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/*
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#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/gfp.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
22#include <linux/vmpressure.h>
23#include <linux/vmstat.h>
24#include <linux/file.h>
25#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29#include <linux/mm_inline.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/compaction.h>
36#include <linux/notifier.h>
37#include <linux/rwsem.h>
38#include <linux/delay.h>
39#include <linux/kthread.h>
40#include <linux/freezer.h>
41#include <linux/memcontrol.h>
42#include <linux/delayacct.h>
43#include <linux/sysctl.h>
44#include <linux/oom.h>
45#include <linux/prefetch.h>
46
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
51#include <linux/balloon_compaction.h>
52
53#include "internal.h"
54
55#define CREATE_TRACE_POINTS
56#include <trace/events/vmscan.h>
57
58struct scan_control {
59 /* Incremented by the number of inactive pages that were scanned */
60 unsigned long nr_scanned;
61
62 /* Number of pages freed so far during a call to shrink_zones() */
63 unsigned long nr_reclaimed;
64
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
68 unsigned long hibernation_mode;
69
70 /* This context's GFP mask */
71 gfp_t gfp_mask;
72
73 int may_writepage;
74
75 /* Can mapped pages be reclaimed? */
76 int may_unmap;
77
78 /* Can pages be swapped as part of reclaim? */
79 int may_swap;
80
81 int order;
82
83 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
86 /*
87 * The memory cgroup that hit its limit and as a result is the
88 * primary target of this reclaim invocation.
89 */
90 struct mem_cgroup *target_mem_cgroup;
91
92 /*
93 * Nodemask of nodes allowed by the caller. If NULL, all nodes
94 * are scanned.
95 */
96 nodemask_t *nodemask;
97};
98
99#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100
101#ifdef ARCH_HAS_PREFETCH
102#define prefetch_prev_lru_page(_page, _base, _field) \
103 do { \
104 if ((_page)->lru.prev != _base) { \
105 struct page *prev; \
106 \
107 prev = lru_to_page(&(_page->lru)); \
108 prefetch(&prev->_field); \
109 } \
110 } while (0)
111#else
112#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113#endif
114
115#ifdef ARCH_HAS_PREFETCHW
116#define prefetchw_prev_lru_page(_page, _base, _field) \
117 do { \
118 if ((_page)->lru.prev != _base) { \
119 struct page *prev; \
120 \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetchw(&prev->_field); \
123 } \
124 } while (0)
125#else
126#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
127#endif
128
129/*
130 * From 0 .. 100. Higher means more swappy.
131 */
132int vm_swappiness = 60;
133unsigned long vm_total_pages; /* The total number of pages which the VM controls */
134
135static LIST_HEAD(shrinker_list);
136static DECLARE_RWSEM(shrinker_rwsem);
137
138#ifdef CONFIG_MEMCG
139static bool global_reclaim(struct scan_control *sc)
140{
141 return !sc->target_mem_cgroup;
142}
143#else
144static bool global_reclaim(struct scan_control *sc)
145{
146 return true;
147}
148#endif
149
150static unsigned long zone_reclaimable_pages(struct zone *zone)
151{
152 int nr;
153
154 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
155 zone_page_state(zone, NR_INACTIVE_FILE);
156
157 if (get_nr_swap_pages() > 0)
158 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
159 zone_page_state(zone, NR_INACTIVE_ANON);
160
161 return nr;
162}
163
164bool zone_reclaimable(struct zone *zone)
165{
166 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
167}
168
169static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
170{
171 if (!mem_cgroup_disabled())
172 return mem_cgroup_get_lru_size(lruvec, lru);
173
174 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
175}
176
177/*
178 * Add a shrinker callback to be called from the vm.
179 */
180int register_shrinker(struct shrinker *shrinker)
181{
182 size_t size = sizeof(*shrinker->nr_deferred);
183
184 /*
185 * If we only have one possible node in the system anyway, save
186 * ourselves the trouble and disable NUMA aware behavior. This way we
187 * will save memory and some small loop time later.
188 */
189 if (nr_node_ids == 1)
190 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
191
192 if (shrinker->flags & SHRINKER_NUMA_AWARE)
193 size *= nr_node_ids;
194
195 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
196 if (!shrinker->nr_deferred)
197 return -ENOMEM;
198
199 down_write(&shrinker_rwsem);
200 list_add_tail(&shrinker->list, &shrinker_list);
201 up_write(&shrinker_rwsem);
202 return 0;
203}
204EXPORT_SYMBOL(register_shrinker);
205
206/*
207 * Remove one
208 */
209void unregister_shrinker(struct shrinker *shrinker)
210{
211 down_write(&shrinker_rwsem);
212 list_del(&shrinker->list);
213 up_write(&shrinker_rwsem);
214 kfree(shrinker->nr_deferred);
215}
216EXPORT_SYMBOL(unregister_shrinker);
217
218#define SHRINK_BATCH 128
219
220static unsigned long
221shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
222 unsigned long nr_pages_scanned, unsigned long lru_pages)
223{
224 unsigned long freed = 0;
225 unsigned long long delta;
226 long total_scan;
227 long freeable;
228 long nr;
229 long new_nr;
230 int nid = shrinkctl->nid;
231 long batch_size = shrinker->batch ? shrinker->batch
232 : SHRINK_BATCH;
233
234 freeable = shrinker->count_objects(shrinker, shrinkctl);
235 if (freeable == 0)
236 return 0;
237
238 /*
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
242 */
243 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
244
245 total_scan = nr;
246 delta = (4 * nr_pages_scanned) / shrinker->seeks;
247 delta *= freeable;
248 do_div(delta, lru_pages + 1);
249 total_scan += delta;
250 if (total_scan < 0) {
251 printk(KERN_ERR
252 "shrink_slab: %pF negative objects to delete nr=%ld\n",
253 shrinker->scan_objects, total_scan);
254 total_scan = freeable;
255 }
256
257 /*
258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
263 * freeable. This is bad for sustaining a working set in
264 * memory.
265 *
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
268 */
269 if (delta < freeable / 4)
270 total_scan = min(total_scan, freeable / 2);
271
272 /*
273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
275 * freeable entries.
276 */
277 if (total_scan > freeable * 2)
278 total_scan = freeable * 2;
279
280 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
281 nr_pages_scanned, lru_pages,
282 freeable, delta, total_scan);
283
284 /*
285 * Normally, we should not scan less than batch_size objects in one
286 * pass to avoid too frequent shrinker calls, but if the slab has less
287 * than batch_size objects in total and we are really tight on memory,
288 * we will try to reclaim all available objects, otherwise we can end
289 * up failing allocations although there are plenty of reclaimable
290 * objects spread over several slabs with usage less than the
291 * batch_size.
292 *
293 * We detect the "tight on memory" situations by looking at the total
294 * number of objects we want to scan (total_scan). If it is greater
295 * than the total number of objects on slab (freeable), we must be
296 * scanning at high prio and therefore should try to reclaim as much as
297 * possible.
298 */
299 while (total_scan >= batch_size ||
300 total_scan >= freeable) {
301 unsigned long ret;
302 unsigned long nr_to_scan = min(batch_size, total_scan);
303
304 shrinkctl->nr_to_scan = nr_to_scan;
305 ret = shrinker->scan_objects(shrinker, shrinkctl);
306 if (ret == SHRINK_STOP)
307 break;
308 freed += ret;
309
310 count_vm_events(SLABS_SCANNED, nr_to_scan);
311 total_scan -= nr_to_scan;
312
313 cond_resched();
314 }
315
316 /*
317 * move the unused scan count back into the shrinker in a
318 * manner that handles concurrent updates. If we exhausted the
319 * scan, there is no need to do an update.
320 */
321 if (total_scan > 0)
322 new_nr = atomic_long_add_return(total_scan,
323 &shrinker->nr_deferred[nid]);
324 else
325 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
326
327 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
328 return freed;
329}
330
331/*
332 * Call the shrink functions to age shrinkable caches
333 *
334 * Here we assume it costs one seek to replace a lru page and that it also
335 * takes a seek to recreate a cache object. With this in mind we age equal
336 * percentages of the lru and ageable caches. This should balance the seeks
337 * generated by these structures.
338 *
339 * If the vm encountered mapped pages on the LRU it increase the pressure on
340 * slab to avoid swapping.
341 *
342 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
343 *
344 * `lru_pages' represents the number of on-LRU pages in all the zones which
345 * are eligible for the caller's allocation attempt. It is used for balancing
346 * slab reclaim versus page reclaim.
347 *
348 * Returns the number of slab objects which we shrunk.
349 */
350unsigned long shrink_slab(struct shrink_control *shrinkctl,
351 unsigned long nr_pages_scanned,
352 unsigned long lru_pages)
353{
354 struct shrinker *shrinker;
355 unsigned long freed = 0;
356
357 if (nr_pages_scanned == 0)
358 nr_pages_scanned = SWAP_CLUSTER_MAX;
359
360 if (!down_read_trylock(&shrinker_rwsem)) {
361 /*
362 * If we would return 0, our callers would understand that we
363 * have nothing else to shrink and give up trying. By returning
364 * 1 we keep it going and assume we'll be able to shrink next
365 * time.
366 */
367 freed = 1;
368 goto out;
369 }
370
371 list_for_each_entry(shrinker, &shrinker_list, list) {
372 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
373 shrinkctl->nid = 0;
374 freed += shrink_slab_node(shrinkctl, shrinker,
375 nr_pages_scanned, lru_pages);
376 continue;
377 }
378
379 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
380 if (node_online(shrinkctl->nid))
381 freed += shrink_slab_node(shrinkctl, shrinker,
382 nr_pages_scanned, lru_pages);
383
384 }
385 }
386 up_read(&shrinker_rwsem);
387out:
388 cond_resched();
389 return freed;
390}
391
392static inline int is_page_cache_freeable(struct page *page)
393{
394 /*
395 * A freeable page cache page is referenced only by the caller
396 * that isolated the page, the page cache radix tree and
397 * optional buffer heads at page->private.
398 */
399 return page_count(page) - page_has_private(page) == 2;
400}
401
402static int may_write_to_queue(struct backing_dev_info *bdi,
403 struct scan_control *sc)
404{
405 if (current->flags & PF_SWAPWRITE)
406 return 1;
407 if (!bdi_write_congested(bdi))
408 return 1;
409 if (bdi == current->backing_dev_info)
410 return 1;
411 return 0;
412}
413
414/*
415 * We detected a synchronous write error writing a page out. Probably
416 * -ENOSPC. We need to propagate that into the address_space for a subsequent
417 * fsync(), msync() or close().
418 *
419 * The tricky part is that after writepage we cannot touch the mapping: nothing
420 * prevents it from being freed up. But we have a ref on the page and once
421 * that page is locked, the mapping is pinned.
422 *
423 * We're allowed to run sleeping lock_page() here because we know the caller has
424 * __GFP_FS.
425 */
426static void handle_write_error(struct address_space *mapping,
427 struct page *page, int error)
428{
429 lock_page(page);
430 if (page_mapping(page) == mapping)
431 mapping_set_error(mapping, error);
432 unlock_page(page);
433}
434
435/* possible outcome of pageout() */
436typedef enum {
437 /* failed to write page out, page is locked */
438 PAGE_KEEP,
439 /* move page to the active list, page is locked */
440 PAGE_ACTIVATE,
441 /* page has been sent to the disk successfully, page is unlocked */
442 PAGE_SUCCESS,
443 /* page is clean and locked */
444 PAGE_CLEAN,
445} pageout_t;
446
447/*
448 * pageout is called by shrink_page_list() for each dirty page.
449 * Calls ->writepage().
450 */
451static pageout_t pageout(struct page *page, struct address_space *mapping,
452 struct scan_control *sc)
453{
454 /*
455 * If the page is dirty, only perform writeback if that write
456 * will be non-blocking. To prevent this allocation from being
457 * stalled by pagecache activity. But note that there may be
458 * stalls if we need to run get_block(). We could test
459 * PagePrivate for that.
460 *
461 * If this process is currently in __generic_file_aio_write() against
462 * this page's queue, we can perform writeback even if that
463 * will block.
464 *
465 * If the page is swapcache, write it back even if that would
466 * block, for some throttling. This happens by accident, because
467 * swap_backing_dev_info is bust: it doesn't reflect the
468 * congestion state of the swapdevs. Easy to fix, if needed.
469 */
470 if (!is_page_cache_freeable(page))
471 return PAGE_KEEP;
472 if (!mapping) {
473 /*
474 * Some data journaling orphaned pages can have
475 * page->mapping == NULL while being dirty with clean buffers.
476 */
477 if (page_has_private(page)) {
478 if (try_to_free_buffers(page)) {
479 ClearPageDirty(page);
480 printk("%s: orphaned page\n", __func__);
481 return PAGE_CLEAN;
482 }
483 }
484 return PAGE_KEEP;
485 }
486 if (mapping->a_ops->writepage == NULL)
487 return PAGE_ACTIVATE;
488 if (!may_write_to_queue(mapping->backing_dev_info, sc))
489 return PAGE_KEEP;
490
491 if (clear_page_dirty_for_io(page)) {
492 int res;
493 struct writeback_control wbc = {
494 .sync_mode = WB_SYNC_NONE,
495 .nr_to_write = SWAP_CLUSTER_MAX,
496 .range_start = 0,
497 .range_end = LLONG_MAX,
498 .for_reclaim = 1,
499 };
500
501 SetPageReclaim(page);
502 res = mapping->a_ops->writepage(page, &wbc);
503 if (res < 0)
504 handle_write_error(mapping, page, res);
505 if (res == AOP_WRITEPAGE_ACTIVATE) {
506 ClearPageReclaim(page);
507 return PAGE_ACTIVATE;
508 }
509
510 if (!PageWriteback(page)) {
511 /* synchronous write or broken a_ops? */
512 ClearPageReclaim(page);
513 }
514 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
515 inc_zone_page_state(page, NR_VMSCAN_WRITE);
516 return PAGE_SUCCESS;
517 }
518
519 return PAGE_CLEAN;
520}
521
522/*
523 * Same as remove_mapping, but if the page is removed from the mapping, it
524 * gets returned with a refcount of 0.
525 */
526static int __remove_mapping(struct address_space *mapping, struct page *page,
527 bool reclaimed)
528{
529 BUG_ON(!PageLocked(page));
530 BUG_ON(mapping != page_mapping(page));
531
532 spin_lock_irq(&mapping->tree_lock);
533 /*
534 * The non racy check for a busy page.
535 *
536 * Must be careful with the order of the tests. When someone has
537 * a ref to the page, it may be possible that they dirty it then
538 * drop the reference. So if PageDirty is tested before page_count
539 * here, then the following race may occur:
540 *
541 * get_user_pages(&page);
542 * [user mapping goes away]
543 * write_to(page);
544 * !PageDirty(page) [good]
545 * SetPageDirty(page);
546 * put_page(page);
547 * !page_count(page) [good, discard it]
548 *
549 * [oops, our write_to data is lost]
550 *
551 * Reversing the order of the tests ensures such a situation cannot
552 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
553 * load is not satisfied before that of page->_count.
554 *
555 * Note that if SetPageDirty is always performed via set_page_dirty,
556 * and thus under tree_lock, then this ordering is not required.
557 */
558 if (!page_freeze_refs(page, 2))
559 goto cannot_free;
560 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
561 if (unlikely(PageDirty(page))) {
562 page_unfreeze_refs(page, 2);
563 goto cannot_free;
564 }
565
566 if (PageSwapCache(page)) {
567 swp_entry_t swap = { .val = page_private(page) };
568 __delete_from_swap_cache(page);
569 spin_unlock_irq(&mapping->tree_lock);
570 swapcache_free(swap, page);
571 } else {
572 void (*freepage)(struct page *);
573 void *shadow = NULL;
574
575 freepage = mapping->a_ops->freepage;
576 /*
577 * Remember a shadow entry for reclaimed file cache in
578 * order to detect refaults, thus thrashing, later on.
579 *
580 * But don't store shadows in an address space that is
581 * already exiting. This is not just an optizimation,
582 * inode reclaim needs to empty out the radix tree or
583 * the nodes are lost. Don't plant shadows behind its
584 * back.
585 */
586 if (reclaimed && page_is_file_cache(page) &&
587 !mapping_exiting(mapping))
588 shadow = workingset_eviction(mapping, page);
589 __delete_from_page_cache(page, shadow);
590 spin_unlock_irq(&mapping->tree_lock);
591 mem_cgroup_uncharge_cache_page(page);
592
593 if (freepage != NULL)
594 freepage(page);
595 }
596
597 return 1;
598
599cannot_free:
600 spin_unlock_irq(&mapping->tree_lock);
601 return 0;
602}
603
604/*
605 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
606 * someone else has a ref on the page, abort and return 0. If it was
607 * successfully detached, return 1. Assumes the caller has a single ref on
608 * this page.
609 */
610int remove_mapping(struct address_space *mapping, struct page *page)
611{
612 if (__remove_mapping(mapping, page, false)) {
613 /*
614 * Unfreezing the refcount with 1 rather than 2 effectively
615 * drops the pagecache ref for us without requiring another
616 * atomic operation.
617 */
618 page_unfreeze_refs(page, 1);
619 return 1;
620 }
621 return 0;
622}
623
624/**
625 * putback_lru_page - put previously isolated page onto appropriate LRU list
626 * @page: page to be put back to appropriate lru list
627 *
628 * Add previously isolated @page to appropriate LRU list.
629 * Page may still be unevictable for other reasons.
630 *
631 * lru_lock must not be held, interrupts must be enabled.
632 */
633void putback_lru_page(struct page *page)
634{
635 bool is_unevictable;
636 int was_unevictable = PageUnevictable(page);
637
638 VM_BUG_ON_PAGE(PageLRU(page), page);
639
640redo:
641 ClearPageUnevictable(page);
642
643 if (page_evictable(page)) {
644 /*
645 * For evictable pages, we can use the cache.
646 * In event of a race, worst case is we end up with an
647 * unevictable page on [in]active list.
648 * We know how to handle that.
649 */
650 is_unevictable = false;
651 lru_cache_add(page);
652 } else {
653 /*
654 * Put unevictable pages directly on zone's unevictable
655 * list.
656 */
657 is_unevictable = true;
658 add_page_to_unevictable_list(page);
659 /*
660 * When racing with an mlock or AS_UNEVICTABLE clearing
661 * (page is unlocked) make sure that if the other thread
662 * does not observe our setting of PG_lru and fails
663 * isolation/check_move_unevictable_pages,
664 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
665 * the page back to the evictable list.
666 *
667 * The other side is TestClearPageMlocked() or shmem_lock().
668 */
669 smp_mb();
670 }
671
672 /*
673 * page's status can change while we move it among lru. If an evictable
674 * page is on unevictable list, it never be freed. To avoid that,
675 * check after we added it to the list, again.
676 */
677 if (is_unevictable && page_evictable(page)) {
678 if (!isolate_lru_page(page)) {
679 put_page(page);
680 goto redo;
681 }
682 /* This means someone else dropped this page from LRU
683 * So, it will be freed or putback to LRU again. There is
684 * nothing to do here.
685 */
686 }
687
688 if (was_unevictable && !is_unevictable)
689 count_vm_event(UNEVICTABLE_PGRESCUED);
690 else if (!was_unevictable && is_unevictable)
691 count_vm_event(UNEVICTABLE_PGCULLED);
692
693 put_page(page); /* drop ref from isolate */
694}
695
696enum page_references {
697 PAGEREF_RECLAIM,
698 PAGEREF_RECLAIM_CLEAN,
699 PAGEREF_KEEP,
700 PAGEREF_ACTIVATE,
701};
702
703static enum page_references page_check_references(struct page *page,
704 struct scan_control *sc)
705{
706 int referenced_ptes, referenced_page;
707 unsigned long vm_flags;
708
709 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
710 &vm_flags);
711 referenced_page = TestClearPageReferenced(page);
712
713 /*
714 * Mlock lost the isolation race with us. Let try_to_unmap()
715 * move the page to the unevictable list.
716 */
717 if (vm_flags & VM_LOCKED)
718 return PAGEREF_RECLAIM;
719
720 if (referenced_ptes) {
721 if (PageSwapBacked(page))
722 return PAGEREF_ACTIVATE;
723 /*
724 * All mapped pages start out with page table
725 * references from the instantiating fault, so we need
726 * to look twice if a mapped file page is used more
727 * than once.
728 *
729 * Mark it and spare it for another trip around the
730 * inactive list. Another page table reference will
731 * lead to its activation.
732 *
733 * Note: the mark is set for activated pages as well
734 * so that recently deactivated but used pages are
735 * quickly recovered.
736 */
737 SetPageReferenced(page);
738
739 if (referenced_page || referenced_ptes > 1)
740 return PAGEREF_ACTIVATE;
741
742 /*
743 * Activate file-backed executable pages after first usage.
744 */
745 if (vm_flags & VM_EXEC)
746 return PAGEREF_ACTIVATE;
747
748 return PAGEREF_KEEP;
749 }
750
751 /* Reclaim if clean, defer dirty pages to writeback */
752 if (referenced_page && !PageSwapBacked(page))
753 return PAGEREF_RECLAIM_CLEAN;
754
755 return PAGEREF_RECLAIM;
756}
757
758/* Check if a page is dirty or under writeback */
759static void page_check_dirty_writeback(struct page *page,
760 bool *dirty, bool *writeback)
761{
762 struct address_space *mapping;
763
764 /*
765 * Anonymous pages are not handled by flushers and must be written
766 * from reclaim context. Do not stall reclaim based on them
767 */
768 if (!page_is_file_cache(page)) {
769 *dirty = false;
770 *writeback = false;
771 return;
772 }
773
774 /* By default assume that the page flags are accurate */
775 *dirty = PageDirty(page);
776 *writeback = PageWriteback(page);
777
778 /* Verify dirty/writeback state if the filesystem supports it */
779 if (!page_has_private(page))
780 return;
781
782 mapping = page_mapping(page);
783 if (mapping && mapping->a_ops->is_dirty_writeback)
784 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
785}
786
787/*
788 * shrink_page_list() returns the number of reclaimed pages
789 */
790static unsigned long shrink_page_list(struct list_head *page_list,
791 struct zone *zone,
792 struct scan_control *sc,
793 enum ttu_flags ttu_flags,
794 unsigned long *ret_nr_dirty,
795 unsigned long *ret_nr_unqueued_dirty,
796 unsigned long *ret_nr_congested,
797 unsigned long *ret_nr_writeback,
798 unsigned long *ret_nr_immediate,
799 bool force_reclaim)
800{
801 LIST_HEAD(ret_pages);
802 LIST_HEAD(free_pages);
803 int pgactivate = 0;
804 unsigned long nr_unqueued_dirty = 0;
805 unsigned long nr_dirty = 0;
806 unsigned long nr_congested = 0;
807 unsigned long nr_reclaimed = 0;
808 unsigned long nr_writeback = 0;
809 unsigned long nr_immediate = 0;
810
811 cond_resched();
812
813 mem_cgroup_uncharge_start();
814 while (!list_empty(page_list)) {
815 struct address_space *mapping;
816 struct page *page;
817 int may_enter_fs;
818 enum page_references references = PAGEREF_RECLAIM_CLEAN;
819 bool dirty, writeback;
820
821 cond_resched();
822
823 page = lru_to_page(page_list);
824 list_del(&page->lru);
825
826 if (!trylock_page(page))
827 goto keep;
828
829 VM_BUG_ON_PAGE(PageActive(page), page);
830 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
831
832 sc->nr_scanned++;
833
834 if (unlikely(!page_evictable(page)))
835 goto cull_mlocked;
836
837 if (!sc->may_unmap && page_mapped(page))
838 goto keep_locked;
839
840 /* Double the slab pressure for mapped and swapcache pages */
841 if (page_mapped(page) || PageSwapCache(page))
842 sc->nr_scanned++;
843
844 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
845 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
846
847 /*
848 * The number of dirty pages determines if a zone is marked
849 * reclaim_congested which affects wait_iff_congested. kswapd
850 * will stall and start writing pages if the tail of the LRU
851 * is all dirty unqueued pages.
852 */
853 page_check_dirty_writeback(page, &dirty, &writeback);
854 if (dirty || writeback)
855 nr_dirty++;
856
857 if (dirty && !writeback)
858 nr_unqueued_dirty++;
859
860 /*
861 * Treat this page as congested if the underlying BDI is or if
862 * pages are cycling through the LRU so quickly that the
863 * pages marked for immediate reclaim are making it to the
864 * end of the LRU a second time.
865 */
866 mapping = page_mapping(page);
867 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
868 (writeback && PageReclaim(page)))
869 nr_congested++;
870
871 /*
872 * If a page at the tail of the LRU is under writeback, there
873 * are three cases to consider.
874 *
875 * 1) If reclaim is encountering an excessive number of pages
876 * under writeback and this page is both under writeback and
877 * PageReclaim then it indicates that pages are being queued
878 * for IO but are being recycled through the LRU before the
879 * IO can complete. Waiting on the page itself risks an
880 * indefinite stall if it is impossible to writeback the
881 * page due to IO error or disconnected storage so instead
882 * note that the LRU is being scanned too quickly and the
883 * caller can stall after page list has been processed.
884 *
885 * 2) Global reclaim encounters a page, memcg encounters a
886 * page that is not marked for immediate reclaim or
887 * the caller does not have __GFP_IO. In this case mark
888 * the page for immediate reclaim and continue scanning.
889 *
890 * __GFP_IO is checked because a loop driver thread might
891 * enter reclaim, and deadlock if it waits on a page for
892 * which it is needed to do the write (loop masks off
893 * __GFP_IO|__GFP_FS for this reason); but more thought
894 * would probably show more reasons.
895 *
896 * Don't require __GFP_FS, since we're not going into the
897 * FS, just waiting on its writeback completion. Worryingly,
898 * ext4 gfs2 and xfs allocate pages with
899 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
900 * may_enter_fs here is liable to OOM on them.
901 *
902 * 3) memcg encounters a page that is not already marked
903 * PageReclaim. memcg does not have any dirty pages
904 * throttling so we could easily OOM just because too many
905 * pages are in writeback and there is nothing else to
906 * reclaim. Wait for the writeback to complete.
907 */
908 if (PageWriteback(page)) {
909 /* Case 1 above */
910 if (current_is_kswapd() &&
911 PageReclaim(page) &&
912 zone_is_reclaim_writeback(zone)) {
913 nr_immediate++;
914 goto keep_locked;
915
916 /* Case 2 above */
917 } else if (global_reclaim(sc) ||
918 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
919 /*
920 * This is slightly racy - end_page_writeback()
921 * might have just cleared PageReclaim, then
922 * setting PageReclaim here end up interpreted
923 * as PageReadahead - but that does not matter
924 * enough to care. What we do want is for this
925 * page to have PageReclaim set next time memcg
926 * reclaim reaches the tests above, so it will
927 * then wait_on_page_writeback() to avoid OOM;
928 * and it's also appropriate in global reclaim.
929 */
930 SetPageReclaim(page);
931 nr_writeback++;
932
933 goto keep_locked;
934
935 /* Case 3 above */
936 } else {
937 wait_on_page_writeback(page);
938 }
939 }
940
941 if (!force_reclaim)
942 references = page_check_references(page, sc);
943
944 switch (references) {
945 case PAGEREF_ACTIVATE:
946 goto activate_locked;
947 case PAGEREF_KEEP:
948 goto keep_locked;
949 case PAGEREF_RECLAIM:
950 case PAGEREF_RECLAIM_CLEAN:
951 ; /* try to reclaim the page below */
952 }
953
954 /*
955 * Anonymous process memory has backing store?
956 * Try to allocate it some swap space here.
957 */
958 if (PageAnon(page) && !PageSwapCache(page)) {
959 if (!(sc->gfp_mask & __GFP_IO))
960 goto keep_locked;
961 if (!add_to_swap(page, page_list))
962 goto activate_locked;
963 may_enter_fs = 1;
964
965 /* Adding to swap updated mapping */
966 mapping = page_mapping(page);
967 }
968
969 /*
970 * The page is mapped into the page tables of one or more
971 * processes. Try to unmap it here.
972 */
973 if (page_mapped(page) && mapping) {
974 switch (try_to_unmap(page, ttu_flags)) {
975 case SWAP_FAIL:
976 goto activate_locked;
977 case SWAP_AGAIN:
978 goto keep_locked;
979 case SWAP_MLOCK:
980 goto cull_mlocked;
981 case SWAP_SUCCESS:
982 ; /* try to free the page below */
983 }
984 }
985
986 if (PageDirty(page)) {
987 /*
988 * Only kswapd can writeback filesystem pages to
989 * avoid risk of stack overflow but only writeback
990 * if many dirty pages have been encountered.
991 */
992 if (page_is_file_cache(page) &&
993 (!current_is_kswapd() ||
994 !zone_is_reclaim_dirty(zone))) {
995 /*
996 * Immediately reclaim when written back.
997 * Similar in principal to deactivate_page()
998 * except we already have the page isolated
999 * and know it's dirty
1000 */
1001 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1002 SetPageReclaim(page);
1003
1004 goto keep_locked;
1005 }
1006
1007 if (references == PAGEREF_RECLAIM_CLEAN)
1008 goto keep_locked;
1009 if (!may_enter_fs)
1010 goto keep_locked;
1011 if (!sc->may_writepage)
1012 goto keep_locked;
1013
1014 /* Page is dirty, try to write it out here */
1015 switch (pageout(page, mapping, sc)) {
1016 case PAGE_KEEP:
1017 goto keep_locked;
1018 case PAGE_ACTIVATE:
1019 goto activate_locked;
1020 case PAGE_SUCCESS:
1021 if (PageWriteback(page))
1022 goto keep;
1023 if (PageDirty(page))
1024 goto keep;
1025
1026 /*
1027 * A synchronous write - probably a ramdisk. Go
1028 * ahead and try to reclaim the page.
1029 */
1030 if (!trylock_page(page))
1031 goto keep;
1032 if (PageDirty(page) || PageWriteback(page))
1033 goto keep_locked;
1034 mapping = page_mapping(page);
1035 case PAGE_CLEAN:
1036 ; /* try to free the page below */
1037 }
1038 }
1039
1040 /*
1041 * If the page has buffers, try to free the buffer mappings
1042 * associated with this page. If we succeed we try to free
1043 * the page as well.
1044 *
1045 * We do this even if the page is PageDirty().
1046 * try_to_release_page() does not perform I/O, but it is
1047 * possible for a page to have PageDirty set, but it is actually
1048 * clean (all its buffers are clean). This happens if the
1049 * buffers were written out directly, with submit_bh(). ext3
1050 * will do this, as well as the blockdev mapping.
1051 * try_to_release_page() will discover that cleanness and will
1052 * drop the buffers and mark the page clean - it can be freed.
1053 *
1054 * Rarely, pages can have buffers and no ->mapping. These are
1055 * the pages which were not successfully invalidated in
1056 * truncate_complete_page(). We try to drop those buffers here
1057 * and if that worked, and the page is no longer mapped into
1058 * process address space (page_count == 1) it can be freed.
1059 * Otherwise, leave the page on the LRU so it is swappable.
1060 */
1061 if (page_has_private(page)) {
1062 if (!try_to_release_page(page, sc->gfp_mask))
1063 goto activate_locked;
1064 if (!mapping && page_count(page) == 1) {
1065 unlock_page(page);
1066 if (put_page_testzero(page))
1067 goto free_it;
1068 else {
1069 /*
1070 * rare race with speculative reference.
1071 * the speculative reference will free
1072 * this page shortly, so we may
1073 * increment nr_reclaimed here (and
1074 * leave it off the LRU).
1075 */
1076 nr_reclaimed++;
1077 continue;
1078 }
1079 }
1080 }
1081
1082 if (!mapping || !__remove_mapping(mapping, page, true))
1083 goto keep_locked;
1084
1085 /*
1086 * At this point, we have no other references and there is
1087 * no way to pick any more up (removed from LRU, removed
1088 * from pagecache). Can use non-atomic bitops now (and
1089 * we obviously don't have to worry about waking up a process
1090 * waiting on the page lock, because there are no references.
1091 */
1092 __clear_page_locked(page);
1093free_it:
1094 nr_reclaimed++;
1095
1096 /*
1097 * Is there need to periodically free_page_list? It would
1098 * appear not as the counts should be low
1099 */
1100 list_add(&page->lru, &free_pages);
1101 continue;
1102
1103cull_mlocked:
1104 if (PageSwapCache(page))
1105 try_to_free_swap(page);
1106 unlock_page(page);
1107 putback_lru_page(page);
1108 continue;
1109
1110activate_locked:
1111 /* Not a candidate for swapping, so reclaim swap space. */
1112 if (PageSwapCache(page) && vm_swap_full())
1113 try_to_free_swap(page);
1114 VM_BUG_ON_PAGE(PageActive(page), page);
1115 SetPageActive(page);
1116 pgactivate++;
1117keep_locked:
1118 unlock_page(page);
1119keep:
1120 list_add(&page->lru, &ret_pages);
1121 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1122 }
1123
1124 free_hot_cold_page_list(&free_pages, 1);
1125
1126 list_splice(&ret_pages, page_list);
1127 count_vm_events(PGACTIVATE, pgactivate);
1128 mem_cgroup_uncharge_end();
1129 *ret_nr_dirty += nr_dirty;
1130 *ret_nr_congested += nr_congested;
1131 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1132 *ret_nr_writeback += nr_writeback;
1133 *ret_nr_immediate += nr_immediate;
1134 return nr_reclaimed;
1135}
1136
1137unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1138 struct list_head *page_list)
1139{
1140 struct scan_control sc = {
1141 .gfp_mask = GFP_KERNEL,
1142 .priority = DEF_PRIORITY,
1143 .may_unmap = 1,
1144 };
1145 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1146 struct page *page, *next;
1147 LIST_HEAD(clean_pages);
1148
1149 list_for_each_entry_safe(page, next, page_list, lru) {
1150 if (page_is_file_cache(page) && !PageDirty(page) &&
1151 !isolated_balloon_page(page)) {
1152 ClearPageActive(page);
1153 list_move(&page->lru, &clean_pages);
1154 }
1155 }
1156
1157 ret = shrink_page_list(&clean_pages, zone, &sc,
1158 TTU_UNMAP|TTU_IGNORE_ACCESS,
1159 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1160 list_splice(&clean_pages, page_list);
1161 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1162 return ret;
1163}
1164
1165/*
1166 * Attempt to remove the specified page from its LRU. Only take this page
1167 * if it is of the appropriate PageActive status. Pages which are being
1168 * freed elsewhere are also ignored.
1169 *
1170 * page: page to consider
1171 * mode: one of the LRU isolation modes defined above
1172 *
1173 * returns 0 on success, -ve errno on failure.
1174 */
1175int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1176{
1177 int ret = -EINVAL;
1178
1179 /* Only take pages on the LRU. */
1180 if (!PageLRU(page))
1181 return ret;
1182
1183 /* Compaction should not handle unevictable pages but CMA can do so */
1184 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1185 return ret;
1186
1187 ret = -EBUSY;
1188
1189 /*
1190 * To minimise LRU disruption, the caller can indicate that it only
1191 * wants to isolate pages it will be able to operate on without
1192 * blocking - clean pages for the most part.
1193 *
1194 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1195 * is used by reclaim when it is cannot write to backing storage
1196 *
1197 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1198 * that it is possible to migrate without blocking
1199 */
1200 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1201 /* All the caller can do on PageWriteback is block */
1202 if (PageWriteback(page))
1203 return ret;
1204
1205 if (PageDirty(page)) {
1206 struct address_space *mapping;
1207
1208 /* ISOLATE_CLEAN means only clean pages */
1209 if (mode & ISOLATE_CLEAN)
1210 return ret;
1211
1212 /*
1213 * Only pages without mappings or that have a
1214 * ->migratepage callback are possible to migrate
1215 * without blocking
1216 */
1217 mapping = page_mapping(page);
1218 if (mapping && !mapping->a_ops->migratepage)
1219 return ret;
1220 }
1221 }
1222
1223 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1224 return ret;
1225
1226 if (likely(get_page_unless_zero(page))) {
1227 /*
1228 * Be careful not to clear PageLRU until after we're
1229 * sure the page is not being freed elsewhere -- the
1230 * page release code relies on it.
1231 */
1232 ClearPageLRU(page);
1233 ret = 0;
1234 }
1235
1236 return ret;
1237}
1238
1239/*
1240 * zone->lru_lock is heavily contended. Some of the functions that
1241 * shrink the lists perform better by taking out a batch of pages
1242 * and working on them outside the LRU lock.
1243 *
1244 * For pagecache intensive workloads, this function is the hottest
1245 * spot in the kernel (apart from copy_*_user functions).
1246 *
1247 * Appropriate locks must be held before calling this function.
1248 *
1249 * @nr_to_scan: The number of pages to look through on the list.
1250 * @lruvec: The LRU vector to pull pages from.
1251 * @dst: The temp list to put pages on to.
1252 * @nr_scanned: The number of pages that were scanned.
1253 * @sc: The scan_control struct for this reclaim session
1254 * @mode: One of the LRU isolation modes
1255 * @lru: LRU list id for isolating
1256 *
1257 * returns how many pages were moved onto *@dst.
1258 */
1259static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1260 struct lruvec *lruvec, struct list_head *dst,
1261 unsigned long *nr_scanned, struct scan_control *sc,
1262 isolate_mode_t mode, enum lru_list lru)
1263{
1264 struct list_head *src = &lruvec->lists[lru];
1265 unsigned long nr_taken = 0;
1266 unsigned long scan;
1267
1268 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1269 struct page *page;
1270 int nr_pages;
1271
1272 page = lru_to_page(src);
1273 prefetchw_prev_lru_page(page, src, flags);
1274
1275 VM_BUG_ON_PAGE(!PageLRU(page), page);
1276
1277 switch (__isolate_lru_page(page, mode)) {
1278 case 0:
1279 nr_pages = hpage_nr_pages(page);
1280 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1281 list_move(&page->lru, dst);
1282 nr_taken += nr_pages;
1283 break;
1284
1285 case -EBUSY:
1286 /* else it is being freed elsewhere */
1287 list_move(&page->lru, src);
1288 continue;
1289
1290 default:
1291 BUG();
1292 }
1293 }
1294
1295 *nr_scanned = scan;
1296 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1297 nr_taken, mode, is_file_lru(lru));
1298 return nr_taken;
1299}
1300
1301/**
1302 * isolate_lru_page - tries to isolate a page from its LRU list
1303 * @page: page to isolate from its LRU list
1304 *
1305 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1306 * vmstat statistic corresponding to whatever LRU list the page was on.
1307 *
1308 * Returns 0 if the page was removed from an LRU list.
1309 * Returns -EBUSY if the page was not on an LRU list.
1310 *
1311 * The returned page will have PageLRU() cleared. If it was found on
1312 * the active list, it will have PageActive set. If it was found on
1313 * the unevictable list, it will have the PageUnevictable bit set. That flag
1314 * may need to be cleared by the caller before letting the page go.
1315 *
1316 * The vmstat statistic corresponding to the list on which the page was
1317 * found will be decremented.
1318 *
1319 * Restrictions:
1320 * (1) Must be called with an elevated refcount on the page. This is a
1321 * fundamentnal difference from isolate_lru_pages (which is called
1322 * without a stable reference).
1323 * (2) the lru_lock must not be held.
1324 * (3) interrupts must be enabled.
1325 */
1326int isolate_lru_page(struct page *page)
1327{
1328 int ret = -EBUSY;
1329
1330 VM_BUG_ON_PAGE(!page_count(page), page);
1331
1332 if (PageLRU(page)) {
1333 struct zone *zone = page_zone(page);
1334 struct lruvec *lruvec;
1335
1336 spin_lock_irq(&zone->lru_lock);
1337 lruvec = mem_cgroup_page_lruvec(page, zone);
1338 if (PageLRU(page)) {
1339 int lru = page_lru(page);
1340 get_page(page);
1341 ClearPageLRU(page);
1342 del_page_from_lru_list(page, lruvec, lru);
1343 ret = 0;
1344 }
1345 spin_unlock_irq(&zone->lru_lock);
1346 }
1347 return ret;
1348}
1349
1350/*
1351 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1352 * then get resheduled. When there are massive number of tasks doing page
1353 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1354 * the LRU list will go small and be scanned faster than necessary, leading to
1355 * unnecessary swapping, thrashing and OOM.
1356 */
1357static int too_many_isolated(struct zone *zone, int file,
1358 struct scan_control *sc)
1359{
1360 unsigned long inactive, isolated;
1361
1362 if (current_is_kswapd())
1363 return 0;
1364
1365 if (!global_reclaim(sc))
1366 return 0;
1367
1368 if (file) {
1369 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1370 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1371 } else {
1372 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1373 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1374 }
1375
1376 /*
1377 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1378 * won't get blocked by normal direct-reclaimers, forming a circular
1379 * deadlock.
1380 */
1381 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1382 inactive >>= 3;
1383
1384 return isolated > inactive;
1385}
1386
1387static noinline_for_stack void
1388putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1389{
1390 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1391 struct zone *zone = lruvec_zone(lruvec);
1392 LIST_HEAD(pages_to_free);
1393
1394 /*
1395 * Put back any unfreeable pages.
1396 */
1397 while (!list_empty(page_list)) {
1398 struct page *page = lru_to_page(page_list);
1399 int lru;
1400
1401 VM_BUG_ON_PAGE(PageLRU(page), page);
1402 list_del(&page->lru);
1403 if (unlikely(!page_evictable(page))) {
1404 spin_unlock_irq(&zone->lru_lock);
1405 putback_lru_page(page);
1406 spin_lock_irq(&zone->lru_lock);
1407 continue;
1408 }
1409
1410 lruvec = mem_cgroup_page_lruvec(page, zone);
1411
1412 SetPageLRU(page);
1413 lru = page_lru(page);
1414 add_page_to_lru_list(page, lruvec, lru);
1415
1416 if (is_active_lru(lru)) {
1417 int file = is_file_lru(lru);
1418 int numpages = hpage_nr_pages(page);
1419 reclaim_stat->recent_rotated[file] += numpages;
1420 }
1421 if (put_page_testzero(page)) {
1422 __ClearPageLRU(page);
1423 __ClearPageActive(page);
1424 del_page_from_lru_list(page, lruvec, lru);
1425
1426 if (unlikely(PageCompound(page))) {
1427 spin_unlock_irq(&zone->lru_lock);
1428 (*get_compound_page_dtor(page))(page);
1429 spin_lock_irq(&zone->lru_lock);
1430 } else
1431 list_add(&page->lru, &pages_to_free);
1432 }
1433 }
1434
1435 /*
1436 * To save our caller's stack, now use input list for pages to free.
1437 */
1438 list_splice(&pages_to_free, page_list);
1439}
1440
1441/*
1442 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1443 * of reclaimed pages
1444 */
1445static noinline_for_stack unsigned long
1446shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1447 struct scan_control *sc, enum lru_list lru)
1448{
1449 LIST_HEAD(page_list);
1450 unsigned long nr_scanned;
1451 unsigned long nr_reclaimed = 0;
1452 unsigned long nr_taken;
1453 unsigned long nr_dirty = 0;
1454 unsigned long nr_congested = 0;
1455 unsigned long nr_unqueued_dirty = 0;
1456 unsigned long nr_writeback = 0;
1457 unsigned long nr_immediate = 0;
1458 isolate_mode_t isolate_mode = 0;
1459 int file = is_file_lru(lru);
1460 struct zone *zone = lruvec_zone(lruvec);
1461 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1462
1463 while (unlikely(too_many_isolated(zone, file, sc))) {
1464 congestion_wait(BLK_RW_ASYNC, HZ/10);
1465
1466 /* We are about to die and free our memory. Return now. */
1467 if (fatal_signal_pending(current))
1468 return SWAP_CLUSTER_MAX;
1469 }
1470
1471 lru_add_drain();
1472
1473 if (!sc->may_unmap)
1474 isolate_mode |= ISOLATE_UNMAPPED;
1475 if (!sc->may_writepage)
1476 isolate_mode |= ISOLATE_CLEAN;
1477
1478 spin_lock_irq(&zone->lru_lock);
1479
1480 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1481 &nr_scanned, sc, isolate_mode, lru);
1482
1483 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1484 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1485
1486 if (global_reclaim(sc)) {
1487 zone->pages_scanned += nr_scanned;
1488 if (current_is_kswapd())
1489 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1490 else
1491 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1492 }
1493 spin_unlock_irq(&zone->lru_lock);
1494
1495 if (nr_taken == 0)
1496 return 0;
1497
1498 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1499 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1500 &nr_writeback, &nr_immediate,
1501 false);
1502
1503 spin_lock_irq(&zone->lru_lock);
1504
1505 reclaim_stat->recent_scanned[file] += nr_taken;
1506
1507 if (global_reclaim(sc)) {
1508 if (current_is_kswapd())
1509 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1510 nr_reclaimed);
1511 else
1512 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1513 nr_reclaimed);
1514 }
1515
1516 putback_inactive_pages(lruvec, &page_list);
1517
1518 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1519
1520 spin_unlock_irq(&zone->lru_lock);
1521
1522 free_hot_cold_page_list(&page_list, 1);
1523
1524 /*
1525 * If reclaim is isolating dirty pages under writeback, it implies
1526 * that the long-lived page allocation rate is exceeding the page
1527 * laundering rate. Either the global limits are not being effective
1528 * at throttling processes due to the page distribution throughout
1529 * zones or there is heavy usage of a slow backing device. The
1530 * only option is to throttle from reclaim context which is not ideal
1531 * as there is no guarantee the dirtying process is throttled in the
1532 * same way balance_dirty_pages() manages.
1533 *
1534 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1535 * of pages under pages flagged for immediate reclaim and stall if any
1536 * are encountered in the nr_immediate check below.
1537 */
1538 if (nr_writeback && nr_writeback == nr_taken)
1539 zone_set_flag(zone, ZONE_WRITEBACK);
1540
1541 /*
1542 * memcg will stall in page writeback so only consider forcibly
1543 * stalling for global reclaim
1544 */
1545 if (global_reclaim(sc)) {
1546 /*
1547 * Tag a zone as congested if all the dirty pages scanned were
1548 * backed by a congested BDI and wait_iff_congested will stall.
1549 */
1550 if (nr_dirty && nr_dirty == nr_congested)
1551 zone_set_flag(zone, ZONE_CONGESTED);
1552
1553 /*
1554 * If dirty pages are scanned that are not queued for IO, it
1555 * implies that flushers are not keeping up. In this case, flag
1556 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1557 * pages from reclaim context. It will forcibly stall in the
1558 * next check.
1559 */
1560 if (nr_unqueued_dirty == nr_taken)
1561 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1562
1563 /*
1564 * In addition, if kswapd scans pages marked marked for
1565 * immediate reclaim and under writeback (nr_immediate), it
1566 * implies that pages are cycling through the LRU faster than
1567 * they are written so also forcibly stall.
1568 */
1569 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1570 congestion_wait(BLK_RW_ASYNC, HZ/10);
1571 }
1572
1573 /*
1574 * Stall direct reclaim for IO completions if underlying BDIs or zone
1575 * is congested. Allow kswapd to continue until it starts encountering
1576 * unqueued dirty pages or cycling through the LRU too quickly.
1577 */
1578 if (!sc->hibernation_mode && !current_is_kswapd())
1579 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1580
1581 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1582 zone_idx(zone),
1583 nr_scanned, nr_reclaimed,
1584 sc->priority,
1585 trace_shrink_flags(file));
1586 return nr_reclaimed;
1587}
1588
1589/*
1590 * This moves pages from the active list to the inactive list.
1591 *
1592 * We move them the other way if the page is referenced by one or more
1593 * processes, from rmap.
1594 *
1595 * If the pages are mostly unmapped, the processing is fast and it is
1596 * appropriate to hold zone->lru_lock across the whole operation. But if
1597 * the pages are mapped, the processing is slow (page_referenced()) so we
1598 * should drop zone->lru_lock around each page. It's impossible to balance
1599 * this, so instead we remove the pages from the LRU while processing them.
1600 * It is safe to rely on PG_active against the non-LRU pages in here because
1601 * nobody will play with that bit on a non-LRU page.
1602 *
1603 * The downside is that we have to touch page->_count against each page.
1604 * But we had to alter page->flags anyway.
1605 */
1606
1607static void move_active_pages_to_lru(struct lruvec *lruvec,
1608 struct list_head *list,
1609 struct list_head *pages_to_free,
1610 enum lru_list lru)
1611{
1612 struct zone *zone = lruvec_zone(lruvec);
1613 unsigned long pgmoved = 0;
1614 struct page *page;
1615 int nr_pages;
1616
1617 while (!list_empty(list)) {
1618 page = lru_to_page(list);
1619 lruvec = mem_cgroup_page_lruvec(page, zone);
1620
1621 VM_BUG_ON_PAGE(PageLRU(page), page);
1622 SetPageLRU(page);
1623
1624 nr_pages = hpage_nr_pages(page);
1625 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1626 list_move(&page->lru, &lruvec->lists[lru]);
1627 pgmoved += nr_pages;
1628
1629 if (put_page_testzero(page)) {
1630 __ClearPageLRU(page);
1631 __ClearPageActive(page);
1632 del_page_from_lru_list(page, lruvec, lru);
1633
1634 if (unlikely(PageCompound(page))) {
1635 spin_unlock_irq(&zone->lru_lock);
1636 (*get_compound_page_dtor(page))(page);
1637 spin_lock_irq(&zone->lru_lock);
1638 } else
1639 list_add(&page->lru, pages_to_free);
1640 }
1641 }
1642 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1643 if (!is_active_lru(lru))
1644 __count_vm_events(PGDEACTIVATE, pgmoved);
1645}
1646
1647static void shrink_active_list(unsigned long nr_to_scan,
1648 struct lruvec *lruvec,
1649 struct scan_control *sc,
1650 enum lru_list lru)
1651{
1652 unsigned long nr_taken;
1653 unsigned long nr_scanned;
1654 unsigned long vm_flags;
1655 LIST_HEAD(l_hold); /* The pages which were snipped off */
1656 LIST_HEAD(l_active);
1657 LIST_HEAD(l_inactive);
1658 struct page *page;
1659 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1660 unsigned long nr_rotated = 0;
1661 isolate_mode_t isolate_mode = 0;
1662 int file = is_file_lru(lru);
1663 struct zone *zone = lruvec_zone(lruvec);
1664
1665 lru_add_drain();
1666
1667 if (!sc->may_unmap)
1668 isolate_mode |= ISOLATE_UNMAPPED;
1669 if (!sc->may_writepage)
1670 isolate_mode |= ISOLATE_CLEAN;
1671
1672 spin_lock_irq(&zone->lru_lock);
1673
1674 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1675 &nr_scanned, sc, isolate_mode, lru);
1676 if (global_reclaim(sc))
1677 zone->pages_scanned += nr_scanned;
1678
1679 reclaim_stat->recent_scanned[file] += nr_taken;
1680
1681 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1682 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1683 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1684 spin_unlock_irq(&zone->lru_lock);
1685
1686 while (!list_empty(&l_hold)) {
1687 cond_resched();
1688 page = lru_to_page(&l_hold);
1689 list_del(&page->lru);
1690
1691 if (unlikely(!page_evictable(page))) {
1692 putback_lru_page(page);
1693 continue;
1694 }
1695
1696 if (unlikely(buffer_heads_over_limit)) {
1697 if (page_has_private(page) && trylock_page(page)) {
1698 if (page_has_private(page))
1699 try_to_release_page(page, 0);
1700 unlock_page(page);
1701 }
1702 }
1703
1704 if (page_referenced(page, 0, sc->target_mem_cgroup,
1705 &vm_flags)) {
1706 nr_rotated += hpage_nr_pages(page);
1707 /*
1708 * Identify referenced, file-backed active pages and
1709 * give them one more trip around the active list. So
1710 * that executable code get better chances to stay in
1711 * memory under moderate memory pressure. Anon pages
1712 * are not likely to be evicted by use-once streaming
1713 * IO, plus JVM can create lots of anon VM_EXEC pages,
1714 * so we ignore them here.
1715 */
1716 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1717 list_add(&page->lru, &l_active);
1718 continue;
1719 }
1720 }
1721
1722 ClearPageActive(page); /* we are de-activating */
1723 list_add(&page->lru, &l_inactive);
1724 }
1725
1726 /*
1727 * Move pages back to the lru list.
1728 */
1729 spin_lock_irq(&zone->lru_lock);
1730 /*
1731 * Count referenced pages from currently used mappings as rotated,
1732 * even though only some of them are actually re-activated. This
1733 * helps balance scan pressure between file and anonymous pages in
1734 * get_scan_ratio.
1735 */
1736 reclaim_stat->recent_rotated[file] += nr_rotated;
1737
1738 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1739 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1740 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1741 spin_unlock_irq(&zone->lru_lock);
1742
1743 free_hot_cold_page_list(&l_hold, 1);
1744}
1745
1746#ifdef CONFIG_SWAP
1747static int inactive_anon_is_low_global(struct zone *zone)
1748{
1749 unsigned long active, inactive;
1750
1751 active = zone_page_state(zone, NR_ACTIVE_ANON);
1752 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1753
1754 if (inactive * zone->inactive_ratio < active)
1755 return 1;
1756
1757 return 0;
1758}
1759
1760/**
1761 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1762 * @lruvec: LRU vector to check
1763 *
1764 * Returns true if the zone does not have enough inactive anon pages,
1765 * meaning some active anon pages need to be deactivated.
1766 */
1767static int inactive_anon_is_low(struct lruvec *lruvec)
1768{
1769 /*
1770 * If we don't have swap space, anonymous page deactivation
1771 * is pointless.
1772 */
1773 if (!total_swap_pages)
1774 return 0;
1775
1776 if (!mem_cgroup_disabled())
1777 return mem_cgroup_inactive_anon_is_low(lruvec);
1778
1779 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1780}
1781#else
1782static inline int inactive_anon_is_low(struct lruvec *lruvec)
1783{
1784 return 0;
1785}
1786#endif
1787
1788/**
1789 * inactive_file_is_low - check if file pages need to be deactivated
1790 * @lruvec: LRU vector to check
1791 *
1792 * When the system is doing streaming IO, memory pressure here
1793 * ensures that active file pages get deactivated, until more
1794 * than half of the file pages are on the inactive list.
1795 *
1796 * Once we get to that situation, protect the system's working
1797 * set from being evicted by disabling active file page aging.
1798 *
1799 * This uses a different ratio than the anonymous pages, because
1800 * the page cache uses a use-once replacement algorithm.
1801 */
1802static int inactive_file_is_low(struct lruvec *lruvec)
1803{
1804 unsigned long inactive;
1805 unsigned long active;
1806
1807 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1808 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1809
1810 return active > inactive;
1811}
1812
1813static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1814{
1815 if (is_file_lru(lru))
1816 return inactive_file_is_low(lruvec);
1817 else
1818 return inactive_anon_is_low(lruvec);
1819}
1820
1821static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1822 struct lruvec *lruvec, struct scan_control *sc)
1823{
1824 if (is_active_lru(lru)) {
1825 if (inactive_list_is_low(lruvec, lru))
1826 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1827 return 0;
1828 }
1829
1830 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1831}
1832
1833static int vmscan_swappiness(struct scan_control *sc)
1834{
1835 if (global_reclaim(sc))
1836 return vm_swappiness;
1837 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1838}
1839
1840enum scan_balance {
1841 SCAN_EQUAL,
1842 SCAN_FRACT,
1843 SCAN_ANON,
1844 SCAN_FILE,
1845};
1846
1847/*
1848 * Determine how aggressively the anon and file LRU lists should be
1849 * scanned. The relative value of each set of LRU lists is determined
1850 * by looking at the fraction of the pages scanned we did rotate back
1851 * onto the active list instead of evict.
1852 *
1853 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1854 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1855 */
1856static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1857 unsigned long *nr)
1858{
1859 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1860 u64 fraction[2];
1861 u64 denominator = 0; /* gcc */
1862 struct zone *zone = lruvec_zone(lruvec);
1863 unsigned long anon_prio, file_prio;
1864 enum scan_balance scan_balance;
1865 unsigned long anon, file;
1866 bool force_scan = false;
1867 unsigned long ap, fp;
1868 enum lru_list lru;
1869
1870 /*
1871 * If the zone or memcg is small, nr[l] can be 0. This
1872 * results in no scanning on this priority and a potential
1873 * priority drop. Global direct reclaim can go to the next
1874 * zone and tends to have no problems. Global kswapd is for
1875 * zone balancing and it needs to scan a minimum amount. When
1876 * reclaiming for a memcg, a priority drop can cause high
1877 * latencies, so it's better to scan a minimum amount there as
1878 * well.
1879 */
1880 if (current_is_kswapd() && !zone_reclaimable(zone))
1881 force_scan = true;
1882 if (!global_reclaim(sc))
1883 force_scan = true;
1884
1885 /* If we have no swap space, do not bother scanning anon pages. */
1886 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1887 scan_balance = SCAN_FILE;
1888 goto out;
1889 }
1890
1891 /*
1892 * Global reclaim will swap to prevent OOM even with no
1893 * swappiness, but memcg users want to use this knob to
1894 * disable swapping for individual groups completely when
1895 * using the memory controller's swap limit feature would be
1896 * too expensive.
1897 */
1898 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1899 scan_balance = SCAN_FILE;
1900 goto out;
1901 }
1902
1903 /*
1904 * Do not apply any pressure balancing cleverness when the
1905 * system is close to OOM, scan both anon and file equally
1906 * (unless the swappiness setting disagrees with swapping).
1907 */
1908 if (!sc->priority && vmscan_swappiness(sc)) {
1909 scan_balance = SCAN_EQUAL;
1910 goto out;
1911 }
1912
1913 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1914 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1915 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1916 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1917
1918 /*
1919 * Prevent the reclaimer from falling into the cache trap: as
1920 * cache pages start out inactive, every cache fault will tip
1921 * the scan balance towards the file LRU. And as the file LRU
1922 * shrinks, so does the window for rotation from references.
1923 * This means we have a runaway feedback loop where a tiny
1924 * thrashing file LRU becomes infinitely more attractive than
1925 * anon pages. Try to detect this based on file LRU size.
1926 */
1927 if (global_reclaim(sc)) {
1928 unsigned long free = zone_page_state(zone, NR_FREE_PAGES);
1929
1930 if (unlikely(file + free <= high_wmark_pages(zone))) {
1931 scan_balance = SCAN_ANON;
1932 goto out;
1933 }
1934 }
1935
1936 /*
1937 * There is enough inactive page cache, do not reclaim
1938 * anything from the anonymous working set right now.
1939 */
1940 if (!inactive_file_is_low(lruvec)) {
1941 scan_balance = SCAN_FILE;
1942 goto out;
1943 }
1944
1945 scan_balance = SCAN_FRACT;
1946
1947 /*
1948 * With swappiness at 100, anonymous and file have the same priority.
1949 * This scanning priority is essentially the inverse of IO cost.
1950 */
1951 anon_prio = vmscan_swappiness(sc);
1952 file_prio = 200 - anon_prio;
1953
1954 /*
1955 * OK, so we have swap space and a fair amount of page cache
1956 * pages. We use the recently rotated / recently scanned
1957 * ratios to determine how valuable each cache is.
1958 *
1959 * Because workloads change over time (and to avoid overflow)
1960 * we keep these statistics as a floating average, which ends
1961 * up weighing recent references more than old ones.
1962 *
1963 * anon in [0], file in [1]
1964 */
1965 spin_lock_irq(&zone->lru_lock);
1966 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1967 reclaim_stat->recent_scanned[0] /= 2;
1968 reclaim_stat->recent_rotated[0] /= 2;
1969 }
1970
1971 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1972 reclaim_stat->recent_scanned[1] /= 2;
1973 reclaim_stat->recent_rotated[1] /= 2;
1974 }
1975
1976 /*
1977 * The amount of pressure on anon vs file pages is inversely
1978 * proportional to the fraction of recently scanned pages on
1979 * each list that were recently referenced and in active use.
1980 */
1981 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1982 ap /= reclaim_stat->recent_rotated[0] + 1;
1983
1984 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1985 fp /= reclaim_stat->recent_rotated[1] + 1;
1986 spin_unlock_irq(&zone->lru_lock);
1987
1988 fraction[0] = ap;
1989 fraction[1] = fp;
1990 denominator = ap + fp + 1;
1991out:
1992 for_each_evictable_lru(lru) {
1993 int file = is_file_lru(lru);
1994 unsigned long size;
1995 unsigned long scan;
1996
1997 size = get_lru_size(lruvec, lru);
1998 scan = size >> sc->priority;
1999
2000 if (!scan && force_scan)
2001 scan = min(size, SWAP_CLUSTER_MAX);
2002
2003 switch (scan_balance) {
2004 case SCAN_EQUAL:
2005 /* Scan lists relative to size */
2006 break;
2007 case SCAN_FRACT:
2008 /*
2009 * Scan types proportional to swappiness and
2010 * their relative recent reclaim efficiency.
2011 */
2012 scan = div64_u64(scan * fraction[file], denominator);
2013 break;
2014 case SCAN_FILE:
2015 case SCAN_ANON:
2016 /* Scan one type exclusively */
2017 if ((scan_balance == SCAN_FILE) != file)
2018 scan = 0;
2019 break;
2020 default:
2021 /* Look ma, no brain */
2022 BUG();
2023 }
2024 nr[lru] = scan;
2025 }
2026}
2027
2028/*
2029 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2030 */
2031static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2032{
2033 unsigned long nr[NR_LRU_LISTS];
2034 unsigned long targets[NR_LRU_LISTS];
2035 unsigned long nr_to_scan;
2036 enum lru_list lru;
2037 unsigned long nr_reclaimed = 0;
2038 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2039 struct blk_plug plug;
2040 bool scan_adjusted = false;
2041
2042 get_scan_count(lruvec, sc, nr);
2043
2044 /* Record the original scan target for proportional adjustments later */
2045 memcpy(targets, nr, sizeof(nr));
2046
2047 blk_start_plug(&plug);
2048 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2049 nr[LRU_INACTIVE_FILE]) {
2050 unsigned long nr_anon, nr_file, percentage;
2051 unsigned long nr_scanned;
2052
2053 for_each_evictable_lru(lru) {
2054 if (nr[lru]) {
2055 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2056 nr[lru] -= nr_to_scan;
2057
2058 nr_reclaimed += shrink_list(lru, nr_to_scan,
2059 lruvec, sc);
2060 }
2061 }
2062
2063 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2064 continue;
2065
2066 /*
2067 * For global direct reclaim, reclaim only the number of pages
2068 * requested. Less care is taken to scan proportionally as it
2069 * is more important to minimise direct reclaim stall latency
2070 * than it is to properly age the LRU lists.
2071 */
2072 if (global_reclaim(sc) && !current_is_kswapd())
2073 break;
2074
2075 /*
2076 * For kswapd and memcg, reclaim at least the number of pages
2077 * requested. Ensure that the anon and file LRUs shrink
2078 * proportionally what was requested by get_scan_count(). We
2079 * stop reclaiming one LRU and reduce the amount scanning
2080 * proportional to the original scan target.
2081 */
2082 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2083 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2084
2085 if (nr_file > nr_anon) {
2086 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2087 targets[LRU_ACTIVE_ANON] + 1;
2088 lru = LRU_BASE;
2089 percentage = nr_anon * 100 / scan_target;
2090 } else {
2091 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2092 targets[LRU_ACTIVE_FILE] + 1;
2093 lru = LRU_FILE;
2094 percentage = nr_file * 100 / scan_target;
2095 }
2096
2097 /* Stop scanning the smaller of the LRU */
2098 nr[lru] = 0;
2099 nr[lru + LRU_ACTIVE] = 0;
2100
2101 /*
2102 * Recalculate the other LRU scan count based on its original
2103 * scan target and the percentage scanning already complete
2104 */
2105 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2106 nr_scanned = targets[lru] - nr[lru];
2107 nr[lru] = targets[lru] * (100 - percentage) / 100;
2108 nr[lru] -= min(nr[lru], nr_scanned);
2109
2110 lru += LRU_ACTIVE;
2111 nr_scanned = targets[lru] - nr[lru];
2112 nr[lru] = targets[lru] * (100 - percentage) / 100;
2113 nr[lru] -= min(nr[lru], nr_scanned);
2114
2115 scan_adjusted = true;
2116 }
2117 blk_finish_plug(&plug);
2118 sc->nr_reclaimed += nr_reclaimed;
2119
2120 /*
2121 * Even if we did not try to evict anon pages at all, we want to
2122 * rebalance the anon lru active/inactive ratio.
2123 */
2124 if (inactive_anon_is_low(lruvec))
2125 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2126 sc, LRU_ACTIVE_ANON);
2127
2128 throttle_vm_writeout(sc->gfp_mask);
2129}
2130
2131/* Use reclaim/compaction for costly allocs or under memory pressure */
2132static bool in_reclaim_compaction(struct scan_control *sc)
2133{
2134 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2135 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2136 sc->priority < DEF_PRIORITY - 2))
2137 return true;
2138
2139 return false;
2140}
2141
2142/*
2143 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2144 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2145 * true if more pages should be reclaimed such that when the page allocator
2146 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2147 * It will give up earlier than that if there is difficulty reclaiming pages.
2148 */
2149static inline bool should_continue_reclaim(struct zone *zone,
2150 unsigned long nr_reclaimed,
2151 unsigned long nr_scanned,
2152 struct scan_control *sc)
2153{
2154 unsigned long pages_for_compaction;
2155 unsigned long inactive_lru_pages;
2156
2157 /* If not in reclaim/compaction mode, stop */
2158 if (!in_reclaim_compaction(sc))
2159 return false;
2160
2161 /* Consider stopping depending on scan and reclaim activity */
2162 if (sc->gfp_mask & __GFP_REPEAT) {
2163 /*
2164 * For __GFP_REPEAT allocations, stop reclaiming if the
2165 * full LRU list has been scanned and we are still failing
2166 * to reclaim pages. This full LRU scan is potentially
2167 * expensive but a __GFP_REPEAT caller really wants to succeed
2168 */
2169 if (!nr_reclaimed && !nr_scanned)
2170 return false;
2171 } else {
2172 /*
2173 * For non-__GFP_REPEAT allocations which can presumably
2174 * fail without consequence, stop if we failed to reclaim
2175 * any pages from the last SWAP_CLUSTER_MAX number of
2176 * pages that were scanned. This will return to the
2177 * caller faster at the risk reclaim/compaction and
2178 * the resulting allocation attempt fails
2179 */
2180 if (!nr_reclaimed)
2181 return false;
2182 }
2183
2184 /*
2185 * If we have not reclaimed enough pages for compaction and the
2186 * inactive lists are large enough, continue reclaiming
2187 */
2188 pages_for_compaction = (2UL << sc->order);
2189 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2190 if (get_nr_swap_pages() > 0)
2191 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2192 if (sc->nr_reclaimed < pages_for_compaction &&
2193 inactive_lru_pages > pages_for_compaction)
2194 return true;
2195
2196 /* If compaction would go ahead or the allocation would succeed, stop */
2197 switch (compaction_suitable(zone, sc->order)) {
2198 case COMPACT_PARTIAL:
2199 case COMPACT_CONTINUE:
2200 return false;
2201 default:
2202 return true;
2203 }
2204}
2205
2206static void shrink_zone(struct zone *zone, struct scan_control *sc)
2207{
2208 unsigned long nr_reclaimed, nr_scanned;
2209
2210 do {
2211 struct mem_cgroup *root = sc->target_mem_cgroup;
2212 struct mem_cgroup_reclaim_cookie reclaim = {
2213 .zone = zone,
2214 .priority = sc->priority,
2215 };
2216 struct mem_cgroup *memcg;
2217
2218 nr_reclaimed = sc->nr_reclaimed;
2219 nr_scanned = sc->nr_scanned;
2220
2221 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2222 do {
2223 struct lruvec *lruvec;
2224
2225 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2226
2227 shrink_lruvec(lruvec, sc);
2228
2229 /*
2230 * Direct reclaim and kswapd have to scan all memory
2231 * cgroups to fulfill the overall scan target for the
2232 * zone.
2233 *
2234 * Limit reclaim, on the other hand, only cares about
2235 * nr_to_reclaim pages to be reclaimed and it will
2236 * retry with decreasing priority if one round over the
2237 * whole hierarchy is not sufficient.
2238 */
2239 if (!global_reclaim(sc) &&
2240 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2241 mem_cgroup_iter_break(root, memcg);
2242 break;
2243 }
2244 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2245 } while (memcg);
2246
2247 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2248 sc->nr_scanned - nr_scanned,
2249 sc->nr_reclaimed - nr_reclaimed);
2250
2251 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2252 sc->nr_scanned - nr_scanned, sc));
2253}
2254
2255/* Returns true if compaction should go ahead for a high-order request */
2256static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2257{
2258 unsigned long balance_gap, watermark;
2259 bool watermark_ok;
2260
2261 /* Do not consider compaction for orders reclaim is meant to satisfy */
2262 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2263 return false;
2264
2265 /*
2266 * Compaction takes time to run and there are potentially other
2267 * callers using the pages just freed. Continue reclaiming until
2268 * there is a buffer of free pages available to give compaction
2269 * a reasonable chance of completing and allocating the page
2270 */
2271 balance_gap = min(low_wmark_pages(zone),
2272 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2273 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2274 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2275 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2276
2277 /*
2278 * If compaction is deferred, reclaim up to a point where
2279 * compaction will have a chance of success when re-enabled
2280 */
2281 if (compaction_deferred(zone, sc->order))
2282 return watermark_ok;
2283
2284 /* If compaction is not ready to start, keep reclaiming */
2285 if (!compaction_suitable(zone, sc->order))
2286 return false;
2287
2288 return watermark_ok;
2289}
2290
2291/*
2292 * This is the direct reclaim path, for page-allocating processes. We only
2293 * try to reclaim pages from zones which will satisfy the caller's allocation
2294 * request.
2295 *
2296 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2297 * Because:
2298 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2299 * allocation or
2300 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2301 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2302 * zone defense algorithm.
2303 *
2304 * If a zone is deemed to be full of pinned pages then just give it a light
2305 * scan then give up on it.
2306 *
2307 * This function returns true if a zone is being reclaimed for a costly
2308 * high-order allocation and compaction is ready to begin. This indicates to
2309 * the caller that it should consider retrying the allocation instead of
2310 * further reclaim.
2311 */
2312static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2313{
2314 struct zoneref *z;
2315 struct zone *zone;
2316 unsigned long nr_soft_reclaimed;
2317 unsigned long nr_soft_scanned;
2318 unsigned long lru_pages = 0;
2319 bool aborted_reclaim = false;
2320 struct reclaim_state *reclaim_state = current->reclaim_state;
2321 gfp_t orig_mask;
2322 struct shrink_control shrink = {
2323 .gfp_mask = sc->gfp_mask,
2324 };
2325 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2326
2327 /*
2328 * If the number of buffer_heads in the machine exceeds the maximum
2329 * allowed level, force direct reclaim to scan the highmem zone as
2330 * highmem pages could be pinning lowmem pages storing buffer_heads
2331 */
2332 orig_mask = sc->gfp_mask;
2333 if (buffer_heads_over_limit)
2334 sc->gfp_mask |= __GFP_HIGHMEM;
2335
2336 nodes_clear(shrink.nodes_to_scan);
2337
2338 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2339 gfp_zone(sc->gfp_mask), sc->nodemask) {
2340 if (!populated_zone(zone))
2341 continue;
2342 /*
2343 * Take care memory controller reclaiming has small influence
2344 * to global LRU.
2345 */
2346 if (global_reclaim(sc)) {
2347 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2348 continue;
2349
2350 lru_pages += zone_reclaimable_pages(zone);
2351 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2352
2353 if (sc->priority != DEF_PRIORITY &&
2354 !zone_reclaimable(zone))
2355 continue; /* Let kswapd poll it */
2356 if (IS_ENABLED(CONFIG_COMPACTION)) {
2357 /*
2358 * If we already have plenty of memory free for
2359 * compaction in this zone, don't free any more.
2360 * Even though compaction is invoked for any
2361 * non-zero order, only frequent costly order
2362 * reclamation is disruptive enough to become a
2363 * noticeable problem, like transparent huge
2364 * page allocations.
2365 */
2366 if ((zonelist_zone_idx(z) <= requested_highidx)
2367 && compaction_ready(zone, sc)) {
2368 aborted_reclaim = true;
2369 continue;
2370 }
2371 }
2372 /*
2373 * This steals pages from memory cgroups over softlimit
2374 * and returns the number of reclaimed pages and
2375 * scanned pages. This works for global memory pressure
2376 * and balancing, not for a memcg's limit.
2377 */
2378 nr_soft_scanned = 0;
2379 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2380 sc->order, sc->gfp_mask,
2381 &nr_soft_scanned);
2382 sc->nr_reclaimed += nr_soft_reclaimed;
2383 sc->nr_scanned += nr_soft_scanned;
2384 /* need some check for avoid more shrink_zone() */
2385 }
2386
2387 shrink_zone(zone, sc);
2388 }
2389
2390 /*
2391 * Don't shrink slabs when reclaiming memory from over limit cgroups
2392 * but do shrink slab at least once when aborting reclaim for
2393 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2394 * pages.
2395 */
2396 if (global_reclaim(sc)) {
2397 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2398 if (reclaim_state) {
2399 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2400 reclaim_state->reclaimed_slab = 0;
2401 }
2402 }
2403
2404 /*
2405 * Restore to original mask to avoid the impact on the caller if we
2406 * promoted it to __GFP_HIGHMEM.
2407 */
2408 sc->gfp_mask = orig_mask;
2409
2410 return aborted_reclaim;
2411}
2412
2413/* All zones in zonelist are unreclaimable? */
2414static bool all_unreclaimable(struct zonelist *zonelist,
2415 struct scan_control *sc)
2416{
2417 struct zoneref *z;
2418 struct zone *zone;
2419
2420 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2421 gfp_zone(sc->gfp_mask), sc->nodemask) {
2422 if (!populated_zone(zone))
2423 continue;
2424 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2425 continue;
2426 if (zone_reclaimable(zone))
2427 return false;
2428 }
2429
2430 return true;
2431}
2432
2433/*
2434 * This is the main entry point to direct page reclaim.
2435 *
2436 * If a full scan of the inactive list fails to free enough memory then we
2437 * are "out of memory" and something needs to be killed.
2438 *
2439 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2440 * high - the zone may be full of dirty or under-writeback pages, which this
2441 * caller can't do much about. We kick the writeback threads and take explicit
2442 * naps in the hope that some of these pages can be written. But if the
2443 * allocating task holds filesystem locks which prevent writeout this might not
2444 * work, and the allocation attempt will fail.
2445 *
2446 * returns: 0, if no pages reclaimed
2447 * else, the number of pages reclaimed
2448 */
2449static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2450 struct scan_control *sc)
2451{
2452 unsigned long total_scanned = 0;
2453 unsigned long writeback_threshold;
2454 bool aborted_reclaim;
2455
2456 delayacct_freepages_start();
2457
2458 if (global_reclaim(sc))
2459 count_vm_event(ALLOCSTALL);
2460
2461 do {
2462 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2463 sc->priority);
2464 sc->nr_scanned = 0;
2465 aborted_reclaim = shrink_zones(zonelist, sc);
2466
2467 total_scanned += sc->nr_scanned;
2468 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2469 goto out;
2470
2471 /*
2472 * If we're getting trouble reclaiming, start doing
2473 * writepage even in laptop mode.
2474 */
2475 if (sc->priority < DEF_PRIORITY - 2)
2476 sc->may_writepage = 1;
2477
2478 /*
2479 * Try to write back as many pages as we just scanned. This
2480 * tends to cause slow streaming writers to write data to the
2481 * disk smoothly, at the dirtying rate, which is nice. But
2482 * that's undesirable in laptop mode, where we *want* lumpy
2483 * writeout. So in laptop mode, write out the whole world.
2484 */
2485 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2486 if (total_scanned > writeback_threshold) {
2487 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2488 WB_REASON_TRY_TO_FREE_PAGES);
2489 sc->may_writepage = 1;
2490 }
2491 } while (--sc->priority >= 0 && !aborted_reclaim);
2492
2493out:
2494 delayacct_freepages_end();
2495
2496 if (sc->nr_reclaimed)
2497 return sc->nr_reclaimed;
2498
2499 /*
2500 * As hibernation is going on, kswapd is freezed so that it can't mark
2501 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2502 * check.
2503 */
2504 if (oom_killer_disabled)
2505 return 0;
2506
2507 /* Aborted reclaim to try compaction? don't OOM, then */
2508 if (aborted_reclaim)
2509 return 1;
2510
2511 /* top priority shrink_zones still had more to do? don't OOM, then */
2512 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2513 return 1;
2514
2515 return 0;
2516}
2517
2518static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2519{
2520 struct zone *zone;
2521 unsigned long pfmemalloc_reserve = 0;
2522 unsigned long free_pages = 0;
2523 int i;
2524 bool wmark_ok;
2525
2526 for (i = 0; i <= ZONE_NORMAL; i++) {
2527 zone = &pgdat->node_zones[i];
2528 pfmemalloc_reserve += min_wmark_pages(zone);
2529 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2530 }
2531
2532 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2533
2534 /* kswapd must be awake if processes are being throttled */
2535 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2536 pgdat->classzone_idx = min(pgdat->classzone_idx,
2537 (enum zone_type)ZONE_NORMAL);
2538 wake_up_interruptible(&pgdat->kswapd_wait);
2539 }
2540
2541 return wmark_ok;
2542}
2543
2544/*
2545 * Throttle direct reclaimers if backing storage is backed by the network
2546 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2547 * depleted. kswapd will continue to make progress and wake the processes
2548 * when the low watermark is reached.
2549 *
2550 * Returns true if a fatal signal was delivered during throttling. If this
2551 * happens, the page allocator should not consider triggering the OOM killer.
2552 */
2553static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2554 nodemask_t *nodemask)
2555{
2556 struct zone *zone;
2557 int high_zoneidx = gfp_zone(gfp_mask);
2558 pg_data_t *pgdat;
2559
2560 /*
2561 * Kernel threads should not be throttled as they may be indirectly
2562 * responsible for cleaning pages necessary for reclaim to make forward
2563 * progress. kjournald for example may enter direct reclaim while
2564 * committing a transaction where throttling it could forcing other
2565 * processes to block on log_wait_commit().
2566 */
2567 if (current->flags & PF_KTHREAD)
2568 goto out;
2569
2570 /*
2571 * If a fatal signal is pending, this process should not throttle.
2572 * It should return quickly so it can exit and free its memory
2573 */
2574 if (fatal_signal_pending(current))
2575 goto out;
2576
2577 /* Check if the pfmemalloc reserves are ok */
2578 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2579 pgdat = zone->zone_pgdat;
2580 if (pfmemalloc_watermark_ok(pgdat))
2581 goto out;
2582
2583 /* Account for the throttling */
2584 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2585
2586 /*
2587 * If the caller cannot enter the filesystem, it's possible that it
2588 * is due to the caller holding an FS lock or performing a journal
2589 * transaction in the case of a filesystem like ext[3|4]. In this case,
2590 * it is not safe to block on pfmemalloc_wait as kswapd could be
2591 * blocked waiting on the same lock. Instead, throttle for up to a
2592 * second before continuing.
2593 */
2594 if (!(gfp_mask & __GFP_FS)) {
2595 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2596 pfmemalloc_watermark_ok(pgdat), HZ);
2597
2598 goto check_pending;
2599 }
2600
2601 /* Throttle until kswapd wakes the process */
2602 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2603 pfmemalloc_watermark_ok(pgdat));
2604
2605check_pending:
2606 if (fatal_signal_pending(current))
2607 return true;
2608
2609out:
2610 return false;
2611}
2612
2613unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2614 gfp_t gfp_mask, nodemask_t *nodemask)
2615{
2616 unsigned long nr_reclaimed;
2617 struct scan_control sc = {
2618 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2619 .may_writepage = !laptop_mode,
2620 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2621 .may_unmap = 1,
2622 .may_swap = 1,
2623 .order = order,
2624 .priority = DEF_PRIORITY,
2625 .target_mem_cgroup = NULL,
2626 .nodemask = nodemask,
2627 };
2628
2629 /*
2630 * Do not enter reclaim if fatal signal was delivered while throttled.
2631 * 1 is returned so that the page allocator does not OOM kill at this
2632 * point.
2633 */
2634 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2635 return 1;
2636
2637 trace_mm_vmscan_direct_reclaim_begin(order,
2638 sc.may_writepage,
2639 gfp_mask);
2640
2641 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2642
2643 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2644
2645 return nr_reclaimed;
2646}
2647
2648#ifdef CONFIG_MEMCG
2649
2650unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2651 gfp_t gfp_mask, bool noswap,
2652 struct zone *zone,
2653 unsigned long *nr_scanned)
2654{
2655 struct scan_control sc = {
2656 .nr_scanned = 0,
2657 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2658 .may_writepage = !laptop_mode,
2659 .may_unmap = 1,
2660 .may_swap = !noswap,
2661 .order = 0,
2662 .priority = 0,
2663 .target_mem_cgroup = memcg,
2664 };
2665 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2666
2667 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2668 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2669
2670 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2671 sc.may_writepage,
2672 sc.gfp_mask);
2673
2674 /*
2675 * NOTE: Although we can get the priority field, using it
2676 * here is not a good idea, since it limits the pages we can scan.
2677 * if we don't reclaim here, the shrink_zone from balance_pgdat
2678 * will pick up pages from other mem cgroup's as well. We hack
2679 * the priority and make it zero.
2680 */
2681 shrink_lruvec(lruvec, &sc);
2682
2683 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2684
2685 *nr_scanned = sc.nr_scanned;
2686 return sc.nr_reclaimed;
2687}
2688
2689unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2690 gfp_t gfp_mask,
2691 bool noswap)
2692{
2693 struct zonelist *zonelist;
2694 unsigned long nr_reclaimed;
2695 int nid;
2696 struct scan_control sc = {
2697 .may_writepage = !laptop_mode,
2698 .may_unmap = 1,
2699 .may_swap = !noswap,
2700 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2701 .order = 0,
2702 .priority = DEF_PRIORITY,
2703 .target_mem_cgroup = memcg,
2704 .nodemask = NULL, /* we don't care the placement */
2705 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2706 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2707 };
2708
2709 /*
2710 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2711 * take care of from where we get pages. So the node where we start the
2712 * scan does not need to be the current node.
2713 */
2714 nid = mem_cgroup_select_victim_node(memcg);
2715
2716 zonelist = NODE_DATA(nid)->node_zonelists;
2717
2718 trace_mm_vmscan_memcg_reclaim_begin(0,
2719 sc.may_writepage,
2720 sc.gfp_mask);
2721
2722 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2723
2724 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2725
2726 return nr_reclaimed;
2727}
2728#endif
2729
2730static void age_active_anon(struct zone *zone, struct scan_control *sc)
2731{
2732 struct mem_cgroup *memcg;
2733
2734 if (!total_swap_pages)
2735 return;
2736
2737 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2738 do {
2739 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2740
2741 if (inactive_anon_is_low(lruvec))
2742 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2743 sc, LRU_ACTIVE_ANON);
2744
2745 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2746 } while (memcg);
2747}
2748
2749static bool zone_balanced(struct zone *zone, int order,
2750 unsigned long balance_gap, int classzone_idx)
2751{
2752 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2753 balance_gap, classzone_idx, 0))
2754 return false;
2755
2756 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2757 !compaction_suitable(zone, order))
2758 return false;
2759
2760 return true;
2761}
2762
2763/*
2764 * pgdat_balanced() is used when checking if a node is balanced.
2765 *
2766 * For order-0, all zones must be balanced!
2767 *
2768 * For high-order allocations only zones that meet watermarks and are in a
2769 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2770 * total of balanced pages must be at least 25% of the zones allowed by
2771 * classzone_idx for the node to be considered balanced. Forcing all zones to
2772 * be balanced for high orders can cause excessive reclaim when there are
2773 * imbalanced zones.
2774 * The choice of 25% is due to
2775 * o a 16M DMA zone that is balanced will not balance a zone on any
2776 * reasonable sized machine
2777 * o On all other machines, the top zone must be at least a reasonable
2778 * percentage of the middle zones. For example, on 32-bit x86, highmem
2779 * would need to be at least 256M for it to be balance a whole node.
2780 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2781 * to balance a node on its own. These seemed like reasonable ratios.
2782 */
2783static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2784{
2785 unsigned long managed_pages = 0;
2786 unsigned long balanced_pages = 0;
2787 int i;
2788
2789 /* Check the watermark levels */
2790 for (i = 0; i <= classzone_idx; i++) {
2791 struct zone *zone = pgdat->node_zones + i;
2792
2793 if (!populated_zone(zone))
2794 continue;
2795
2796 managed_pages += zone->managed_pages;
2797
2798 /*
2799 * A special case here:
2800 *
2801 * balance_pgdat() skips over all_unreclaimable after
2802 * DEF_PRIORITY. Effectively, it considers them balanced so
2803 * they must be considered balanced here as well!
2804 */
2805 if (!zone_reclaimable(zone)) {
2806 balanced_pages += zone->managed_pages;
2807 continue;
2808 }
2809
2810 if (zone_balanced(zone, order, 0, i))
2811 balanced_pages += zone->managed_pages;
2812 else if (!order)
2813 return false;
2814 }
2815
2816 if (order)
2817 return balanced_pages >= (managed_pages >> 2);
2818 else
2819 return true;
2820}
2821
2822/*
2823 * Prepare kswapd for sleeping. This verifies that there are no processes
2824 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2825 *
2826 * Returns true if kswapd is ready to sleep
2827 */
2828static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2829 int classzone_idx)
2830{
2831 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2832 if (remaining)
2833 return false;
2834
2835 /*
2836 * There is a potential race between when kswapd checks its watermarks
2837 * and a process gets throttled. There is also a potential race if
2838 * processes get throttled, kswapd wakes, a large process exits therby
2839 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2840 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2841 * so wake them now if necessary. If necessary, processes will wake
2842 * kswapd and get throttled again
2843 */
2844 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2845 wake_up(&pgdat->pfmemalloc_wait);
2846 return false;
2847 }
2848
2849 return pgdat_balanced(pgdat, order, classzone_idx);
2850}
2851
2852/*
2853 * kswapd shrinks the zone by the number of pages required to reach
2854 * the high watermark.
2855 *
2856 * Returns true if kswapd scanned at least the requested number of pages to
2857 * reclaim or if the lack of progress was due to pages under writeback.
2858 * This is used to determine if the scanning priority needs to be raised.
2859 */
2860static bool kswapd_shrink_zone(struct zone *zone,
2861 int classzone_idx,
2862 struct scan_control *sc,
2863 unsigned long lru_pages,
2864 unsigned long *nr_attempted)
2865{
2866 int testorder = sc->order;
2867 unsigned long balance_gap;
2868 struct reclaim_state *reclaim_state = current->reclaim_state;
2869 struct shrink_control shrink = {
2870 .gfp_mask = sc->gfp_mask,
2871 };
2872 bool lowmem_pressure;
2873
2874 /* Reclaim above the high watermark. */
2875 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2876
2877 /*
2878 * Kswapd reclaims only single pages with compaction enabled. Trying
2879 * too hard to reclaim until contiguous free pages have become
2880 * available can hurt performance by evicting too much useful data
2881 * from memory. Do not reclaim more than needed for compaction.
2882 */
2883 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2884 compaction_suitable(zone, sc->order) !=
2885 COMPACT_SKIPPED)
2886 testorder = 0;
2887
2888 /*
2889 * We put equal pressure on every zone, unless one zone has way too
2890 * many pages free already. The "too many pages" is defined as the
2891 * high wmark plus a "gap" where the gap is either the low
2892 * watermark or 1% of the zone, whichever is smaller.
2893 */
2894 balance_gap = min(low_wmark_pages(zone),
2895 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2896 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2897
2898 /*
2899 * If there is no low memory pressure or the zone is balanced then no
2900 * reclaim is necessary
2901 */
2902 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2903 if (!lowmem_pressure && zone_balanced(zone, testorder,
2904 balance_gap, classzone_idx))
2905 return true;
2906
2907 shrink_zone(zone, sc);
2908 nodes_clear(shrink.nodes_to_scan);
2909 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2910
2911 reclaim_state->reclaimed_slab = 0;
2912 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2913 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2914
2915 /* Account for the number of pages attempted to reclaim */
2916 *nr_attempted += sc->nr_to_reclaim;
2917
2918 zone_clear_flag(zone, ZONE_WRITEBACK);
2919
2920 /*
2921 * If a zone reaches its high watermark, consider it to be no longer
2922 * congested. It's possible there are dirty pages backed by congested
2923 * BDIs but as pressure is relieved, speculatively avoid congestion
2924 * waits.
2925 */
2926 if (zone_reclaimable(zone) &&
2927 zone_balanced(zone, testorder, 0, classzone_idx)) {
2928 zone_clear_flag(zone, ZONE_CONGESTED);
2929 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2930 }
2931
2932 return sc->nr_scanned >= sc->nr_to_reclaim;
2933}
2934
2935/*
2936 * For kswapd, balance_pgdat() will work across all this node's zones until
2937 * they are all at high_wmark_pages(zone).
2938 *
2939 * Returns the final order kswapd was reclaiming at
2940 *
2941 * There is special handling here for zones which are full of pinned pages.
2942 * This can happen if the pages are all mlocked, or if they are all used by
2943 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2944 * What we do is to detect the case where all pages in the zone have been
2945 * scanned twice and there has been zero successful reclaim. Mark the zone as
2946 * dead and from now on, only perform a short scan. Basically we're polling
2947 * the zone for when the problem goes away.
2948 *
2949 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2950 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2951 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2952 * lower zones regardless of the number of free pages in the lower zones. This
2953 * interoperates with the page allocator fallback scheme to ensure that aging
2954 * of pages is balanced across the zones.
2955 */
2956static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2957 int *classzone_idx)
2958{
2959 int i;
2960 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2961 unsigned long nr_soft_reclaimed;
2962 unsigned long nr_soft_scanned;
2963 struct scan_control sc = {
2964 .gfp_mask = GFP_KERNEL,
2965 .priority = DEF_PRIORITY,
2966 .may_unmap = 1,
2967 .may_swap = 1,
2968 .may_writepage = !laptop_mode,
2969 .order = order,
2970 .target_mem_cgroup = NULL,
2971 };
2972 count_vm_event(PAGEOUTRUN);
2973
2974 do {
2975 unsigned long lru_pages = 0;
2976 unsigned long nr_attempted = 0;
2977 bool raise_priority = true;
2978 bool pgdat_needs_compaction = (order > 0);
2979
2980 sc.nr_reclaimed = 0;
2981
2982 /*
2983 * Scan in the highmem->dma direction for the highest
2984 * zone which needs scanning
2985 */
2986 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2987 struct zone *zone = pgdat->node_zones + i;
2988
2989 if (!populated_zone(zone))
2990 continue;
2991
2992 if (sc.priority != DEF_PRIORITY &&
2993 !zone_reclaimable(zone))
2994 continue;
2995
2996 /*
2997 * Do some background aging of the anon list, to give
2998 * pages a chance to be referenced before reclaiming.
2999 */
3000 age_active_anon(zone, &sc);
3001
3002 /*
3003 * If the number of buffer_heads in the machine
3004 * exceeds the maximum allowed level and this node
3005 * has a highmem zone, force kswapd to reclaim from
3006 * it to relieve lowmem pressure.
3007 */
3008 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3009 end_zone = i;
3010 break;
3011 }
3012
3013 if (!zone_balanced(zone, order, 0, 0)) {
3014 end_zone = i;
3015 break;
3016 } else {
3017 /*
3018 * If balanced, clear the dirty and congested
3019 * flags
3020 */
3021 zone_clear_flag(zone, ZONE_CONGESTED);
3022 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3023 }
3024 }
3025
3026 if (i < 0)
3027 goto out;
3028
3029 for (i = 0; i <= end_zone; i++) {
3030 struct zone *zone = pgdat->node_zones + i;
3031
3032 if (!populated_zone(zone))
3033 continue;
3034
3035 lru_pages += zone_reclaimable_pages(zone);
3036
3037 /*
3038 * If any zone is currently balanced then kswapd will
3039 * not call compaction as it is expected that the
3040 * necessary pages are already available.
3041 */
3042 if (pgdat_needs_compaction &&
3043 zone_watermark_ok(zone, order,
3044 low_wmark_pages(zone),
3045 *classzone_idx, 0))
3046 pgdat_needs_compaction = false;
3047 }
3048
3049 /*
3050 * If we're getting trouble reclaiming, start doing writepage
3051 * even in laptop mode.
3052 */
3053 if (sc.priority < DEF_PRIORITY - 2)
3054 sc.may_writepage = 1;
3055
3056 /*
3057 * Now scan the zone in the dma->highmem direction, stopping
3058 * at the last zone which needs scanning.
3059 *
3060 * We do this because the page allocator works in the opposite
3061 * direction. This prevents the page allocator from allocating
3062 * pages behind kswapd's direction of progress, which would
3063 * cause too much scanning of the lower zones.
3064 */
3065 for (i = 0; i <= end_zone; i++) {
3066 struct zone *zone = pgdat->node_zones + i;
3067
3068 if (!populated_zone(zone))
3069 continue;
3070
3071 if (sc.priority != DEF_PRIORITY &&
3072 !zone_reclaimable(zone))
3073 continue;
3074
3075 sc.nr_scanned = 0;
3076
3077 nr_soft_scanned = 0;
3078 /*
3079 * Call soft limit reclaim before calling shrink_zone.
3080 */
3081 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3082 order, sc.gfp_mask,
3083 &nr_soft_scanned);
3084 sc.nr_reclaimed += nr_soft_reclaimed;
3085
3086 /*
3087 * There should be no need to raise the scanning
3088 * priority if enough pages are already being scanned
3089 * that that high watermark would be met at 100%
3090 * efficiency.
3091 */
3092 if (kswapd_shrink_zone(zone, end_zone, &sc,
3093 lru_pages, &nr_attempted))
3094 raise_priority = false;
3095 }
3096
3097 /*
3098 * If the low watermark is met there is no need for processes
3099 * to be throttled on pfmemalloc_wait as they should not be
3100 * able to safely make forward progress. Wake them
3101 */
3102 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3103 pfmemalloc_watermark_ok(pgdat))
3104 wake_up(&pgdat->pfmemalloc_wait);
3105
3106 /*
3107 * Fragmentation may mean that the system cannot be rebalanced
3108 * for high-order allocations in all zones. If twice the
3109 * allocation size has been reclaimed and the zones are still
3110 * not balanced then recheck the watermarks at order-0 to
3111 * prevent kswapd reclaiming excessively. Assume that a
3112 * process requested a high-order can direct reclaim/compact.
3113 */
3114 if (order && sc.nr_reclaimed >= 2UL << order)
3115 order = sc.order = 0;
3116
3117 /* Check if kswapd should be suspending */
3118 if (try_to_freeze() || kthread_should_stop())
3119 break;
3120
3121 /*
3122 * Compact if necessary and kswapd is reclaiming at least the
3123 * high watermark number of pages as requsted
3124 */
3125 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3126 compact_pgdat(pgdat, order);
3127
3128 /*
3129 * Raise priority if scanning rate is too low or there was no
3130 * progress in reclaiming pages
3131 */
3132 if (raise_priority || !sc.nr_reclaimed)
3133 sc.priority--;
3134 } while (sc.priority >= 1 &&
3135 !pgdat_balanced(pgdat, order, *classzone_idx));
3136
3137out:
3138 /*
3139 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3140 * makes a decision on the order we were last reclaiming at. However,
3141 * if another caller entered the allocator slow path while kswapd
3142 * was awake, order will remain at the higher level
3143 */
3144 *classzone_idx = end_zone;
3145 return order;
3146}
3147
3148static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3149{
3150 long remaining = 0;
3151 DEFINE_WAIT(wait);
3152
3153 if (freezing(current) || kthread_should_stop())
3154 return;
3155
3156 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3157
3158 /* Try to sleep for a short interval */
3159 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3160 remaining = schedule_timeout(HZ/10);
3161 finish_wait(&pgdat->kswapd_wait, &wait);
3162 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3163 }
3164
3165 /*
3166 * After a short sleep, check if it was a premature sleep. If not, then
3167 * go fully to sleep until explicitly woken up.
3168 */
3169 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3170 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3171
3172 /*
3173 * vmstat counters are not perfectly accurate and the estimated
3174 * value for counters such as NR_FREE_PAGES can deviate from the
3175 * true value by nr_online_cpus * threshold. To avoid the zone
3176 * watermarks being breached while under pressure, we reduce the
3177 * per-cpu vmstat threshold while kswapd is awake and restore
3178 * them before going back to sleep.
3179 */
3180 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3181
3182 /*
3183 * Compaction records what page blocks it recently failed to
3184 * isolate pages from and skips them in the future scanning.
3185 * When kswapd is going to sleep, it is reasonable to assume
3186 * that pages and compaction may succeed so reset the cache.
3187 */
3188 reset_isolation_suitable(pgdat);
3189
3190 if (!kthread_should_stop())
3191 schedule();
3192
3193 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3194 } else {
3195 if (remaining)
3196 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3197 else
3198 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3199 }
3200 finish_wait(&pgdat->kswapd_wait, &wait);
3201}
3202
3203/*
3204 * The background pageout daemon, started as a kernel thread
3205 * from the init process.
3206 *
3207 * This basically trickles out pages so that we have _some_
3208 * free memory available even if there is no other activity
3209 * that frees anything up. This is needed for things like routing
3210 * etc, where we otherwise might have all activity going on in
3211 * asynchronous contexts that cannot page things out.
3212 *
3213 * If there are applications that are active memory-allocators
3214 * (most normal use), this basically shouldn't matter.
3215 */
3216static int kswapd(void *p)
3217{
3218 unsigned long order, new_order;
3219 unsigned balanced_order;
3220 int classzone_idx, new_classzone_idx;
3221 int balanced_classzone_idx;
3222 pg_data_t *pgdat = (pg_data_t*)p;
3223 struct task_struct *tsk = current;
3224
3225 struct reclaim_state reclaim_state = {
3226 .reclaimed_slab = 0,
3227 };
3228 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3229
3230 lockdep_set_current_reclaim_state(GFP_KERNEL);
3231
3232 if (!cpumask_empty(cpumask))
3233 set_cpus_allowed_ptr(tsk, cpumask);
3234 current->reclaim_state = &reclaim_state;
3235
3236 /*
3237 * Tell the memory management that we're a "memory allocator",
3238 * and that if we need more memory we should get access to it
3239 * regardless (see "__alloc_pages()"). "kswapd" should
3240 * never get caught in the normal page freeing logic.
3241 *
3242 * (Kswapd normally doesn't need memory anyway, but sometimes
3243 * you need a small amount of memory in order to be able to
3244 * page out something else, and this flag essentially protects
3245 * us from recursively trying to free more memory as we're
3246 * trying to free the first piece of memory in the first place).
3247 */
3248 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3249 set_freezable();
3250
3251 order = new_order = 0;
3252 balanced_order = 0;
3253 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3254 balanced_classzone_idx = classzone_idx;
3255 for ( ; ; ) {
3256 bool ret;
3257
3258 /*
3259 * If the last balance_pgdat was unsuccessful it's unlikely a
3260 * new request of a similar or harder type will succeed soon
3261 * so consider going to sleep on the basis we reclaimed at
3262 */
3263 if (balanced_classzone_idx >= new_classzone_idx &&
3264 balanced_order == new_order) {
3265 new_order = pgdat->kswapd_max_order;
3266 new_classzone_idx = pgdat->classzone_idx;
3267 pgdat->kswapd_max_order = 0;
3268 pgdat->classzone_idx = pgdat->nr_zones - 1;
3269 }
3270
3271 if (order < new_order || classzone_idx > new_classzone_idx) {
3272 /*
3273 * Don't sleep if someone wants a larger 'order'
3274 * allocation or has tigher zone constraints
3275 */
3276 order = new_order;
3277 classzone_idx = new_classzone_idx;
3278 } else {
3279 kswapd_try_to_sleep(pgdat, balanced_order,
3280 balanced_classzone_idx);
3281 order = pgdat->kswapd_max_order;
3282 classzone_idx = pgdat->classzone_idx;
3283 new_order = order;
3284 new_classzone_idx = classzone_idx;
3285 pgdat->kswapd_max_order = 0;
3286 pgdat->classzone_idx = pgdat->nr_zones - 1;
3287 }
3288
3289 ret = try_to_freeze();
3290 if (kthread_should_stop())
3291 break;
3292
3293 /*
3294 * We can speed up thawing tasks if we don't call balance_pgdat
3295 * after returning from the refrigerator
3296 */
3297 if (!ret) {
3298 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3299 balanced_classzone_idx = classzone_idx;
3300 balanced_order = balance_pgdat(pgdat, order,
3301 &balanced_classzone_idx);
3302 }
3303 }
3304
3305 current->reclaim_state = NULL;
3306 return 0;
3307}
3308
3309/*
3310 * A zone is low on free memory, so wake its kswapd task to service it.
3311 */
3312void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3313{
3314 pg_data_t *pgdat;
3315
3316 if (!populated_zone(zone))
3317 return;
3318
3319 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3320 return;
3321 pgdat = zone->zone_pgdat;
3322 if (pgdat->kswapd_max_order < order) {
3323 pgdat->kswapd_max_order = order;
3324 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3325 }
3326 if (!waitqueue_active(&pgdat->kswapd_wait))
3327 return;
3328 if (zone_balanced(zone, order, 0, 0))
3329 return;
3330
3331 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3332 wake_up_interruptible(&pgdat->kswapd_wait);
3333}
3334
3335#ifdef CONFIG_HIBERNATION
3336/*
3337 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3338 * freed pages.
3339 *
3340 * Rather than trying to age LRUs the aim is to preserve the overall
3341 * LRU order by reclaiming preferentially
3342 * inactive > active > active referenced > active mapped
3343 */
3344unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3345{
3346 struct reclaim_state reclaim_state;
3347 struct scan_control sc = {
3348 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3349 .may_swap = 1,
3350 .may_unmap = 1,
3351 .may_writepage = 1,
3352 .nr_to_reclaim = nr_to_reclaim,
3353 .hibernation_mode = 1,
3354 .order = 0,
3355 .priority = DEF_PRIORITY,
3356 };
3357 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3358 struct task_struct *p = current;
3359 unsigned long nr_reclaimed;
3360
3361 p->flags |= PF_MEMALLOC;
3362 lockdep_set_current_reclaim_state(sc.gfp_mask);
3363 reclaim_state.reclaimed_slab = 0;
3364 p->reclaim_state = &reclaim_state;
3365
3366 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3367
3368 p->reclaim_state = NULL;
3369 lockdep_clear_current_reclaim_state();
3370 p->flags &= ~PF_MEMALLOC;
3371
3372 return nr_reclaimed;
3373}
3374#endif /* CONFIG_HIBERNATION */
3375
3376/* It's optimal to keep kswapds on the same CPUs as their memory, but
3377 not required for correctness. So if the last cpu in a node goes
3378 away, we get changed to run anywhere: as the first one comes back,
3379 restore their cpu bindings. */
3380static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3381 void *hcpu)
3382{
3383 int nid;
3384
3385 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3386 for_each_node_state(nid, N_MEMORY) {
3387 pg_data_t *pgdat = NODE_DATA(nid);
3388 const struct cpumask *mask;
3389
3390 mask = cpumask_of_node(pgdat->node_id);
3391
3392 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3393 /* One of our CPUs online: restore mask */
3394 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3395 }
3396 }
3397 return NOTIFY_OK;
3398}
3399
3400/*
3401 * This kswapd start function will be called by init and node-hot-add.
3402 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3403 */
3404int kswapd_run(int nid)
3405{
3406 pg_data_t *pgdat = NODE_DATA(nid);
3407 int ret = 0;
3408
3409 if (pgdat->kswapd)
3410 return 0;
3411
3412 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3413 if (IS_ERR(pgdat->kswapd)) {
3414 /* failure at boot is fatal */
3415 BUG_ON(system_state == SYSTEM_BOOTING);
3416 pr_err("Failed to start kswapd on node %d\n", nid);
3417 ret = PTR_ERR(pgdat->kswapd);
3418 pgdat->kswapd = NULL;
3419 }
3420 return ret;
3421}
3422
3423/*
3424 * Called by memory hotplug when all memory in a node is offlined. Caller must
3425 * hold lock_memory_hotplug().
3426 */
3427void kswapd_stop(int nid)
3428{
3429 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3430
3431 if (kswapd) {
3432 kthread_stop(kswapd);
3433 NODE_DATA(nid)->kswapd = NULL;
3434 }
3435}
3436
3437static int __init kswapd_init(void)
3438{
3439 int nid;
3440
3441 swap_setup();
3442 for_each_node_state(nid, N_MEMORY)
3443 kswapd_run(nid);
3444 hotcpu_notifier(cpu_callback, 0);
3445 return 0;
3446}
3447
3448module_init(kswapd_init)
3449
3450#ifdef CONFIG_NUMA
3451/*
3452 * Zone reclaim mode
3453 *
3454 * If non-zero call zone_reclaim when the number of free pages falls below
3455 * the watermarks.
3456 */
3457int zone_reclaim_mode __read_mostly;
3458
3459#define RECLAIM_OFF 0
3460#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3461#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3462#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3463
3464/*
3465 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3466 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3467 * a zone.
3468 */
3469#define ZONE_RECLAIM_PRIORITY 4
3470
3471/*
3472 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3473 * occur.
3474 */
3475int sysctl_min_unmapped_ratio = 1;
3476
3477/*
3478 * If the number of slab pages in a zone grows beyond this percentage then
3479 * slab reclaim needs to occur.
3480 */
3481int sysctl_min_slab_ratio = 5;
3482
3483static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3484{
3485 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3486 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3487 zone_page_state(zone, NR_ACTIVE_FILE);
3488
3489 /*
3490 * It's possible for there to be more file mapped pages than
3491 * accounted for by the pages on the file LRU lists because
3492 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3493 */
3494 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3495}
3496
3497/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3498static long zone_pagecache_reclaimable(struct zone *zone)
3499{
3500 long nr_pagecache_reclaimable;
3501 long delta = 0;
3502
3503 /*
3504 * If RECLAIM_SWAP is set, then all file pages are considered
3505 * potentially reclaimable. Otherwise, we have to worry about
3506 * pages like swapcache and zone_unmapped_file_pages() provides
3507 * a better estimate
3508 */
3509 if (zone_reclaim_mode & RECLAIM_SWAP)
3510 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3511 else
3512 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3513
3514 /* If we can't clean pages, remove dirty pages from consideration */
3515 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3516 delta += zone_page_state(zone, NR_FILE_DIRTY);
3517
3518 /* Watch for any possible underflows due to delta */
3519 if (unlikely(delta > nr_pagecache_reclaimable))
3520 delta = nr_pagecache_reclaimable;
3521
3522 return nr_pagecache_reclaimable - delta;
3523}
3524
3525/*
3526 * Try to free up some pages from this zone through reclaim.
3527 */
3528static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3529{
3530 /* Minimum pages needed in order to stay on node */
3531 const unsigned long nr_pages = 1 << order;
3532 struct task_struct *p = current;
3533 struct reclaim_state reclaim_state;
3534 struct scan_control sc = {
3535 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3536 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3537 .may_swap = 1,
3538 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3539 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3540 .order = order,
3541 .priority = ZONE_RECLAIM_PRIORITY,
3542 };
3543 struct shrink_control shrink = {
3544 .gfp_mask = sc.gfp_mask,
3545 };
3546 unsigned long nr_slab_pages0, nr_slab_pages1;
3547
3548 cond_resched();
3549 /*
3550 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3551 * and we also need to be able to write out pages for RECLAIM_WRITE
3552 * and RECLAIM_SWAP.
3553 */
3554 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3555 lockdep_set_current_reclaim_state(gfp_mask);
3556 reclaim_state.reclaimed_slab = 0;
3557 p->reclaim_state = &reclaim_state;
3558
3559 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3560 /*
3561 * Free memory by calling shrink zone with increasing
3562 * priorities until we have enough memory freed.
3563 */
3564 do {
3565 shrink_zone(zone, &sc);
3566 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3567 }
3568
3569 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3570 if (nr_slab_pages0 > zone->min_slab_pages) {
3571 /*
3572 * shrink_slab() does not currently allow us to determine how
3573 * many pages were freed in this zone. So we take the current
3574 * number of slab pages and shake the slab until it is reduced
3575 * by the same nr_pages that we used for reclaiming unmapped
3576 * pages.
3577 */
3578 nodes_clear(shrink.nodes_to_scan);
3579 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3580 for (;;) {
3581 unsigned long lru_pages = zone_reclaimable_pages(zone);
3582
3583 /* No reclaimable slab or very low memory pressure */
3584 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3585 break;
3586
3587 /* Freed enough memory */
3588 nr_slab_pages1 = zone_page_state(zone,
3589 NR_SLAB_RECLAIMABLE);
3590 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3591 break;
3592 }
3593
3594 /*
3595 * Update nr_reclaimed by the number of slab pages we
3596 * reclaimed from this zone.
3597 */
3598 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3599 if (nr_slab_pages1 < nr_slab_pages0)
3600 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3601 }
3602
3603 p->reclaim_state = NULL;
3604 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3605 lockdep_clear_current_reclaim_state();
3606 return sc.nr_reclaimed >= nr_pages;
3607}
3608
3609int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3610{
3611 int node_id;
3612 int ret;
3613
3614 /*
3615 * Zone reclaim reclaims unmapped file backed pages and
3616 * slab pages if we are over the defined limits.
3617 *
3618 * A small portion of unmapped file backed pages is needed for
3619 * file I/O otherwise pages read by file I/O will be immediately
3620 * thrown out if the zone is overallocated. So we do not reclaim
3621 * if less than a specified percentage of the zone is used by
3622 * unmapped file backed pages.
3623 */
3624 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3625 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3626 return ZONE_RECLAIM_FULL;
3627
3628 if (!zone_reclaimable(zone))
3629 return ZONE_RECLAIM_FULL;
3630
3631 /*
3632 * Do not scan if the allocation should not be delayed.
3633 */
3634 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3635 return ZONE_RECLAIM_NOSCAN;
3636
3637 /*
3638 * Only run zone reclaim on the local zone or on zones that do not
3639 * have associated processors. This will favor the local processor
3640 * over remote processors and spread off node memory allocations
3641 * as wide as possible.
3642 */
3643 node_id = zone_to_nid(zone);
3644 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3645 return ZONE_RECLAIM_NOSCAN;
3646
3647 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3648 return ZONE_RECLAIM_NOSCAN;
3649
3650 ret = __zone_reclaim(zone, gfp_mask, order);
3651 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3652
3653 if (!ret)
3654 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3655
3656 return ret;
3657}
3658#endif
3659
3660/*
3661 * page_evictable - test whether a page is evictable
3662 * @page: the page to test
3663 *
3664 * Test whether page is evictable--i.e., should be placed on active/inactive
3665 * lists vs unevictable list.
3666 *
3667 * Reasons page might not be evictable:
3668 * (1) page's mapping marked unevictable
3669 * (2) page is part of an mlocked VMA
3670 *
3671 */
3672int page_evictable(struct page *page)
3673{
3674 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3675}
3676
3677#ifdef CONFIG_SHMEM
3678/**
3679 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3680 * @pages: array of pages to check
3681 * @nr_pages: number of pages to check
3682 *
3683 * Checks pages for evictability and moves them to the appropriate lru list.
3684 *
3685 * This function is only used for SysV IPC SHM_UNLOCK.
3686 */
3687void check_move_unevictable_pages(struct page **pages, int nr_pages)
3688{
3689 struct lruvec *lruvec;
3690 struct zone *zone = NULL;
3691 int pgscanned = 0;
3692 int pgrescued = 0;
3693 int i;
3694
3695 for (i = 0; i < nr_pages; i++) {
3696 struct page *page = pages[i];
3697 struct zone *pagezone;
3698
3699 pgscanned++;
3700 pagezone = page_zone(page);
3701 if (pagezone != zone) {
3702 if (zone)
3703 spin_unlock_irq(&zone->lru_lock);
3704 zone = pagezone;
3705 spin_lock_irq(&zone->lru_lock);
3706 }
3707 lruvec = mem_cgroup_page_lruvec(page, zone);
3708
3709 if (!PageLRU(page) || !PageUnevictable(page))
3710 continue;
3711
3712 if (page_evictable(page)) {
3713 enum lru_list lru = page_lru_base_type(page);
3714
3715 VM_BUG_ON_PAGE(PageActive(page), page);
3716 ClearPageUnevictable(page);
3717 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3718 add_page_to_lru_list(page, lruvec, lru);
3719 pgrescued++;
3720 }
3721 }
3722
3723 if (zone) {
3724 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3725 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3726 spin_unlock_irq(&zone->lru_lock);
3727 }
3728}
3729#endif /* CONFIG_SHMEM */
3730
3731static void warn_scan_unevictable_pages(void)
3732{
3733 printk_once(KERN_WARNING
3734 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3735 "disabled for lack of a legitimate use case. If you have "
3736 "one, please send an email to linux-mm@kvack.org.\n",
3737 current->comm);
3738}
3739
3740/*
3741 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3742 * all nodes' unevictable lists for evictable pages
3743 */
3744unsigned long scan_unevictable_pages;
3745
3746int scan_unevictable_handler(struct ctl_table *table, int write,
3747 void __user *buffer,
3748 size_t *length, loff_t *ppos)
3749{
3750 warn_scan_unevictable_pages();
3751 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3752 scan_unevictable_pages = 0;
3753 return 0;
3754}
3755
3756#ifdef CONFIG_NUMA
3757/*
3758 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3759 * a specified node's per zone unevictable lists for evictable pages.
3760 */
3761
3762static ssize_t read_scan_unevictable_node(struct device *dev,
3763 struct device_attribute *attr,
3764 char *buf)
3765{
3766 warn_scan_unevictable_pages();
3767 return sprintf(buf, "0\n"); /* always zero; should fit... */
3768}
3769
3770static ssize_t write_scan_unevictable_node(struct device *dev,
3771 struct device_attribute *attr,
3772 const char *buf, size_t count)
3773{
3774 warn_scan_unevictable_pages();
3775 return 1;
3776}
3777
3778
3779static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3780 read_scan_unevictable_node,
3781 write_scan_unevictable_node);
3782
3783int scan_unevictable_register_node(struct node *node)
3784{
3785 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3786}
3787
3788void scan_unevictable_unregister_node(struct node *node)
3789{
3790 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3791}
3792#endif