWriteback(page)) {
1234 /* Case 1 above */
1235 if (current_is_kswapd() &&
1236 PageReclaim(page) &&
1237 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1238 stat->nr_immediate++;
1239 goto activate_locked;
1240
1241 /* Case 2 above */
1242 } else if (sane_reclaim(sc) ||
1243 !PageReclaim(page) || !may_enter_fs) {
1244 /*
1245 * This is slightly racy - end_page_writeback()
1246 * might have just cleared PageReclaim, then
1247 * setting PageReclaim here end up interpreted
1248 * as PageReadahead - but that does not matter
1249 * enough to care. What we do want is for this
1250 * page to have PageReclaim set next time memcg
1251 * reclaim reaches the tests above, so it will
1252 * then wait_on_page_writeback() to avoid OOM;
1253 * and it's also appropriate in global reclaim.
1254 */
1255 SetPageReclaim(page);
1256 stat->nr_writeback++;
1257 goto activate_locked;
1258
1259 /* Case 3 above */
1260 } else {
1261 unlock_page(page);
1262 wait_on_page_writeback(page);
1263 /* then go back and try same page again */
1264 list_add_tail(&page->lru, page_list);
1265 continue;
1266 }
1267 }
1268
1269 if (!ignore_references)
1270 references = page_check_references(page, sc);
1271
1272 switch (references) {
1273 case PAGEREF_ACTIVATE:
1274 goto activate_locked;
1275 case PAGEREF_KEEP:
1276 stat->nr_ref_keep += nr_pages;
1277 goto keep_locked;
1278 case PAGEREF_RECLAIM:
1279 case PAGEREF_RECLAIM_CLEAN:
1280 ; /* try to reclaim the page below */
1281 }
1282
1283 /*
1284 * Anonymous process memory has backing store?
1285 * Try to allocate it some swap space here.
1286 * Lazyfree page could be freed directly
1287 */
1288 if (PageAnon(page) && PageSwapBacked(page)) {
1289 if (!PageSwapCache(page)) {
1290 if (!(sc->gfp_mask & __GFP_IO))
1291 goto keep_locked;
1292 if (PageTransHuge(page)) {
1293 /* cannot split THP, skip it */
1294 if (!can_split_huge_page(page, NULL))
1295 goto activate_locked;
1296 /*
1297 * Split pages without a PMD map right
1298 * away. Chances are some or all of the
1299 * tail pages can be freed without IO.
1300 */
1301 if (!compound_mapcount(page) &&
1302 split_huge_page_to_list(page,
1303 page_list))
1304 goto activate_locked;
1305 }
1306 if (!add_to_swap(page)) {
1307 if (!PageTransHuge(page))
1308 goto activate_locked_split;
1309 /* Fallback to swap normal pages */
1310 if (split_huge_page_to_list(page,
1311 page_list))
1312 goto activate_locked;
1313#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314 count_vm_event(THP_SWPOUT_FALLBACK);
1315#endif
1316 if (!add_to_swap(page))
1317 goto activate_locked_split;
1318 }
1319
1320 may_enter_fs = 1;
1321
1322 /* Adding to swap updated mapping */
1323 mapping = page_mapping(page);
1324 }
1325 } else if (unlikely(PageTransHuge(page))) {
1326 /* Split file THP */
1327 if (split_huge_page_to_list(page, page_list))
1328 goto keep_locked;
1329 }
1330
1331 /*
1332 * THP may get split above, need minus tail pages and update
1333 * nr_pages to avoid accounting tail pages twice.
1334 *
1335 * The tail pages that are added into swap cache successfully
1336 * reach here.
1337 */
1338 if ((nr_pages > 1) && !PageTransHuge(page)) {
1339 sc->nr_scanned -= (nr_pages - 1);
1340 nr_pages = 1;
1341 }
1342
1343 /*
1344 * The page is mapped into the page tables of one or more
1345 * processes. Try to unmap it here.
1346 */
1347 if (page_mapped(page)) {
1348 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1349
1350 if (unlikely(PageTransHuge(page)))
1351 flags |= TTU_SPLIT_HUGE_PMD;
1352 if (!try_to_unmap(page, flags)) {
1353 stat->nr_unmap_fail += nr_pages;
1354 goto activate_locked;
1355 }
1356 }
1357
1358 if (PageDirty(page)) {
1359 /*
1360 * Only kswapd can writeback filesystem pages
1361 * to avoid risk of stack overflow. But avoid
1362 * injecting inefficient single-page IO into
1363 * flusher writeback as much as possible: only
1364 * write pages when we've encountered many
1365 * dirty pages, and when we've already scanned
1366 * the rest of the LRU for clean pages and see
1367 * the same dirty pages again (PageReclaim).
1368 */
1369 if (page_is_file_cache(page) &&
1370 (!current_is_kswapd() || !PageReclaim(page) ||
1371 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1372 /*
1373 * Immediately reclaim when written back.
1374 * Similar in principal to deactivate_page()
1375 * except we already have the page isolated
1376 * and know it's dirty
1377 */
1378 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1379 SetPageReclaim(page);
1380
1381 goto activate_locked;
1382 }
1383
1384 if (references == PAGEREF_RECLAIM_CLEAN)
1385 goto keep_locked;
1386 if (!may_enter_fs)
1387 goto keep_locked;
1388 if (!sc->may_writepage)
1389 goto keep_locked;
1390
1391 /*
1392 * Page is dirty. Flush the TLB if a writable entry
1393 * potentially exists to avoid CPU writes after IO
1394 * starts and then write it out here.
1395 */
1396 try_to_unmap_flush_dirty();
1397 switch (pageout(page, mapping, sc)) {
1398 case PAGE_KEEP:
1399 goto keep_locked;
1400 case PAGE_ACTIVATE:
1401 goto activate_locked;
1402 case PAGE_SUCCESS:
1403 if (PageWriteback(page))
1404 goto keep;
1405 if (PageDirty(page))
1406 goto keep;
1407
1408 /*
1409 * A synchronous write - probably a ramdisk. Go
1410 * ahead and try to reclaim the page.
1411 */
1412 if (!trylock_page(page))
1413 goto keep;
1414 if (PageDirty(page) || PageWriteback(page))
1415 goto keep_locked;
1416 mapping = page_mapping(page);
1417 case PAGE_CLEAN:
1418 ; /* try to free the page below */
1419 }
1420 }
1421
1422 /*
1423 * If the page has buffers, try to free the buffer mappings
1424 * associated with this page. If we succeed we try to free
1425 * the page as well.
1426 *
1427 * We do this even if the page is PageDirty().
1428 * try_to_release_page() does not perform I/O, but it is
1429 * possible for a page to have PageDirty set, but it is actually
1430 * clean (all its buffers are clean). This happens if the
1431 * buffers were written out directly, with submit_bh(). ext3
1432 * will do this, as well as the blockdev mapping.
1433 * try_to_release_page() will discover that cleanness and will
1434 * drop the buffers and mark the page clean - it can be freed.
1435 *
1436 * Rarely, pages can have buffers and no ->mapping. These are
1437 * the pages which were not successfully invalidated in
1438 * truncate_complete_page(). We try to drop those buffers here
1439 * and if that worked, and the page is no longer mapped into
1440 * process address space (page_count == 1) it can be freed.
1441 * Otherwise, leave the page on the LRU so it is swappable.
1442 */
1443 if (page_has_private(page)) {
1444 if (!try_to_release_page(page, sc->gfp_mask))
1445 goto activate_locked;
1446 if (!mapping && page_count(page) == 1) {
1447 unlock_page(page);
1448 if (put_page_testzero(page))
1449 goto free_it;
1450 else {
1451 /*
1452 * rare race with speculative reference.
1453 * the speculative reference will free
1454 * this page shortly, so we may
1455 * increment nr_reclaimed here (and
1456 * leave it off the LRU).
1457 */
1458 nr_reclaimed++;
1459 continue;
1460 }
1461 }
1462 }
1463
1464 if (PageAnon(page) && !PageSwapBacked(page)) {
1465 /* follow __remove_mapping for reference */
1466 if (!page_ref_freeze(page, 1))
1467 goto keep_locked;
1468 if (PageDirty(page)) {
1469 page_ref_unfreeze(page, 1);
1470 goto keep_locked;
1471 }
1472
1473 count_vm_event(PGLAZYFREED);
1474 count_memcg_page_event(page, PGLAZYFREED);
1475 } else if (!mapping || !__remove_mapping(mapping, page, true))
1476 goto keep_locked;
1477
1478 unlock_page(page);
1479free_it:
1480 /*
1481 * THP may get swapped out in a whole, need account
1482 * all base pages.
1483 */
1484 nr_reclaimed += nr_pages;
1485
1486 /*
1487 * Is there need to periodically free_page_list? It would
1488 * appear not as the counts should be low
1489 */
1490 if (unlikely(PageTransHuge(page)))
1491 (*get_compound_page_dtor(page))(page);
1492 else
1493 list_add(&page->lru, &free_pages);
1494 continue;
1495
1496activate_locked_split:
1497 /*
1498 * The tail pages that are failed to add into swap cache
1499 * reach here. Fixup nr_scanned and nr_pages.
1500 */
1501 if (nr_pages > 1) {
1502 sc->nr_scanned -= (nr_pages - 1);
1503 nr_pages = 1;
1504 }
1505activate_locked:
1506 /* Not a candidate for swapping, so reclaim swap space. */
1507 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1508 PageMlocked(page)))
1509 try_to_free_swap(page);
1510 VM_BUG_ON_PAGE(PageActive(page), page);
1511 if (!PageMlocked(page)) {
1512 int type = page_is_file_cache(page);
1513 SetPageActive(page);
1514 stat->nr_activate[type] += nr_pages;
1515 count_memcg_page_event(page, PGACTIVATE);
1516 }
1517keep_locked:
1518 unlock_page(page);
1519keep:
1520 list_add(&page->lru, &ret_pages);
1521 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1522 }
1523
1524 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1525
1526 mem_cgroup_uncharge_list(&free_pages);
1527 try_to_unmap_flush();
1528 free_unref_page_list(&free_pages);
1529
1530 list_splice(&ret_pages, page_list);
1531 count_vm_events(PGACTIVATE, pgactivate);
1532
1533 return nr_reclaimed;
1534}
1535
1536unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1537 struct list_head *page_list)
1538{
1539 struct scan_control sc = {
1540 .gfp_mask = GFP_KERNEL,
1541 .priority = DEF_PRIORITY,
1542 .may_unmap = 1,
1543 };
1544 struct reclaim_stat dummy_stat;
1545 unsigned long ret;
1546 struct page *page, *next;
1547 LIST_HEAD(clean_pages);
1548
1549 list_for_each_entry_safe(page, next, page_list, lru) {
1550 if (page_is_file_cache(page) && !PageDirty(page) &&
1551 !__PageMovable(page) && !PageUnevictable(page)) {
1552 ClearPageActive(page);
1553 list_move(&page->lru, &clean_pages);
1554 }
1555 }
1556
1557 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1558 TTU_IGNORE_ACCESS, &dummy_stat, true);
1559 list_splice(&clean_pages, page_list);
1560 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1561 return ret;
1562}
1563
1564/*
1565 * Attempt to remove the specified page from its LRU. Only take this page
1566 * if it is of the appropriate PageActive status. Pages which are being
1567 * freed elsewhere are also ignored.
1568 *
1569 * page: page to consider
1570 * mode: one of the LRU isolation modes defined above
1571 *
1572 * returns 0 on success, -ve errno on failure.
1573 */
1574int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1575{
1576 int ret = -EINVAL;
1577
1578 /* Only take pages on the LRU. */
1579 if (!PageLRU(page))
1580 return ret;
1581
1582 /* Compaction should not handle unevictable pages but CMA can do so */
1583 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1584 return ret;
1585
1586 ret = -EBUSY;
1587
1588 /*
1589 * To minimise LRU disruption, the caller can indicate that it only
1590 * wants to isolate pages it will be able to operate on without
1591 * blocking - clean pages for the most part.
1592 *
1593 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594 * that it is possible to migrate without blocking
1595 */
1596 if (mode & ISOLATE_ASYNC_MIGRATE) {
1597 /* All the caller can do on PageWriteback is block */
1598 if (PageWriteback(page))
1599 return ret;
1600
1601 if (PageDirty(page)) {
1602 struct address_space *mapping;
1603 bool migrate_dirty;
1604
1605 /*
1606 * Only pages without mappings or that have a
1607 * ->migratepage callback are possible to migrate
1608 * without blocking. However, we can be racing with
1609 * truncation so it's necessary to lock the page
1610 * to stabilise the mapping as truncation holds
1611 * the page lock until after the page is removed
1612 * from the page cache.
1613 */
1614 if (!trylock_page(page))
1615 return ret;
1616
1617 mapping = page_mapping(page);
1618 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1619 unlock_page(page);
1620 if (!migrate_dirty)
1621 return ret;
1622 }
1623 }
1624
1625 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1626 return ret;
1627
1628 if (likely(get_page_unless_zero(page))) {
1629 /*
1630 * Be careful not to clear PageLRU until after we're
1631 * sure the page is not being freed elsewhere -- the
1632 * page release code relies on it.
1633 */
1634 ClearPageLRU(page);
1635 ret = 0;
1636 }
1637
1638 return ret;
1639}
1640
1641
1642/*
1643 * Update LRU sizes after isolating pages. The LRU size updates must
1644 * be complete before mem_cgroup_update_lru_size due to a santity check.
1645 */
1646static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1647 enum lru_list lru, unsigned long *nr_zone_taken)
1648{
1649 int zid;
1650
1651 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1652 if (!nr_zone_taken[zid])
1653 continue;
1654
1655 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1656#ifdef CONFIG_MEMCG
1657 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1658#endif
1659 }
1660
1661}
1662
1663/**
1664 * pgdat->lru_lock is heavily contended. Some of the functions that
1665 * shrink the lists perform better by taking out a batch of pages
1666 * and working on them outside the LRU lock.
1667 *
1668 * For pagecache intensive workloads, this function is the hottest
1669 * spot in the kernel (apart from copy_*_user functions).
1670 *
1671 * Appropriate locks must be held before calling this function.
1672 *
1673 * @nr_to_scan: The number of eligible pages to look through on the list.
1674 * @lruvec: The LRU vector to pull pages from.
1675 * @dst: The temp list to put pages on to.
1676 * @nr_scanned: The number of pages that were scanned.
1677 * @sc: The scan_control struct for this reclaim session
1678 * @mode: One of the LRU isolation modes
1679 * @lru: LRU list id for isolating
1680 *
1681 * returns how many pages were moved onto *@dst.
1682 */
1683static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1684 struct lruvec *lruvec, struct list_head *dst,
1685 unsigned long *nr_scanned, struct scan_control *sc,
1686 enum lru_list lru)
1687{
1688 struct list_head *src = &lruvec->lists[lru];
1689 unsigned long nr_taken = 0;
1690 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1691 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1692 unsigned long skipped = 0;
1693 unsigned long scan, total_scan, nr_pages;
1694 LIST_HEAD(pages_skipped);
1695 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1696
1697 total_scan = 0;
1698 scan = 0;
1699 while (scan < nr_to_scan && !list_empty(src)) {
1700 struct page *page;
1701
1702 page = lru_to_page(src);
1703 prefetchw_prev_lru_page(page, src, flags);
1704
1705 VM_BUG_ON_PAGE(!PageLRU(page), page);
1706
1707 nr_pages = compound_nr(page);
1708 total_scan += nr_pages;
1709
1710 if (page_zonenum(page) > sc->reclaim_idx) {
1711 list_move(&page->lru, &pages_skipped);
1712 nr_skipped[page_zonenum(page)] += nr_pages;
1713 continue;
1714 }
1715
1716 /*
1717 * Do not count skipped pages because that makes the function
1718 * return with no isolated pages if the LRU mostly contains
1719 * ineligible pages. This causes the VM to not reclaim any
1720 * pages, triggering a premature OOM.
1721 *
1722 * Account all tail pages of THP. This would not cause
1723 * premature OOM since __isolate_lru_page() returns -EBUSY
1724 * only when the page is being freed somewhere else.
1725 */
1726 scan += nr_pages;
1727 switch (__isolate_lru_page(page, mode)) {
1728 case 0:
1729 nr_taken += nr_pages;
1730 nr_zone_taken[page_zonenum(page)] += nr_pages;
1731 list_move(&page->lru, dst);
1732 break;
1733
1734 case -EBUSY:
1735 /* else it is being freed elsewhere */
1736 list_move(&page->lru, src);
1737 continue;
1738
1739 default:
1740 BUG();
1741 }
1742 }
1743
1744 /*
1745 * Splice any skipped pages to the start of the LRU list. Note that
1746 * this disrupts the LRU order when reclaiming for lower zones but
1747 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748 * scanning would soon rescan the same pages to skip and put the
1749 * system at risk of premature OOM.
1750 */
1751 if (!list_empty(&pages_skipped)) {
1752 int zid;
1753
1754 list_splice(&pages_skipped, src);
1755 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1756 if (!nr_skipped[zid])
1757 continue;
1758
1759 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1760 skipped += nr_skipped[zid];
1761 }
1762 }
1763 *nr_scanned = total_scan;
1764 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1765 total_scan, skipped, nr_taken, mode, lru);
1766 update_lru_sizes(lruvec, lru, nr_zone_taken);
1767 return nr_taken;
1768}
1769
1770/**
1771 * isolate_lru_page - tries to isolate a page from its LRU list
1772 * @page: page to isolate from its LRU list
1773 *
1774 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 * vmstat statistic corresponding to whatever LRU list the page was on.
1776 *
1777 * Returns 0 if the page was removed from an LRU list.
1778 * Returns -EBUSY if the page was not on an LRU list.
1779 *
1780 * The returned page will have PageLRU() cleared. If it was found on
1781 * the active list, it will have PageActive set. If it was found on
1782 * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 * may need to be cleared by the caller before letting the page go.
1784 *
1785 * The vmstat statistic corresponding to the list on which the page was
1786 * found will be decremented.
1787 *
1788 * Restrictions:
1789 *
1790 * (1) Must be called with an elevated refcount on the page. This is a
1791 * fundamentnal difference from isolate_lru_pages (which is called
1792 * without a stable reference).
1793 * (2) the lru_lock must not be held.
1794 * (3) interrupts must be enabled.
1795 */
1796int isolate_lru_page(struct page *page)
1797{
1798 int ret = -EBUSY;
1799
1800 VM_BUG_ON_PAGE(!page_count(page), page);
1801 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1802
1803 if (PageLRU(page)) {
1804 pg_data_t *pgdat = page_pgdat(page);
1805 struct lruvec *lruvec;
1806
1807 spin_lock_irq(&pgdat->lru_lock);
1808 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1809 if (PageLRU(page)) {
1810 int lru = page_lru(page);
1811 get_page(page);
1812 ClearPageLRU(page);
1813 del_page_from_lru_list(page, lruvec, lru);
1814 ret = 0;
1815 }
1816 spin_unlock_irq(&pgdat->lru_lock);
1817 }
1818 return ret;
1819}
1820
1821/*
1822 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 * then get resheduled. When there are massive number of tasks doing page
1824 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 * the LRU list will go small and be scanned faster than necessary, leading to
1826 * unnecessary swapping, thrashing and OOM.
1827 */
1828static int too_many_isolated(struct pglist_data *pgdat, int file,
1829 struct scan_control *sc)
1830{
1831 unsigned long inactive, isolated;
1832
1833 if (current_is_kswapd())
1834 return 0;
1835
1836 if (!sane_reclaim(sc))
1837 return 0;
1838
1839 if (file) {
1840 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1841 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1842 } else {
1843 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1844 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1845 }
1846
1847 /*
1848 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849 * won't get blocked by normal direct-reclaimers, forming a circular
1850 * deadlock.
1851 */
1852 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1853 inactive >>= 3;
1854
1855 return isolated > inactive;
1856}
1857
1858/*
1859 * This moves pages from @list to corresponding LRU list.
1860 *
1861 * We move them the other way if the page is referenced by one or more
1862 * processes, from rmap.
1863 *
1864 * If the pages are mostly unmapped, the processing is fast and it is
1865 * appropriate to hold zone_lru_lock across the whole operation. But if
1866 * the pages are mapped, the processing is slow (page_referenced()) so we
1867 * should drop zone_lru_lock around each page. It's impossible to balance
1868 * this, so instead we remove the pages from the LRU while processing them.
1869 * It is safe to rely on PG_active against the non-LRU pages in here because
1870 * nobody will play with that bit on a non-LRU page.
1871 *
1872 * The downside is that we have to touch page->_refcount against each page.
1873 * But we had to alter page->flags anyway.
1874 *
1875 * Returns the number of pages moved to the given lruvec.
1876 */
1877
1878static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1879 struct list_head *list)
1880{
1881 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1882 int nr_pages, nr_moved = 0;
1883 LIST_HEAD(pages_to_free);
1884 struct page *page;
1885 enum lru_list lru;
1886
1887 while (!list_empty(list)) {
1888 page = lru_to_page(list);
1889 VM_BUG_ON_PAGE(PageLRU(page), page);
1890 if (unlikely(!page_evictable(page))) {
1891 list_del(&page->lru);
1892 spin_unlock_irq(&pgdat->lru_lock);
1893 putback_lru_page(page);
1894 spin_lock_irq(&pgdat->lru_lock);
1895 continue;
1896 }
1897 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1898
1899 SetPageLRU(page);
1900 lru = page_lru(page);
1901
1902 nr_pages = hpage_nr_pages(page);
1903 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1904 list_move(&page->lru, &lruvec->lists[lru]);
1905
1906 if (put_page_testzero(page)) {
1907 __ClearPageLRU(page);
1908 __ClearPageActive(page);
1909 del_page_from_lru_list(page, lruvec, lru);
1910
1911 if (unlikely(PageCompound(page))) {
1912 spin_unlock_irq(&pgdat->lru_lock);
1913 (*get_compound_page_dtor(page))(page);
1914 spin_lock_irq(&pgdat->lru_lock);
1915 } else
1916 list_add(&page->lru, &pages_to_free);
1917 } else {
1918 nr_moved += nr_pages;
1919 }
1920 }
1921
1922 /*
1923 * To save our caller's stack, now use input list for pages to free.
1924 */
1925 list_splice(&pages_to_free, list);
1926
1927 return nr_moved;
1928}
1929
1930/*
1931 * If a kernel thread (such as nfsd for loop-back mounts) services
1932 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 * In that case we should only throttle if the backing device it is
1934 * writing to is congested. In other cases it is safe to throttle.
1935 */
1936static int current_may_throttle(void)
1937{
1938 return !(current->flags & PF_LESS_THROTTLE) ||
1939 current->backing_dev_info == NULL ||
1940 bdi_write_congested(current->backing_dev_info);
1941}
1942
1943/*
1944 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1945 * of reclaimed pages
1946 */
1947static noinline_for_stack unsigned long
1948shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1949 struct scan_control *sc, enum lru_list lru)
1950{
1951 LIST_HEAD(page_list);
1952 unsigned long nr_scanned;
1953 unsigned long nr_reclaimed = 0;
1954 unsigned long nr_taken;
1955 struct reclaim_stat stat;
1956 int file = is_file_lru(lru);
1957 enum vm_event_item item;
1958 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1959 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1960 bool stalled = false;
1961
1962 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1963 if (stalled)
1964 return 0;
1965
1966 /* wait a bit for the reclaimer. */
1967 msleep(100);
1968 stalled = true;
1969
1970 /* We are about to die and free our memory. Return now. */
1971 if (fatal_signal_pending(current))
1972 return SWAP_CLUSTER_MAX;
1973 }
1974
1975 lru_add_drain();
1976
1977 spin_lock_irq(&pgdat->lru_lock);
1978
1979 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1980 &nr_scanned, sc, lru);
1981
1982 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983 reclaim_stat->recent_scanned[file] += nr_taken;
1984
1985 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1986 if (global_reclaim(sc))
1987 __count_vm_events(item, nr_scanned);
1988 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1989 spin_unlock_irq(&pgdat->lru_lock);
1990
1991 if (nr_taken == 0)
1992 return 0;
1993
1994 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1995 &stat, false);
1996
1997 spin_lock_irq(&pgdat->lru_lock);
1998
1999 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2000 if (global_reclaim(sc))
2001 __count_vm_events(item, nr_reclaimed);
2002 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2003 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2004 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2005
2006 move_pages_to_lru(lruvec, &page_list);
2007
2008 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2009
2010 spin_unlock_irq(&pgdat->lru_lock);
2011
2012 mem_cgroup_uncharge_list(&page_list);
2013 free_unref_page_list(&page_list);
2014
2015 /*
2016 * If dirty pages are scanned that are not queued for IO, it
2017 * implies that flushers are not doing their job. This can
2018 * happen when memory pressure pushes dirty pages to the end of
2019 * the LRU before the dirty limits are breached and the dirty
2020 * data has expired. It can also happen when the proportion of
2021 * dirty pages grows not through writes but through memory
2022 * pressure reclaiming all the clean cache. And in some cases,
2023 * the flushers simply cannot keep up with the allocation
2024 * rate. Nudge the flusher threads in case they are asleep.
2025 */
2026 if (stat.nr_unqueued_dirty == nr_taken)
2027 wakeup_flusher_threads(WB_REASON_VMSCAN);
2028
2029 sc->nr.dirty += stat.nr_dirty;
2030 sc->nr.congested += stat.nr_congested;
2031 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2032 sc->nr.writeback += stat.nr_writeback;
2033 sc->nr.immediate += stat.nr_immediate;
2034 sc->nr.taken += nr_taken;
2035 if (file)
2036 sc->nr.file_taken += nr_taken;
2037
2038 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2039 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2040 return nr_reclaimed;
2041}
2042
2043static void shrink_active_list(unsigned long nr_to_scan,
2044 struct lruvec *lruvec,
2045 struct scan_control *sc,
2046 enum lru_list lru)
2047{
2048 unsigned long nr_taken;
2049 unsigned long nr_scanned;
2050 unsigned long vm_flags;
2051 LIST_HEAD(l_hold); /* The pages which were snipped off */
2052 LIST_HEAD(l_active);
2053 LIST_HEAD(l_inactive);
2054 struct page *page;
2055 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2056 unsigned nr_deactivate, nr_activate;
2057 unsigned nr_rotated = 0;
2058 int file = is_file_lru(lru);
2059 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2060
2061 lru_add_drain();
2062
2063 spin_lock_irq(&pgdat->lru_lock);
2064
2065 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2066 &nr_scanned, sc, lru);
2067
2068 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2069 reclaim_stat->recent_scanned[file] += nr_taken;
2070
2071 __count_vm_events(PGREFILL, nr_scanned);
2072 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2073
2074 spin_unlock_irq(&pgdat->lru_lock);
2075
2076 while (!list_empty(&l_hold)) {
2077 cond_resched();
2078 page = lru_to_page(&l_hold);
2079 list_del(&page->lru);
2080
2081 if (unlikely(!page_evictable(page))) {
2082 putback_lru_page(page);
2083 continue;
2084 }
2085
2086 if (unlikely(buffer_heads_over_limit)) {
2087 if (page_has_private(page) && trylock_page(page)) {
2088 if (page_has_private(page))
2089 try_to_release_page(page, 0);
2090 unlock_page(page);
2091 }
2092 }
2093
2094 if (page_referenced(page, 0, sc->target_mem_cgroup,
2095 &vm_flags)) {
2096 nr_rotated += hpage_nr_pages(page);
2097 /*
2098 * Identify referenced, file-backed active pages and
2099 * give them one more trip around the active list. So
2100 * that executable code get better chances to stay in
2101 * memory under moderate memory pressure. Anon pages
2102 * are not likely to be evicted by use-once streaming
2103 * IO, plus JVM can create lots of anon VM_EXEC pages,
2104 * so we ignore them here.
2105 */
2106 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2107 list_add(&page->lru, &l_active);
2108 continue;
2109 }
2110 }
2111
2112 ClearPageActive(page); /* we are de-activating */
2113 SetPageWorkingset(page);
2114 list_add(&page->lru, &l_inactive);
2115 }
2116
2117 /*
2118 * Move pages back to the lru list.
2119 */
2120 spin_lock_irq(&pgdat->lru_lock);
2121 /*
2122 * Count referenced pages from currently used mappings as rotated,
2123 * even though only some of them are actually re-activated. This
2124 * helps balance scan pressure between file and anonymous pages in
2125 * get_scan_count.
2126 */
2127 reclaim_stat->recent_rotated[file] += nr_rotated;
2128
2129 nr_activate = move_pages_to_lru(lruvec, &l_active);
2130 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2131 /* Keep all free pages in l_active list */
2132 list_splice(&l_inactive, &l_active);
2133
2134 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2135 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2136
2137 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2138 spin_unlock_irq(&pgdat->lru_lock);
2139
2140 mem_cgroup_uncharge_list(&l_active);
2141 free_unref_page_list(&l_active);
2142 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2143 nr_deactivate, nr_rotated, sc->priority, file);
2144}
2145
2146unsigned long reclaim_pages(struct list_head *page_list)
2147{
2148 int nid = -1;
2149 unsigned long nr_reclaimed = 0;
2150 LIST_HEAD(node_page_list);
2151 struct reclaim_stat dummy_stat;
2152 struct page *page;
2153 struct scan_control sc = {
2154 .gfp_mask = GFP_KERNEL,
2155 .priority = DEF_PRIORITY,
2156 .may_writepage = 1,
2157 .may_unmap = 1,
2158 .may_swap = 1,
2159 };
2160
2161 while (!list_empty(page_list)) {
2162 page = lru_to_page(page_list);
2163 if (nid == -1) {
2164 nid = page_to_nid(page);
2165 INIT_LIST_HEAD(&node_page_list);
2166 }
2167
2168 if (nid == page_to_nid(page)) {
2169 ClearPageActive(page);
2170 list_move(&page->lru, &node_page_list);
2171 continue;
2172 }
2173
2174 nr_reclaimed += shrink_page_list(&node_page_list,
2175 NODE_DATA(nid),
2176 &sc, 0,
2177 &dummy_stat, false);
2178 while (!list_empty(&node_page_list)) {
2179 page = lru_to_page(&node_page_list);
2180 list_del(&page->lru);
2181 putback_lru_page(page);
2182 }
2183
2184 nid = -1;
2185 }
2186
2187 if (!list_empty(&node_page_list)) {
2188 nr_reclaimed += shrink_page_list(&node_page_list,
2189 NODE_DATA(nid),
2190 &sc, 0,
2191 &dummy_stat, false);
2192 while (!list_empty(&node_page_list)) {
2193 page = lru_to_page(&node_page_list);
2194 list_del(&page->lru);
2195 putback_lru_page(page);
2196 }
2197 }
2198
2199 return nr_reclaimed;
2200}
2201
2202/*
2203 * The inactive anon list should be small enough that the VM never has
2204 * to do too much work.
2205 *
2206 * The inactive file list should be small enough to leave most memory
2207 * to the established workingset on the scan-resistant active list,
2208 * but large enough to avoid thrashing the aggregate readahead window.
2209 *
2210 * Both inactive lists should also be large enough that each inactive
2211 * page has a chance to be referenced again before it is reclaimed.
2212 *
2213 * If that fails and refaulting is observed, the inactive list grows.
2214 *
2215 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2216 * on this LRU, maintained by the pageout code. An inactive_ratio
2217 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2218 *
2219 * total target max
2220 * memory ratio inactive
2221 * -------------------------------------
2222 * 10MB 1 5MB
2223 * 100MB 1 50MB
2224 * 1GB 3 250MB
2225 * 10GB 10 0.9GB
2226 * 100GB 31 3GB
2227 * 1TB 101 10GB
2228 * 10TB 320 32GB
2229 */
2230static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2231 struct scan_control *sc, bool trace)
2232{
2233 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2234 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2235 enum lru_list inactive_lru = file * LRU_FILE;
2236 unsigned long inactive, active;
2237 unsigned long inactive_ratio;
2238 unsigned long refaults;
2239 unsigned long gb;
2240
2241 /*
2242 * If we don't have swap space, anonymous page deactivation
2243 * is pointless.
2244 */
2245 if (!file && !total_swap_pages)
2246 return false;
2247
2248 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2249 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2250
2251 /*
2252 * When refaults are being observed, it means a new workingset
2253 * is being established. Disable active list protection to get
2254 * rid of the stale workingset quickly.
2255 */
2256 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2257 if (file && lruvec->refaults != refaults) {
2258 inactive_ratio = 0;
2259 } else {
2260 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2261 if (gb)
2262 inactive_ratio = int_sqrt(10 * gb);
2263 else
2264 inactive_ratio = 1;
2265 }
2266
2267 if (trace)
2268 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2269 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2270 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2271 inactive_ratio, file);
2272
2273 return inactive * inactive_ratio < active;
2274}
2275
2276static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2277 struct lruvec *lruvec, struct scan_control *sc)
2278{
2279 if (is_active_lru(lru)) {
2280 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2281 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2282 return 0;
2283 }
2284
2285 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2286}
2287
2288enum scan_balance {
2289 SCAN_EQUAL,
2290 SCAN_FRACT,
2291 SCAN_ANON,
2292 SCAN_FILE,
2293};
2294
2295/*
2296 * Determine how aggressively the anon and file LRU lists should be
2297 * scanned. The relative value of each set of LRU lists is determined
2298 * by looking at the fraction of the pages scanned we did rotate back
2299 * onto the active list instead of evict.
2300 *
2301 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2302 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2303 */
2304static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2305 struct scan_control *sc, unsigned long *nr,
2306 unsigned long *lru_pages)
2307{
2308 int swappiness = mem_cgroup_swappiness(memcg);
2309 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2310 u64 fraction[2];
2311 u64 denominator = 0; /* gcc */
2312 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2313 unsigned long anon_prio, file_prio;
2314 enum scan_balance scan_balance;
2315 unsigned long anon, file;
2316 unsigned long ap, fp;
2317 enum lru_list lru;
2318
2319 /* If we have no swap space, do not bother scanning anon pages. */
2320 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2321 scan_balance = SCAN_FILE;
2322 goto out;
2323 }
2324
2325 /*
2326 * Global reclaim will swap to prevent OOM even with no
2327 * swappiness, but memcg users want to use this knob to
2328 * disable swapping for individual groups completely when
2329 * using the memory controller's swap limit feature would be
2330 * too expensive.
2331 */
2332 if (!global_reclaim(sc) && !swappiness) {
2333 scan_balance = SCAN_FILE;
2334 goto out;
2335 }
2336
2337 /*
2338 * Do not apply any pressure balancing cleverness when the
2339 * system is close to OOM, scan both anon and file equally
2340 * (unless the swappiness setting disagrees with swapping).
2341 */
2342 if (!sc->priority && swappiness) {
2343 scan_balance = SCAN_EQUAL;
2344 goto out;
2345 }
2346
2347 /*
2348 * Prevent the reclaimer from falling into the cache trap: as
2349 * cache pages start out inactive, every cache fault will tip
2350 * the scan balance towards the file LRU. And as the file LRU
2351 * shrinks, so does the window for rotation from references.
2352 * This means we have a runaway feedback loop where a tiny
2353 * thrashing file LRU becomes infinitely more attractive than
2354 * anon pages. Try to detect this based on file LRU size.
2355 */
2356 if (global_reclaim(sc)) {
2357 unsigned long pgdatfile;
2358 unsigned long pgdatfree;
2359 int z;
2360 unsigned long total_high_wmark = 0;
2361
2362 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2363 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2364 node_page_state(pgdat, NR_INACTIVE_FILE);
2365
2366 for (z = 0; z < MAX_NR_ZONES; z++) {
2367 struct zone *zone = &pgdat->node_zones[z];
2368 if (!managed_zone(zone))
2369 continue;
2370
2371 total_high_wmark += high_wmark_pages(zone);
2372 }
2373
2374 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2375 /*
2376 * Force SCAN_ANON if there are enough inactive
2377 * anonymous pages on the LRU in eligible zones.
2378 * Otherwise, the small LRU gets thrashed.
2379 */
2380 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2381 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2382 >> sc->priority) {
2383 scan_balance = SCAN_ANON;
2384 goto out;
2385 }
2386 }
2387 }
2388
2389 /*
2390 * If there is enough inactive page cache, i.e. if the size of the
2391 * inactive list is greater than that of the active list *and* the
2392 * inactive list actually has some pages to scan on this priority, we
2393 * do not reclaim anything from the anonymous working set right now.
2394 * Without the second condition we could end up never scanning an
2395 * lruvec even if it has plenty of old anonymous pages unless the
2396 * system is under heavy pressure.
2397 */
2398 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2399 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2400 scan_balance = SCAN_FILE;
2401 goto out;
2402 }
2403
2404 scan_balance = SCAN_FRACT;
2405
2406 /*
2407 * With swappiness at 100, anonymous and file have the same priority.
2408 * This scanning priority is essentially the inverse of IO cost.
2409 */
2410 anon_prio = swappiness;
2411 file_prio = 200 - anon_prio;
2412
2413 /*
2414 * OK, so we have swap space and a fair amount of page cache
2415 * pages. We use the recently rotated / recently scanned
2416 * ratios to determine how valuable each cache is.
2417 *
2418 * Because workloads change over time (and to avoid overflow)
2419 * we keep these statistics as a floating average, which ends
2420 * up weighing recent references more than old ones.
2421 *
2422 * anon in [0], file in [1]
2423 */
2424
2425 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2426 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2427 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2428 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2429
2430 spin_lock_irq(&pgdat->lru_lock);
2431 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2432 reclaim_stat->recent_scanned[0] /= 2;
2433 reclaim_stat->recent_rotated[0] /= 2;
2434 }
2435
2436 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2437 reclaim_stat->recent_scanned[1] /= 2;
2438 reclaim_stat->recent_rotated[1] /= 2;
2439 }
2440
2441 /*
2442 * The amount of pressure on anon vs file pages is inversely
2443 * proportional to the fraction of recently scanned pages on
2444 * each list that were recently referenced and in active use.
2445 */
2446 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2447 ap /= reclaim_stat->recent_rotated[0] + 1;
2448
2449 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2450 fp /= reclaim_stat->recent_rotated[1] + 1;
2451 spin_unlock_irq(&pgdat->lru_lock);
2452
2453 fraction[0] = ap;
2454 fraction[1] = fp;
2455 denominator = ap + fp + 1;
2456out:
2457 *lru_pages = 0;
2458 for_each_evictable_lru(lru) {
2459 int file = is_file_lru(lru);
2460 unsigned long lruvec_size;
2461 unsigned long scan;
2462 unsigned long protection;
2463
2464 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2465 protection = mem_cgroup_protection(memcg,
2466 sc->memcg_low_reclaim);
2467
2468 if (protection) {
2469 /*
2470 * Scale a cgroup's reclaim pressure by proportioning
2471 * its current usage to its memory.low or memory.min
2472 * setting.
2473 *
2474 * This is important, as otherwise scanning aggression
2475 * becomes extremely binary -- from nothing as we
2476 * approach the memory protection threshold, to totally
2477 * nominal as we exceed it. This results in requiring
2478 * setting extremely liberal protection thresholds. It
2479 * also means we simply get no protection at all if we
2480 * set it too low, which is not ideal.
2481 *
2482 * If there is any protection in place, we reduce scan
2483 * pressure by how much of the total memory used is
2484 * within protection thresholds.
2485 *
2486 * There is one special case: in the first reclaim pass,
2487 * we skip over all groups that are within their low
2488 * protection. If that fails to reclaim enough pages to
2489 * satisfy the reclaim goal, we come back and override
2490 * the best-effort low protection. However, we still
2491 * ideally want to honor how well-behaved groups are in
2492 * that case instead of simply punishing them all
2493 * equally. As such, we reclaim them based on how much
2494 * memory they are using, reducing the scan pressure
2495 * again by how much of the total memory used is under
2496 * hard protection.
2497 */
2498 unsigned long cgroup_size = mem_cgroup_size(memcg);
2499
2500 /* Avoid TOCTOU with earlier protection check */
2501 cgroup_size = max(cgroup_size, protection);
2502
2503 scan = lruvec_size - lruvec_size * protection /
2504 cgroup_size;
2505
2506 /*
2507 * Minimally target SWAP_CLUSTER_MAX pages to keep
2508 * reclaim moving forwards, avoiding decremeting
2509 * sc->priority further than desirable.
2510 */
2511 scan = max(scan, SWAP_CLUSTER_MAX);
2512 } else {
2513 scan = lruvec_size;
2514 }
2515
2516 scan >>= sc->priority;
2517
2518 /*
2519 * If the cgroup's already been deleted, make sure to
2520 * scrape out the remaining cache.
2521 */
2522 if (!scan && !mem_cgroup_online(memcg))
2523 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2524
2525 switch (scan_balance) {
2526 case SCAN_EQUAL:
2527 /* Scan lists relative to size */
2528 break;
2529 case SCAN_FRACT:
2530 /*
2531 * Scan types proportional to swappiness and
2532 * their relative recent reclaim efficiency.
2533 * Make sure we don't miss the last page
2534 * because of a round-off error.
2535 */
2536 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2537 denominator);
2538 break;
2539 case SCAN_FILE:
2540 case SCAN_ANON:
2541 /* Scan one type exclusively */
2542 if ((scan_balance == SCAN_FILE) != file) {
2543 lruvec_size = 0;
2544 scan = 0;
2545 }
2546 break;
2547 default:
2548 /* Look ma, no brain */
2549 BUG();
2550 }
2551
2552 *lru_pages += lruvec_size;
2553 nr[lru] = scan;
2554 }
2555}
2556
2557/*
2558 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2559 */
2560static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2561 struct scan_control *sc, unsigned long *lru_pages)
2562{
2563 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2564 unsigned long nr[NR_LRU_LISTS];
2565 unsigned long targets[NR_LRU_LISTS];
2566 unsigned long nr_to_scan;
2567 enum lru_list lru;
2568 unsigned long nr_reclaimed = 0;
2569 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2570 struct blk_plug plug;
2571 bool scan_adjusted;
2572
2573 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2574
2575 /* Record the original scan target for proportional adjustments later */
2576 memcpy(targets, nr, sizeof(nr));
2577
2578 /*
2579 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2580 * event that can occur when there is little memory pressure e.g.
2581 * multiple streaming readers/writers. Hence, we do not abort scanning
2582 * when the requested number of pages are reclaimed when scanning at
2583 * DEF_PRIORITY on the assumption that the fact we are direct
2584 * reclaiming implies that kswapd is not keeping up and it is best to
2585 * do a batch of work at once. For memcg reclaim one check is made to
2586 * abort proportional reclaim if either the file or anon lru has already
2587 * dropped to zero at the first pass.
2588 */
2589 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2590 sc->priority == DEF_PRIORITY);
2591
2592 blk_start_plug(&plug);
2593 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2594 nr[LRU_INACTIVE_FILE]) {
2595 unsigned long nr_anon, nr_file, percentage;
2596 unsigned long nr_scanned;
2597
2598 for_each_evictable_lru(lru) {
2599 if (nr[lru]) {
2600 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2601 nr[lru] -= nr_to_scan;
2602
2603 nr_reclaimed += shrink_list(lru, nr_to_scan,
2604 lruvec, sc);
2605 }
2606 }
2607
2608 cond_resched();
2609
2610 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2611 continue;
2612
2613 /*
2614 * For kswapd and memcg, reclaim at least the number of pages
2615 * requested. Ensure that the anon and file LRUs are scanned
2616 * proportionally what was requested by get_scan_count(). We
2617 * stop reclaiming one LRU and reduce the amount scanning
2618 * proportional to the original scan target.
2619 */
2620 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2621 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2622
2623 /*
2624 * It's just vindictive to attack the larger once the smaller
2625 * has gone to zero. And given the way we stop scanning the
2626 * smaller below, this makes sure that we only make one nudge
2627 * towards proportionality once we've got nr_to_reclaim.
2628 */
2629 if (!nr_file || !nr_anon)
2630 break;
2631
2632 if (nr_file > nr_anon) {
2633 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2634 targets[LRU_ACTIVE_ANON] + 1;
2635 lru = LRU_BASE;
2636 percentage = nr_anon * 100 / scan_target;
2637 } else {
2638 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2639 targets[LRU_ACTIVE_FILE] + 1;
2640 lru = LRU_FILE;
2641 percentage = nr_file * 100 / scan_target;
2642 }
2643
2644 /* Stop scanning the smaller of the LRU */
2645 nr[lru] = 0;
2646 nr[lru + LRU_ACTIVE] = 0;
2647
2648 /*
2649 * Recalculate the other LRU scan count based on its original
2650 * scan target and the percentage scanning already complete
2651 */
2652 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2653 nr_scanned = targets[lru] - nr[lru];
2654 nr[lru] = targets[lru] * (100 - percentage) / 100;
2655 nr[lru] -= min(nr[lru], nr_scanned);
2656
2657 lru += LRU_ACTIVE;
2658 nr_scanned = targets[lru] - nr[lru];
2659 nr[lru] = targets[lru] * (100 - percentage) / 100;
2660 nr[lru] -= min(nr[lru], nr_scanned);
2661
2662 scan_adjusted = true;
2663 }
2664 blk_finish_plug(&plug);
2665 sc->nr_reclaimed += nr_reclaimed;
2666
2667 /*
2668 * Even if we did not try to evict anon pages at all, we want to
2669 * rebalance the anon lru active/inactive ratio.
2670 */
2671 if (inactive_list_is_low(lruvec, false, sc, true))
2672 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2673 sc, LRU_ACTIVE_ANON);
2674}
2675
2676/* Use reclaim/compaction for costly allocs or under memory pressure */
2677static bool in_reclaim_compaction(struct scan_control *sc)
2678{
2679 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2680 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2681 sc->priority < DEF_PRIORITY - 2))
2682 return true;
2683
2684 return false;
2685}
2686
2687/*
2688 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2689 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2690 * true if more pages should be reclaimed such that when the page allocator
2691 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2692 * It will give up earlier than that if there is difficulty reclaiming pages.
2693 */
2694static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2695 unsigned long nr_reclaimed,
2696 struct scan_control *sc)
2697{
2698 unsigned long pages_for_compaction;
2699 unsigned long inactive_lru_pages;
2700 int z;
2701
2702 /* If not in reclaim/compaction mode, stop */
2703 if (!in_reclaim_compaction(sc))
2704 return false;
2705
2706 /*
2707 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2708 * number of pages that were scanned. This will return to the caller
2709 * with the risk reclaim/compaction and the resulting allocation attempt
2710 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2711 * allocations through requiring that the full LRU list has been scanned
2712 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2713 * scan, but that approximation was wrong, and there were corner cases
2714 * where always a non-zero amount of pages were scanned.
2715 */
2716 if (!nr_reclaimed)
2717 return false;
2718
2719 /* If compaction would go ahead or the allocation would succeed, stop */
2720 for (z = 0; z <= sc->reclaim_idx; z++) {
2721 struct zone *zone = &pgdat->node_zones[z];
2722 if (!managed_zone(zone))
2723 continue;
2724
2725 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2726 case COMPACT_SUCCESS:
2727 case COMPACT_CONTINUE:
2728 return false;
2729 default:
2730 /* check next zone */
2731 ;
2732 }
2733 }
2734
2735 /*
2736 * If we have not reclaimed enough pages for compaction and the
2737 * inactive lists are large enough, continue reclaiming
2738 */
2739 pages_for_compaction = compact_gap(sc->order);
2740 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2741 if (get_nr_swap_pages() > 0)
2742 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2743
2744 return inactive_lru_pages > pages_for_compaction;
2745}
2746
2747static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2748{
2749 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2750 (memcg && memcg_congested(pgdat, memcg));
2751}
2752
2753static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2754{
2755 struct reclaim_state *reclaim_state = current->reclaim_state;
2756 unsigned long nr_reclaimed, nr_scanned;
2757 bool reclaimable = false;
2758
2759 do {
2760 struct mem_cgroup *root = sc->target_mem_cgroup;
2761 unsigned long node_lru_pages = 0;
2762 struct mem_cgroup *memcg;
2763
2764 memset(&sc->nr, 0, sizeof(sc->nr));
2765
2766 nr_reclaimed = sc->nr_reclaimed;
2767 nr_scanned = sc->nr_scanned;
2768
2769 memcg = mem_cgroup_iter(root, NULL, NULL);
2770 do {
2771 unsigned long lru_pages;
2772 unsigned long reclaimed;
2773 unsigned long scanned;
2774
2775 switch (mem_cgroup_protected(root, memcg)) {
2776 case MEMCG_PROT_MIN:
2777 /*
2778 * Hard protection.
2779 * If there is no reclaimable memory, OOM.
2780 */
2781 continue;
2782 case MEMCG_PROT_LOW:
2783 /*
2784 * Soft protection.
2785 * Respect the protection only as long as
2786 * there is an unprotected supply
2787 * of reclaimable memory from other cgroups.
2788 */
2789 if (!sc->memcg_low_reclaim) {
2790 sc->memcg_low_skipped = 1;
2791 continue;
2792 }
2793 memcg_memory_event(memcg, MEMCG_LOW);
2794 break;
2795 case MEMCG_PROT_NONE:
2796 /*
2797 * All protection thresholds breached. We may
2798 * still choose to vary the scan pressure
2799 * applied based on by how much the cgroup in
2800 * question has exceeded its protection
2801 * thresholds (see get_scan_count).
2802 */
2803 break;
2804 }
2805
2806 reclaimed = sc->nr_reclaimed;
2807 scanned = sc->nr_scanned;
2808 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2809 node_lru_pages += lru_pages;
2810
2811 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2812 sc->priority);
2813
2814 /* Record the group's reclaim efficiency */
2815 vmpressure(sc->gfp_mask, memcg, false,
2816 sc->nr_scanned - scanned,
2817 sc->nr_reclaimed - reclaimed);
2818
2819 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2820
2821 if (reclaim_state) {
2822 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2823 reclaim_state->reclaimed_slab = 0;
2824 }
2825
2826 /* Record the subtree's reclaim efficiency */
2827 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2828 sc->nr_scanned - nr_scanned,
2829 sc->nr_reclaimed - nr_reclaimed);
2830
2831 if (sc->nr_reclaimed - nr_reclaimed)
2832 reclaimable = true;
2833
2834 if (current_is_kswapd()) {
2835 /*
2836 * If reclaim is isolating dirty pages under writeback,
2837 * it implies that the long-lived page allocation rate
2838 * is exceeding the page laundering rate. Either the
2839 * global limits are not being effective at throttling
2840 * processes due to the page distribution throughout
2841 * zones or there is heavy usage of a slow backing
2842 * device. The only option is to throttle from reclaim
2843 * context which is not ideal as there is no guarantee
2844 * the dirtying process is throttled in the same way
2845 * balance_dirty_pages() manages.
2846 *
2847 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2848 * count the number of pages under pages flagged for
2849 * immediate reclaim and stall if any are encountered
2850 * in the nr_immediate check below.
2851 */
2852 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2853 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2854
2855 /*
2856 * Tag a node as congested if all the dirty pages
2857 * scanned were backed by a congested BDI and
2858 * wait_iff_congested will stall.
2859 */
2860 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2861 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2862
2863 /* Allow kswapd to start writing pages during reclaim.*/
2864 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2865 set_bit(PGDAT_DIRTY, &pgdat->flags);
2866
2867 /*
2868 * If kswapd scans pages marked marked for immediate
2869 * reclaim and under writeback (nr_immediate), it
2870 * implies that pages are cycling through the LRU
2871 * faster than they are written so also forcibly stall.
2872 */
2873 if (sc->nr.immediate)
2874 congestion_wait(BLK_RW_ASYNC, HZ/10);
2875 }
2876
2877 /*
2878 * Legacy memcg will stall in page writeback so avoid forcibly
2879 * stalling in wait_iff_congested().
2880 */
2881 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2882 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2883 set_memcg_congestion(pgdat, root, true);
2884
2885 /*
2886 * Stall direct reclaim for IO completions if underlying BDIs
2887 * and node is congested. Allow kswapd to continue until it
2888 * starts encountering unqueued dirty pages or cycling through
2889 * the LRU too quickly.
2890 */
2891 if (!sc->hibernation_mode && !current_is_kswapd() &&
2892 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2893 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2894
2895 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2896 sc));
2897
2898 /*
2899 * Kswapd gives up on balancing particular nodes after too
2900 * many failures to reclaim anything from them and goes to
2901 * sleep. On reclaim progress, reset the failure counter. A
2902 * successful direct reclaim run will revive a dormant kswapd.
2903 */
2904 if (reclaimable)
2905 pgdat->kswapd_failures = 0;
2906
2907 return reclaimable;
2908}
2909
2910/*
2911 * Returns true if compaction should go ahead for a costly-order request, or
2912 * the allocation would already succeed without compaction. Return false if we
2913 * should reclaim first.
2914 */
2915static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2916{
2917 unsigned long watermark;
2918 enum compact_result suitable;
2919
2920 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2921 if (suitable == COMPACT_SUCCESS)
2922 /* Allocation should succeed already. Don't reclaim. */
2923 return true;
2924 if (suitable == COMPACT_SKIPPED)
2925 /* Compaction cannot yet proceed. Do reclaim. */
2926 return false;
2927
2928 /*
2929 * Compaction is already possible, but it takes time to run and there
2930 * are potentially other callers using the pages just freed. So proceed
2931 * with reclaim to make a buffer of free pages available to give
2932 * compaction a reasonable chance of completing and allocating the page.
2933 * Note that we won't actually reclaim the whole buffer in one attempt
2934 * as the target watermark in should_continue_reclaim() is lower. But if
2935 * we are already above the high+gap watermark, don't reclaim at all.
2936 */
2937 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2938
2939 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2940}
2941
2942/*
2943 * This is the direct reclaim path, for page-allocating processes. We only
2944 * try to reclaim pages from zones which will satisfy the caller's allocation
2945 * request.
2946 *
2947 * If a zone is deemed to be full of pinned pages then just give it a light
2948 * scan then give up on it.
2949 */
2950static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2951{
2952 struct zoneref *z;
2953 struct zone *zone;
2954 unsigned long nr_soft_reclaimed;
2955 unsigned long nr_soft_scanned;
2956 gfp_t orig_mask;
2957 pg_data_t *last_pgdat = NULL;
2958
2959 /*
2960 * If the number of buffer_heads in the machine exceeds the maximum
2961 * allowed level, force direct reclaim to scan the highmem zone as
2962 * highmem pages could be pinning lowmem pages storing buffer_heads
2963 */
2964 orig_mask = sc->gfp_mask;
2965 if (buffer_heads_over_limit) {
2966 sc->gfp_mask |= __GFP_HIGHMEM;
2967 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2968 }
2969
2970 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2971 sc->reclaim_idx, sc->nodemask) {
2972 /*
2973 * Take care memory controller reclaiming has small influence
2974 * to global LRU.
2975 */
2976 if (global_reclaim(sc)) {
2977 if (!cpuset_zone_allowed(zone,
2978 GFP_KERNEL | __GFP_HARDWALL))
2979 continue;
2980
2981 /*
2982 * If we already have plenty of memory free for
2983 * compaction in this zone, don't free any more.
2984 * Even though compaction is invoked for any
2985 * non-zero order, only frequent costly order
2986 * reclamation is disruptive enough to become a
2987 * noticeable problem, like transparent huge
2988 * page allocations.
2989 */
2990 if (IS_ENABLED(CONFIG_COMPACTION) &&
2991 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2992 compaction_ready(zone, sc)) {
2993 sc->compaction_ready = true;
2994 continue;
2995 }
2996
2997 /*
2998 * Shrink each node in the zonelist once. If the
2999 * zonelist is ordered by zone (not the default) then a
3000 * node may be shrunk multiple times but in that case
3001 * the user prefers lower zones being preserved.
3002 */
3003 if (zone->zone_pgdat == last_pgdat)
3004 continue;
3005
3006 /*
3007 * This steals pages from memory cgroups over softlimit
3008 * and returns the number of reclaimed pages and
3009 * scanned pages. This works for global memory pressure
3010 * and balancing, not for a memcg's limit.
3011 */
3012 nr_soft_scanned = 0;
3013 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3014 sc->order, sc->gfp_mask,
3015 &nr_soft_scanned);
3016 sc->nr_reclaimed += nr_soft_reclaimed;
3017 sc->nr_scanned += nr_soft_scanned;
3018 /* need some check for avoid more shrink_zone() */
3019 }
3020
3021 /* See comment about same check for global reclaim above */
3022 if (zone->zone_pgdat == last_pgdat)
3023 continue;
3024 last_pgdat = zone->zone_pgdat;
3025 shrink_node(zone->zone_pgdat, sc);
3026 }
3027
3028 /*
3029 * Restore to original mask to avoid the impact on the caller if we
3030 * promoted it to __GFP_HIGHMEM.
3031 */
3032 sc->gfp_mask = orig_mask;
3033}
3034
3035static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3036{
3037 struct mem_cgroup *memcg;
3038
3039 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3040 do {
3041 unsigned long refaults;
3042 struct lruvec *lruvec;
3043
3044 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3045 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3046 lruvec->refaults = refaults;
3047 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3048}
3049
3050/*
3051 * This is the main entry point to direct page reclaim.
3052 *
3053 * If a full scan of the inactive list fails to free enough memory then we
3054 * are "out of memory" and something needs to be killed.
3055 *
3056 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3057 * high - the zone may be full of dirty or under-writeback pages, which this
3058 * caller can't do much about. We kick the writeback threads and take explicit
3059 * naps in the hope that some of these pages can be written. But if the
3060 * allocating task holds filesystem locks which prevent writeout this might not
3061 * work, and the allocation attempt will fail.
3062 *
3063 * returns: 0, if no pages reclaimed
3064 * else, the number of pages reclaimed
3065 */
3066static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3067 struct scan_control *sc)
3068{
3069 int initial_priority = sc->priority;
3070 pg_data_t *last_pgdat;
3071 struct zoneref *z;
3072 struct zone *zone;
3073retry:
3074 delayacct_freepages_start();
3075
3076 if (global_reclaim(sc))
3077 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3078
3079 do {
3080 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3081 sc->priority);
3082 sc->nr_scanned = 0;
3083 shrink_zones(zonelist, sc);
3084
3085 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3086 break;
3087
3088 if (sc->compaction_ready)
3089 break;
3090
3091 /*
3092 * If we're getting trouble reclaiming, start doing
3093 * writepage even in laptop mode.
3094 */
3095 if (sc->priority < DEF_PRIORITY - 2)
3096 sc->may_writepage = 1;
3097 } while (--sc->priority >= 0);
3098
3099 last_pgdat = NULL;
3100 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3101 sc->nodemask) {
3102 if (zone->zone_pgdat == last_pgdat)
3103 continue;
3104 last_pgdat = zone->zone_pgdat;
3105 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3106 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3107 }
3108
3109 delayacct_freepages_end();
3110
3111 if (sc->nr_reclaimed)
3112 return sc->nr_reclaimed;
3113
3114 /* Aborted reclaim to try compaction? don't OOM, then */
3115 if (sc->compaction_ready)
3116 return 1;
3117
3118 /* Untapped cgroup reserves? Don't OOM, retry. */
3119 if (sc->memcg_low_skipped) {
3120 sc->priority = initial_priority;
3121 sc->memcg_low_reclaim = 1;
3122 sc->memcg_low_skipped = 0;
3123 goto retry;
3124 }
3125
3126 return 0;
3127}
3128
3129static bool allow_direct_reclaim(pg_data_t *pgdat)
3130{
3131 struct zone *zone;
3132 unsigned long pfmemalloc_reserve = 0;
3133 unsigned long free_pages = 0;
3134 int i;
3135 bool wmark_ok;
3136
3137 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3138 return true;
3139
3140 for (i = 0; i <= ZONE_NORMAL; i++) {
3141 zone = &pgdat->node_zones[i];
3142 if (!managed_zone(zone))
3143 continue;
3144
3145 if (!zone_reclaimable_pages(zone))
3146 continue;
3147
3148 pfmemalloc_reserve += min_wmark_pages(zone);
3149 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3150 }
3151
3152 /* If there are no reserves (unexpected config) then do not throttle */
3153 if (!pfmemalloc_reserve)
3154 return true;
3155
3156 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3157
3158 /* kswapd must be awake if processes are being throttled */
3159 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3160 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3161 (enum zone_type)ZONE_NORMAL);
3162 wake_up_interruptible(&pgdat->kswapd_wait);
3163 }
3164
3165 return wmark_ok;
3166}
3167
3168/*
3169 * Throttle direct reclaimers if backing storage is backed by the network
3170 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3171 * depleted. kswapd will continue to make progress and wake the processes
3172 * when the low watermark is reached.
3173 *
3174 * Returns true if a fatal signal was delivered during throttling. If this
3175 * happens, the page allocator should not consider triggering the OOM killer.
3176 */
3177static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3178 nodemask_t *nodemask)
3179{
3180 struct zoneref *z;
3181 struct zone *zone;
3182 pg_data_t *pgdat = NULL;
3183
3184 /*
3185 * Kernel threads should not be throttled as they may be indirectly
3186 * responsible for cleaning pages necessary for reclaim to make forward
3187 * progress. kjournald for example may enter direct reclaim while
3188 * committing a transaction where throttling it could forcing other
3189 * processes to block on log_wait_commit().
3190 */
3191 if (current->flags & PF_KTHREAD)
3192 goto out;
3193
3194 /*
3195 * If a fatal signal is pending, this process should not throttle.
3196 * It should return quickly so it can exit and free its memory
3197 */
3198 if (fatal_signal_pending(current))
3199 goto out;
3200
3201 /*
3202 * Check if the pfmemalloc reserves are ok by finding the first node
3203 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3204 * GFP_KERNEL will be required for allocating network buffers when
3205 * swapping over the network so ZONE_HIGHMEM is unusable.
3206 *
3207 * Throttling is based on the first usable node and throttled processes
3208 * wait on a queue until kswapd makes progress and wakes them. There
3209 * is an affinity then between processes waking up and where reclaim
3210 * progress has been made assuming the process wakes on the same node.
3211 * More importantly, processes running on remote nodes will not compete
3212 * for remote pfmemalloc reserves and processes on different nodes
3213 * should make reasonable progress.
3214 */
3215 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3216 gfp_zone(gfp_mask), nodemask) {
3217 if (zone_idx(zone) > ZONE_NORMAL)
3218 continue;
3219
3220 /* Throttle based on the first usable node */
3221 pgdat = zone->zone_pgdat;
3222 if (allow_direct_reclaim(pgdat))
3223 goto out;
3224 break;
3225 }
3226
3227 /* If no zone was usable by the allocation flags then do not throttle */
3228 if (!pgdat)
3229 goto out;
3230
3231 /* Account for the throttling */
3232 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3233
3234 /*
3235 * If the caller cannot enter the filesystem, it's possible that it
3236 * is due to the caller holding an FS lock or performing a journal
3237 * transaction in the case of a filesystem like ext[3|4]. In this case,
3238 * it is not safe to block on pfmemalloc_wait as kswapd could be
3239 * blocked waiting on the same lock. Instead, throttle for up to a
3240 * second before continuing.
3241 */
3242 if (!(gfp_mask & __GFP_FS)) {
3243 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3244 allow_direct_reclaim(pgdat), HZ);
3245
3246 goto check_pending;
3247 }
3248
3249 /* Throttle until kswapd wakes the process */
3250 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3251 allow_direct_reclaim(pgdat));
3252
3253check_pending:
3254 if (fatal_signal_pending(current))
3255 return true;
3256
3257out:
3258 return false;
3259}
3260
3261unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3262 gfp_t gfp_mask, nodemask_t *nodemask)
3263{
3264 unsigned long nr_reclaimed;
3265 struct scan_control sc = {
3266 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3267 .gfp_mask = current_gfp_context(gfp_mask),
3268 .reclaim_idx = gfp_zone(gfp_mask),
3269 .order = order,
3270 .nodemask = nodemask,
3271 .priority = DEF_PRIORITY,
3272 .may_writepage = !laptop_mode,
3273 .may_unmap = 1,
3274 .may_swap = 1,
3275 };
3276
3277 /*
3278 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3279 * Confirm they are large enough for max values.
3280 */
3281 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3282 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3283 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3284
3285 /*
3286 * Do not enter reclaim if fatal signal was delivered while throttled.
3287 * 1 is returned so that the page allocator does not OOM kill at this
3288 * point.
3289 */
3290 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3291 return 1;
3292
3293 set_task_reclaim_state(current, &sc.reclaim_state);
3294 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3295
3296 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3297
3298 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3299 set_task_reclaim_state(current, NULL);
3300
3301 return nr_reclaimed;
3302}
3303
3304#ifdef CONFIG_MEMCG
3305
3306/* Only used by soft limit reclaim. Do not reuse for anything else. */
3307unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3308 gfp_t gfp_mask, bool noswap,
3309 pg_data_t *pgdat,
3310 unsigned long *nr_scanned)
3311{
3312 struct scan_control sc = {
3313 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3314 .target_mem_cgroup = memcg,
3315 .may_writepage = !laptop_mode,
3316 .may_unmap = 1,
3317 .reclaim_idx = MAX_NR_ZONES - 1,
3318 .may_swap = !noswap,
3319 };
3320 unsigned long lru_pages;
3321
3322 WARN_ON_ONCE(!current->reclaim_state);
3323
3324 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3325 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3326
3327 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3328 sc.gfp_mask);
3329
3330 /*
3331 * NOTE: Although we can get the priority field, using it
3332 * here is not a good idea, since it limits the pages we can scan.
3333 * if we don't reclaim here, the shrink_node from balance_pgdat
3334 * will pick up pages from other mem cgroup's as well. We hack
3335 * the priority and make it zero.
3336 */
3337 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3338
3339 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3340
3341 *nr_scanned = sc.nr_scanned;
3342
3343 return sc.nr_reclaimed;
3344}
3345
3346unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3347 unsigned long nr_pages,
3348 gfp_t gfp_mask,
3349 bool may_swap)
3350{
3351 struct zonelist *zonelist;
3352 unsigned long nr_reclaimed;
3353 unsigned long pflags;
3354 int nid;
3355 unsigned int noreclaim_flag;
3356 struct scan_control sc = {
3357 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3358 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3359 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3360 .reclaim_idx = MAX_NR_ZONES - 1,
3361 .target_mem_cgroup = memcg,
3362 .priority = DEF_PRIORITY,
3363 .may_writepage = !laptop_mode,
3364 .may_unmap = 1,
3365 .may_swap = may_swap,
3366 };
3367
3368 set_task_reclaim_state(current, &sc.reclaim_state);
3369 /*
3370 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3371 * take care of from where we get pages. So the node where we start the
3372 * scan does not need to be the current node.
3373 */
3374 nid = mem_cgroup_select_victim_node(memcg);
3375
3376 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3377
3378 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3379
3380 psi_memstall_enter(&pflags);
3381 noreclaim_flag = memalloc_noreclaim_save();
3382
3383 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3384
3385 memalloc_noreclaim_restore(noreclaim_flag);
3386 psi_memstall_leave(&pflags);
3387
3388 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3389 set_task_reclaim_state(current, NULL);
3390
3391 return nr_reclaimed;
3392}
3393#endif
3394
3395static void age_active_anon(struct pglist_data *pgdat,
3396 struct scan_control *sc)
3397{
3398 struct mem_cgroup *memcg;
3399
3400 if (!total_swap_pages)
3401 return;
3402
3403 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3404 do {
3405 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3406
3407 if (inactive_list_is_low(lruvec, false, sc, true))
3408 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3409 sc, LRU_ACTIVE_ANON);
3410
3411 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3412 } while (memcg);
3413}
3414
3415static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3416{
3417 int i;
3418 struct zone *zone;
3419
3420 /*
3421 * Check for watermark boosts top-down as the higher zones
3422 * are more likely to be boosted. Both watermarks and boosts
3423 * should not be checked at the time time as reclaim would
3424 * start prematurely when there is no boosting and a lower
3425 * zone is balanced.
3426 */
3427 for (i = classzone_idx; i >= 0; i--) {
3428 zone = pgdat->node_zones + i;
3429 if (!managed_zone(zone))
3430 continue;
3431
3432 if (zone->watermark_boost)
3433 return true;
3434 }
3435
3436 return false;
3437}
3438
3439/*
3440 * Returns true if there is an eligible zone balanced for the request order
3441 * and classzone_idx
3442 */
3443static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3444{
3445 int i;
3446 unsigned long mark = -1;
3447 struct zone *zone;
3448
3449 /*
3450 * Check watermarks bottom-up as lower zones are more likely to
3451 * meet watermarks.
3452 */
3453 for (i = 0; i <= classzone_idx; i++) {
3454 zone = pgdat->node_zones + i;
3455
3456 if (!managed_zone(zone))
3457 continue;
3458
3459 mark = high_wmark_pages(zone);
3460 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3461 return true;
3462 }
3463
3464 /*
3465 * If a node has no populated zone within classzone_idx, it does not
3466 * need balancing by definition. This can happen if a zone-restricted
3467 * allocation tries to wake a remote kswapd.
3468 */
3469 if (mark == -1)
3470 return true;
3471
3472 return false;
3473}
3474
3475/* Clear pgdat state for congested, dirty or under writeback. */
3476static void clear_pgdat_congested(pg_data_t *pgdat)
3477{
3478 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3479 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3480 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3481}
3482
3483/*
3484 * Prepare kswapd for sleeping. This verifies that there are no processes
3485 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3486 *
3487 * Returns true if kswapd is ready to sleep
3488 */
3489static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3490{
3491 /*
3492 * The throttled processes are normally woken up in balance_pgdat() as
3493 * soon as allow_direct_reclaim() is true. But there is a potential
3494 * race between when kswapd checks the watermarks and a process gets
3495 * throttled. There is also a potential race if processes get
3496 * throttled, kswapd wakes, a large process exits thereby balancing the
3497 * zones, which causes kswapd to exit balance_pgdat() before reaching
3498 * the wake up checks. If kswapd is going to sleep, no process should
3499 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3500 * the wake up is premature, processes will wake kswapd and get
3501 * throttled again. The difference from wake ups in balance_pgdat() is
3502 * that here we are under prepare_to_wait().
3503 */
3504 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3505 wake_up_all(&pgdat->pfmemalloc_wait);
3506
3507 /* Hopeless node, leave it to direct reclaim */
3508 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3509 return true;
3510
3511 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3512 clear_pgdat_congested(pgdat);
3513 return true;
3514 }
3515
3516 return false;
3517}
3518
3519/*
3520 * kswapd shrinks a node of pages that are at or below the highest usable
3521 * zone that is currently unbalanced.
3522 *
3523 * Returns true if kswapd scanned at least the requested number of pages to
3524 * reclaim or if the lack of progress was due to pages under writeback.
3525 * This is used to determine if the scanning priority needs to be raised.
3526 */
3527static bool kswapd_shrink_node(pg_data_t *pgdat,
3528 struct scan_control *sc)
3529{
3530 struct zone *zone;
3531 int z;
3532
3533 /* Reclaim a number of pages proportional to the number of zones */
3534 sc->nr_to_reclaim = 0;
3535 for (z = 0; z <= sc->reclaim_idx; z++) {
3536 zone = pgdat->node_zones + z;
3537 if (!managed_zone(zone))
3538 continue;
3539
3540 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3541 }
3542
3543 /*
3544 * Historically care was taken to put equal pressure on all zones but
3545 * now pressure is applied based on node LRU order.
3546 */
3547 shrink_node(pgdat, sc);
3548
3549 /*
3550 * Fragmentation may mean that the system cannot be rebalanced for
3551 * high-order allocations. If twice the allocation size has been
3552 * reclaimed then recheck watermarks only at order-0 to prevent
3553 * excessive reclaim. Assume that a process requested a high-order
3554 * can direct reclaim/compact.
3555 */
3556 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3557 sc->order = 0;
3558
3559 return sc->nr_scanned >= sc->nr_to_reclaim;
3560}
3561
3562/*
3563 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3564 * that are eligible for use by the caller until at least one zone is
3565 * balanced.
3566 *
3567 * Returns the order kswapd finished reclaiming at.
3568 *
3569 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3570 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3571 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3572 * or lower is eligible for reclaim until at least one usable zone is
3573 * balanced.
3574 */
3575static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3576{
3577 int i;
3578 unsigned long nr_soft_reclaimed;
3579 unsigned long nr_soft_scanned;
3580 unsigned long pflags;
3581 unsigned long nr_boost_reclaim;
3582 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3583 bool boosted;
3584 struct zone *zone;
3585 struct scan_control sc = {
3586 .gfp_mask = GFP_KERNEL,
3587 .order = order,
3588 .may_unmap = 1,
3589 };
3590
3591 set_task_reclaim_state(current, &sc.reclaim_state);
3592 psi_memstall_enter(&pflags);
3593 __fs_reclaim_acquire();
3594
3595 count_vm_event(PAGEOUTRUN);
3596
3597 /*
3598 * Account for the reclaim boost. Note that the zone boost is left in
3599 * place so that parallel allocations that are near the watermark will
3600 * stall or direct reclaim until kswapd is finished.
3601 */
3602 nr_boost_reclaim = 0;
3603 for (i = 0; i <= classzone_idx; i++) {
3604 zone = pgdat->node_zones + i;
3605 if (!managed_zone(zone))
3606 continue;
3607
3608 nr_boost_reclaim += zone->watermark_boost;
3609 zone_boosts[i] = zone->watermark_boost;
3610 }
3611 boosted = nr_boost_reclaim;
3612
3613restart:
3614 sc.priority = DEF_PRIORITY;
3615 do {
3616 unsigned long nr_reclaimed = sc.nr_reclaimed;
3617 bool raise_priority = true;
3618 bool balanced;
3619 bool ret;
3620
3621 sc.reclaim_idx = classzone_idx;
3622
3623 /*
3624 * If the number of buffer_heads exceeds the maximum allowed
3625 * then consider reclaiming from all zones. This has a dual
3626 * purpose -- on 64-bit systems it is expected that
3627 * buffer_heads are stripped during active rotation. On 32-bit
3628 * systems, highmem pages can pin lowmem memory and shrinking
3629 * buffers can relieve lowmem pressure. Reclaim may still not
3630 * go ahead if all eligible zones for the original allocation
3631 * request are balanced to avoid excessive reclaim from kswapd.
3632 */
3633 if (buffer_heads_over_limit) {
3634 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3635 zone = pgdat->node_zones + i;
3636 if (!managed_zone(zone))
3637 continue;
3638
3639 sc.reclaim_idx = i;
3640 break;
3641 }
3642 }
3643
3644 /*
3645 * If the pgdat is imbalanced then ignore boosting and preserve
3646 * the watermarks for a later time and restart. Note that the
3647 * zone watermarks will be still reset at the end of balancing
3648 * on the grounds that the normal reclaim should be enough to
3649 * re-evaluate if boosting is required when kswapd next wakes.
3650 */
3651 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3652 if (!balanced && nr_boost_reclaim) {
3653 nr_boost_reclaim = 0;
3654 goto restart;
3655 }
3656
3657 /*
3658 * If boosting is not active then only reclaim if there are no
3659 * eligible zones. Note that sc.reclaim_idx is not used as
3660 * buffer_heads_over_limit may have adjusted it.
3661 */
3662 if (!nr_boost_reclaim && balanced)
3663 goto out;
3664
3665 /* Limit the priority of boosting to avoid reclaim writeback */
3666 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3667 raise_priority = false;
3668
3669 /*
3670 * Do not writeback or swap pages for boosted reclaim. The
3671 * intent is to relieve pressure not issue sub-optimal IO
3672 * from reclaim context. If no pages are reclaimed, the
3673 * reclaim will be aborted.
3674 */
3675 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3676 sc.may_swap = !nr_boost_reclaim;
3677
3678 /*
3679 * Do some background aging of the anon list, to give
3680 * pages a chance to be referenced before reclaiming. All
3681 * pages are rotated regardless of classzone as this is
3682 * about consistent aging.
3683 */
3684 age_active_anon(pgdat, &sc);
3685
3686 /*
3687 * If we're getting trouble reclaiming, start doing writepage
3688 * even in laptop mode.
3689 */
3690 if (sc.priority < DEF_PRIORITY - 2)
3691 sc.may_writepage = 1;
3692
3693 /* Call soft limit reclaim before calling shrink_node. */
3694 sc.nr_scanned = 0;
3695 nr_soft_scanned = 0;
3696 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3697 sc.gfp_mask, &nr_soft_scanned);
3698 sc.nr_reclaimed += nr_soft_reclaimed;
3699
3700 /*
3701 * There should be no need to raise the scanning priority if
3702 * enough pages are already being scanned that that high
3703 * watermark would be met at 100% efficiency.
3704 */
3705 if (kswapd_shrink_node(pgdat, &sc))
3706 raise_priority = false;
3707
3708 /*
3709 * If the low watermark is met there is no need for processes
3710 * to be throttled on pfmemalloc_wait as they should not be
3711 * able to safely make forward progress. Wake them
3712 */
3713 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3714 allow_direct_reclaim(pgdat))
3715 wake_up_all(&pgdat->pfmemalloc_wait);
3716
3717 /* Check if kswapd should be suspending */
3718 __fs_reclaim_release();
3719 ret = try_to_freeze();
3720 __fs_reclaim_acquire();
3721 if (ret || kthread_should_stop())
3722 break;
3723
3724 /*
3725 * Raise priority if scanning rate is too low or there was no
3726 * progress in reclaiming pages
3727 */
3728 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3729 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3730
3731 /*
3732 * If reclaim made no progress for a boost, stop reclaim as
3733 * IO cannot be queued and it could be an infinite loop in
3734 * extreme circumstances.
3735 */
3736 if (nr_boost_reclaim && !nr_reclaimed)
3737 break;
3738
3739 if (raise_priority || !nr_reclaimed)
3740 sc.priority--;
3741 } while (sc.priority >= 1);
3742
3743 if (!sc.nr_reclaimed)
3744 pgdat->kswapd_failures++;
3745
3746out:
3747 /* If reclaim was boosted, account for the reclaim done in this pass */
3748 if (boosted) {
3749 unsigned long flags;
3750
3751 for (i = 0; i <= classzone_idx; i++) {
3752 if (!zone_boosts[i])
3753 continue;
3754
3755 /* Increments are under the zone lock */
3756 zone = pgdat->node_zones + i;
3757 spin_lock_irqsave(&zone->lock, flags);
3758 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3759 spin_unlock_irqrestore(&zone->lock, flags);
3760 }
3761
3762 /*
3763 * As there is now likely space, wakeup kcompact to defragment
3764 * pageblocks.
3765 */
3766 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3767 }
3768
3769 snapshot_refaults(NULL, pgdat);
3770 __fs_reclaim_release();
3771 psi_memstall_leave(&pflags);
3772 set_task_reclaim_state(current, NULL);
3773
3774 /*
3775 * Return the order kswapd stopped reclaiming at as
3776 * prepare_kswapd_sleep() takes it into account. If another caller
3777 * entered the allocator slow path while kswapd was awake, order will
3778 * remain at the higher level.
3779 */
3780 return sc.order;
3781}
3782
3783/*
3784 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3785 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3786 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3787 * after previous reclaim attempt (node is still unbalanced). In that case
3788 * return the zone index of the previous kswapd reclaim cycle.
3789 */
3790static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3791 enum zone_type prev_classzone_idx)
3792{
3793 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3794 return prev_classzone_idx;
3795 return pgdat->kswapd_classzone_idx;
3796}
3797
3798static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3799 unsigned int classzone_idx)
3800{
3801 long remaining = 0;
3802 DEFINE_WAIT(wait);
3803
3804 if (freezing(current) || kthread_should_stop())
3805 return;
3806
3807 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3808
3809 /*
3810 * Try to sleep for a short interval. Note that kcompactd will only be
3811 * woken if it is possible to sleep for a short interval. This is
3812 * deliberate on the assumption that if reclaim cannot keep an
3813 * eligible zone balanced that it's also unlikely that compaction will
3814 * succeed.
3815 */
3816 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3817 /*
3818 * Compaction records what page blocks it recently failed to
3819 * isolate pages from and skips them in the future scanning.
3820 * When kswapd is going to sleep, it is reasonable to assume
3821 * that pages and compaction may succeed so reset the cache.
3822 */
3823 reset_isolation_suitable(pgdat);
3824
3825 /*
3826 * We have freed the memory, now we should compact it to make
3827 * allocation of the requested order possible.
3828 */
3829 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3830
3831 remaining = schedule_timeout(HZ/10);
3832
3833 /*
3834 * If woken prematurely then reset kswapd_classzone_idx and
3835 * order. The values will either be from a wakeup request or
3836 * the previous request that slept prematurely.
3837 */
3838 if (remaining) {
3839 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3840 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3841 }
3842
3843 finish_wait(&pgdat->kswapd_wait, &wait);
3844 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3845 }
3846
3847 /*
3848 * After a short sleep, check if it was a premature sleep. If not, then
3849 * go fully to sleep until explicitly woken up.
3850 */
3851 if (!remaining &&
3852 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3853 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3854
3855 /*
3856 * vmstat counters are not perfectly accurate and the estimated
3857 * value for counters such as NR_FREE_PAGES can deviate from the
3858 * true value by nr_online_cpus * threshold. To avoid the zone
3859 * watermarks being breached while under pressure, we reduce the
3860 * per-cpu vmstat threshold while kswapd is awake and restore
3861 * them before going back to sleep.
3862 */
3863 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3864
3865 if (!kthread_should_stop())
3866 schedule();
3867
3868 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3869 } else {
3870 if (remaining)
3871 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3872 else
3873 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3874 }
3875 finish_wait(&pgdat->kswapd_wait, &wait);
3876}
3877
3878/*
3879 * The background pageout daemon, started as a kernel thread
3880 * from the init process.
3881 *
3882 * This basically trickles out pages so that we have _some_
3883 * free memory available even if there is no other activity
3884 * that frees anything up. This is needed for things like routing
3885 * etc, where we otherwise might have all activity going on in
3886 * asynchronous contexts that cannot page things out.
3887 *
3888 * If there are applications that are active memory-allocators
3889 * (most normal use), this basically shouldn't matter.
3890 */
3891static int kswapd(void *p)
3892{
3893 unsigned int alloc_order, reclaim_order;
3894 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3895 pg_data_t *pgdat = (pg_data_t*)p;
3896 struct task_struct *tsk = current;
3897 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3898
3899 if (!cpumask_empty(cpumask))
3900 set_cpus_allowed_ptr(tsk, cpumask);
3901
3902 /*
3903 * Tell the memory management that we're a "memory allocator",
3904 * and that if we need more memory we should get access to it
3905 * regardless (see "__alloc_pages()"). "kswapd" should
3906 * never get caught in the normal page freeing logic.
3907 *
3908 * (Kswapd normally doesn't need memory anyway, but sometimes
3909 * you need a small amount of memory in order to be able to
3910 * page out something else, and this flag essentially protects
3911 * us from recursively trying to free more memory as we're
3912 * trying to free the first piece of memory in the first place).
3913 */
3914 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3915 set_freezable();
3916
3917 pgdat->kswapd_order = 0;
3918 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3919 for ( ; ; ) {
3920 bool ret;
3921
3922 alloc_order = reclaim_order = pgdat->kswapd_order;
3923 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3924
3925kswapd_try_sleep:
3926 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3927 classzone_idx);
3928
3929 /* Read the new order and classzone_idx */
3930 alloc_order = reclaim_order = pgdat->kswapd_order;
3931 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3932 pgdat->kswapd_order = 0;
3933 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3934
3935 ret = try_to_freeze();
3936 if (kthread_should_stop())
3937 break;
3938
3939 /*
3940 * We can speed up thawing tasks if we don't call balance_pgdat
3941 * after returning from the refrigerator
3942 */
3943 if (ret)
3944 continue;
3945
3946 /*
3947 * Reclaim begins at the requested order but if a high-order
3948 * reclaim fails then kswapd falls back to reclaiming for
3949 * order-0. If that happens, kswapd will consider sleeping
3950 * for the order it finished reclaiming at (reclaim_order)
3951 * but kcompactd is woken to compact for the original
3952 * request (alloc_order).
3953 */
3954 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3955 alloc_order);
3956 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3957 if (reclaim_order < alloc_order)
3958 goto kswapd_try_sleep;
3959 }
3960
3961 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3962
3963 return 0;
3964}
3965
3966/*
3967 * A zone is low on free memory or too fragmented for high-order memory. If
3968 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3969 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3970 * has failed or is not needed, still wake up kcompactd if only compaction is
3971 * needed.
3972 */
3973void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3974 enum zone_type classzone_idx)
3975{
3976 pg_data_t *pgdat;
3977
3978 if (!managed_zone(zone))
3979 return;
3980
3981 if (!cpuset_zone_allowed(zone, gfp_flags))
3982 return;
3983 pgdat = zone->zone_pgdat;
3984
3985 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3986 pgdat->kswapd_classzone_idx = classzone_idx;
3987 else
3988 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3989 classzone_idx);
3990 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3991 if (!waitqueue_active(&pgdat->kswapd_wait))
3992 return;
3993
3994 /* Hopeless node, leave it to direct reclaim if possible */
3995 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3996 (pgdat_balanced(pgdat, order, classzone_idx) &&
3997 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3998 /*
3999 * There may be plenty of free memory available, but it's too
4000 * fragmented for high-order allocations. Wake up kcompactd
4001 * and rely on compaction_suitable() to determine if it's
4002 * needed. If it fails, it will defer subsequent attempts to
4003 * ratelimit its work.
4004 */
4005 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4006 wakeup_kcompactd(pgdat, order, classzone_idx);
4007 return;
4008 }
4009
4010 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
4011 gfp_flags);
4012 wake_up_interruptible(&pgdat->kswapd_wait);
4013}
4014
4015#ifdef CONFIG_HIBERNATION
4016/*
4017 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4018 * freed pages.
4019 *
4020 * Rather than trying to age LRUs the aim is to preserve the overall
4021 * LRU order by reclaiming preferentially
4022 * inactive > active > active referenced > active mapped
4023 */
4024unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4025{
4026 struct scan_control sc = {
4027 .nr_to_reclaim = nr_to_reclaim,
4028 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4029 .reclaim_idx = MAX_NR_ZONES - 1,
4030 .priority = DEF_PRIORITY,
4031 .may_writepage = 1,
4032 .may_unmap = 1,
4033 .may_swap = 1,
4034 .hibernation_mode = 1,
4035 };
4036 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4037 unsigned long nr_reclaimed;
4038 unsigned int noreclaim_flag;
4039
4040 fs_reclaim_acquire(sc.gfp_mask);
4041 noreclaim_flag = memalloc_noreclaim_save();
4042 set_task_reclaim_state(current, &sc.reclaim_state);
4043
4044 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4045
4046 set_task_reclaim_state(current, NULL);
4047 memalloc_noreclaim_restore(noreclaim_flag);
4048 fs_reclaim_release(sc.gfp_mask);
4049
4050 return nr_reclaimed;
4051}
4052#endif /* CONFIG_HIBERNATION */
4053
4054/* It's optimal to keep kswapds on the same CPUs as their memory, but
4055 not required for correctness. So if the last cpu in a node goes
4056 away, we get changed to run anywhere: as the first one comes back,
4057 restore their cpu bindings. */
4058static int kswapd_cpu_online(unsigned int cpu)
4059{
4060 int nid;
4061
4062 for_each_node_state(nid, N_MEMORY) {
4063 pg_data_t *pgdat = NODE_DATA(nid);
4064 const struct cpumask *mask;
4065
4066 mask = cpumask_of_node(pgdat->node_id);
4067
4068 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4069 /* One of our CPUs online: restore mask */
4070 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4071 }
4072 return 0;
4073}
4074
4075/*
4076 * This kswapd start function will be called by init and node-hot-add.
4077 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4078 */
4079int kswapd_run(int nid)
4080{
4081 pg_data_t *pgdat = NODE_DATA(nid);
4082 int ret = 0;
4083
4084 if (pgdat->kswapd)
4085 return 0;
4086
4087 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4088 if (IS_ERR(pgdat->kswapd)) {
4089 /* failure at boot is fatal */
4090 BUG_ON(system_state < SYSTEM_RUNNING);
4091 pr_err("Failed to start kswapd on node %d\n", nid);
4092 ret = PTR_ERR(pgdat->kswapd);
4093 pgdat->kswapd = NULL;
4094 }
4095 return ret;
4096}
4097
4098/*
4099 * Called by memory hotplug when all memory in a node is offlined. Caller must
4100 * hold mem_hotplug_begin/end().
4101 */
4102void kswapd_stop(int nid)
4103{
4104 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4105
4106 if (kswapd) {
4107 kthread_stop(kswapd);
4108 NODE_DATA(nid)->kswapd = NULL;
4109 }
4110}
4111
4112static int __init kswapd_init(void)
4113{
4114 int nid, ret;
4115
4116 swap_setup();
4117 for_each_node_state(nid, N_MEMORY)
4118 kswapd_run(nid);
4119 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4120 "mm/vmscan:online", kswapd_cpu_online,
4121 NULL);
4122 WARN_ON(ret < 0);
4123 return 0;
4124}
4125
4126module_init(kswapd_init)
4127
4128#ifdef CONFIG_NUMA
4129/*
4130 * Node reclaim mode
4131 *
4132 * If non-zero call node_reclaim when the number of free pages falls below
4133 * the watermarks.
4134 */
4135int node_reclaim_mode __read_mostly;
4136
4137#define RECLAIM_OFF 0
4138#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4139#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4140#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4141
4142/*
4143 * Priority for NODE_RECLAIM. This determines the fraction of pages
4144 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4145 * a zone.
4146 */
4147#define NODE_RECLAIM_PRIORITY 4
4148
4149/*
4150 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4151 * occur.
4152 */
4153int sysctl_min_unmapped_ratio = 1;
4154
4155/*
4156 * If the number of slab pages in a zone grows beyond this percentage then
4157 * slab reclaim needs to occur.
4158 */
4159int sysctl_min_slab_ratio = 5;
4160
4161static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4162{
4163 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4164 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4165 node_page_state(pgdat, NR_ACTIVE_FILE);
4166
4167 /*
4168 * It's possible for there to be more file mapped pages than
4169 * accounted for by the pages on the file LRU lists because
4170 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4171 */
4172 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4173}
4174
4175/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4176static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4177{
4178 unsigned long nr_pagecache_reclaimable;
4179 unsigned long delta = 0;
4180
4181 /*
4182 * If RECLAIM_UNMAP is set, then all file pages are considered
4183 * potentially reclaimable. Otherwise, we have to worry about
4184 * pages like swapcache and node_unmapped_file_pages() provides
4185 * a better estimate
4186 */
4187 if (node_reclaim_mode & RECLAIM_UNMAP)
4188 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4189 else
4190 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4191
4192 /* If we can't clean pages, remove dirty pages from consideration */
4193 if (!(node_reclaim_mode & RECLAIM_WRITE))
4194 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4195
4196 /* Watch for any possible underflows due to delta */
4197 if (unlikely(delta > nr_pagecache_reclaimable))
4198 delta = nr_pagecache_reclaimable;
4199
4200 return nr_pagecache_reclaimable - delta;
4201}
4202
4203/*
4204 * Try to free up some pages from this node through reclaim.
4205 */
4206static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4207{
4208 /* Minimum pages needed in order to stay on node */
4209 const unsigned long nr_pages = 1 << order;
4210 struct task_struct *p = current;
4211 unsigned int noreclaim_flag;
4212 struct scan_control sc = {
4213 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4214 .gfp_mask = current_gfp_context(gfp_mask),
4215 .order = order,
4216 .priority = NODE_RECLAIM_PRIORITY,
4217 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4218 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4219 .may_swap = 1,
4220 .reclaim_idx = gfp_zone(gfp_mask),
4221 };
4222
4223 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4224 sc.gfp_mask);
4225
4226 cond_resched();
4227 fs_reclaim_acquire(sc.gfp_mask);
4228 /*
4229 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4230 * and we also need to be able to write out pages for RECLAIM_WRITE
4231 * and RECLAIM_UNMAP.
4232 */
4233 noreclaim_flag = memalloc_noreclaim_save();
4234 p->flags |= PF_SWAPWRITE;
4235 set_task_reclaim_state(p, &sc.reclaim_state);
4236
4237 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4238 /*
4239 * Free memory by calling shrink node with increasing
4240 * priorities until we have enough memory freed.
4241 */
4242 do {
4243 shrink_node(pgdat, &sc);
4244 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4245 }
4246
4247 set_task_reclaim_state(p, NULL);
4248 current->flags &= ~PF_SWAPWRITE;
4249 memalloc_noreclaim_restore(noreclaim_flag);
4250 fs_reclaim_release(sc.gfp_mask);
4251
4252 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4253
4254 return sc.nr_reclaimed >= nr_pages;
4255}
4256
4257int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4258{
4259 int ret;
4260
4261 /*
4262 * Node reclaim reclaims unmapped file backed pages and
4263 * slab pages if we are over the defined limits.
4264 *
4265 * A small portion of unmapped file backed pages is needed for
4266 * file I/O otherwise pages read by file I/O will be immediately
4267 * thrown out if the node is overallocated. So we do not reclaim
4268 * if less than a specified percentage of the node is used by
4269 * unmapped file backed pages.
4270 */
4271 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4272 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4273 return NODE_RECLAIM_FULL;
4274
4275 /*
4276 * Do not scan if the allocation should not be delayed.
4277 */
4278 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4279 return NODE_RECLAIM_NOSCAN;
4280
4281 /*
4282 * Only run node reclaim on the local node or on nodes that do not
4283 * have associated processors. This will favor the local processor
4284 * over remote processors and spread off node memory allocations
4285 * as wide as possible.
4286 */
4287 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4288 return NODE_RECLAIM_NOSCAN;
4289
4290 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4291 return NODE_RECLAIM_NOSCAN;
4292
4293 ret = __node_reclaim(pgdat, gfp_mask, order);
4294 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4295
4296 if (!ret)
4297 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4298
4299 return ret;
4300}
4301#endif
4302
4303/*
4304 * page_evictable - test whether a page is evictable
4305 * @page: the page to test
4306 *
4307 * Test whether page is evictable--i.e., should be placed on active/inactive
4308 * lists vs unevictable list.
4309 *
4310 * Reasons page might not be evictable:
4311 * (1) page's mapping marked unevictable
4312 * (2) page is part of an mlocked VMA
4313 *
4314 */
4315int page_evictable(struct page *page)
4316{
4317 int ret;
4318
4319 /* Prevent address_space of inode and swap cache from being freed */
4320 rcu_read_lock();
4321 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4322 rcu_read_unlock();
4323 return ret;
4324}
4325
4326/**
4327 * check_move_unevictable_pages - check pages for evictability and move to
4328 * appropriate zone lru list
4329 * @pvec: pagevec with lru pages to check
4330 *
4331 * Checks pages for evictability, if an evictable page is in the unevictable
4332 * lru list, moves it to the appropriate evictable lru list. This function
4333 * should be only used for lru pages.
4334 */
4335void check_move_unevictable_pages(struct pagevec *pvec)
4336{
4337 struct lruvec *lruvec;
4338 struct pglist_data *pgdat = NULL;
4339 int pgscanned = 0;
4340 int pgrescued = 0;
4341 int i;
4342
4343 for (i = 0; i < pvec->nr; i++) {
4344 struct page *page = pvec->pages[i];
4345 struct pglist_data *pagepgdat = page_pgdat(page);
4346
4347 pgscanned++;
4348 if (pagepgdat != pgdat) {
4349 if (pgdat)
4350 spin_unlock_irq(&pgdat->lru_lock);
4351 pgdat = pagepgdat;
4352 spin_lock_irq(&pgdat->lru_lock);
4353 }
4354 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4355
4356 if (!PageLRU(page) || !PageUnevictable(page))
4357 continue;
4358
4359 if (page_evictable(page)) {
4360 enum lru_list lru = page_lru_base_type(page);
4361
4362 VM_BUG_ON_PAGE(PageActive(page), page);
4363 ClearPageUnevictable(page);
4364 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4365 add_page_to_lru_list(page, lruvec, lru);
4366 pgrescued++;
4367 }
4368 }
4369
4370 if (pgdat) {
4371 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4372 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4373 spin_unlock_irq(&pgdat->lru_lock);
4374 }
4375}
4376EXPORT_SYMBOL_GPL(check_move_unevictable_pages);