]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/vmscan.c
Merge branch 'stable-4.8' of git://git.infradead.org/users/pcmoore/audit
[mirror_ubuntu-artful-kernel.git] / mm / vmscan.c
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
62 struct 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 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
89
90 unsigned int may_writepage:1;
91
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
94
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
97
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
100
101 unsigned int hibernation_mode:1;
102
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
105
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
108
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
111 };
112
113 #ifdef ARCH_HAS_PREFETCH
114 #define prefetch_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
121 } \
122 } while (0)
123 #else
124 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 #ifdef ARCH_HAS_PREFETCHW
128 #define prefetchw_prev_lru_page(_page, _base, _field) \
129 do { \
130 if ((_page)->lru.prev != _base) { \
131 struct page *prev; \
132 \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
135 } \
136 } while (0)
137 #else
138 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139 #endif
140
141 /*
142 * From 0 .. 100. Higher means more swappy.
143 */
144 int vm_swappiness = 60;
145 /*
146 * The total number of pages which are beyond the high watermark within all
147 * zones.
148 */
149 unsigned long vm_total_pages;
150
151 static LIST_HEAD(shrinker_list);
152 static DECLARE_RWSEM(shrinker_rwsem);
153
154 #ifdef CONFIG_MEMCG
155 static bool global_reclaim(struct scan_control *sc)
156 {
157 return !sc->target_mem_cgroup;
158 }
159
160 /**
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
163 *
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
169 *
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
172 */
173 static bool sane_reclaim(struct scan_control *sc)
174 {
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
176
177 if (!memcg)
178 return true;
179 #ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
181 return true;
182 #endif
183 return false;
184 }
185 #else
186 static bool global_reclaim(struct scan_control *sc)
187 {
188 return true;
189 }
190
191 static bool sane_reclaim(struct scan_control *sc)
192 {
193 return true;
194 }
195 #endif
196
197 /*
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
201 */
202 unsigned long zone_reclaimable_pages(struct zone *zone)
203 {
204 unsigned long nr;
205
206 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
207 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
208 if (get_nr_swap_pages() > 0)
209 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
210 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
211
212 return nr;
213 }
214
215 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
216 {
217 unsigned long nr;
218
219 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
220 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
222
223 if (get_nr_swap_pages() > 0)
224 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
225 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
227
228 return nr;
229 }
230
231 bool pgdat_reclaimable(struct pglist_data *pgdat)
232 {
233 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
234 pgdat_reclaimable_pages(pgdat) * 6;
235 }
236
237 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
238 {
239 if (!mem_cgroup_disabled())
240 return mem_cgroup_get_lru_size(lruvec, lru);
241
242 return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
243 }
244
245 /*
246 * Add a shrinker callback to be called from the vm.
247 */
248 int register_shrinker(struct shrinker *shrinker)
249 {
250 size_t size = sizeof(*shrinker->nr_deferred);
251
252 if (shrinker->flags & SHRINKER_NUMA_AWARE)
253 size *= nr_node_ids;
254
255 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
256 if (!shrinker->nr_deferred)
257 return -ENOMEM;
258
259 down_write(&shrinker_rwsem);
260 list_add_tail(&shrinker->list, &shrinker_list);
261 up_write(&shrinker_rwsem);
262 return 0;
263 }
264 EXPORT_SYMBOL(register_shrinker);
265
266 /*
267 * Remove one
268 */
269 void unregister_shrinker(struct shrinker *shrinker)
270 {
271 down_write(&shrinker_rwsem);
272 list_del(&shrinker->list);
273 up_write(&shrinker_rwsem);
274 kfree(shrinker->nr_deferred);
275 }
276 EXPORT_SYMBOL(unregister_shrinker);
277
278 #define SHRINK_BATCH 128
279
280 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
281 struct shrinker *shrinker,
282 unsigned long nr_scanned,
283 unsigned long nr_eligible)
284 {
285 unsigned long freed = 0;
286 unsigned long long delta;
287 long total_scan;
288 long freeable;
289 long nr;
290 long new_nr;
291 int nid = shrinkctl->nid;
292 long batch_size = shrinker->batch ? shrinker->batch
293 : SHRINK_BATCH;
294
295 freeable = shrinker->count_objects(shrinker, shrinkctl);
296 if (freeable == 0)
297 return 0;
298
299 /*
300 * copy the current shrinker scan count into a local variable
301 * and zero it so that other concurrent shrinker invocations
302 * don't also do this scanning work.
303 */
304 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
305
306 total_scan = nr;
307 delta = (4 * nr_scanned) / shrinker->seeks;
308 delta *= freeable;
309 do_div(delta, nr_eligible + 1);
310 total_scan += delta;
311 if (total_scan < 0) {
312 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
313 shrinker->scan_objects, total_scan);
314 total_scan = freeable;
315 }
316
317 /*
318 * We need to avoid excessive windup on filesystem shrinkers
319 * due to large numbers of GFP_NOFS allocations causing the
320 * shrinkers to return -1 all the time. This results in a large
321 * nr being built up so when a shrink that can do some work
322 * comes along it empties the entire cache due to nr >>>
323 * freeable. This is bad for sustaining a working set in
324 * memory.
325 *
326 * Hence only allow the shrinker to scan the entire cache when
327 * a large delta change is calculated directly.
328 */
329 if (delta < freeable / 4)
330 total_scan = min(total_scan, freeable / 2);
331
332 /*
333 * Avoid risking looping forever due to too large nr value:
334 * never try to free more than twice the estimate number of
335 * freeable entries.
336 */
337 if (total_scan > freeable * 2)
338 total_scan = freeable * 2;
339
340 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
341 nr_scanned, nr_eligible,
342 freeable, delta, total_scan);
343
344 /*
345 * Normally, we should not scan less than batch_size objects in one
346 * pass to avoid too frequent shrinker calls, but if the slab has less
347 * than batch_size objects in total and we are really tight on memory,
348 * we will try to reclaim all available objects, otherwise we can end
349 * up failing allocations although there are plenty of reclaimable
350 * objects spread over several slabs with usage less than the
351 * batch_size.
352 *
353 * We detect the "tight on memory" situations by looking at the total
354 * number of objects we want to scan (total_scan). If it is greater
355 * than the total number of objects on slab (freeable), we must be
356 * scanning at high prio and therefore should try to reclaim as much as
357 * possible.
358 */
359 while (total_scan >= batch_size ||
360 total_scan >= freeable) {
361 unsigned long ret;
362 unsigned long nr_to_scan = min(batch_size, total_scan);
363
364 shrinkctl->nr_to_scan = nr_to_scan;
365 ret = shrinker->scan_objects(shrinker, shrinkctl);
366 if (ret == SHRINK_STOP)
367 break;
368 freed += ret;
369
370 count_vm_events(SLABS_SCANNED, nr_to_scan);
371 total_scan -= nr_to_scan;
372
373 cond_resched();
374 }
375
376 /*
377 * move the unused scan count back into the shrinker in a
378 * manner that handles concurrent updates. If we exhausted the
379 * scan, there is no need to do an update.
380 */
381 if (total_scan > 0)
382 new_nr = atomic_long_add_return(total_scan,
383 &shrinker->nr_deferred[nid]);
384 else
385 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
386
387 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
388 return freed;
389 }
390
391 /**
392 * shrink_slab - shrink slab caches
393 * @gfp_mask: allocation context
394 * @nid: node whose slab caches to target
395 * @memcg: memory cgroup whose slab caches to target
396 * @nr_scanned: pressure numerator
397 * @nr_eligible: pressure denominator
398 *
399 * Call the shrink functions to age shrinkable caches.
400 *
401 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
402 * unaware shrinkers will receive a node id of 0 instead.
403 *
404 * @memcg specifies the memory cgroup to target. If it is not NULL,
405 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
406 * objects from the memory cgroup specified. Otherwise, only unaware
407 * shrinkers are called.
408 *
409 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
410 * the available objects should be scanned. Page reclaim for example
411 * passes the number of pages scanned and the number of pages on the
412 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
413 * when it encountered mapped pages. The ratio is further biased by
414 * the ->seeks setting of the shrink function, which indicates the
415 * cost to recreate an object relative to that of an LRU page.
416 *
417 * Returns the number of reclaimed slab objects.
418 */
419 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
420 struct mem_cgroup *memcg,
421 unsigned long nr_scanned,
422 unsigned long nr_eligible)
423 {
424 struct shrinker *shrinker;
425 unsigned long freed = 0;
426
427 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
428 return 0;
429
430 if (nr_scanned == 0)
431 nr_scanned = SWAP_CLUSTER_MAX;
432
433 if (!down_read_trylock(&shrinker_rwsem)) {
434 /*
435 * If we would return 0, our callers would understand that we
436 * have nothing else to shrink and give up trying. By returning
437 * 1 we keep it going and assume we'll be able to shrink next
438 * time.
439 */
440 freed = 1;
441 goto out;
442 }
443
444 list_for_each_entry(shrinker, &shrinker_list, list) {
445 struct shrink_control sc = {
446 .gfp_mask = gfp_mask,
447 .nid = nid,
448 .memcg = memcg,
449 };
450
451 /*
452 * If kernel memory accounting is disabled, we ignore
453 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
454 * passing NULL for memcg.
455 */
456 if (memcg_kmem_enabled() &&
457 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
458 continue;
459
460 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
461 sc.nid = 0;
462
463 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
464 }
465
466 up_read(&shrinker_rwsem);
467 out:
468 cond_resched();
469 return freed;
470 }
471
472 void drop_slab_node(int nid)
473 {
474 unsigned long freed;
475
476 do {
477 struct mem_cgroup *memcg = NULL;
478
479 freed = 0;
480 do {
481 freed += shrink_slab(GFP_KERNEL, nid, memcg,
482 1000, 1000);
483 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
484 } while (freed > 10);
485 }
486
487 void drop_slab(void)
488 {
489 int nid;
490
491 for_each_online_node(nid)
492 drop_slab_node(nid);
493 }
494
495 static inline int is_page_cache_freeable(struct page *page)
496 {
497 /*
498 * A freeable page cache page is referenced only by the caller
499 * that isolated the page, the page cache radix tree and
500 * optional buffer heads at page->private.
501 */
502 return page_count(page) - page_has_private(page) == 2;
503 }
504
505 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
506 {
507 if (current->flags & PF_SWAPWRITE)
508 return 1;
509 if (!inode_write_congested(inode))
510 return 1;
511 if (inode_to_bdi(inode) == current->backing_dev_info)
512 return 1;
513 return 0;
514 }
515
516 /*
517 * We detected a synchronous write error writing a page out. Probably
518 * -ENOSPC. We need to propagate that into the address_space for a subsequent
519 * fsync(), msync() or close().
520 *
521 * The tricky part is that after writepage we cannot touch the mapping: nothing
522 * prevents it from being freed up. But we have a ref on the page and once
523 * that page is locked, the mapping is pinned.
524 *
525 * We're allowed to run sleeping lock_page() here because we know the caller has
526 * __GFP_FS.
527 */
528 static void handle_write_error(struct address_space *mapping,
529 struct page *page, int error)
530 {
531 lock_page(page);
532 if (page_mapping(page) == mapping)
533 mapping_set_error(mapping, error);
534 unlock_page(page);
535 }
536
537 /* possible outcome of pageout() */
538 typedef enum {
539 /* failed to write page out, page is locked */
540 PAGE_KEEP,
541 /* move page to the active list, page is locked */
542 PAGE_ACTIVATE,
543 /* page has been sent to the disk successfully, page is unlocked */
544 PAGE_SUCCESS,
545 /* page is clean and locked */
546 PAGE_CLEAN,
547 } pageout_t;
548
549 /*
550 * pageout is called by shrink_page_list() for each dirty page.
551 * Calls ->writepage().
552 */
553 static pageout_t pageout(struct page *page, struct address_space *mapping,
554 struct scan_control *sc)
555 {
556 /*
557 * If the page is dirty, only perform writeback if that write
558 * will be non-blocking. To prevent this allocation from being
559 * stalled by pagecache activity. But note that there may be
560 * stalls if we need to run get_block(). We could test
561 * PagePrivate for that.
562 *
563 * If this process is currently in __generic_file_write_iter() against
564 * this page's queue, we can perform writeback even if that
565 * will block.
566 *
567 * If the page is swapcache, write it back even if that would
568 * block, for some throttling. This happens by accident, because
569 * swap_backing_dev_info is bust: it doesn't reflect the
570 * congestion state of the swapdevs. Easy to fix, if needed.
571 */
572 if (!is_page_cache_freeable(page))
573 return PAGE_KEEP;
574 if (!mapping) {
575 /*
576 * Some data journaling orphaned pages can have
577 * page->mapping == NULL while being dirty with clean buffers.
578 */
579 if (page_has_private(page)) {
580 if (try_to_free_buffers(page)) {
581 ClearPageDirty(page);
582 pr_info("%s: orphaned page\n", __func__);
583 return PAGE_CLEAN;
584 }
585 }
586 return PAGE_KEEP;
587 }
588 if (mapping->a_ops->writepage == NULL)
589 return PAGE_ACTIVATE;
590 if (!may_write_to_inode(mapping->host, sc))
591 return PAGE_KEEP;
592
593 if (clear_page_dirty_for_io(page)) {
594 int res;
595 struct writeback_control wbc = {
596 .sync_mode = WB_SYNC_NONE,
597 .nr_to_write = SWAP_CLUSTER_MAX,
598 .range_start = 0,
599 .range_end = LLONG_MAX,
600 .for_reclaim = 1,
601 };
602
603 SetPageReclaim(page);
604 res = mapping->a_ops->writepage(page, &wbc);
605 if (res < 0)
606 handle_write_error(mapping, page, res);
607 if (res == AOP_WRITEPAGE_ACTIVATE) {
608 ClearPageReclaim(page);
609 return PAGE_ACTIVATE;
610 }
611
612 if (!PageWriteback(page)) {
613 /* synchronous write or broken a_ops? */
614 ClearPageReclaim(page);
615 }
616 trace_mm_vmscan_writepage(page);
617 inc_node_page_state(page, NR_VMSCAN_WRITE);
618 return PAGE_SUCCESS;
619 }
620
621 return PAGE_CLEAN;
622 }
623
624 /*
625 * Same as remove_mapping, but if the page is removed from the mapping, it
626 * gets returned with a refcount of 0.
627 */
628 static int __remove_mapping(struct address_space *mapping, struct page *page,
629 bool reclaimed)
630 {
631 unsigned long flags;
632
633 BUG_ON(!PageLocked(page));
634 BUG_ON(mapping != page_mapping(page));
635
636 spin_lock_irqsave(&mapping->tree_lock, flags);
637 /*
638 * The non racy check for a busy page.
639 *
640 * Must be careful with the order of the tests. When someone has
641 * a ref to the page, it may be possible that they dirty it then
642 * drop the reference. So if PageDirty is tested before page_count
643 * here, then the following race may occur:
644 *
645 * get_user_pages(&page);
646 * [user mapping goes away]
647 * write_to(page);
648 * !PageDirty(page) [good]
649 * SetPageDirty(page);
650 * put_page(page);
651 * !page_count(page) [good, discard it]
652 *
653 * [oops, our write_to data is lost]
654 *
655 * Reversing the order of the tests ensures such a situation cannot
656 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
657 * load is not satisfied before that of page->_refcount.
658 *
659 * Note that if SetPageDirty is always performed via set_page_dirty,
660 * and thus under tree_lock, then this ordering is not required.
661 */
662 if (!page_ref_freeze(page, 2))
663 goto cannot_free;
664 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
665 if (unlikely(PageDirty(page))) {
666 page_ref_unfreeze(page, 2);
667 goto cannot_free;
668 }
669
670 if (PageSwapCache(page)) {
671 swp_entry_t swap = { .val = page_private(page) };
672 mem_cgroup_swapout(page, swap);
673 __delete_from_swap_cache(page);
674 spin_unlock_irqrestore(&mapping->tree_lock, flags);
675 swapcache_free(swap);
676 } else {
677 void (*freepage)(struct page *);
678 void *shadow = NULL;
679
680 freepage = mapping->a_ops->freepage;
681 /*
682 * Remember a shadow entry for reclaimed file cache in
683 * order to detect refaults, thus thrashing, later on.
684 *
685 * But don't store shadows in an address space that is
686 * already exiting. This is not just an optizimation,
687 * inode reclaim needs to empty out the radix tree or
688 * the nodes are lost. Don't plant shadows behind its
689 * back.
690 *
691 * We also don't store shadows for DAX mappings because the
692 * only page cache pages found in these are zero pages
693 * covering holes, and because we don't want to mix DAX
694 * exceptional entries and shadow exceptional entries in the
695 * same page_tree.
696 */
697 if (reclaimed && page_is_file_cache(page) &&
698 !mapping_exiting(mapping) && !dax_mapping(mapping))
699 shadow = workingset_eviction(mapping, page);
700 __delete_from_page_cache(page, shadow);
701 spin_unlock_irqrestore(&mapping->tree_lock, flags);
702
703 if (freepage != NULL)
704 freepage(page);
705 }
706
707 return 1;
708
709 cannot_free:
710 spin_unlock_irqrestore(&mapping->tree_lock, flags);
711 return 0;
712 }
713
714 /*
715 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
716 * someone else has a ref on the page, abort and return 0. If it was
717 * successfully detached, return 1. Assumes the caller has a single ref on
718 * this page.
719 */
720 int remove_mapping(struct address_space *mapping, struct page *page)
721 {
722 if (__remove_mapping(mapping, page, false)) {
723 /*
724 * Unfreezing the refcount with 1 rather than 2 effectively
725 * drops the pagecache ref for us without requiring another
726 * atomic operation.
727 */
728 page_ref_unfreeze(page, 1);
729 return 1;
730 }
731 return 0;
732 }
733
734 /**
735 * putback_lru_page - put previously isolated page onto appropriate LRU list
736 * @page: page to be put back to appropriate lru list
737 *
738 * Add previously isolated @page to appropriate LRU list.
739 * Page may still be unevictable for other reasons.
740 *
741 * lru_lock must not be held, interrupts must be enabled.
742 */
743 void putback_lru_page(struct page *page)
744 {
745 bool is_unevictable;
746 int was_unevictable = PageUnevictable(page);
747
748 VM_BUG_ON_PAGE(PageLRU(page), page);
749
750 redo:
751 ClearPageUnevictable(page);
752
753 if (page_evictable(page)) {
754 /*
755 * For evictable pages, we can use the cache.
756 * In event of a race, worst case is we end up with an
757 * unevictable page on [in]active list.
758 * We know how to handle that.
759 */
760 is_unevictable = false;
761 lru_cache_add(page);
762 } else {
763 /*
764 * Put unevictable pages directly on zone's unevictable
765 * list.
766 */
767 is_unevictable = true;
768 add_page_to_unevictable_list(page);
769 /*
770 * When racing with an mlock or AS_UNEVICTABLE clearing
771 * (page is unlocked) make sure that if the other thread
772 * does not observe our setting of PG_lru and fails
773 * isolation/check_move_unevictable_pages,
774 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
775 * the page back to the evictable list.
776 *
777 * The other side is TestClearPageMlocked() or shmem_lock().
778 */
779 smp_mb();
780 }
781
782 /*
783 * page's status can change while we move it among lru. If an evictable
784 * page is on unevictable list, it never be freed. To avoid that,
785 * check after we added it to the list, again.
786 */
787 if (is_unevictable && page_evictable(page)) {
788 if (!isolate_lru_page(page)) {
789 put_page(page);
790 goto redo;
791 }
792 /* This means someone else dropped this page from LRU
793 * So, it will be freed or putback to LRU again. There is
794 * nothing to do here.
795 */
796 }
797
798 if (was_unevictable && !is_unevictable)
799 count_vm_event(UNEVICTABLE_PGRESCUED);
800 else if (!was_unevictable && is_unevictable)
801 count_vm_event(UNEVICTABLE_PGCULLED);
802
803 put_page(page); /* drop ref from isolate */
804 }
805
806 enum page_references {
807 PAGEREF_RECLAIM,
808 PAGEREF_RECLAIM_CLEAN,
809 PAGEREF_KEEP,
810 PAGEREF_ACTIVATE,
811 };
812
813 static enum page_references page_check_references(struct page *page,
814 struct scan_control *sc)
815 {
816 int referenced_ptes, referenced_page;
817 unsigned long vm_flags;
818
819 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
820 &vm_flags);
821 referenced_page = TestClearPageReferenced(page);
822
823 /*
824 * Mlock lost the isolation race with us. Let try_to_unmap()
825 * move the page to the unevictable list.
826 */
827 if (vm_flags & VM_LOCKED)
828 return PAGEREF_RECLAIM;
829
830 if (referenced_ptes) {
831 if (PageSwapBacked(page))
832 return PAGEREF_ACTIVATE;
833 /*
834 * All mapped pages start out with page table
835 * references from the instantiating fault, so we need
836 * to look twice if a mapped file page is used more
837 * than once.
838 *
839 * Mark it and spare it for another trip around the
840 * inactive list. Another page table reference will
841 * lead to its activation.
842 *
843 * Note: the mark is set for activated pages as well
844 * so that recently deactivated but used pages are
845 * quickly recovered.
846 */
847 SetPageReferenced(page);
848
849 if (referenced_page || referenced_ptes > 1)
850 return PAGEREF_ACTIVATE;
851
852 /*
853 * Activate file-backed executable pages after first usage.
854 */
855 if (vm_flags & VM_EXEC)
856 return PAGEREF_ACTIVATE;
857
858 return PAGEREF_KEEP;
859 }
860
861 /* Reclaim if clean, defer dirty pages to writeback */
862 if (referenced_page && !PageSwapBacked(page))
863 return PAGEREF_RECLAIM_CLEAN;
864
865 return PAGEREF_RECLAIM;
866 }
867
868 /* Check if a page is dirty or under writeback */
869 static void page_check_dirty_writeback(struct page *page,
870 bool *dirty, bool *writeback)
871 {
872 struct address_space *mapping;
873
874 /*
875 * Anonymous pages are not handled by flushers and must be written
876 * from reclaim context. Do not stall reclaim based on them
877 */
878 if (!page_is_file_cache(page)) {
879 *dirty = false;
880 *writeback = false;
881 return;
882 }
883
884 /* By default assume that the page flags are accurate */
885 *dirty = PageDirty(page);
886 *writeback = PageWriteback(page);
887
888 /* Verify dirty/writeback state if the filesystem supports it */
889 if (!page_has_private(page))
890 return;
891
892 mapping = page_mapping(page);
893 if (mapping && mapping->a_ops->is_dirty_writeback)
894 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
895 }
896
897 /*
898 * shrink_page_list() returns the number of reclaimed pages
899 */
900 static unsigned long shrink_page_list(struct list_head *page_list,
901 struct pglist_data *pgdat,
902 struct scan_control *sc,
903 enum ttu_flags ttu_flags,
904 unsigned long *ret_nr_dirty,
905 unsigned long *ret_nr_unqueued_dirty,
906 unsigned long *ret_nr_congested,
907 unsigned long *ret_nr_writeback,
908 unsigned long *ret_nr_immediate,
909 bool force_reclaim)
910 {
911 LIST_HEAD(ret_pages);
912 LIST_HEAD(free_pages);
913 int pgactivate = 0;
914 unsigned long nr_unqueued_dirty = 0;
915 unsigned long nr_dirty = 0;
916 unsigned long nr_congested = 0;
917 unsigned long nr_reclaimed = 0;
918 unsigned long nr_writeback = 0;
919 unsigned long nr_immediate = 0;
920
921 cond_resched();
922
923 while (!list_empty(page_list)) {
924 struct address_space *mapping;
925 struct page *page;
926 int may_enter_fs;
927 enum page_references references = PAGEREF_RECLAIM_CLEAN;
928 bool dirty, writeback;
929 bool lazyfree = false;
930 int ret = SWAP_SUCCESS;
931
932 cond_resched();
933
934 page = lru_to_page(page_list);
935 list_del(&page->lru);
936
937 if (!trylock_page(page))
938 goto keep;
939
940 VM_BUG_ON_PAGE(PageActive(page), page);
941
942 sc->nr_scanned++;
943
944 if (unlikely(!page_evictable(page)))
945 goto cull_mlocked;
946
947 if (!sc->may_unmap && page_mapped(page))
948 goto keep_locked;
949
950 /* Double the slab pressure for mapped and swapcache pages */
951 if (page_mapped(page) || PageSwapCache(page))
952 sc->nr_scanned++;
953
954 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
955 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
956
957 /*
958 * The number of dirty pages determines if a zone is marked
959 * reclaim_congested which affects wait_iff_congested. kswapd
960 * will stall and start writing pages if the tail of the LRU
961 * is all dirty unqueued pages.
962 */
963 page_check_dirty_writeback(page, &dirty, &writeback);
964 if (dirty || writeback)
965 nr_dirty++;
966
967 if (dirty && !writeback)
968 nr_unqueued_dirty++;
969
970 /*
971 * Treat this page as congested if the underlying BDI is or if
972 * pages are cycling through the LRU so quickly that the
973 * pages marked for immediate reclaim are making it to the
974 * end of the LRU a second time.
975 */
976 mapping = page_mapping(page);
977 if (((dirty || writeback) && mapping &&
978 inode_write_congested(mapping->host)) ||
979 (writeback && PageReclaim(page)))
980 nr_congested++;
981
982 /*
983 * If a page at the tail of the LRU is under writeback, there
984 * are three cases to consider.
985 *
986 * 1) If reclaim is encountering an excessive number of pages
987 * under writeback and this page is both under writeback and
988 * PageReclaim then it indicates that pages are being queued
989 * for IO but are being recycled through the LRU before the
990 * IO can complete. Waiting on the page itself risks an
991 * indefinite stall if it is impossible to writeback the
992 * page due to IO error or disconnected storage so instead
993 * note that the LRU is being scanned too quickly and the
994 * caller can stall after page list has been processed.
995 *
996 * 2) Global or new memcg reclaim encounters a page that is
997 * not marked for immediate reclaim, or the caller does not
998 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
999 * not to fs). In this case mark the page for immediate
1000 * reclaim and continue scanning.
1001 *
1002 * Require may_enter_fs because we would wait on fs, which
1003 * may not have submitted IO yet. And the loop driver might
1004 * enter reclaim, and deadlock if it waits on a page for
1005 * which it is needed to do the write (loop masks off
1006 * __GFP_IO|__GFP_FS for this reason); but more thought
1007 * would probably show more reasons.
1008 *
1009 * 3) Legacy memcg encounters a page that is already marked
1010 * PageReclaim. memcg does not have any dirty pages
1011 * throttling so we could easily OOM just because too many
1012 * pages are in writeback and there is nothing else to
1013 * reclaim. Wait for the writeback to complete.
1014 */
1015 if (PageWriteback(page)) {
1016 /* Case 1 above */
1017 if (current_is_kswapd() &&
1018 PageReclaim(page) &&
1019 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1020 nr_immediate++;
1021 goto keep_locked;
1022
1023 /* Case 2 above */
1024 } else if (sane_reclaim(sc) ||
1025 !PageReclaim(page) || !may_enter_fs) {
1026 /*
1027 * This is slightly racy - end_page_writeback()
1028 * might have just cleared PageReclaim, then
1029 * setting PageReclaim here end up interpreted
1030 * as PageReadahead - but that does not matter
1031 * enough to care. What we do want is for this
1032 * page to have PageReclaim set next time memcg
1033 * reclaim reaches the tests above, so it will
1034 * then wait_on_page_writeback() to avoid OOM;
1035 * and it's also appropriate in global reclaim.
1036 */
1037 SetPageReclaim(page);
1038 nr_writeback++;
1039 goto keep_locked;
1040
1041 /* Case 3 above */
1042 } else {
1043 unlock_page(page);
1044 wait_on_page_writeback(page);
1045 /* then go back and try same page again */
1046 list_add_tail(&page->lru, page_list);
1047 continue;
1048 }
1049 }
1050
1051 if (!force_reclaim)
1052 references = page_check_references(page, sc);
1053
1054 switch (references) {
1055 case PAGEREF_ACTIVATE:
1056 goto activate_locked;
1057 case PAGEREF_KEEP:
1058 goto keep_locked;
1059 case PAGEREF_RECLAIM:
1060 case PAGEREF_RECLAIM_CLEAN:
1061 ; /* try to reclaim the page below */
1062 }
1063
1064 /*
1065 * Anonymous process memory has backing store?
1066 * Try to allocate it some swap space here.
1067 */
1068 if (PageAnon(page) && !PageSwapCache(page)) {
1069 if (!(sc->gfp_mask & __GFP_IO))
1070 goto keep_locked;
1071 if (!add_to_swap(page, page_list))
1072 goto activate_locked;
1073 lazyfree = true;
1074 may_enter_fs = 1;
1075
1076 /* Adding to swap updated mapping */
1077 mapping = page_mapping(page);
1078 } else if (unlikely(PageTransHuge(page))) {
1079 /* Split file THP */
1080 if (split_huge_page_to_list(page, page_list))
1081 goto keep_locked;
1082 }
1083
1084 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1085
1086 /*
1087 * The page is mapped into the page tables of one or more
1088 * processes. Try to unmap it here.
1089 */
1090 if (page_mapped(page) && mapping) {
1091 switch (ret = try_to_unmap(page, lazyfree ?
1092 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1093 (ttu_flags | TTU_BATCH_FLUSH))) {
1094 case SWAP_FAIL:
1095 goto activate_locked;
1096 case SWAP_AGAIN:
1097 goto keep_locked;
1098 case SWAP_MLOCK:
1099 goto cull_mlocked;
1100 case SWAP_LZFREE:
1101 goto lazyfree;
1102 case SWAP_SUCCESS:
1103 ; /* try to free the page below */
1104 }
1105 }
1106
1107 if (PageDirty(page)) {
1108 /*
1109 * Only kswapd can writeback filesystem pages to
1110 * avoid risk of stack overflow but only writeback
1111 * if many dirty pages have been encountered.
1112 */
1113 if (page_is_file_cache(page) &&
1114 (!current_is_kswapd() ||
1115 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1116 /*
1117 * Immediately reclaim when written back.
1118 * Similar in principal to deactivate_page()
1119 * except we already have the page isolated
1120 * and know it's dirty
1121 */
1122 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1123 SetPageReclaim(page);
1124
1125 goto keep_locked;
1126 }
1127
1128 if (references == PAGEREF_RECLAIM_CLEAN)
1129 goto keep_locked;
1130 if (!may_enter_fs)
1131 goto keep_locked;
1132 if (!sc->may_writepage)
1133 goto keep_locked;
1134
1135 /*
1136 * Page is dirty. Flush the TLB if a writable entry
1137 * potentially exists to avoid CPU writes after IO
1138 * starts and then write it out here.
1139 */
1140 try_to_unmap_flush_dirty();
1141 switch (pageout(page, mapping, sc)) {
1142 case PAGE_KEEP:
1143 goto keep_locked;
1144 case PAGE_ACTIVATE:
1145 goto activate_locked;
1146 case PAGE_SUCCESS:
1147 if (PageWriteback(page))
1148 goto keep;
1149 if (PageDirty(page))
1150 goto keep;
1151
1152 /*
1153 * A synchronous write - probably a ramdisk. Go
1154 * ahead and try to reclaim the page.
1155 */
1156 if (!trylock_page(page))
1157 goto keep;
1158 if (PageDirty(page) || PageWriteback(page))
1159 goto keep_locked;
1160 mapping = page_mapping(page);
1161 case PAGE_CLEAN:
1162 ; /* try to free the page below */
1163 }
1164 }
1165
1166 /*
1167 * If the page has buffers, try to free the buffer mappings
1168 * associated with this page. If we succeed we try to free
1169 * the page as well.
1170 *
1171 * We do this even if the page is PageDirty().
1172 * try_to_release_page() does not perform I/O, but it is
1173 * possible for a page to have PageDirty set, but it is actually
1174 * clean (all its buffers are clean). This happens if the
1175 * buffers were written out directly, with submit_bh(). ext3
1176 * will do this, as well as the blockdev mapping.
1177 * try_to_release_page() will discover that cleanness and will
1178 * drop the buffers and mark the page clean - it can be freed.
1179 *
1180 * Rarely, pages can have buffers and no ->mapping. These are
1181 * the pages which were not successfully invalidated in
1182 * truncate_complete_page(). We try to drop those buffers here
1183 * and if that worked, and the page is no longer mapped into
1184 * process address space (page_count == 1) it can be freed.
1185 * Otherwise, leave the page on the LRU so it is swappable.
1186 */
1187 if (page_has_private(page)) {
1188 if (!try_to_release_page(page, sc->gfp_mask))
1189 goto activate_locked;
1190 if (!mapping && page_count(page) == 1) {
1191 unlock_page(page);
1192 if (put_page_testzero(page))
1193 goto free_it;
1194 else {
1195 /*
1196 * rare race with speculative reference.
1197 * the speculative reference will free
1198 * this page shortly, so we may
1199 * increment nr_reclaimed here (and
1200 * leave it off the LRU).
1201 */
1202 nr_reclaimed++;
1203 continue;
1204 }
1205 }
1206 }
1207
1208 lazyfree:
1209 if (!mapping || !__remove_mapping(mapping, page, true))
1210 goto keep_locked;
1211
1212 /*
1213 * At this point, we have no other references and there is
1214 * no way to pick any more up (removed from LRU, removed
1215 * from pagecache). Can use non-atomic bitops now (and
1216 * we obviously don't have to worry about waking up a process
1217 * waiting on the page lock, because there are no references.
1218 */
1219 __ClearPageLocked(page);
1220 free_it:
1221 if (ret == SWAP_LZFREE)
1222 count_vm_event(PGLAZYFREED);
1223
1224 nr_reclaimed++;
1225
1226 /*
1227 * Is there need to periodically free_page_list? It would
1228 * appear not as the counts should be low
1229 */
1230 list_add(&page->lru, &free_pages);
1231 continue;
1232
1233 cull_mlocked:
1234 if (PageSwapCache(page))
1235 try_to_free_swap(page);
1236 unlock_page(page);
1237 list_add(&page->lru, &ret_pages);
1238 continue;
1239
1240 activate_locked:
1241 /* Not a candidate for swapping, so reclaim swap space. */
1242 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1243 try_to_free_swap(page);
1244 VM_BUG_ON_PAGE(PageActive(page), page);
1245 SetPageActive(page);
1246 pgactivate++;
1247 keep_locked:
1248 unlock_page(page);
1249 keep:
1250 list_add(&page->lru, &ret_pages);
1251 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1252 }
1253
1254 mem_cgroup_uncharge_list(&free_pages);
1255 try_to_unmap_flush();
1256 free_hot_cold_page_list(&free_pages, true);
1257
1258 list_splice(&ret_pages, page_list);
1259 count_vm_events(PGACTIVATE, pgactivate);
1260
1261 *ret_nr_dirty += nr_dirty;
1262 *ret_nr_congested += nr_congested;
1263 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1264 *ret_nr_writeback += nr_writeback;
1265 *ret_nr_immediate += nr_immediate;
1266 return nr_reclaimed;
1267 }
1268
1269 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1270 struct list_head *page_list)
1271 {
1272 struct scan_control sc = {
1273 .gfp_mask = GFP_KERNEL,
1274 .priority = DEF_PRIORITY,
1275 .may_unmap = 1,
1276 };
1277 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1278 struct page *page, *next;
1279 LIST_HEAD(clean_pages);
1280
1281 list_for_each_entry_safe(page, next, page_list, lru) {
1282 if (page_is_file_cache(page) && !PageDirty(page) &&
1283 !__PageMovable(page)) {
1284 ClearPageActive(page);
1285 list_move(&page->lru, &clean_pages);
1286 }
1287 }
1288
1289 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1290 TTU_UNMAP|TTU_IGNORE_ACCESS,
1291 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1292 list_splice(&clean_pages, page_list);
1293 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1294 return ret;
1295 }
1296
1297 /*
1298 * Attempt to remove the specified page from its LRU. Only take this page
1299 * if it is of the appropriate PageActive status. Pages which are being
1300 * freed elsewhere are also ignored.
1301 *
1302 * page: page to consider
1303 * mode: one of the LRU isolation modes defined above
1304 *
1305 * returns 0 on success, -ve errno on failure.
1306 */
1307 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1308 {
1309 int ret = -EINVAL;
1310
1311 /* Only take pages on the LRU. */
1312 if (!PageLRU(page))
1313 return ret;
1314
1315 /* Compaction should not handle unevictable pages but CMA can do so */
1316 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1317 return ret;
1318
1319 ret = -EBUSY;
1320
1321 /*
1322 * To minimise LRU disruption, the caller can indicate that it only
1323 * wants to isolate pages it will be able to operate on without
1324 * blocking - clean pages for the most part.
1325 *
1326 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1327 * is used by reclaim when it is cannot write to backing storage
1328 *
1329 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1330 * that it is possible to migrate without blocking
1331 */
1332 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1333 /* All the caller can do on PageWriteback is block */
1334 if (PageWriteback(page))
1335 return ret;
1336
1337 if (PageDirty(page)) {
1338 struct address_space *mapping;
1339
1340 /* ISOLATE_CLEAN means only clean pages */
1341 if (mode & ISOLATE_CLEAN)
1342 return ret;
1343
1344 /*
1345 * Only pages without mappings or that have a
1346 * ->migratepage callback are possible to migrate
1347 * without blocking
1348 */
1349 mapping = page_mapping(page);
1350 if (mapping && !mapping->a_ops->migratepage)
1351 return ret;
1352 }
1353 }
1354
1355 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1356 return ret;
1357
1358 if (likely(get_page_unless_zero(page))) {
1359 /*
1360 * Be careful not to clear PageLRU until after we're
1361 * sure the page is not being freed elsewhere -- the
1362 * page release code relies on it.
1363 */
1364 ClearPageLRU(page);
1365 ret = 0;
1366 }
1367
1368 return ret;
1369 }
1370
1371
1372 /*
1373 * Update LRU sizes after isolating pages. The LRU size updates must
1374 * be complete before mem_cgroup_update_lru_size due to a santity check.
1375 */
1376 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1377 enum lru_list lru, unsigned long *nr_zone_taken,
1378 unsigned long nr_taken)
1379 {
1380 int zid;
1381
1382 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1383 if (!nr_zone_taken[zid])
1384 continue;
1385
1386 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1387 }
1388
1389 #ifdef CONFIG_MEMCG
1390 mem_cgroup_update_lru_size(lruvec, lru, -nr_taken);
1391 #endif
1392 }
1393
1394 /*
1395 * zone_lru_lock is heavily contended. Some of the functions that
1396 * shrink the lists perform better by taking out a batch of pages
1397 * and working on them outside the LRU lock.
1398 *
1399 * For pagecache intensive workloads, this function is the hottest
1400 * spot in the kernel (apart from copy_*_user functions).
1401 *
1402 * Appropriate locks must be held before calling this function.
1403 *
1404 * @nr_to_scan: The number of pages to look through on the list.
1405 * @lruvec: The LRU vector to pull pages from.
1406 * @dst: The temp list to put pages on to.
1407 * @nr_scanned: The number of pages that were scanned.
1408 * @sc: The scan_control struct for this reclaim session
1409 * @mode: One of the LRU isolation modes
1410 * @lru: LRU list id for isolating
1411 *
1412 * returns how many pages were moved onto *@dst.
1413 */
1414 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1415 struct lruvec *lruvec, struct list_head *dst,
1416 unsigned long *nr_scanned, struct scan_control *sc,
1417 isolate_mode_t mode, enum lru_list lru)
1418 {
1419 struct list_head *src = &lruvec->lists[lru];
1420 unsigned long nr_taken = 0;
1421 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1422 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1423 unsigned long scan, nr_pages;
1424 LIST_HEAD(pages_skipped);
1425
1426 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1427 !list_empty(src);) {
1428 struct page *page;
1429
1430 page = lru_to_page(src);
1431 prefetchw_prev_lru_page(page, src, flags);
1432
1433 VM_BUG_ON_PAGE(!PageLRU(page), page);
1434
1435 if (page_zonenum(page) > sc->reclaim_idx) {
1436 list_move(&page->lru, &pages_skipped);
1437 nr_skipped[page_zonenum(page)]++;
1438 continue;
1439 }
1440
1441 /*
1442 * Account for scanned and skipped separetly to avoid the pgdat
1443 * being prematurely marked unreclaimable by pgdat_reclaimable.
1444 */
1445 scan++;
1446
1447 switch (__isolate_lru_page(page, mode)) {
1448 case 0:
1449 nr_pages = hpage_nr_pages(page);
1450 nr_taken += nr_pages;
1451 nr_zone_taken[page_zonenum(page)] += nr_pages;
1452 list_move(&page->lru, dst);
1453 break;
1454
1455 case -EBUSY:
1456 /* else it is being freed elsewhere */
1457 list_move(&page->lru, src);
1458 continue;
1459
1460 default:
1461 BUG();
1462 }
1463 }
1464
1465 /*
1466 * Splice any skipped pages to the start of the LRU list. Note that
1467 * this disrupts the LRU order when reclaiming for lower zones but
1468 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1469 * scanning would soon rescan the same pages to skip and put the
1470 * system at risk of premature OOM.
1471 */
1472 if (!list_empty(&pages_skipped)) {
1473 int zid;
1474 unsigned long total_skipped = 0;
1475
1476 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1477 if (!nr_skipped[zid])
1478 continue;
1479
1480 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1481 total_skipped += nr_skipped[zid];
1482 }
1483
1484 /*
1485 * Account skipped pages as a partial scan as the pgdat may be
1486 * close to unreclaimable. If the LRU list is empty, account
1487 * skipped pages as a full scan.
1488 */
1489 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1490
1491 list_splice(&pages_skipped, src);
1492 }
1493 *nr_scanned = scan;
1494 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
1495 nr_taken, mode, is_file_lru(lru));
1496 update_lru_sizes(lruvec, lru, nr_zone_taken, nr_taken);
1497 return nr_taken;
1498 }
1499
1500 /**
1501 * isolate_lru_page - tries to isolate a page from its LRU list
1502 * @page: page to isolate from its LRU list
1503 *
1504 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1505 * vmstat statistic corresponding to whatever LRU list the page was on.
1506 *
1507 * Returns 0 if the page was removed from an LRU list.
1508 * Returns -EBUSY if the page was not on an LRU list.
1509 *
1510 * The returned page will have PageLRU() cleared. If it was found on
1511 * the active list, it will have PageActive set. If it was found on
1512 * the unevictable list, it will have the PageUnevictable bit set. That flag
1513 * may need to be cleared by the caller before letting the page go.
1514 *
1515 * The vmstat statistic corresponding to the list on which the page was
1516 * found will be decremented.
1517 *
1518 * Restrictions:
1519 * (1) Must be called with an elevated refcount on the page. This is a
1520 * fundamentnal difference from isolate_lru_pages (which is called
1521 * without a stable reference).
1522 * (2) the lru_lock must not be held.
1523 * (3) interrupts must be enabled.
1524 */
1525 int isolate_lru_page(struct page *page)
1526 {
1527 int ret = -EBUSY;
1528
1529 VM_BUG_ON_PAGE(!page_count(page), page);
1530 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1531
1532 if (PageLRU(page)) {
1533 struct zone *zone = page_zone(page);
1534 struct lruvec *lruvec;
1535
1536 spin_lock_irq(zone_lru_lock(zone));
1537 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1538 if (PageLRU(page)) {
1539 int lru = page_lru(page);
1540 get_page(page);
1541 ClearPageLRU(page);
1542 del_page_from_lru_list(page, lruvec, lru);
1543 ret = 0;
1544 }
1545 spin_unlock_irq(zone_lru_lock(zone));
1546 }
1547 return ret;
1548 }
1549
1550 /*
1551 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1552 * then get resheduled. When there are massive number of tasks doing page
1553 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1554 * the LRU list will go small and be scanned faster than necessary, leading to
1555 * unnecessary swapping, thrashing and OOM.
1556 */
1557 static int too_many_isolated(struct pglist_data *pgdat, int file,
1558 struct scan_control *sc)
1559 {
1560 unsigned long inactive, isolated;
1561
1562 if (current_is_kswapd())
1563 return 0;
1564
1565 if (!sane_reclaim(sc))
1566 return 0;
1567
1568 if (file) {
1569 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1570 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1571 } else {
1572 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1573 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1574 }
1575
1576 /*
1577 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1578 * won't get blocked by normal direct-reclaimers, forming a circular
1579 * deadlock.
1580 */
1581 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1582 inactive >>= 3;
1583
1584 return isolated > inactive;
1585 }
1586
1587 static noinline_for_stack void
1588 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1589 {
1590 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1591 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1592 LIST_HEAD(pages_to_free);
1593
1594 /*
1595 * Put back any unfreeable pages.
1596 */
1597 while (!list_empty(page_list)) {
1598 struct page *page = lru_to_page(page_list);
1599 int lru;
1600
1601 VM_BUG_ON_PAGE(PageLRU(page), page);
1602 list_del(&page->lru);
1603 if (unlikely(!page_evictable(page))) {
1604 spin_unlock_irq(&pgdat->lru_lock);
1605 putback_lru_page(page);
1606 spin_lock_irq(&pgdat->lru_lock);
1607 continue;
1608 }
1609
1610 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1611
1612 SetPageLRU(page);
1613 lru = page_lru(page);
1614 add_page_to_lru_list(page, lruvec, lru);
1615
1616 if (is_active_lru(lru)) {
1617 int file = is_file_lru(lru);
1618 int numpages = hpage_nr_pages(page);
1619 reclaim_stat->recent_rotated[file] += numpages;
1620 }
1621 if (put_page_testzero(page)) {
1622 __ClearPageLRU(page);
1623 __ClearPageActive(page);
1624 del_page_from_lru_list(page, lruvec, lru);
1625
1626 if (unlikely(PageCompound(page))) {
1627 spin_unlock_irq(&pgdat->lru_lock);
1628 mem_cgroup_uncharge(page);
1629 (*get_compound_page_dtor(page))(page);
1630 spin_lock_irq(&pgdat->lru_lock);
1631 } else
1632 list_add(&page->lru, &pages_to_free);
1633 }
1634 }
1635
1636 /*
1637 * To save our caller's stack, now use input list for pages to free.
1638 */
1639 list_splice(&pages_to_free, page_list);
1640 }
1641
1642 /*
1643 * If a kernel thread (such as nfsd for loop-back mounts) services
1644 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1645 * In that case we should only throttle if the backing device it is
1646 * writing to is congested. In other cases it is safe to throttle.
1647 */
1648 static int current_may_throttle(void)
1649 {
1650 return !(current->flags & PF_LESS_THROTTLE) ||
1651 current->backing_dev_info == NULL ||
1652 bdi_write_congested(current->backing_dev_info);
1653 }
1654
1655 static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1656 struct scan_control *sc, enum lru_list lru)
1657 {
1658 int zid;
1659 struct zone *zone;
1660 int file = is_file_lru(lru);
1661 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1662
1663 if (!global_reclaim(sc))
1664 return true;
1665
1666 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1667 zone = &pgdat->node_zones[zid];
1668 if (!populated_zone(zone))
1669 continue;
1670
1671 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1672 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1673 return true;
1674 }
1675
1676 return false;
1677 }
1678
1679 /*
1680 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1681 * of reclaimed pages
1682 */
1683 static noinline_for_stack unsigned long
1684 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1685 struct scan_control *sc, enum lru_list lru)
1686 {
1687 LIST_HEAD(page_list);
1688 unsigned long nr_scanned;
1689 unsigned long nr_reclaimed = 0;
1690 unsigned long nr_taken;
1691 unsigned long nr_dirty = 0;
1692 unsigned long nr_congested = 0;
1693 unsigned long nr_unqueued_dirty = 0;
1694 unsigned long nr_writeback = 0;
1695 unsigned long nr_immediate = 0;
1696 isolate_mode_t isolate_mode = 0;
1697 int file = is_file_lru(lru);
1698 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1699 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1700
1701 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1702 return 0;
1703
1704 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1705 congestion_wait(BLK_RW_ASYNC, HZ/10);
1706
1707 /* We are about to die and free our memory. Return now. */
1708 if (fatal_signal_pending(current))
1709 return SWAP_CLUSTER_MAX;
1710 }
1711
1712 lru_add_drain();
1713
1714 if (!sc->may_unmap)
1715 isolate_mode |= ISOLATE_UNMAPPED;
1716 if (!sc->may_writepage)
1717 isolate_mode |= ISOLATE_CLEAN;
1718
1719 spin_lock_irq(&pgdat->lru_lock);
1720
1721 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1722 &nr_scanned, sc, isolate_mode, lru);
1723
1724 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1725 reclaim_stat->recent_scanned[file] += nr_taken;
1726
1727 if (global_reclaim(sc)) {
1728 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1729 if (current_is_kswapd())
1730 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1731 else
1732 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1733 }
1734 spin_unlock_irq(&pgdat->lru_lock);
1735
1736 if (nr_taken == 0)
1737 return 0;
1738
1739 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1740 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1741 &nr_writeback, &nr_immediate,
1742 false);
1743
1744 spin_lock_irq(&pgdat->lru_lock);
1745
1746 if (global_reclaim(sc)) {
1747 if (current_is_kswapd())
1748 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1749 else
1750 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1751 }
1752
1753 putback_inactive_pages(lruvec, &page_list);
1754
1755 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1756
1757 spin_unlock_irq(&pgdat->lru_lock);
1758
1759 mem_cgroup_uncharge_list(&page_list);
1760 free_hot_cold_page_list(&page_list, true);
1761
1762 /*
1763 * If reclaim is isolating dirty pages under writeback, it implies
1764 * that the long-lived page allocation rate is exceeding the page
1765 * laundering rate. Either the global limits are not being effective
1766 * at throttling processes due to the page distribution throughout
1767 * zones or there is heavy usage of a slow backing device. The
1768 * only option is to throttle from reclaim context which is not ideal
1769 * as there is no guarantee the dirtying process is throttled in the
1770 * same way balance_dirty_pages() manages.
1771 *
1772 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1773 * of pages under pages flagged for immediate reclaim and stall if any
1774 * are encountered in the nr_immediate check below.
1775 */
1776 if (nr_writeback && nr_writeback == nr_taken)
1777 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1778
1779 /*
1780 * Legacy memcg will stall in page writeback so avoid forcibly
1781 * stalling here.
1782 */
1783 if (sane_reclaim(sc)) {
1784 /*
1785 * Tag a zone as congested if all the dirty pages scanned were
1786 * backed by a congested BDI and wait_iff_congested will stall.
1787 */
1788 if (nr_dirty && nr_dirty == nr_congested)
1789 set_bit(PGDAT_CONGESTED, &pgdat->flags);
1790
1791 /*
1792 * If dirty pages are scanned that are not queued for IO, it
1793 * implies that flushers are not keeping up. In this case, flag
1794 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1795 * reclaim context.
1796 */
1797 if (nr_unqueued_dirty == nr_taken)
1798 set_bit(PGDAT_DIRTY, &pgdat->flags);
1799
1800 /*
1801 * If kswapd scans pages marked marked for immediate
1802 * reclaim and under writeback (nr_immediate), it implies
1803 * that pages are cycling through the LRU faster than
1804 * they are written so also forcibly stall.
1805 */
1806 if (nr_immediate && current_may_throttle())
1807 congestion_wait(BLK_RW_ASYNC, HZ/10);
1808 }
1809
1810 /*
1811 * Stall direct reclaim for IO completions if underlying BDIs or zone
1812 * is congested. Allow kswapd to continue until it starts encountering
1813 * unqueued dirty pages or cycling through the LRU too quickly.
1814 */
1815 if (!sc->hibernation_mode && !current_is_kswapd() &&
1816 current_may_throttle())
1817 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1818
1819 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1820 nr_scanned, nr_reclaimed,
1821 sc->priority, file);
1822 return nr_reclaimed;
1823 }
1824
1825 /*
1826 * This moves pages from the active list to the inactive list.
1827 *
1828 * We move them the other way if the page is referenced by one or more
1829 * processes, from rmap.
1830 *
1831 * If the pages are mostly unmapped, the processing is fast and it is
1832 * appropriate to hold zone_lru_lock across the whole operation. But if
1833 * the pages are mapped, the processing is slow (page_referenced()) so we
1834 * should drop zone_lru_lock around each page. It's impossible to balance
1835 * this, so instead we remove the pages from the LRU while processing them.
1836 * It is safe to rely on PG_active against the non-LRU pages in here because
1837 * nobody will play with that bit on a non-LRU page.
1838 *
1839 * The downside is that we have to touch page->_refcount against each page.
1840 * But we had to alter page->flags anyway.
1841 */
1842
1843 static void move_active_pages_to_lru(struct lruvec *lruvec,
1844 struct list_head *list,
1845 struct list_head *pages_to_free,
1846 enum lru_list lru)
1847 {
1848 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1849 unsigned long pgmoved = 0;
1850 struct page *page;
1851 int nr_pages;
1852
1853 while (!list_empty(list)) {
1854 page = lru_to_page(list);
1855 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1856
1857 VM_BUG_ON_PAGE(PageLRU(page), page);
1858 SetPageLRU(page);
1859
1860 nr_pages = hpage_nr_pages(page);
1861 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1862 list_move(&page->lru, &lruvec->lists[lru]);
1863 pgmoved += nr_pages;
1864
1865 if (put_page_testzero(page)) {
1866 __ClearPageLRU(page);
1867 __ClearPageActive(page);
1868 del_page_from_lru_list(page, lruvec, lru);
1869
1870 if (unlikely(PageCompound(page))) {
1871 spin_unlock_irq(&pgdat->lru_lock);
1872 mem_cgroup_uncharge(page);
1873 (*get_compound_page_dtor(page))(page);
1874 spin_lock_irq(&pgdat->lru_lock);
1875 } else
1876 list_add(&page->lru, pages_to_free);
1877 }
1878 }
1879
1880 if (!is_active_lru(lru))
1881 __count_vm_events(PGDEACTIVATE, pgmoved);
1882 }
1883
1884 static void shrink_active_list(unsigned long nr_to_scan,
1885 struct lruvec *lruvec,
1886 struct scan_control *sc,
1887 enum lru_list lru)
1888 {
1889 unsigned long nr_taken;
1890 unsigned long nr_scanned;
1891 unsigned long vm_flags;
1892 LIST_HEAD(l_hold); /* The pages which were snipped off */
1893 LIST_HEAD(l_active);
1894 LIST_HEAD(l_inactive);
1895 struct page *page;
1896 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1897 unsigned long nr_rotated = 0;
1898 isolate_mode_t isolate_mode = 0;
1899 int file = is_file_lru(lru);
1900 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1901
1902 lru_add_drain();
1903
1904 if (!sc->may_unmap)
1905 isolate_mode |= ISOLATE_UNMAPPED;
1906 if (!sc->may_writepage)
1907 isolate_mode |= ISOLATE_CLEAN;
1908
1909 spin_lock_irq(&pgdat->lru_lock);
1910
1911 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1912 &nr_scanned, sc, isolate_mode, lru);
1913
1914 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1915 reclaim_stat->recent_scanned[file] += nr_taken;
1916
1917 if (global_reclaim(sc))
1918 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1919 __count_vm_events(PGREFILL, nr_scanned);
1920
1921 spin_unlock_irq(&pgdat->lru_lock);
1922
1923 while (!list_empty(&l_hold)) {
1924 cond_resched();
1925 page = lru_to_page(&l_hold);
1926 list_del(&page->lru);
1927
1928 if (unlikely(!page_evictable(page))) {
1929 putback_lru_page(page);
1930 continue;
1931 }
1932
1933 if (unlikely(buffer_heads_over_limit)) {
1934 if (page_has_private(page) && trylock_page(page)) {
1935 if (page_has_private(page))
1936 try_to_release_page(page, 0);
1937 unlock_page(page);
1938 }
1939 }
1940
1941 if (page_referenced(page, 0, sc->target_mem_cgroup,
1942 &vm_flags)) {
1943 nr_rotated += hpage_nr_pages(page);
1944 /*
1945 * Identify referenced, file-backed active pages and
1946 * give them one more trip around the active list. So
1947 * that executable code get better chances to stay in
1948 * memory under moderate memory pressure. Anon pages
1949 * are not likely to be evicted by use-once streaming
1950 * IO, plus JVM can create lots of anon VM_EXEC pages,
1951 * so we ignore them here.
1952 */
1953 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1954 list_add(&page->lru, &l_active);
1955 continue;
1956 }
1957 }
1958
1959 ClearPageActive(page); /* we are de-activating */
1960 list_add(&page->lru, &l_inactive);
1961 }
1962
1963 /*
1964 * Move pages back to the lru list.
1965 */
1966 spin_lock_irq(&pgdat->lru_lock);
1967 /*
1968 * Count referenced pages from currently used mappings as rotated,
1969 * even though only some of them are actually re-activated. This
1970 * helps balance scan pressure between file and anonymous pages in
1971 * get_scan_count.
1972 */
1973 reclaim_stat->recent_rotated[file] += nr_rotated;
1974
1975 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1976 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1977 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1978 spin_unlock_irq(&pgdat->lru_lock);
1979
1980 mem_cgroup_uncharge_list(&l_hold);
1981 free_hot_cold_page_list(&l_hold, true);
1982 }
1983
1984 /*
1985 * The inactive anon list should be small enough that the VM never has
1986 * to do too much work.
1987 *
1988 * The inactive file list should be small enough to leave most memory
1989 * to the established workingset on the scan-resistant active list,
1990 * but large enough to avoid thrashing the aggregate readahead window.
1991 *
1992 * Both inactive lists should also be large enough that each inactive
1993 * page has a chance to be referenced again before it is reclaimed.
1994 *
1995 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1996 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1997 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1998 *
1999 * total target max
2000 * memory ratio inactive
2001 * -------------------------------------
2002 * 10MB 1 5MB
2003 * 100MB 1 50MB
2004 * 1GB 3 250MB
2005 * 10GB 10 0.9GB
2006 * 100GB 31 3GB
2007 * 1TB 101 10GB
2008 * 10TB 320 32GB
2009 */
2010 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2011 struct scan_control *sc)
2012 {
2013 unsigned long inactive_ratio;
2014 unsigned long inactive;
2015 unsigned long active;
2016 unsigned long gb;
2017 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2018 int zid;
2019
2020 /*
2021 * If we don't have swap space, anonymous page deactivation
2022 * is pointless.
2023 */
2024 if (!file && !total_swap_pages)
2025 return false;
2026
2027 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
2028 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
2029
2030 /*
2031 * For zone-constrained allocations, it is necessary to check if
2032 * deactivations are required for lowmem to be reclaimed. This
2033 * calculates the inactive/active pages available in eligible zones.
2034 */
2035 for (zid = sc->reclaim_idx + 1; zid < MAX_NR_ZONES; zid++) {
2036 struct zone *zone = &pgdat->node_zones[zid];
2037 unsigned long inactive_zone, active_zone;
2038
2039 if (!populated_zone(zone))
2040 continue;
2041
2042 inactive_zone = zone_page_state(zone,
2043 NR_ZONE_LRU_BASE + (file * LRU_FILE));
2044 active_zone = zone_page_state(zone,
2045 NR_ZONE_LRU_BASE + (file * LRU_FILE) + LRU_ACTIVE);
2046
2047 inactive -= min(inactive, inactive_zone);
2048 active -= min(active, active_zone);
2049 }
2050
2051 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2052 if (gb)
2053 inactive_ratio = int_sqrt(10 * gb);
2054 else
2055 inactive_ratio = 1;
2056
2057 return inactive * inactive_ratio < active;
2058 }
2059
2060 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2061 struct lruvec *lruvec, struct scan_control *sc)
2062 {
2063 if (is_active_lru(lru)) {
2064 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
2065 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2066 return 0;
2067 }
2068
2069 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2070 }
2071
2072 enum scan_balance {
2073 SCAN_EQUAL,
2074 SCAN_FRACT,
2075 SCAN_ANON,
2076 SCAN_FILE,
2077 };
2078
2079 /*
2080 * Determine how aggressively the anon and file LRU lists should be
2081 * scanned. The relative value of each set of LRU lists is determined
2082 * by looking at the fraction of the pages scanned we did rotate back
2083 * onto the active list instead of evict.
2084 *
2085 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2086 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2087 */
2088 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2089 struct scan_control *sc, unsigned long *nr,
2090 unsigned long *lru_pages)
2091 {
2092 int swappiness = mem_cgroup_swappiness(memcg);
2093 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2094 u64 fraction[2];
2095 u64 denominator = 0; /* gcc */
2096 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2097 unsigned long anon_prio, file_prio;
2098 enum scan_balance scan_balance;
2099 unsigned long anon, file;
2100 bool force_scan = false;
2101 unsigned long ap, fp;
2102 enum lru_list lru;
2103 bool some_scanned;
2104 int pass;
2105
2106 /*
2107 * If the zone or memcg is small, nr[l] can be 0. This
2108 * results in no scanning on this priority and a potential
2109 * priority drop. Global direct reclaim can go to the next
2110 * zone and tends to have no problems. Global kswapd is for
2111 * zone balancing and it needs to scan a minimum amount. When
2112 * reclaiming for a memcg, a priority drop can cause high
2113 * latencies, so it's better to scan a minimum amount there as
2114 * well.
2115 */
2116 if (current_is_kswapd()) {
2117 if (!pgdat_reclaimable(pgdat))
2118 force_scan = true;
2119 if (!mem_cgroup_online(memcg))
2120 force_scan = true;
2121 }
2122 if (!global_reclaim(sc))
2123 force_scan = true;
2124
2125 /* If we have no swap space, do not bother scanning anon pages. */
2126 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2127 scan_balance = SCAN_FILE;
2128 goto out;
2129 }
2130
2131 /*
2132 * Global reclaim will swap to prevent OOM even with no
2133 * swappiness, but memcg users want to use this knob to
2134 * disable swapping for individual groups completely when
2135 * using the memory controller's swap limit feature would be
2136 * too expensive.
2137 */
2138 if (!global_reclaim(sc) && !swappiness) {
2139 scan_balance = SCAN_FILE;
2140 goto out;
2141 }
2142
2143 /*
2144 * Do not apply any pressure balancing cleverness when the
2145 * system is close to OOM, scan both anon and file equally
2146 * (unless the swappiness setting disagrees with swapping).
2147 */
2148 if (!sc->priority && swappiness) {
2149 scan_balance = SCAN_EQUAL;
2150 goto out;
2151 }
2152
2153 /*
2154 * Prevent the reclaimer from falling into the cache trap: as
2155 * cache pages start out inactive, every cache fault will tip
2156 * the scan balance towards the file LRU. And as the file LRU
2157 * shrinks, so does the window for rotation from references.
2158 * This means we have a runaway feedback loop where a tiny
2159 * thrashing file LRU becomes infinitely more attractive than
2160 * anon pages. Try to detect this based on file LRU size.
2161 */
2162 if (global_reclaim(sc)) {
2163 unsigned long pgdatfile;
2164 unsigned long pgdatfree;
2165 int z;
2166 unsigned long total_high_wmark = 0;
2167
2168 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2169 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2170 node_page_state(pgdat, NR_INACTIVE_FILE);
2171
2172 for (z = 0; z < MAX_NR_ZONES; z++) {
2173 struct zone *zone = &pgdat->node_zones[z];
2174 if (!populated_zone(zone))
2175 continue;
2176
2177 total_high_wmark += high_wmark_pages(zone);
2178 }
2179
2180 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2181 scan_balance = SCAN_ANON;
2182 goto out;
2183 }
2184 }
2185
2186 /*
2187 * If there is enough inactive page cache, i.e. if the size of the
2188 * inactive list is greater than that of the active list *and* the
2189 * inactive list actually has some pages to scan on this priority, we
2190 * do not reclaim anything from the anonymous working set right now.
2191 * Without the second condition we could end up never scanning an
2192 * lruvec even if it has plenty of old anonymous pages unless the
2193 * system is under heavy pressure.
2194 */
2195 if (!inactive_list_is_low(lruvec, true, sc) &&
2196 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2197 scan_balance = SCAN_FILE;
2198 goto out;
2199 }
2200
2201 scan_balance = SCAN_FRACT;
2202
2203 /*
2204 * With swappiness at 100, anonymous and file have the same priority.
2205 * This scanning priority is essentially the inverse of IO cost.
2206 */
2207 anon_prio = swappiness;
2208 file_prio = 200 - anon_prio;
2209
2210 /*
2211 * OK, so we have swap space and a fair amount of page cache
2212 * pages. We use the recently rotated / recently scanned
2213 * ratios to determine how valuable each cache is.
2214 *
2215 * Because workloads change over time (and to avoid overflow)
2216 * we keep these statistics as a floating average, which ends
2217 * up weighing recent references more than old ones.
2218 *
2219 * anon in [0], file in [1]
2220 */
2221
2222 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2223 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2224 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2225 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2226
2227 spin_lock_irq(&pgdat->lru_lock);
2228 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2229 reclaim_stat->recent_scanned[0] /= 2;
2230 reclaim_stat->recent_rotated[0] /= 2;
2231 }
2232
2233 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2234 reclaim_stat->recent_scanned[1] /= 2;
2235 reclaim_stat->recent_rotated[1] /= 2;
2236 }
2237
2238 /*
2239 * The amount of pressure on anon vs file pages is inversely
2240 * proportional to the fraction of recently scanned pages on
2241 * each list that were recently referenced and in active use.
2242 */
2243 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2244 ap /= reclaim_stat->recent_rotated[0] + 1;
2245
2246 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2247 fp /= reclaim_stat->recent_rotated[1] + 1;
2248 spin_unlock_irq(&pgdat->lru_lock);
2249
2250 fraction[0] = ap;
2251 fraction[1] = fp;
2252 denominator = ap + fp + 1;
2253 out:
2254 some_scanned = false;
2255 /* Only use force_scan on second pass. */
2256 for (pass = 0; !some_scanned && pass < 2; pass++) {
2257 *lru_pages = 0;
2258 for_each_evictable_lru(lru) {
2259 int file = is_file_lru(lru);
2260 unsigned long size;
2261 unsigned long scan;
2262
2263 size = lruvec_lru_size(lruvec, lru);
2264 scan = size >> sc->priority;
2265
2266 if (!scan && pass && force_scan)
2267 scan = min(size, SWAP_CLUSTER_MAX);
2268
2269 switch (scan_balance) {
2270 case SCAN_EQUAL:
2271 /* Scan lists relative to size */
2272 break;
2273 case SCAN_FRACT:
2274 /*
2275 * Scan types proportional to swappiness and
2276 * their relative recent reclaim efficiency.
2277 */
2278 scan = div64_u64(scan * fraction[file],
2279 denominator);
2280 break;
2281 case SCAN_FILE:
2282 case SCAN_ANON:
2283 /* Scan one type exclusively */
2284 if ((scan_balance == SCAN_FILE) != file) {
2285 size = 0;
2286 scan = 0;
2287 }
2288 break;
2289 default:
2290 /* Look ma, no brain */
2291 BUG();
2292 }
2293
2294 *lru_pages += size;
2295 nr[lru] = scan;
2296
2297 /*
2298 * Skip the second pass and don't force_scan,
2299 * if we found something to scan.
2300 */
2301 some_scanned |= !!scan;
2302 }
2303 }
2304 }
2305
2306 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2307 static void init_tlb_ubc(void)
2308 {
2309 /*
2310 * This deliberately does not clear the cpumask as it's expensive
2311 * and unnecessary. If there happens to be data in there then the
2312 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2313 * then will be cleared.
2314 */
2315 current->tlb_ubc.flush_required = false;
2316 }
2317 #else
2318 static inline void init_tlb_ubc(void)
2319 {
2320 }
2321 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2322
2323 /*
2324 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2325 */
2326 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2327 struct scan_control *sc, unsigned long *lru_pages)
2328 {
2329 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2330 unsigned long nr[NR_LRU_LISTS];
2331 unsigned long targets[NR_LRU_LISTS];
2332 unsigned long nr_to_scan;
2333 enum lru_list lru;
2334 unsigned long nr_reclaimed = 0;
2335 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2336 struct blk_plug plug;
2337 bool scan_adjusted;
2338
2339 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2340
2341 /* Record the original scan target for proportional adjustments later */
2342 memcpy(targets, nr, sizeof(nr));
2343
2344 /*
2345 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2346 * event that can occur when there is little memory pressure e.g.
2347 * multiple streaming readers/writers. Hence, we do not abort scanning
2348 * when the requested number of pages are reclaimed when scanning at
2349 * DEF_PRIORITY on the assumption that the fact we are direct
2350 * reclaiming implies that kswapd is not keeping up and it is best to
2351 * do a batch of work at once. For memcg reclaim one check is made to
2352 * abort proportional reclaim if either the file or anon lru has already
2353 * dropped to zero at the first pass.
2354 */
2355 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2356 sc->priority == DEF_PRIORITY);
2357
2358 init_tlb_ubc();
2359
2360 blk_start_plug(&plug);
2361 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2362 nr[LRU_INACTIVE_FILE]) {
2363 unsigned long nr_anon, nr_file, percentage;
2364 unsigned long nr_scanned;
2365
2366 for_each_evictable_lru(lru) {
2367 if (nr[lru]) {
2368 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2369 nr[lru] -= nr_to_scan;
2370
2371 nr_reclaimed += shrink_list(lru, nr_to_scan,
2372 lruvec, sc);
2373 }
2374 }
2375
2376 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2377 continue;
2378
2379 /*
2380 * For kswapd and memcg, reclaim at least the number of pages
2381 * requested. Ensure that the anon and file LRUs are scanned
2382 * proportionally what was requested by get_scan_count(). We
2383 * stop reclaiming one LRU and reduce the amount scanning
2384 * proportional to the original scan target.
2385 */
2386 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2387 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2388
2389 /*
2390 * It's just vindictive to attack the larger once the smaller
2391 * has gone to zero. And given the way we stop scanning the
2392 * smaller below, this makes sure that we only make one nudge
2393 * towards proportionality once we've got nr_to_reclaim.
2394 */
2395 if (!nr_file || !nr_anon)
2396 break;
2397
2398 if (nr_file > nr_anon) {
2399 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2400 targets[LRU_ACTIVE_ANON] + 1;
2401 lru = LRU_BASE;
2402 percentage = nr_anon * 100 / scan_target;
2403 } else {
2404 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2405 targets[LRU_ACTIVE_FILE] + 1;
2406 lru = LRU_FILE;
2407 percentage = nr_file * 100 / scan_target;
2408 }
2409
2410 /* Stop scanning the smaller of the LRU */
2411 nr[lru] = 0;
2412 nr[lru + LRU_ACTIVE] = 0;
2413
2414 /*
2415 * Recalculate the other LRU scan count based on its original
2416 * scan target and the percentage scanning already complete
2417 */
2418 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2419 nr_scanned = targets[lru] - nr[lru];
2420 nr[lru] = targets[lru] * (100 - percentage) / 100;
2421 nr[lru] -= min(nr[lru], nr_scanned);
2422
2423 lru += LRU_ACTIVE;
2424 nr_scanned = targets[lru] - nr[lru];
2425 nr[lru] = targets[lru] * (100 - percentage) / 100;
2426 nr[lru] -= min(nr[lru], nr_scanned);
2427
2428 scan_adjusted = true;
2429 }
2430 blk_finish_plug(&plug);
2431 sc->nr_reclaimed += nr_reclaimed;
2432
2433 /*
2434 * Even if we did not try to evict anon pages at all, we want to
2435 * rebalance the anon lru active/inactive ratio.
2436 */
2437 if (inactive_list_is_low(lruvec, false, sc))
2438 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2439 sc, LRU_ACTIVE_ANON);
2440
2441 throttle_vm_writeout(sc->gfp_mask);
2442 }
2443
2444 /* Use reclaim/compaction for costly allocs or under memory pressure */
2445 static bool in_reclaim_compaction(struct scan_control *sc)
2446 {
2447 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2448 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2449 sc->priority < DEF_PRIORITY - 2))
2450 return true;
2451
2452 return false;
2453 }
2454
2455 /*
2456 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2457 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2458 * true if more pages should be reclaimed such that when the page allocator
2459 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2460 * It will give up earlier than that if there is difficulty reclaiming pages.
2461 */
2462 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2463 unsigned long nr_reclaimed,
2464 unsigned long nr_scanned,
2465 struct scan_control *sc)
2466 {
2467 unsigned long pages_for_compaction;
2468 unsigned long inactive_lru_pages;
2469 int z;
2470
2471 /* If not in reclaim/compaction mode, stop */
2472 if (!in_reclaim_compaction(sc))
2473 return false;
2474
2475 /* Consider stopping depending on scan and reclaim activity */
2476 if (sc->gfp_mask & __GFP_REPEAT) {
2477 /*
2478 * For __GFP_REPEAT allocations, stop reclaiming if the
2479 * full LRU list has been scanned and we are still failing
2480 * to reclaim pages. This full LRU scan is potentially
2481 * expensive but a __GFP_REPEAT caller really wants to succeed
2482 */
2483 if (!nr_reclaimed && !nr_scanned)
2484 return false;
2485 } else {
2486 /*
2487 * For non-__GFP_REPEAT allocations which can presumably
2488 * fail without consequence, stop if we failed to reclaim
2489 * any pages from the last SWAP_CLUSTER_MAX number of
2490 * pages that were scanned. This will return to the
2491 * caller faster at the risk reclaim/compaction and
2492 * the resulting allocation attempt fails
2493 */
2494 if (!nr_reclaimed)
2495 return false;
2496 }
2497
2498 /*
2499 * If we have not reclaimed enough pages for compaction and the
2500 * inactive lists are large enough, continue reclaiming
2501 */
2502 pages_for_compaction = (2UL << sc->order);
2503 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2504 if (get_nr_swap_pages() > 0)
2505 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2506 if (sc->nr_reclaimed < pages_for_compaction &&
2507 inactive_lru_pages > pages_for_compaction)
2508 return true;
2509
2510 /* If compaction would go ahead or the allocation would succeed, stop */
2511 for (z = 0; z <= sc->reclaim_idx; z++) {
2512 struct zone *zone = &pgdat->node_zones[z];
2513 if (!populated_zone(zone))
2514 continue;
2515
2516 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2517 case COMPACT_PARTIAL:
2518 case COMPACT_CONTINUE:
2519 return false;
2520 default:
2521 /* check next zone */
2522 ;
2523 }
2524 }
2525 return true;
2526 }
2527
2528 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2529 {
2530 struct reclaim_state *reclaim_state = current->reclaim_state;
2531 unsigned long nr_reclaimed, nr_scanned;
2532 bool reclaimable = false;
2533
2534 do {
2535 struct mem_cgroup *root = sc->target_mem_cgroup;
2536 struct mem_cgroup_reclaim_cookie reclaim = {
2537 .pgdat = pgdat,
2538 .priority = sc->priority,
2539 };
2540 unsigned long node_lru_pages = 0;
2541 struct mem_cgroup *memcg;
2542
2543 nr_reclaimed = sc->nr_reclaimed;
2544 nr_scanned = sc->nr_scanned;
2545
2546 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2547 do {
2548 unsigned long lru_pages;
2549 unsigned long reclaimed;
2550 unsigned long scanned;
2551
2552 if (mem_cgroup_low(root, memcg)) {
2553 if (!sc->may_thrash)
2554 continue;
2555 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2556 }
2557
2558 reclaimed = sc->nr_reclaimed;
2559 scanned = sc->nr_scanned;
2560
2561 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2562 node_lru_pages += lru_pages;
2563
2564 if (!global_reclaim(sc))
2565 shrink_slab(sc->gfp_mask, pgdat->node_id,
2566 memcg, sc->nr_scanned - scanned,
2567 lru_pages);
2568
2569 /* Record the group's reclaim efficiency */
2570 vmpressure(sc->gfp_mask, memcg, false,
2571 sc->nr_scanned - scanned,
2572 sc->nr_reclaimed - reclaimed);
2573
2574 /*
2575 * Direct reclaim and kswapd have to scan all memory
2576 * cgroups to fulfill the overall scan target for the
2577 * node.
2578 *
2579 * Limit reclaim, on the other hand, only cares about
2580 * nr_to_reclaim pages to be reclaimed and it will
2581 * retry with decreasing priority if one round over the
2582 * whole hierarchy is not sufficient.
2583 */
2584 if (!global_reclaim(sc) &&
2585 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2586 mem_cgroup_iter_break(root, memcg);
2587 break;
2588 }
2589 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2590
2591 /*
2592 * Shrink the slab caches in the same proportion that
2593 * the eligible LRU pages were scanned.
2594 */
2595 if (global_reclaim(sc))
2596 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2597 sc->nr_scanned - nr_scanned,
2598 node_lru_pages);
2599
2600 if (reclaim_state) {
2601 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2602 reclaim_state->reclaimed_slab = 0;
2603 }
2604
2605 /* Record the subtree's reclaim efficiency */
2606 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2607 sc->nr_scanned - nr_scanned,
2608 sc->nr_reclaimed - nr_reclaimed);
2609
2610 if (sc->nr_reclaimed - nr_reclaimed)
2611 reclaimable = true;
2612
2613 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2614 sc->nr_scanned - nr_scanned, sc));
2615
2616 return reclaimable;
2617 }
2618
2619 /*
2620 * Returns true if compaction should go ahead for a high-order request, or
2621 * the high-order allocation would succeed without compaction.
2622 */
2623 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2624 {
2625 unsigned long watermark;
2626 bool watermark_ok;
2627
2628 /*
2629 * Compaction takes time to run and there are potentially other
2630 * callers using the pages just freed. Continue reclaiming until
2631 * there is a buffer of free pages available to give compaction
2632 * a reasonable chance of completing and allocating the page
2633 */
2634 watermark = high_wmark_pages(zone) + (2UL << sc->order);
2635 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2636
2637 /*
2638 * If compaction is deferred, reclaim up to a point where
2639 * compaction will have a chance of success when re-enabled
2640 */
2641 if (compaction_deferred(zone, sc->order))
2642 return watermark_ok;
2643
2644 /*
2645 * If compaction is not ready to start and allocation is not likely
2646 * to succeed without it, then keep reclaiming.
2647 */
2648 if (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx) == COMPACT_SKIPPED)
2649 return false;
2650
2651 return watermark_ok;
2652 }
2653
2654 /*
2655 * This is the direct reclaim path, for page-allocating processes. We only
2656 * try to reclaim pages from zones which will satisfy the caller's allocation
2657 * request.
2658 *
2659 * If a zone is deemed to be full of pinned pages then just give it a light
2660 * scan then give up on it.
2661 */
2662 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2663 {
2664 struct zoneref *z;
2665 struct zone *zone;
2666 unsigned long nr_soft_reclaimed;
2667 unsigned long nr_soft_scanned;
2668 gfp_t orig_mask;
2669 pg_data_t *last_pgdat = NULL;
2670
2671 /*
2672 * If the number of buffer_heads in the machine exceeds the maximum
2673 * allowed level, force direct reclaim to scan the highmem zone as
2674 * highmem pages could be pinning lowmem pages storing buffer_heads
2675 */
2676 orig_mask = sc->gfp_mask;
2677 if (buffer_heads_over_limit) {
2678 sc->gfp_mask |= __GFP_HIGHMEM;
2679 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2680 }
2681
2682 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2683 sc->reclaim_idx, sc->nodemask) {
2684 /*
2685 * Take care memory controller reclaiming has small influence
2686 * to global LRU.
2687 */
2688 if (global_reclaim(sc)) {
2689 if (!cpuset_zone_allowed(zone,
2690 GFP_KERNEL | __GFP_HARDWALL))
2691 continue;
2692
2693 if (sc->priority != DEF_PRIORITY &&
2694 !pgdat_reclaimable(zone->zone_pgdat))
2695 continue; /* Let kswapd poll it */
2696
2697 /*
2698 * If we already have plenty of memory free for
2699 * compaction in this zone, don't free any more.
2700 * Even though compaction is invoked for any
2701 * non-zero order, only frequent costly order
2702 * reclamation is disruptive enough to become a
2703 * noticeable problem, like transparent huge
2704 * page allocations.
2705 */
2706 if (IS_ENABLED(CONFIG_COMPACTION) &&
2707 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2708 compaction_ready(zone, sc)) {
2709 sc->compaction_ready = true;
2710 continue;
2711 }
2712
2713 /*
2714 * Shrink each node in the zonelist once. If the
2715 * zonelist is ordered by zone (not the default) then a
2716 * node may be shrunk multiple times but in that case
2717 * the user prefers lower zones being preserved.
2718 */
2719 if (zone->zone_pgdat == last_pgdat)
2720 continue;
2721
2722 /*
2723 * This steals pages from memory cgroups over softlimit
2724 * and returns the number of reclaimed pages and
2725 * scanned pages. This works for global memory pressure
2726 * and balancing, not for a memcg's limit.
2727 */
2728 nr_soft_scanned = 0;
2729 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2730 sc->order, sc->gfp_mask,
2731 &nr_soft_scanned);
2732 sc->nr_reclaimed += nr_soft_reclaimed;
2733 sc->nr_scanned += nr_soft_scanned;
2734 /* need some check for avoid more shrink_zone() */
2735 }
2736
2737 /* See comment about same check for global reclaim above */
2738 if (zone->zone_pgdat == last_pgdat)
2739 continue;
2740 last_pgdat = zone->zone_pgdat;
2741 shrink_node(zone->zone_pgdat, sc);
2742 }
2743
2744 /*
2745 * Restore to original mask to avoid the impact on the caller if we
2746 * promoted it to __GFP_HIGHMEM.
2747 */
2748 sc->gfp_mask = orig_mask;
2749 }
2750
2751 /*
2752 * This is the main entry point to direct page reclaim.
2753 *
2754 * If a full scan of the inactive list fails to free enough memory then we
2755 * are "out of memory" and something needs to be killed.
2756 *
2757 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2758 * high - the zone may be full of dirty or under-writeback pages, which this
2759 * caller can't do much about. We kick the writeback threads and take explicit
2760 * naps in the hope that some of these pages can be written. But if the
2761 * allocating task holds filesystem locks which prevent writeout this might not
2762 * work, and the allocation attempt will fail.
2763 *
2764 * returns: 0, if no pages reclaimed
2765 * else, the number of pages reclaimed
2766 */
2767 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2768 struct scan_control *sc)
2769 {
2770 int initial_priority = sc->priority;
2771 unsigned long total_scanned = 0;
2772 unsigned long writeback_threshold;
2773 retry:
2774 delayacct_freepages_start();
2775
2776 if (global_reclaim(sc))
2777 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2778
2779 do {
2780 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2781 sc->priority);
2782 sc->nr_scanned = 0;
2783 shrink_zones(zonelist, sc);
2784
2785 total_scanned += sc->nr_scanned;
2786 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2787 break;
2788
2789 if (sc->compaction_ready)
2790 break;
2791
2792 /*
2793 * If we're getting trouble reclaiming, start doing
2794 * writepage even in laptop mode.
2795 */
2796 if (sc->priority < DEF_PRIORITY - 2)
2797 sc->may_writepage = 1;
2798
2799 /*
2800 * Try to write back as many pages as we just scanned. This
2801 * tends to cause slow streaming writers to write data to the
2802 * disk smoothly, at the dirtying rate, which is nice. But
2803 * that's undesirable in laptop mode, where we *want* lumpy
2804 * writeout. So in laptop mode, write out the whole world.
2805 */
2806 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2807 if (total_scanned > writeback_threshold) {
2808 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2809 WB_REASON_TRY_TO_FREE_PAGES);
2810 sc->may_writepage = 1;
2811 }
2812 } while (--sc->priority >= 0);
2813
2814 delayacct_freepages_end();
2815
2816 if (sc->nr_reclaimed)
2817 return sc->nr_reclaimed;
2818
2819 /* Aborted reclaim to try compaction? don't OOM, then */
2820 if (sc->compaction_ready)
2821 return 1;
2822
2823 /* Untapped cgroup reserves? Don't OOM, retry. */
2824 if (!sc->may_thrash) {
2825 sc->priority = initial_priority;
2826 sc->may_thrash = 1;
2827 goto retry;
2828 }
2829
2830 return 0;
2831 }
2832
2833 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2834 {
2835 struct zone *zone;
2836 unsigned long pfmemalloc_reserve = 0;
2837 unsigned long free_pages = 0;
2838 int i;
2839 bool wmark_ok;
2840
2841 for (i = 0; i <= ZONE_NORMAL; i++) {
2842 zone = &pgdat->node_zones[i];
2843 if (!populated_zone(zone) ||
2844 pgdat_reclaimable_pages(pgdat) == 0)
2845 continue;
2846
2847 pfmemalloc_reserve += min_wmark_pages(zone);
2848 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2849 }
2850
2851 /* If there are no reserves (unexpected config) then do not throttle */
2852 if (!pfmemalloc_reserve)
2853 return true;
2854
2855 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2856
2857 /* kswapd must be awake if processes are being throttled */
2858 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2859 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2860 (enum zone_type)ZONE_NORMAL);
2861 wake_up_interruptible(&pgdat->kswapd_wait);
2862 }
2863
2864 return wmark_ok;
2865 }
2866
2867 /*
2868 * Throttle direct reclaimers if backing storage is backed by the network
2869 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2870 * depleted. kswapd will continue to make progress and wake the processes
2871 * when the low watermark is reached.
2872 *
2873 * Returns true if a fatal signal was delivered during throttling. If this
2874 * happens, the page allocator should not consider triggering the OOM killer.
2875 */
2876 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2877 nodemask_t *nodemask)
2878 {
2879 struct zoneref *z;
2880 struct zone *zone;
2881 pg_data_t *pgdat = NULL;
2882
2883 /*
2884 * Kernel threads should not be throttled as they may be indirectly
2885 * responsible for cleaning pages necessary for reclaim to make forward
2886 * progress. kjournald for example may enter direct reclaim while
2887 * committing a transaction where throttling it could forcing other
2888 * processes to block on log_wait_commit().
2889 */
2890 if (current->flags & PF_KTHREAD)
2891 goto out;
2892
2893 /*
2894 * If a fatal signal is pending, this process should not throttle.
2895 * It should return quickly so it can exit and free its memory
2896 */
2897 if (fatal_signal_pending(current))
2898 goto out;
2899
2900 /*
2901 * Check if the pfmemalloc reserves are ok by finding the first node
2902 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2903 * GFP_KERNEL will be required for allocating network buffers when
2904 * swapping over the network so ZONE_HIGHMEM is unusable.
2905 *
2906 * Throttling is based on the first usable node and throttled processes
2907 * wait on a queue until kswapd makes progress and wakes them. There
2908 * is an affinity then between processes waking up and where reclaim
2909 * progress has been made assuming the process wakes on the same node.
2910 * More importantly, processes running on remote nodes will not compete
2911 * for remote pfmemalloc reserves and processes on different nodes
2912 * should make reasonable progress.
2913 */
2914 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2915 gfp_zone(gfp_mask), nodemask) {
2916 if (zone_idx(zone) > ZONE_NORMAL)
2917 continue;
2918
2919 /* Throttle based on the first usable node */
2920 pgdat = zone->zone_pgdat;
2921 if (pfmemalloc_watermark_ok(pgdat))
2922 goto out;
2923 break;
2924 }
2925
2926 /* If no zone was usable by the allocation flags then do not throttle */
2927 if (!pgdat)
2928 goto out;
2929
2930 /* Account for the throttling */
2931 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2932
2933 /*
2934 * If the caller cannot enter the filesystem, it's possible that it
2935 * is due to the caller holding an FS lock or performing a journal
2936 * transaction in the case of a filesystem like ext[3|4]. In this case,
2937 * it is not safe to block on pfmemalloc_wait as kswapd could be
2938 * blocked waiting on the same lock. Instead, throttle for up to a
2939 * second before continuing.
2940 */
2941 if (!(gfp_mask & __GFP_FS)) {
2942 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2943 pfmemalloc_watermark_ok(pgdat), HZ);
2944
2945 goto check_pending;
2946 }
2947
2948 /* Throttle until kswapd wakes the process */
2949 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2950 pfmemalloc_watermark_ok(pgdat));
2951
2952 check_pending:
2953 if (fatal_signal_pending(current))
2954 return true;
2955
2956 out:
2957 return false;
2958 }
2959
2960 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2961 gfp_t gfp_mask, nodemask_t *nodemask)
2962 {
2963 unsigned long nr_reclaimed;
2964 struct scan_control sc = {
2965 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2966 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2967 .reclaim_idx = gfp_zone(gfp_mask),
2968 .order = order,
2969 .nodemask = nodemask,
2970 .priority = DEF_PRIORITY,
2971 .may_writepage = !laptop_mode,
2972 .may_unmap = 1,
2973 .may_swap = 1,
2974 };
2975
2976 /*
2977 * Do not enter reclaim if fatal signal was delivered while throttled.
2978 * 1 is returned so that the page allocator does not OOM kill at this
2979 * point.
2980 */
2981 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2982 return 1;
2983
2984 trace_mm_vmscan_direct_reclaim_begin(order,
2985 sc.may_writepage,
2986 gfp_mask,
2987 sc.reclaim_idx);
2988
2989 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2990
2991 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2992
2993 return nr_reclaimed;
2994 }
2995
2996 #ifdef CONFIG_MEMCG
2997
2998 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
2999 gfp_t gfp_mask, bool noswap,
3000 pg_data_t *pgdat,
3001 unsigned long *nr_scanned)
3002 {
3003 struct scan_control sc = {
3004 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3005 .target_mem_cgroup = memcg,
3006 .may_writepage = !laptop_mode,
3007 .may_unmap = 1,
3008 .reclaim_idx = MAX_NR_ZONES - 1,
3009 .may_swap = !noswap,
3010 };
3011 unsigned long lru_pages;
3012
3013 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3014 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3015
3016 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3017 sc.may_writepage,
3018 sc.gfp_mask,
3019 sc.reclaim_idx);
3020
3021 /*
3022 * NOTE: Although we can get the priority field, using it
3023 * here is not a good idea, since it limits the pages we can scan.
3024 * if we don't reclaim here, the shrink_node from balance_pgdat
3025 * will pick up pages from other mem cgroup's as well. We hack
3026 * the priority and make it zero.
3027 */
3028 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3029
3030 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3031
3032 *nr_scanned = sc.nr_scanned;
3033 return sc.nr_reclaimed;
3034 }
3035
3036 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3037 unsigned long nr_pages,
3038 gfp_t gfp_mask,
3039 bool may_swap)
3040 {
3041 struct zonelist *zonelist;
3042 unsigned long nr_reclaimed;
3043 int nid;
3044 struct scan_control sc = {
3045 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3046 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3047 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3048 .reclaim_idx = MAX_NR_ZONES - 1,
3049 .target_mem_cgroup = memcg,
3050 .priority = DEF_PRIORITY,
3051 .may_writepage = !laptop_mode,
3052 .may_unmap = 1,
3053 .may_swap = may_swap,
3054 };
3055
3056 /*
3057 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3058 * take care of from where we get pages. So the node where we start the
3059 * scan does not need to be the current node.
3060 */
3061 nid = mem_cgroup_select_victim_node(memcg);
3062
3063 zonelist = NODE_DATA(nid)->node_zonelists;
3064
3065 trace_mm_vmscan_memcg_reclaim_begin(0,
3066 sc.may_writepage,
3067 sc.gfp_mask,
3068 sc.reclaim_idx);
3069
3070 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3071
3072 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3073
3074 return nr_reclaimed;
3075 }
3076 #endif
3077
3078 static void age_active_anon(struct pglist_data *pgdat,
3079 struct scan_control *sc)
3080 {
3081 struct mem_cgroup *memcg;
3082
3083 if (!total_swap_pages)
3084 return;
3085
3086 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3087 do {
3088 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3089
3090 if (inactive_list_is_low(lruvec, false, sc))
3091 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3092 sc, LRU_ACTIVE_ANON);
3093
3094 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3095 } while (memcg);
3096 }
3097
3098 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3099 {
3100 unsigned long mark = high_wmark_pages(zone);
3101
3102 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3103 return false;
3104
3105 /*
3106 * If any eligible zone is balanced then the node is not considered
3107 * to be congested or dirty
3108 */
3109 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3110 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3111
3112 return true;
3113 }
3114
3115 /*
3116 * Prepare kswapd for sleeping. This verifies that there are no processes
3117 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3118 *
3119 * Returns true if kswapd is ready to sleep
3120 */
3121 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3122 {
3123 int i;
3124
3125 /*
3126 * The throttled processes are normally woken up in balance_pgdat() as
3127 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3128 * race between when kswapd checks the watermarks and a process gets
3129 * throttled. There is also a potential race if processes get
3130 * throttled, kswapd wakes, a large process exits thereby balancing the
3131 * zones, which causes kswapd to exit balance_pgdat() before reaching
3132 * the wake up checks. If kswapd is going to sleep, no process should
3133 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3134 * the wake up is premature, processes will wake kswapd and get
3135 * throttled again. The difference from wake ups in balance_pgdat() is
3136 * that here we are under prepare_to_wait().
3137 */
3138 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3139 wake_up_all(&pgdat->pfmemalloc_wait);
3140
3141 for (i = 0; i <= classzone_idx; i++) {
3142 struct zone *zone = pgdat->node_zones + i;
3143
3144 if (!populated_zone(zone))
3145 continue;
3146
3147 if (!zone_balanced(zone, order, classzone_idx))
3148 return false;
3149 }
3150
3151 return true;
3152 }
3153
3154 /*
3155 * kswapd shrinks a node of pages that are at or below the highest usable
3156 * zone that is currently unbalanced.
3157 *
3158 * Returns true if kswapd scanned at least the requested number of pages to
3159 * reclaim or if the lack of progress was due to pages under writeback.
3160 * This is used to determine if the scanning priority needs to be raised.
3161 */
3162 static bool kswapd_shrink_node(pg_data_t *pgdat,
3163 struct scan_control *sc)
3164 {
3165 struct zone *zone;
3166 int z;
3167
3168 /* Reclaim a number of pages proportional to the number of zones */
3169 sc->nr_to_reclaim = 0;
3170 for (z = 0; z <= sc->reclaim_idx; z++) {
3171 zone = pgdat->node_zones + z;
3172 if (!populated_zone(zone))
3173 continue;
3174
3175 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3176 }
3177
3178 /*
3179 * Historically care was taken to put equal pressure on all zones but
3180 * now pressure is applied based on node LRU order.
3181 */
3182 shrink_node(pgdat, sc);
3183
3184 /*
3185 * Fragmentation may mean that the system cannot be rebalanced for
3186 * high-order allocations. If twice the allocation size has been
3187 * reclaimed then recheck watermarks only at order-0 to prevent
3188 * excessive reclaim. Assume that a process requested a high-order
3189 * can direct reclaim/compact.
3190 */
3191 if (sc->order && sc->nr_reclaimed >= 2UL << sc->order)
3192 sc->order = 0;
3193
3194 return sc->nr_scanned >= sc->nr_to_reclaim;
3195 }
3196
3197 /*
3198 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3199 * that are eligible for use by the caller until at least one zone is
3200 * balanced.
3201 *
3202 * Returns the order kswapd finished reclaiming at.
3203 *
3204 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3205 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3206 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3207 * or lower is eligible for reclaim until at least one usable zone is
3208 * balanced.
3209 */
3210 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3211 {
3212 int i;
3213 unsigned long nr_soft_reclaimed;
3214 unsigned long nr_soft_scanned;
3215 struct zone *zone;
3216 struct scan_control sc = {
3217 .gfp_mask = GFP_KERNEL,
3218 .order = order,
3219 .priority = DEF_PRIORITY,
3220 .may_writepage = !laptop_mode,
3221 .may_unmap = 1,
3222 .may_swap = 1,
3223 };
3224 count_vm_event(PAGEOUTRUN);
3225
3226 do {
3227 bool raise_priority = true;
3228
3229 sc.nr_reclaimed = 0;
3230 sc.reclaim_idx = classzone_idx;
3231
3232 /*
3233 * If the number of buffer_heads exceeds the maximum allowed
3234 * then consider reclaiming from all zones. This has a dual
3235 * purpose -- on 64-bit systems it is expected that
3236 * buffer_heads are stripped during active rotation. On 32-bit
3237 * systems, highmem pages can pin lowmem memory and shrinking
3238 * buffers can relieve lowmem pressure. Reclaim may still not
3239 * go ahead if all eligible zones for the original allocation
3240 * request are balanced to avoid excessive reclaim from kswapd.
3241 */
3242 if (buffer_heads_over_limit) {
3243 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3244 zone = pgdat->node_zones + i;
3245 if (!populated_zone(zone))
3246 continue;
3247
3248 sc.reclaim_idx = i;
3249 break;
3250 }
3251 }
3252
3253 /*
3254 * Only reclaim if there are no eligible zones. Check from
3255 * high to low zone as allocations prefer higher zones.
3256 * Scanning from low to high zone would allow congestion to be
3257 * cleared during a very small window when a small low
3258 * zone was balanced even under extreme pressure when the
3259 * overall node may be congested. Note that sc.reclaim_idx
3260 * is not used as buffer_heads_over_limit may have adjusted
3261 * it.
3262 */
3263 for (i = classzone_idx; i >= 0; i--) {
3264 zone = pgdat->node_zones + i;
3265 if (!populated_zone(zone))
3266 continue;
3267
3268 if (zone_balanced(zone, sc.order, classzone_idx))
3269 goto out;
3270 }
3271
3272 /*
3273 * Do some background aging of the anon list, to give
3274 * pages a chance to be referenced before reclaiming. All
3275 * pages are rotated regardless of classzone as this is
3276 * about consistent aging.
3277 */
3278 age_active_anon(pgdat, &sc);
3279
3280 /*
3281 * If we're getting trouble reclaiming, start doing writepage
3282 * even in laptop mode.
3283 */
3284 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3285 sc.may_writepage = 1;
3286
3287 /* Call soft limit reclaim before calling shrink_node. */
3288 sc.nr_scanned = 0;
3289 nr_soft_scanned = 0;
3290 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3291 sc.gfp_mask, &nr_soft_scanned);
3292 sc.nr_reclaimed += nr_soft_reclaimed;
3293
3294 /*
3295 * There should be no need to raise the scanning priority if
3296 * enough pages are already being scanned that that high
3297 * watermark would be met at 100% efficiency.
3298 */
3299 if (kswapd_shrink_node(pgdat, &sc))
3300 raise_priority = false;
3301
3302 /*
3303 * If the low watermark is met there is no need for processes
3304 * to be throttled on pfmemalloc_wait as they should not be
3305 * able to safely make forward progress. Wake them
3306 */
3307 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3308 pfmemalloc_watermark_ok(pgdat))
3309 wake_up_all(&pgdat->pfmemalloc_wait);
3310
3311 /* Check if kswapd should be suspending */
3312 if (try_to_freeze() || kthread_should_stop())
3313 break;
3314
3315 /*
3316 * Raise priority if scanning rate is too low or there was no
3317 * progress in reclaiming pages
3318 */
3319 if (raise_priority || !sc.nr_reclaimed)
3320 sc.priority--;
3321 } while (sc.priority >= 1);
3322
3323 out:
3324 /*
3325 * Return the order kswapd stopped reclaiming at as
3326 * prepare_kswapd_sleep() takes it into account. If another caller
3327 * entered the allocator slow path while kswapd was awake, order will
3328 * remain at the higher level.
3329 */
3330 return sc.order;
3331 }
3332
3333 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3334 unsigned int classzone_idx)
3335 {
3336 long remaining = 0;
3337 DEFINE_WAIT(wait);
3338
3339 if (freezing(current) || kthread_should_stop())
3340 return;
3341
3342 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3343
3344 /* Try to sleep for a short interval */
3345 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3346 /*
3347 * Compaction records what page blocks it recently failed to
3348 * isolate pages from and skips them in the future scanning.
3349 * When kswapd is going to sleep, it is reasonable to assume
3350 * that pages and compaction may succeed so reset the cache.
3351 */
3352 reset_isolation_suitable(pgdat);
3353
3354 /*
3355 * We have freed the memory, now we should compact it to make
3356 * allocation of the requested order possible.
3357 */
3358 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3359
3360 remaining = schedule_timeout(HZ/10);
3361
3362 /*
3363 * If woken prematurely then reset kswapd_classzone_idx and
3364 * order. The values will either be from a wakeup request or
3365 * the previous request that slept prematurely.
3366 */
3367 if (remaining) {
3368 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3369 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3370 }
3371
3372 finish_wait(&pgdat->kswapd_wait, &wait);
3373 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3374 }
3375
3376 /*
3377 * After a short sleep, check if it was a premature sleep. If not, then
3378 * go fully to sleep until explicitly woken up.
3379 */
3380 if (!remaining &&
3381 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3382 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3383
3384 /*
3385 * vmstat counters are not perfectly accurate and the estimated
3386 * value for counters such as NR_FREE_PAGES can deviate from the
3387 * true value by nr_online_cpus * threshold. To avoid the zone
3388 * watermarks being breached while under pressure, we reduce the
3389 * per-cpu vmstat threshold while kswapd is awake and restore
3390 * them before going back to sleep.
3391 */
3392 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3393
3394 if (!kthread_should_stop())
3395 schedule();
3396
3397 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3398 } else {
3399 if (remaining)
3400 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3401 else
3402 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3403 }
3404 finish_wait(&pgdat->kswapd_wait, &wait);
3405 }
3406
3407 /*
3408 * The background pageout daemon, started as a kernel thread
3409 * from the init process.
3410 *
3411 * This basically trickles out pages so that we have _some_
3412 * free memory available even if there is no other activity
3413 * that frees anything up. This is needed for things like routing
3414 * etc, where we otherwise might have all activity going on in
3415 * asynchronous contexts that cannot page things out.
3416 *
3417 * If there are applications that are active memory-allocators
3418 * (most normal use), this basically shouldn't matter.
3419 */
3420 static int kswapd(void *p)
3421 {
3422 unsigned int alloc_order, reclaim_order, classzone_idx;
3423 pg_data_t *pgdat = (pg_data_t*)p;
3424 struct task_struct *tsk = current;
3425
3426 struct reclaim_state reclaim_state = {
3427 .reclaimed_slab = 0,
3428 };
3429 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3430
3431 lockdep_set_current_reclaim_state(GFP_KERNEL);
3432
3433 if (!cpumask_empty(cpumask))
3434 set_cpus_allowed_ptr(tsk, cpumask);
3435 current->reclaim_state = &reclaim_state;
3436
3437 /*
3438 * Tell the memory management that we're a "memory allocator",
3439 * and that if we need more memory we should get access to it
3440 * regardless (see "__alloc_pages()"). "kswapd" should
3441 * never get caught in the normal page freeing logic.
3442 *
3443 * (Kswapd normally doesn't need memory anyway, but sometimes
3444 * you need a small amount of memory in order to be able to
3445 * page out something else, and this flag essentially protects
3446 * us from recursively trying to free more memory as we're
3447 * trying to free the first piece of memory in the first place).
3448 */
3449 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3450 set_freezable();
3451
3452 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3453 pgdat->kswapd_classzone_idx = classzone_idx = 0;
3454 for ( ; ; ) {
3455 bool ret;
3456
3457 kswapd_try_sleep:
3458 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3459 classzone_idx);
3460
3461 /* Read the new order and classzone_idx */
3462 alloc_order = reclaim_order = pgdat->kswapd_order;
3463 classzone_idx = pgdat->kswapd_classzone_idx;
3464 pgdat->kswapd_order = 0;
3465 pgdat->kswapd_classzone_idx = 0;
3466
3467 ret = try_to_freeze();
3468 if (kthread_should_stop())
3469 break;
3470
3471 /*
3472 * We can speed up thawing tasks if we don't call balance_pgdat
3473 * after returning from the refrigerator
3474 */
3475 if (ret)
3476 continue;
3477
3478 /*
3479 * Reclaim begins at the requested order but if a high-order
3480 * reclaim fails then kswapd falls back to reclaiming for
3481 * order-0. If that happens, kswapd will consider sleeping
3482 * for the order it finished reclaiming at (reclaim_order)
3483 * but kcompactd is woken to compact for the original
3484 * request (alloc_order).
3485 */
3486 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3487 alloc_order);
3488 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3489 if (reclaim_order < alloc_order)
3490 goto kswapd_try_sleep;
3491
3492 alloc_order = reclaim_order = pgdat->kswapd_order;
3493 classzone_idx = pgdat->kswapd_classzone_idx;
3494 }
3495
3496 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3497 current->reclaim_state = NULL;
3498 lockdep_clear_current_reclaim_state();
3499
3500 return 0;
3501 }
3502
3503 /*
3504 * A zone is low on free memory, so wake its kswapd task to service it.
3505 */
3506 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3507 {
3508 pg_data_t *pgdat;
3509 int z;
3510
3511 if (!populated_zone(zone))
3512 return;
3513
3514 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3515 return;
3516 pgdat = zone->zone_pgdat;
3517 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3518 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3519 if (!waitqueue_active(&pgdat->kswapd_wait))
3520 return;
3521
3522 /* Only wake kswapd if all zones are unbalanced */
3523 for (z = 0; z <= classzone_idx; z++) {
3524 zone = pgdat->node_zones + z;
3525 if (!populated_zone(zone))
3526 continue;
3527
3528 if (zone_balanced(zone, order, classzone_idx))
3529 return;
3530 }
3531
3532 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3533 wake_up_interruptible(&pgdat->kswapd_wait);
3534 }
3535
3536 #ifdef CONFIG_HIBERNATION
3537 /*
3538 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3539 * freed pages.
3540 *
3541 * Rather than trying to age LRUs the aim is to preserve the overall
3542 * LRU order by reclaiming preferentially
3543 * inactive > active > active referenced > active mapped
3544 */
3545 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3546 {
3547 struct reclaim_state reclaim_state;
3548 struct scan_control sc = {
3549 .nr_to_reclaim = nr_to_reclaim,
3550 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3551 .reclaim_idx = MAX_NR_ZONES - 1,
3552 .priority = DEF_PRIORITY,
3553 .may_writepage = 1,
3554 .may_unmap = 1,
3555 .may_swap = 1,
3556 .hibernation_mode = 1,
3557 };
3558 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3559 struct task_struct *p = current;
3560 unsigned long nr_reclaimed;
3561
3562 p->flags |= PF_MEMALLOC;
3563 lockdep_set_current_reclaim_state(sc.gfp_mask);
3564 reclaim_state.reclaimed_slab = 0;
3565 p->reclaim_state = &reclaim_state;
3566
3567 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3568
3569 p->reclaim_state = NULL;
3570 lockdep_clear_current_reclaim_state();
3571 p->flags &= ~PF_MEMALLOC;
3572
3573 return nr_reclaimed;
3574 }
3575 #endif /* CONFIG_HIBERNATION */
3576
3577 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3578 not required for correctness. So if the last cpu in a node goes
3579 away, we get changed to run anywhere: as the first one comes back,
3580 restore their cpu bindings. */
3581 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3582 void *hcpu)
3583 {
3584 int nid;
3585
3586 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3587 for_each_node_state(nid, N_MEMORY) {
3588 pg_data_t *pgdat = NODE_DATA(nid);
3589 const struct cpumask *mask;
3590
3591 mask = cpumask_of_node(pgdat->node_id);
3592
3593 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3594 /* One of our CPUs online: restore mask */
3595 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3596 }
3597 }
3598 return NOTIFY_OK;
3599 }
3600
3601 /*
3602 * This kswapd start function will be called by init and node-hot-add.
3603 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3604 */
3605 int kswapd_run(int nid)
3606 {
3607 pg_data_t *pgdat = NODE_DATA(nid);
3608 int ret = 0;
3609
3610 if (pgdat->kswapd)
3611 return 0;
3612
3613 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3614 if (IS_ERR(pgdat->kswapd)) {
3615 /* failure at boot is fatal */
3616 BUG_ON(system_state == SYSTEM_BOOTING);
3617 pr_err("Failed to start kswapd on node %d\n", nid);
3618 ret = PTR_ERR(pgdat->kswapd);
3619 pgdat->kswapd = NULL;
3620 }
3621 return ret;
3622 }
3623
3624 /*
3625 * Called by memory hotplug when all memory in a node is offlined. Caller must
3626 * hold mem_hotplug_begin/end().
3627 */
3628 void kswapd_stop(int nid)
3629 {
3630 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3631
3632 if (kswapd) {
3633 kthread_stop(kswapd);
3634 NODE_DATA(nid)->kswapd = NULL;
3635 }
3636 }
3637
3638 static int __init kswapd_init(void)
3639 {
3640 int nid;
3641
3642 swap_setup();
3643 for_each_node_state(nid, N_MEMORY)
3644 kswapd_run(nid);
3645 hotcpu_notifier(cpu_callback, 0);
3646 return 0;
3647 }
3648
3649 module_init(kswapd_init)
3650
3651 #ifdef CONFIG_NUMA
3652 /*
3653 * Node reclaim mode
3654 *
3655 * If non-zero call node_reclaim when the number of free pages falls below
3656 * the watermarks.
3657 */
3658 int node_reclaim_mode __read_mostly;
3659
3660 #define RECLAIM_OFF 0
3661 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3662 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3663 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3664
3665 /*
3666 * Priority for NODE_RECLAIM. This determines the fraction of pages
3667 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3668 * a zone.
3669 */
3670 #define NODE_RECLAIM_PRIORITY 4
3671
3672 /*
3673 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3674 * occur.
3675 */
3676 int sysctl_min_unmapped_ratio = 1;
3677
3678 /*
3679 * If the number of slab pages in a zone grows beyond this percentage then
3680 * slab reclaim needs to occur.
3681 */
3682 int sysctl_min_slab_ratio = 5;
3683
3684 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3685 {
3686 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3687 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3688 node_page_state(pgdat, NR_ACTIVE_FILE);
3689
3690 /*
3691 * It's possible for there to be more file mapped pages than
3692 * accounted for by the pages on the file LRU lists because
3693 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3694 */
3695 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3696 }
3697
3698 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3699 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3700 {
3701 unsigned long nr_pagecache_reclaimable;
3702 unsigned long delta = 0;
3703
3704 /*
3705 * If RECLAIM_UNMAP is set, then all file pages are considered
3706 * potentially reclaimable. Otherwise, we have to worry about
3707 * pages like swapcache and node_unmapped_file_pages() provides
3708 * a better estimate
3709 */
3710 if (node_reclaim_mode & RECLAIM_UNMAP)
3711 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3712 else
3713 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3714
3715 /* If we can't clean pages, remove dirty pages from consideration */
3716 if (!(node_reclaim_mode & RECLAIM_WRITE))
3717 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3718
3719 /* Watch for any possible underflows due to delta */
3720 if (unlikely(delta > nr_pagecache_reclaimable))
3721 delta = nr_pagecache_reclaimable;
3722
3723 return nr_pagecache_reclaimable - delta;
3724 }
3725
3726 /*
3727 * Try to free up some pages from this node through reclaim.
3728 */
3729 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3730 {
3731 /* Minimum pages needed in order to stay on node */
3732 const unsigned long nr_pages = 1 << order;
3733 struct task_struct *p = current;
3734 struct reclaim_state reclaim_state;
3735 int classzone_idx = gfp_zone(gfp_mask);
3736 struct scan_control sc = {
3737 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3738 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3739 .order = order,
3740 .priority = NODE_RECLAIM_PRIORITY,
3741 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3742 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3743 .may_swap = 1,
3744 .reclaim_idx = classzone_idx,
3745 };
3746
3747 cond_resched();
3748 /*
3749 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3750 * and we also need to be able to write out pages for RECLAIM_WRITE
3751 * and RECLAIM_UNMAP.
3752 */
3753 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3754 lockdep_set_current_reclaim_state(gfp_mask);
3755 reclaim_state.reclaimed_slab = 0;
3756 p->reclaim_state = &reclaim_state;
3757
3758 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3759 /*
3760 * Free memory by calling shrink zone with increasing
3761 * priorities until we have enough memory freed.
3762 */
3763 do {
3764 shrink_node(pgdat, &sc);
3765 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3766 }
3767
3768 p->reclaim_state = NULL;
3769 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3770 lockdep_clear_current_reclaim_state();
3771 return sc.nr_reclaimed >= nr_pages;
3772 }
3773
3774 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3775 {
3776 int ret;
3777
3778 /*
3779 * Node reclaim reclaims unmapped file backed pages and
3780 * slab pages if we are over the defined limits.
3781 *
3782 * A small portion of unmapped file backed pages is needed for
3783 * file I/O otherwise pages read by file I/O will be immediately
3784 * thrown out if the node is overallocated. So we do not reclaim
3785 * if less than a specified percentage of the node is used by
3786 * unmapped file backed pages.
3787 */
3788 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3789 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3790 return NODE_RECLAIM_FULL;
3791
3792 if (!pgdat_reclaimable(pgdat))
3793 return NODE_RECLAIM_FULL;
3794
3795 /*
3796 * Do not scan if the allocation should not be delayed.
3797 */
3798 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3799 return NODE_RECLAIM_NOSCAN;
3800
3801 /*
3802 * Only run node reclaim on the local node or on nodes that do not
3803 * have associated processors. This will favor the local processor
3804 * over remote processors and spread off node memory allocations
3805 * as wide as possible.
3806 */
3807 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3808 return NODE_RECLAIM_NOSCAN;
3809
3810 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3811 return NODE_RECLAIM_NOSCAN;
3812
3813 ret = __node_reclaim(pgdat, gfp_mask, order);
3814 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3815
3816 if (!ret)
3817 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3818
3819 return ret;
3820 }
3821 #endif
3822
3823 /*
3824 * page_evictable - test whether a page is evictable
3825 * @page: the page to test
3826 *
3827 * Test whether page is evictable--i.e., should be placed on active/inactive
3828 * lists vs unevictable list.
3829 *
3830 * Reasons page might not be evictable:
3831 * (1) page's mapping marked unevictable
3832 * (2) page is part of an mlocked VMA
3833 *
3834 */
3835 int page_evictable(struct page *page)
3836 {
3837 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3838 }
3839
3840 #ifdef CONFIG_SHMEM
3841 /**
3842 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3843 * @pages: array of pages to check
3844 * @nr_pages: number of pages to check
3845 *
3846 * Checks pages for evictability and moves them to the appropriate lru list.
3847 *
3848 * This function is only used for SysV IPC SHM_UNLOCK.
3849 */
3850 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3851 {
3852 struct lruvec *lruvec;
3853 struct pglist_data *pgdat = NULL;
3854 int pgscanned = 0;
3855 int pgrescued = 0;
3856 int i;
3857
3858 for (i = 0; i < nr_pages; i++) {
3859 struct page *page = pages[i];
3860 struct pglist_data *pagepgdat = page_pgdat(page);
3861
3862 pgscanned++;
3863 if (pagepgdat != pgdat) {
3864 if (pgdat)
3865 spin_unlock_irq(&pgdat->lru_lock);
3866 pgdat = pagepgdat;
3867 spin_lock_irq(&pgdat->lru_lock);
3868 }
3869 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3870
3871 if (!PageLRU(page) || !PageUnevictable(page))
3872 continue;
3873
3874 if (page_evictable(page)) {
3875 enum lru_list lru = page_lru_base_type(page);
3876
3877 VM_BUG_ON_PAGE(PageActive(page), page);
3878 ClearPageUnevictable(page);
3879 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3880 add_page_to_lru_list(page, lruvec, lru);
3881 pgrescued++;
3882 }
3883 }
3884
3885 if (pgdat) {
3886 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3887 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3888 spin_unlock_irq(&pgdat->lru_lock);
3889 }
3890 }
3891 #endif /* CONFIG_SHMEM */