Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swap.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * This file contains the default values for the operation of the
10 * Linux VM subsystem. Fine-tuning documentation can be found in
11 * Documentation/admin-guide/sysctl/vm.rst.
12 * Started 18.12.91
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
15 */
16
17#include <linux/mm.h>
18#include <linux/sched.h>
19#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/mman.h>
22#include <linux/pagemap.h>
23#include <linux/pagevec.h>
24#include <linux/init.h>
25#include <linux/export.h>
26#include <linux/mm_inline.h>
27#include <linux/percpu_counter.h>
28#include <linux/memremap.h>
29#include <linux/percpu.h>
30#include <linux/cpu.h>
31#include <linux/notifier.h>
32#include <linux/backing-dev.h>
33#include <linux/memcontrol.h>
34#include <linux/gfp.h>
35#include <linux/uio.h>
36#include <linux/hugetlb.h>
37#include <linux/page_idle.h>
38#include <linux/local_lock.h>
39#include <linux/buffer_head.h>
40
41#include "internal.h"
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/pagemap.h>
45
46/* How many pages do we try to swap or page in/out together? As a power of 2 */
47int page_cluster;
48const int page_cluster_max = 31;
49
50/* Protecting only lru_rotate.fbatch which requires disabling interrupts */
51struct lru_rotate {
52 local_lock_t lock;
53 struct folio_batch fbatch;
54};
55static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
56 .lock = INIT_LOCAL_LOCK(lock),
57};
58
59/*
60 * The following folio batches are grouped together because they are protected
61 * by disabling preemption (and interrupts remain enabled).
62 */
63struct cpu_fbatches {
64 local_lock_t lock;
65 struct folio_batch lru_add;
66 struct folio_batch lru_deactivate_file;
67 struct folio_batch lru_deactivate;
68 struct folio_batch lru_lazyfree;
69#ifdef CONFIG_SMP
70 struct folio_batch activate;
71#endif
72};
73static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
74 .lock = INIT_LOCAL_LOCK(lock),
75};
76
77/*
78 * This path almost never happens for VM activity - pages are normally freed
79 * in batches. But it gets used by networking - and for compound pages.
80 */
81static void __page_cache_release(struct folio *folio)
82{
83 if (folio_test_lru(folio)) {
84 struct lruvec *lruvec;
85 unsigned long flags;
86
87 lruvec = folio_lruvec_lock_irqsave(folio, &flags);
88 lruvec_del_folio(lruvec, folio);
89 __folio_clear_lru_flags(folio);
90 unlock_page_lruvec_irqrestore(lruvec, flags);
91 }
92 /* See comment on folio_test_mlocked in release_pages() */
93 if (unlikely(folio_test_mlocked(folio))) {
94 long nr_pages = folio_nr_pages(folio);
95
96 __folio_clear_mlocked(folio);
97 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
98 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
99 }
100}
101
102static void __folio_put_small(struct folio *folio)
103{
104 __page_cache_release(folio);
105 mem_cgroup_uncharge(folio);
106 free_unref_page(&folio->page, 0);
107}
108
109static void __folio_put_large(struct folio *folio)
110{
111 /*
112 * __page_cache_release() is supposed to be called for thp, not for
113 * hugetlb. This is because hugetlb page does never have PageLRU set
114 * (it's never listed to any LRU lists) and no memcg routines should
115 * be called for hugetlb (it has a separate hugetlb_cgroup.)
116 */
117 if (!folio_test_hugetlb(folio))
118 __page_cache_release(folio);
119 destroy_large_folio(folio);
120}
121
122void __folio_put(struct folio *folio)
123{
124 if (unlikely(folio_is_zone_device(folio)))
125 free_zone_device_page(&folio->page);
126 else if (unlikely(folio_test_large(folio)))
127 __folio_put_large(folio);
128 else
129 __folio_put_small(folio);
130}
131EXPORT_SYMBOL(__folio_put);
132
133/**
134 * put_pages_list() - release a list of pages
135 * @pages: list of pages threaded on page->lru
136 *
137 * Release a list of pages which are strung together on page.lru.
138 */
139void put_pages_list(struct list_head *pages)
140{
141 struct folio *folio, *next;
142
143 list_for_each_entry_safe(folio, next, pages, lru) {
144 if (!folio_put_testzero(folio)) {
145 list_del(&folio->lru);
146 continue;
147 }
148 if (folio_test_large(folio)) {
149 list_del(&folio->lru);
150 __folio_put_large(folio);
151 continue;
152 }
153 /* LRU flag must be clear because it's passed using the lru */
154 }
155
156 free_unref_page_list(pages);
157 INIT_LIST_HEAD(pages);
158}
159EXPORT_SYMBOL(put_pages_list);
160
161typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
162
163static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
164{
165 int was_unevictable = folio_test_clear_unevictable(folio);
166 long nr_pages = folio_nr_pages(folio);
167
168 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
169
170 /*
171 * Is an smp_mb__after_atomic() still required here, before
172 * folio_evictable() tests the mlocked flag, to rule out the possibility
173 * of stranding an evictable folio on an unevictable LRU? I think
174 * not, because __munlock_folio() only clears the mlocked flag
175 * while the LRU lock is held.
176 *
177 * (That is not true of __page_cache_release(), and not necessarily
178 * true of release_pages(): but those only clear the mlocked flag after
179 * folio_put_testzero() has excluded any other users of the folio.)
180 */
181 if (folio_evictable(folio)) {
182 if (was_unevictable)
183 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
184 } else {
185 folio_clear_active(folio);
186 folio_set_unevictable(folio);
187 /*
188 * folio->mlock_count = !!folio_test_mlocked(folio)?
189 * But that leaves __mlock_folio() in doubt whether another
190 * actor has already counted the mlock or not. Err on the
191 * safe side, underestimate, let page reclaim fix it, rather
192 * than leaving a page on the unevictable LRU indefinitely.
193 */
194 folio->mlock_count = 0;
195 if (!was_unevictable)
196 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
197 }
198
199 lruvec_add_folio(lruvec, folio);
200 trace_mm_lru_insertion(folio);
201}
202
203static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
204{
205 int i;
206 struct lruvec *lruvec = NULL;
207 unsigned long flags = 0;
208
209 for (i = 0; i < folio_batch_count(fbatch); i++) {
210 struct folio *folio = fbatch->folios[i];
211
212 /* block memcg migration while the folio moves between lru */
213 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
214 continue;
215
216 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
217 move_fn(lruvec, folio);
218
219 folio_set_lru(folio);
220 }
221
222 if (lruvec)
223 unlock_page_lruvec_irqrestore(lruvec, flags);
224 folios_put(fbatch->folios, folio_batch_count(fbatch));
225 folio_batch_reinit(fbatch);
226}
227
228static void folio_batch_add_and_move(struct folio_batch *fbatch,
229 struct folio *folio, move_fn_t move_fn)
230{
231 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
232 !lru_cache_disabled())
233 return;
234 folio_batch_move_lru(fbatch, move_fn);
235}
236
237static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
238{
239 if (!folio_test_unevictable(folio)) {
240 lruvec_del_folio(lruvec, folio);
241 folio_clear_active(folio);
242 lruvec_add_folio_tail(lruvec, folio);
243 __count_vm_events(PGROTATED, folio_nr_pages(folio));
244 }
245}
246
247/*
248 * Writeback is about to end against a folio which has been marked for
249 * immediate reclaim. If it still appears to be reclaimable, move it
250 * to the tail of the inactive list.
251 *
252 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
253 */
254void folio_rotate_reclaimable(struct folio *folio)
255{
256 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
257 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
258 struct folio_batch *fbatch;
259 unsigned long flags;
260
261 folio_get(folio);
262 local_lock_irqsave(&lru_rotate.lock, flags);
263 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
264 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
265 local_unlock_irqrestore(&lru_rotate.lock, flags);
266 }
267}
268
269void lru_note_cost(struct lruvec *lruvec, bool file,
270 unsigned int nr_io, unsigned int nr_rotated)
271{
272 unsigned long cost;
273
274 /*
275 * Reflect the relative cost of incurring IO and spending CPU
276 * time on rotations. This doesn't attempt to make a precise
277 * comparison, it just says: if reloads are about comparable
278 * between the LRU lists, or rotations are overwhelmingly
279 * different between them, adjust scan balance for CPU work.
280 */
281 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
282
283 do {
284 unsigned long lrusize;
285
286 /*
287 * Hold lruvec->lru_lock is safe here, since
288 * 1) The pinned lruvec in reclaim, or
289 * 2) From a pre-LRU page during refault (which also holds the
290 * rcu lock, so would be safe even if the page was on the LRU
291 * and could move simultaneously to a new lruvec).
292 */
293 spin_lock_irq(&lruvec->lru_lock);
294 /* Record cost event */
295 if (file)
296 lruvec->file_cost += cost;
297 else
298 lruvec->anon_cost += cost;
299
300 /*
301 * Decay previous events
302 *
303 * Because workloads change over time (and to avoid
304 * overflow) we keep these statistics as a floating
305 * average, which ends up weighing recent refaults
306 * more than old ones.
307 */
308 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
309 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
310 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
311 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
312
313 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
314 lruvec->file_cost /= 2;
315 lruvec->anon_cost /= 2;
316 }
317 spin_unlock_irq(&lruvec->lru_lock);
318 } while ((lruvec = parent_lruvec(lruvec)));
319}
320
321void lru_note_cost_refault(struct folio *folio)
322{
323 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
324 folio_nr_pages(folio), 0);
325}
326
327static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
328{
329 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
330 long nr_pages = folio_nr_pages(folio);
331
332 lruvec_del_folio(lruvec, folio);
333 folio_set_active(folio);
334 lruvec_add_folio(lruvec, folio);
335 trace_mm_lru_activate(folio);
336
337 __count_vm_events(PGACTIVATE, nr_pages);
338 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
339 nr_pages);
340 }
341}
342
343#ifdef CONFIG_SMP
344static void folio_activate_drain(int cpu)
345{
346 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
347
348 if (folio_batch_count(fbatch))
349 folio_batch_move_lru(fbatch, folio_activate_fn);
350}
351
352void folio_activate(struct folio *folio)
353{
354 if (folio_test_lru(folio) && !folio_test_active(folio) &&
355 !folio_test_unevictable(folio)) {
356 struct folio_batch *fbatch;
357
358 folio_get(folio);
359 local_lock(&cpu_fbatches.lock);
360 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
361 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
362 local_unlock(&cpu_fbatches.lock);
363 }
364}
365
366#else
367static inline void folio_activate_drain(int cpu)
368{
369}
370
371void folio_activate(struct folio *folio)
372{
373 struct lruvec *lruvec;
374
375 if (folio_test_clear_lru(folio)) {
376 lruvec = folio_lruvec_lock_irq(folio);
377 folio_activate_fn(lruvec, folio);
378 unlock_page_lruvec_irq(lruvec);
379 folio_set_lru(folio);
380 }
381}
382#endif
383
384static void __lru_cache_activate_folio(struct folio *folio)
385{
386 struct folio_batch *fbatch;
387 int i;
388
389 local_lock(&cpu_fbatches.lock);
390 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
391
392 /*
393 * Search backwards on the optimistic assumption that the folio being
394 * activated has just been added to this batch. Note that only
395 * the local batch is examined as a !LRU folio could be in the
396 * process of being released, reclaimed, migrated or on a remote
397 * batch that is currently being drained. Furthermore, marking
398 * a remote batch's folio active potentially hits a race where
399 * a folio is marked active just after it is added to the inactive
400 * list causing accounting errors and BUG_ON checks to trigger.
401 */
402 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
403 struct folio *batch_folio = fbatch->folios[i];
404
405 if (batch_folio == folio) {
406 folio_set_active(folio);
407 break;
408 }
409 }
410
411 local_unlock(&cpu_fbatches.lock);
412}
413
414#ifdef CONFIG_LRU_GEN
415static void folio_inc_refs(struct folio *folio)
416{
417 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
418
419 if (folio_test_unevictable(folio))
420 return;
421
422 if (!folio_test_referenced(folio)) {
423 folio_set_referenced(folio);
424 return;
425 }
426
427 if (!folio_test_workingset(folio)) {
428 folio_set_workingset(folio);
429 return;
430 }
431
432 /* see the comment on MAX_NR_TIERS */
433 do {
434 new_flags = old_flags & LRU_REFS_MASK;
435 if (new_flags == LRU_REFS_MASK)
436 break;
437
438 new_flags += BIT(LRU_REFS_PGOFF);
439 new_flags |= old_flags & ~LRU_REFS_MASK;
440 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
441}
442#else
443static void folio_inc_refs(struct folio *folio)
444{
445}
446#endif /* CONFIG_LRU_GEN */
447
448/*
449 * Mark a page as having seen activity.
450 *
451 * inactive,unreferenced -> inactive,referenced
452 * inactive,referenced -> active,unreferenced
453 * active,unreferenced -> active,referenced
454 *
455 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
456 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
457 */
458void folio_mark_accessed(struct folio *folio)
459{
460 if (lru_gen_enabled()) {
461 folio_inc_refs(folio);
462 return;
463 }
464
465 if (!folio_test_referenced(folio)) {
466 folio_set_referenced(folio);
467 } else if (folio_test_unevictable(folio)) {
468 /*
469 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
470 * this list is never rotated or maintained, so marking an
471 * unevictable page accessed has no effect.
472 */
473 } else if (!folio_test_active(folio)) {
474 /*
475 * If the folio is on the LRU, queue it for activation via
476 * cpu_fbatches.activate. Otherwise, assume the folio is in a
477 * folio_batch, mark it active and it'll be moved to the active
478 * LRU on the next drain.
479 */
480 if (folio_test_lru(folio))
481 folio_activate(folio);
482 else
483 __lru_cache_activate_folio(folio);
484 folio_clear_referenced(folio);
485 workingset_activation(folio);
486 }
487 if (folio_test_idle(folio))
488 folio_clear_idle(folio);
489}
490EXPORT_SYMBOL(folio_mark_accessed);
491
492/**
493 * folio_add_lru - Add a folio to an LRU list.
494 * @folio: The folio to be added to the LRU.
495 *
496 * Queue the folio for addition to the LRU. The decision on whether
497 * to add the page to the [in]active [file|anon] list is deferred until the
498 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
499 * have the folio added to the active list using folio_mark_accessed().
500 */
501void folio_add_lru(struct folio *folio)
502{
503 struct folio_batch *fbatch;
504
505 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
506 folio_test_unevictable(folio), folio);
507 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
508
509 /* see the comment in lru_gen_add_folio() */
510 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
511 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
512 folio_set_active(folio);
513
514 folio_get(folio);
515 local_lock(&cpu_fbatches.lock);
516 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
517 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
518 local_unlock(&cpu_fbatches.lock);
519}
520EXPORT_SYMBOL(folio_add_lru);
521
522/**
523 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
524 * @folio: The folio to be added to the LRU.
525 * @vma: VMA in which the folio is mapped.
526 *
527 * If the VMA is mlocked, @folio is added to the unevictable list.
528 * Otherwise, it is treated the same way as folio_add_lru().
529 */
530void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
531{
532 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
533
534 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
535 mlock_new_folio(folio);
536 else
537 folio_add_lru(folio);
538}
539
540/*
541 * If the folio cannot be invalidated, it is moved to the
542 * inactive list to speed up its reclaim. It is moved to the
543 * head of the list, rather than the tail, to give the flusher
544 * threads some time to write it out, as this is much more
545 * effective than the single-page writeout from reclaim.
546 *
547 * If the folio isn't mapped and dirty/writeback, the folio
548 * could be reclaimed asap using the reclaim flag.
549 *
550 * 1. active, mapped folio -> none
551 * 2. active, dirty/writeback folio -> inactive, head, reclaim
552 * 3. inactive, mapped folio -> none
553 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
554 * 5. inactive, clean -> inactive, tail
555 * 6. Others -> none
556 *
557 * In 4, it moves to the head of the inactive list so the folio is
558 * written out by flusher threads as this is much more efficient
559 * than the single-page writeout from reclaim.
560 */
561static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
562{
563 bool active = folio_test_active(folio);
564 long nr_pages = folio_nr_pages(folio);
565
566 if (folio_test_unevictable(folio))
567 return;
568
569 /* Some processes are using the folio */
570 if (folio_mapped(folio))
571 return;
572
573 lruvec_del_folio(lruvec, folio);
574 folio_clear_active(folio);
575 folio_clear_referenced(folio);
576
577 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
578 /*
579 * Setting the reclaim flag could race with
580 * folio_end_writeback() and confuse readahead. But the
581 * race window is _really_ small and it's not a critical
582 * problem.
583 */
584 lruvec_add_folio(lruvec, folio);
585 folio_set_reclaim(folio);
586 } else {
587 /*
588 * The folio's writeback ended while it was in the batch.
589 * We move that folio to the tail of the inactive list.
590 */
591 lruvec_add_folio_tail(lruvec, folio);
592 __count_vm_events(PGROTATED, nr_pages);
593 }
594
595 if (active) {
596 __count_vm_events(PGDEACTIVATE, nr_pages);
597 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
598 nr_pages);
599 }
600}
601
602static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
603{
604 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
605 long nr_pages = folio_nr_pages(folio);
606
607 lruvec_del_folio(lruvec, folio);
608 folio_clear_active(folio);
609 folio_clear_referenced(folio);
610 lruvec_add_folio(lruvec, folio);
611
612 __count_vm_events(PGDEACTIVATE, nr_pages);
613 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
614 nr_pages);
615 }
616}
617
618static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
619{
620 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
621 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
622 long nr_pages = folio_nr_pages(folio);
623
624 lruvec_del_folio(lruvec, folio);
625 folio_clear_active(folio);
626 folio_clear_referenced(folio);
627 /*
628 * Lazyfree folios are clean anonymous folios. They have
629 * the swapbacked flag cleared, to distinguish them from normal
630 * anonymous folios
631 */
632 folio_clear_swapbacked(folio);
633 lruvec_add_folio(lruvec, folio);
634
635 __count_vm_events(PGLAZYFREE, nr_pages);
636 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
637 nr_pages);
638 }
639}
640
641/*
642 * Drain pages out of the cpu's folio_batch.
643 * Either "cpu" is the current CPU, and preemption has already been
644 * disabled; or "cpu" is being hot-unplugged, and is already dead.
645 */
646void lru_add_drain_cpu(int cpu)
647{
648 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
649 struct folio_batch *fbatch = &fbatches->lru_add;
650
651 if (folio_batch_count(fbatch))
652 folio_batch_move_lru(fbatch, lru_add_fn);
653
654 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
655 /* Disabling interrupts below acts as a compiler barrier. */
656 if (data_race(folio_batch_count(fbatch))) {
657 unsigned long flags;
658
659 /* No harm done if a racing interrupt already did this */
660 local_lock_irqsave(&lru_rotate.lock, flags);
661 folio_batch_move_lru(fbatch, lru_move_tail_fn);
662 local_unlock_irqrestore(&lru_rotate.lock, flags);
663 }
664
665 fbatch = &fbatches->lru_deactivate_file;
666 if (folio_batch_count(fbatch))
667 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
668
669 fbatch = &fbatches->lru_deactivate;
670 if (folio_batch_count(fbatch))
671 folio_batch_move_lru(fbatch, lru_deactivate_fn);
672
673 fbatch = &fbatches->lru_lazyfree;
674 if (folio_batch_count(fbatch))
675 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
676
677 folio_activate_drain(cpu);
678}
679
680/**
681 * deactivate_file_folio() - Deactivate a file folio.
682 * @folio: Folio to deactivate.
683 *
684 * This function hints to the VM that @folio is a good reclaim candidate,
685 * for example if its invalidation fails due to the folio being dirty
686 * or under writeback.
687 *
688 * Context: Caller holds a reference on the folio.
689 */
690void deactivate_file_folio(struct folio *folio)
691{
692 struct folio_batch *fbatch;
693
694 /* Deactivating an unevictable folio will not accelerate reclaim */
695 if (folio_test_unevictable(folio))
696 return;
697
698 folio_get(folio);
699 local_lock(&cpu_fbatches.lock);
700 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
701 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
702 local_unlock(&cpu_fbatches.lock);
703}
704
705/*
706 * folio_deactivate - deactivate a folio
707 * @folio: folio to deactivate
708 *
709 * folio_deactivate() moves @folio to the inactive list if @folio was on the
710 * active list and was not unevictable. This is done to accelerate the
711 * reclaim of @folio.
712 */
713void folio_deactivate(struct folio *folio)
714{
715 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
716 (folio_test_active(folio) || lru_gen_enabled())) {
717 struct folio_batch *fbatch;
718
719 folio_get(folio);
720 local_lock(&cpu_fbatches.lock);
721 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
722 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
723 local_unlock(&cpu_fbatches.lock);
724 }
725}
726
727/**
728 * folio_mark_lazyfree - make an anon folio lazyfree
729 * @folio: folio to deactivate
730 *
731 * folio_mark_lazyfree() moves @folio to the inactive file list.
732 * This is done to accelerate the reclaim of @folio.
733 */
734void folio_mark_lazyfree(struct folio *folio)
735{
736 if (folio_test_lru(folio) && folio_test_anon(folio) &&
737 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
738 !folio_test_unevictable(folio)) {
739 struct folio_batch *fbatch;
740
741 folio_get(folio);
742 local_lock(&cpu_fbatches.lock);
743 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
744 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
745 local_unlock(&cpu_fbatches.lock);
746 }
747}
748
749void lru_add_drain(void)
750{
751 local_lock(&cpu_fbatches.lock);
752 lru_add_drain_cpu(smp_processor_id());
753 local_unlock(&cpu_fbatches.lock);
754 mlock_drain_local();
755}
756
757/*
758 * It's called from per-cpu workqueue context in SMP case so
759 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
760 * the same cpu. It shouldn't be a problem in !SMP case since
761 * the core is only one and the locks will disable preemption.
762 */
763static void lru_add_and_bh_lrus_drain(void)
764{
765 local_lock(&cpu_fbatches.lock);
766 lru_add_drain_cpu(smp_processor_id());
767 local_unlock(&cpu_fbatches.lock);
768 invalidate_bh_lrus_cpu();
769 mlock_drain_local();
770}
771
772void lru_add_drain_cpu_zone(struct zone *zone)
773{
774 local_lock(&cpu_fbatches.lock);
775 lru_add_drain_cpu(smp_processor_id());
776 drain_local_pages(zone);
777 local_unlock(&cpu_fbatches.lock);
778 mlock_drain_local();
779}
780
781#ifdef CONFIG_SMP
782
783static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
784
785static void lru_add_drain_per_cpu(struct work_struct *dummy)
786{
787 lru_add_and_bh_lrus_drain();
788}
789
790static bool cpu_needs_drain(unsigned int cpu)
791{
792 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
793
794 /* Check these in order of likelihood that they're not zero */
795 return folio_batch_count(&fbatches->lru_add) ||
796 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
797 folio_batch_count(&fbatches->lru_deactivate_file) ||
798 folio_batch_count(&fbatches->lru_deactivate) ||
799 folio_batch_count(&fbatches->lru_lazyfree) ||
800 folio_batch_count(&fbatches->activate) ||
801 need_mlock_drain(cpu) ||
802 has_bh_in_lru(cpu, NULL);
803}
804
805/*
806 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
807 * kworkers being shut down before our page_alloc_cpu_dead callback is
808 * executed on the offlined cpu.
809 * Calling this function with cpu hotplug locks held can actually lead
810 * to obscure indirect dependencies via WQ context.
811 */
812static inline void __lru_add_drain_all(bool force_all_cpus)
813{
814 /*
815 * lru_drain_gen - Global pages generation number
816 *
817 * (A) Definition: global lru_drain_gen = x implies that all generations
818 * 0 < n <= x are already *scheduled* for draining.
819 *
820 * This is an optimization for the highly-contended use case where a
821 * user space workload keeps constantly generating a flow of pages for
822 * each CPU.
823 */
824 static unsigned int lru_drain_gen;
825 static struct cpumask has_work;
826 static DEFINE_MUTEX(lock);
827 unsigned cpu, this_gen;
828
829 /*
830 * Make sure nobody triggers this path before mm_percpu_wq is fully
831 * initialized.
832 */
833 if (WARN_ON(!mm_percpu_wq))
834 return;
835
836 /*
837 * Guarantee folio_batch counter stores visible by this CPU
838 * are visible to other CPUs before loading the current drain
839 * generation.
840 */
841 smp_mb();
842
843 /*
844 * (B) Locally cache global LRU draining generation number
845 *
846 * The read barrier ensures that the counter is loaded before the mutex
847 * is taken. It pairs with smp_mb() inside the mutex critical section
848 * at (D).
849 */
850 this_gen = smp_load_acquire(&lru_drain_gen);
851
852 mutex_lock(&lock);
853
854 /*
855 * (C) Exit the draining operation if a newer generation, from another
856 * lru_add_drain_all(), was already scheduled for draining. Check (A).
857 */
858 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
859 goto done;
860
861 /*
862 * (D) Increment global generation number
863 *
864 * Pairs with smp_load_acquire() at (B), outside of the critical
865 * section. Use a full memory barrier to guarantee that the
866 * new global drain generation number is stored before loading
867 * folio_batch counters.
868 *
869 * This pairing must be done here, before the for_each_online_cpu loop
870 * below which drains the page vectors.
871 *
872 * Let x, y, and z represent some system CPU numbers, where x < y < z.
873 * Assume CPU #z is in the middle of the for_each_online_cpu loop
874 * below and has already reached CPU #y's per-cpu data. CPU #x comes
875 * along, adds some pages to its per-cpu vectors, then calls
876 * lru_add_drain_all().
877 *
878 * If the paired barrier is done at any later step, e.g. after the
879 * loop, CPU #x will just exit at (C) and miss flushing out all of its
880 * added pages.
881 */
882 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
883 smp_mb();
884
885 cpumask_clear(&has_work);
886 for_each_online_cpu(cpu) {
887 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
888
889 if (cpu_needs_drain(cpu)) {
890 INIT_WORK(work, lru_add_drain_per_cpu);
891 queue_work_on(cpu, mm_percpu_wq, work);
892 __cpumask_set_cpu(cpu, &has_work);
893 }
894 }
895
896 for_each_cpu(cpu, &has_work)
897 flush_work(&per_cpu(lru_add_drain_work, cpu));
898
899done:
900 mutex_unlock(&lock);
901}
902
903void lru_add_drain_all(void)
904{
905 __lru_add_drain_all(false);
906}
907#else
908void lru_add_drain_all(void)
909{
910 lru_add_drain();
911}
912#endif /* CONFIG_SMP */
913
914atomic_t lru_disable_count = ATOMIC_INIT(0);
915
916/*
917 * lru_cache_disable() needs to be called before we start compiling
918 * a list of pages to be migrated using isolate_lru_page().
919 * It drains pages on LRU cache and then disable on all cpus until
920 * lru_cache_enable is called.
921 *
922 * Must be paired with a call to lru_cache_enable().
923 */
924void lru_cache_disable(void)
925{
926 atomic_inc(&lru_disable_count);
927 /*
928 * Readers of lru_disable_count are protected by either disabling
929 * preemption or rcu_read_lock:
930 *
931 * preempt_disable, local_irq_disable [bh_lru_lock()]
932 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
933 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
934 *
935 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
936 * preempt_disable() regions of code. So any CPU which sees
937 * lru_disable_count = 0 will have exited the critical
938 * section when synchronize_rcu() returns.
939 */
940 synchronize_rcu_expedited();
941#ifdef CONFIG_SMP
942 __lru_add_drain_all(true);
943#else
944 lru_add_and_bh_lrus_drain();
945#endif
946}
947
948/**
949 * release_pages - batched put_page()
950 * @arg: array of pages to release
951 * @nr: number of pages
952 *
953 * Decrement the reference count on all the pages in @arg. If it
954 * fell to zero, remove the page from the LRU and free it.
955 *
956 * Note that the argument can be an array of pages, encoded pages,
957 * or folio pointers. We ignore any encoded bits, and turn any of
958 * them into just a folio that gets free'd.
959 */
960void release_pages(release_pages_arg arg, int nr)
961{
962 int i;
963 struct encoded_page **encoded = arg.encoded_pages;
964 LIST_HEAD(pages_to_free);
965 struct lruvec *lruvec = NULL;
966 unsigned long flags = 0;
967 unsigned int lock_batch;
968
969 for (i = 0; i < nr; i++) {
970 struct folio *folio;
971
972 /* Turn any of the argument types into a folio */
973 folio = page_folio(encoded_page_ptr(encoded[i]));
974
975 /*
976 * Make sure the IRQ-safe lock-holding time does not get
977 * excessive with a continuous string of pages from the
978 * same lruvec. The lock is held only if lruvec != NULL.
979 */
980 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
981 unlock_page_lruvec_irqrestore(lruvec, flags);
982 lruvec = NULL;
983 }
984
985 if (is_huge_zero_page(&folio->page))
986 continue;
987
988 if (folio_is_zone_device(folio)) {
989 if (lruvec) {
990 unlock_page_lruvec_irqrestore(lruvec, flags);
991 lruvec = NULL;
992 }
993 if (put_devmap_managed_page(&folio->page))
994 continue;
995 if (folio_put_testzero(folio))
996 free_zone_device_page(&folio->page);
997 continue;
998 }
999
1000 if (!folio_put_testzero(folio))
1001 continue;
1002
1003 if (folio_test_large(folio)) {
1004 if (lruvec) {
1005 unlock_page_lruvec_irqrestore(lruvec, flags);
1006 lruvec = NULL;
1007 }
1008 __folio_put_large(folio);
1009 continue;
1010 }
1011
1012 if (folio_test_lru(folio)) {
1013 struct lruvec *prev_lruvec = lruvec;
1014
1015 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1016 &flags);
1017 if (prev_lruvec != lruvec)
1018 lock_batch = 0;
1019
1020 lruvec_del_folio(lruvec, folio);
1021 __folio_clear_lru_flags(folio);
1022 }
1023
1024 /*
1025 * In rare cases, when truncation or holepunching raced with
1026 * munlock after VM_LOCKED was cleared, Mlocked may still be
1027 * found set here. This does not indicate a problem, unless
1028 * "unevictable_pgs_cleared" appears worryingly large.
1029 */
1030 if (unlikely(folio_test_mlocked(folio))) {
1031 __folio_clear_mlocked(folio);
1032 zone_stat_sub_folio(folio, NR_MLOCK);
1033 count_vm_event(UNEVICTABLE_PGCLEARED);
1034 }
1035
1036 list_add(&folio->lru, &pages_to_free);
1037 }
1038 if (lruvec)
1039 unlock_page_lruvec_irqrestore(lruvec, flags);
1040
1041 mem_cgroup_uncharge_list(&pages_to_free);
1042 free_unref_page_list(&pages_to_free);
1043}
1044EXPORT_SYMBOL(release_pages);
1045
1046/*
1047 * The folios which we're about to release may be in the deferred lru-addition
1048 * queues. That would prevent them from really being freed right now. That's
1049 * OK from a correctness point of view but is inefficient - those folios may be
1050 * cache-warm and we want to give them back to the page allocator ASAP.
1051 *
1052 * So __folio_batch_release() will drain those queues here.
1053 * folio_batch_move_lru() calls folios_put() directly to avoid
1054 * mutual recursion.
1055 */
1056void __folio_batch_release(struct folio_batch *fbatch)
1057{
1058 if (!fbatch->percpu_pvec_drained) {
1059 lru_add_drain();
1060 fbatch->percpu_pvec_drained = true;
1061 }
1062 release_pages(fbatch->folios, folio_batch_count(fbatch));
1063 folio_batch_reinit(fbatch);
1064}
1065EXPORT_SYMBOL(__folio_batch_release);
1066
1067/**
1068 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1069 * @fbatch: The batch to prune
1070 *
1071 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1072 * entries. This function prunes all the non-folio entries from @fbatch
1073 * without leaving holes, so that it can be passed on to folio-only batch
1074 * operations.
1075 */
1076void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1077{
1078 unsigned int i, j;
1079
1080 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1081 struct folio *folio = fbatch->folios[i];
1082 if (!xa_is_value(folio))
1083 fbatch->folios[j++] = folio;
1084 }
1085 fbatch->nr = j;
1086}
1087
1088/*
1089 * Perform any setup for the swap system
1090 */
1091void __init swap_setup(void)
1092{
1093 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1094
1095 /* Use a smaller cluster for small-memory machines */
1096 if (megs < 16)
1097 page_cluster = 2;
1098 else
1099 page_cluster = 3;
1100 /*
1101 * Right now other parts of the system means that we
1102 * _really_ don't want to cluster much more
1103 */
1104}
1/*
2 * linux/mm/swap.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * This file contains the default values for the operation of the
9 * Linux VM subsystem. Fine-tuning documentation can be found in
10 * Documentation/sysctl/vm.txt.
11 * Started 18.12.91
12 * Swap aging added 23.2.95, Stephen Tweedie.
13 * Buffermem limits added 12.3.98, Rik van Riel.
14 */
15
16#include <linux/mm.h>
17#include <linux/sched.h>
18#include <linux/kernel_stat.h>
19#include <linux/swap.h>
20#include <linux/mman.h>
21#include <linux/pagemap.h>
22#include <linux/pagevec.h>
23#include <linux/init.h>
24#include <linux/export.h>
25#include <linux/mm_inline.h>
26#include <linux/percpu_counter.h>
27#include <linux/memremap.h>
28#include <linux/percpu.h>
29#include <linux/cpu.h>
30#include <linux/notifier.h>
31#include <linux/backing-dev.h>
32#include <linux/memcontrol.h>
33#include <linux/gfp.h>
34#include <linux/uio.h>
35#include <linux/hugetlb.h>
36#include <linux/page_idle.h>
37
38#include "internal.h"
39
40#define CREATE_TRACE_POINTS
41#include <trace/events/pagemap.h>
42
43/* How many pages do we try to swap or page in/out together? */
44int page_cluster;
45
46static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
47static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
48static DEFINE_PER_CPU(struct pagevec, lru_deactivate_file_pvecs);
49static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
50
51/*
52 * This path almost never happens for VM activity - pages are normally
53 * freed via pagevecs. But it gets used by networking.
54 */
55static void __page_cache_release(struct page *page)
56{
57 if (PageLRU(page)) {
58 struct zone *zone = page_zone(page);
59 struct lruvec *lruvec;
60 unsigned long flags;
61
62 spin_lock_irqsave(&zone->lru_lock, flags);
63 lruvec = mem_cgroup_page_lruvec(page, zone);
64 VM_BUG_ON_PAGE(!PageLRU(page), page);
65 __ClearPageLRU(page);
66 del_page_from_lru_list(page, lruvec, page_off_lru(page));
67 spin_unlock_irqrestore(&zone->lru_lock, flags);
68 }
69 mem_cgroup_uncharge(page);
70}
71
72static void __put_single_page(struct page *page)
73{
74 __page_cache_release(page);
75 free_hot_cold_page(page, false);
76}
77
78static void __put_compound_page(struct page *page)
79{
80 compound_page_dtor *dtor;
81
82 /*
83 * __page_cache_release() is supposed to be called for thp, not for
84 * hugetlb. This is because hugetlb page does never have PageLRU set
85 * (it's never listed to any LRU lists) and no memcg routines should
86 * be called for hugetlb (it has a separate hugetlb_cgroup.)
87 */
88 if (!PageHuge(page))
89 __page_cache_release(page);
90 dtor = get_compound_page_dtor(page);
91 (*dtor)(page);
92}
93
94void __put_page(struct page *page)
95{
96 if (unlikely(PageCompound(page)))
97 __put_compound_page(page);
98 else
99 __put_single_page(page);
100}
101EXPORT_SYMBOL(__put_page);
102
103/**
104 * put_pages_list() - release a list of pages
105 * @pages: list of pages threaded on page->lru
106 *
107 * Release a list of pages which are strung together on page.lru. Currently
108 * used by read_cache_pages() and related error recovery code.
109 */
110void put_pages_list(struct list_head *pages)
111{
112 while (!list_empty(pages)) {
113 struct page *victim;
114
115 victim = list_entry(pages->prev, struct page, lru);
116 list_del(&victim->lru);
117 put_page(victim);
118 }
119}
120EXPORT_SYMBOL(put_pages_list);
121
122/*
123 * get_kernel_pages() - pin kernel pages in memory
124 * @kiov: An array of struct kvec structures
125 * @nr_segs: number of segments to pin
126 * @write: pinning for read/write, currently ignored
127 * @pages: array that receives pointers to the pages pinned.
128 * Should be at least nr_segs long.
129 *
130 * Returns number of pages pinned. This may be fewer than the number
131 * requested. If nr_pages is 0 or negative, returns 0. If no pages
132 * were pinned, returns -errno. Each page returned must be released
133 * with a put_page() call when it is finished with.
134 */
135int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
136 struct page **pages)
137{
138 int seg;
139
140 for (seg = 0; seg < nr_segs; seg++) {
141 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
142 return seg;
143
144 pages[seg] = kmap_to_page(kiov[seg].iov_base);
145 get_page(pages[seg]);
146 }
147
148 return seg;
149}
150EXPORT_SYMBOL_GPL(get_kernel_pages);
151
152/*
153 * get_kernel_page() - pin a kernel page in memory
154 * @start: starting kernel address
155 * @write: pinning for read/write, currently ignored
156 * @pages: array that receives pointer to the page pinned.
157 * Must be at least nr_segs long.
158 *
159 * Returns 1 if page is pinned. If the page was not pinned, returns
160 * -errno. The page returned must be released with a put_page() call
161 * when it is finished with.
162 */
163int get_kernel_page(unsigned long start, int write, struct page **pages)
164{
165 const struct kvec kiov = {
166 .iov_base = (void *)start,
167 .iov_len = PAGE_SIZE
168 };
169
170 return get_kernel_pages(&kiov, 1, write, pages);
171}
172EXPORT_SYMBOL_GPL(get_kernel_page);
173
174static void pagevec_lru_move_fn(struct pagevec *pvec,
175 void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
176 void *arg)
177{
178 int i;
179 struct zone *zone = NULL;
180 struct lruvec *lruvec;
181 unsigned long flags = 0;
182
183 for (i = 0; i < pagevec_count(pvec); i++) {
184 struct page *page = pvec->pages[i];
185 struct zone *pagezone = page_zone(page);
186
187 if (pagezone != zone) {
188 if (zone)
189 spin_unlock_irqrestore(&zone->lru_lock, flags);
190 zone = pagezone;
191 spin_lock_irqsave(&zone->lru_lock, flags);
192 }
193
194 lruvec = mem_cgroup_page_lruvec(page, zone);
195 (*move_fn)(page, lruvec, arg);
196 }
197 if (zone)
198 spin_unlock_irqrestore(&zone->lru_lock, flags);
199 release_pages(pvec->pages, pvec->nr, pvec->cold);
200 pagevec_reinit(pvec);
201}
202
203static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
204 void *arg)
205{
206 int *pgmoved = arg;
207
208 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
209 enum lru_list lru = page_lru_base_type(page);
210 list_move_tail(&page->lru, &lruvec->lists[lru]);
211 (*pgmoved)++;
212 }
213}
214
215/*
216 * pagevec_move_tail() must be called with IRQ disabled.
217 * Otherwise this may cause nasty races.
218 */
219static void pagevec_move_tail(struct pagevec *pvec)
220{
221 int pgmoved = 0;
222
223 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
224 __count_vm_events(PGROTATED, pgmoved);
225}
226
227/*
228 * Writeback is about to end against a page which has been marked for immediate
229 * reclaim. If it still appears to be reclaimable, move it to the tail of the
230 * inactive list.
231 */
232void rotate_reclaimable_page(struct page *page)
233{
234 if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
235 !PageUnevictable(page) && PageLRU(page)) {
236 struct pagevec *pvec;
237 unsigned long flags;
238
239 get_page(page);
240 local_irq_save(flags);
241 pvec = this_cpu_ptr(&lru_rotate_pvecs);
242 if (!pagevec_add(pvec, page))
243 pagevec_move_tail(pvec);
244 local_irq_restore(flags);
245 }
246}
247
248static void update_page_reclaim_stat(struct lruvec *lruvec,
249 int file, int rotated)
250{
251 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
252
253 reclaim_stat->recent_scanned[file]++;
254 if (rotated)
255 reclaim_stat->recent_rotated[file]++;
256}
257
258static void __activate_page(struct page *page, struct lruvec *lruvec,
259 void *arg)
260{
261 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
262 int file = page_is_file_cache(page);
263 int lru = page_lru_base_type(page);
264
265 del_page_from_lru_list(page, lruvec, lru);
266 SetPageActive(page);
267 lru += LRU_ACTIVE;
268 add_page_to_lru_list(page, lruvec, lru);
269 trace_mm_lru_activate(page);
270
271 __count_vm_event(PGACTIVATE);
272 update_page_reclaim_stat(lruvec, file, 1);
273 }
274}
275
276#ifdef CONFIG_SMP
277static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
278
279static void activate_page_drain(int cpu)
280{
281 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
282
283 if (pagevec_count(pvec))
284 pagevec_lru_move_fn(pvec, __activate_page, NULL);
285}
286
287static bool need_activate_page_drain(int cpu)
288{
289 return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
290}
291
292void activate_page(struct page *page)
293{
294 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
295 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
296
297 get_page(page);
298 if (!pagevec_add(pvec, page))
299 pagevec_lru_move_fn(pvec, __activate_page, NULL);
300 put_cpu_var(activate_page_pvecs);
301 }
302}
303
304#else
305static inline void activate_page_drain(int cpu)
306{
307}
308
309static bool need_activate_page_drain(int cpu)
310{
311 return false;
312}
313
314void activate_page(struct page *page)
315{
316 struct zone *zone = page_zone(page);
317
318 spin_lock_irq(&zone->lru_lock);
319 __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
320 spin_unlock_irq(&zone->lru_lock);
321}
322#endif
323
324static void __lru_cache_activate_page(struct page *page)
325{
326 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
327 int i;
328
329 /*
330 * Search backwards on the optimistic assumption that the page being
331 * activated has just been added to this pagevec. Note that only
332 * the local pagevec is examined as a !PageLRU page could be in the
333 * process of being released, reclaimed, migrated or on a remote
334 * pagevec that is currently being drained. Furthermore, marking
335 * a remote pagevec's page PageActive potentially hits a race where
336 * a page is marked PageActive just after it is added to the inactive
337 * list causing accounting errors and BUG_ON checks to trigger.
338 */
339 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
340 struct page *pagevec_page = pvec->pages[i];
341
342 if (pagevec_page == page) {
343 SetPageActive(page);
344 break;
345 }
346 }
347
348 put_cpu_var(lru_add_pvec);
349}
350
351/*
352 * Mark a page as having seen activity.
353 *
354 * inactive,unreferenced -> inactive,referenced
355 * inactive,referenced -> active,unreferenced
356 * active,unreferenced -> active,referenced
357 *
358 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
359 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
360 */
361void mark_page_accessed(struct page *page)
362{
363 page = compound_head(page);
364 if (!PageActive(page) && !PageUnevictable(page) &&
365 PageReferenced(page)) {
366
367 /*
368 * If the page is on the LRU, queue it for activation via
369 * activate_page_pvecs. Otherwise, assume the page is on a
370 * pagevec, mark it active and it'll be moved to the active
371 * LRU on the next drain.
372 */
373 if (PageLRU(page))
374 activate_page(page);
375 else
376 __lru_cache_activate_page(page);
377 ClearPageReferenced(page);
378 if (page_is_file_cache(page))
379 workingset_activation(page);
380 } else if (!PageReferenced(page)) {
381 SetPageReferenced(page);
382 }
383 if (page_is_idle(page))
384 clear_page_idle(page);
385}
386EXPORT_SYMBOL(mark_page_accessed);
387
388static void __lru_cache_add(struct page *page)
389{
390 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
391
392 get_page(page);
393 if (!pagevec_space(pvec))
394 __pagevec_lru_add(pvec);
395 pagevec_add(pvec, page);
396 put_cpu_var(lru_add_pvec);
397}
398
399/**
400 * lru_cache_add: add a page to the page lists
401 * @page: the page to add
402 */
403void lru_cache_add_anon(struct page *page)
404{
405 if (PageActive(page))
406 ClearPageActive(page);
407 __lru_cache_add(page);
408}
409
410void lru_cache_add_file(struct page *page)
411{
412 if (PageActive(page))
413 ClearPageActive(page);
414 __lru_cache_add(page);
415}
416EXPORT_SYMBOL(lru_cache_add_file);
417
418/**
419 * lru_cache_add - add a page to a page list
420 * @page: the page to be added to the LRU.
421 *
422 * Queue the page for addition to the LRU via pagevec. The decision on whether
423 * to add the page to the [in]active [file|anon] list is deferred until the
424 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
425 * have the page added to the active list using mark_page_accessed().
426 */
427void lru_cache_add(struct page *page)
428{
429 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
430 VM_BUG_ON_PAGE(PageLRU(page), page);
431 __lru_cache_add(page);
432}
433
434/**
435 * add_page_to_unevictable_list - add a page to the unevictable list
436 * @page: the page to be added to the unevictable list
437 *
438 * Add page directly to its zone's unevictable list. To avoid races with
439 * tasks that might be making the page evictable, through eg. munlock,
440 * munmap or exit, while it's not on the lru, we want to add the page
441 * while it's locked or otherwise "invisible" to other tasks. This is
442 * difficult to do when using the pagevec cache, so bypass that.
443 */
444void add_page_to_unevictable_list(struct page *page)
445{
446 struct zone *zone = page_zone(page);
447 struct lruvec *lruvec;
448
449 spin_lock_irq(&zone->lru_lock);
450 lruvec = mem_cgroup_page_lruvec(page, zone);
451 ClearPageActive(page);
452 SetPageUnevictable(page);
453 SetPageLRU(page);
454 add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
455 spin_unlock_irq(&zone->lru_lock);
456}
457
458/**
459 * lru_cache_add_active_or_unevictable
460 * @page: the page to be added to LRU
461 * @vma: vma in which page is mapped for determining reclaimability
462 *
463 * Place @page on the active or unevictable LRU list, depending on its
464 * evictability. Note that if the page is not evictable, it goes
465 * directly back onto it's zone's unevictable list, it does NOT use a
466 * per cpu pagevec.
467 */
468void lru_cache_add_active_or_unevictable(struct page *page,
469 struct vm_area_struct *vma)
470{
471 VM_BUG_ON_PAGE(PageLRU(page), page);
472
473 if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
474 SetPageActive(page);
475 lru_cache_add(page);
476 return;
477 }
478
479 if (!TestSetPageMlocked(page)) {
480 /*
481 * We use the irq-unsafe __mod_zone_page_stat because this
482 * counter is not modified from interrupt context, and the pte
483 * lock is held(spinlock), which implies preemption disabled.
484 */
485 __mod_zone_page_state(page_zone(page), NR_MLOCK,
486 hpage_nr_pages(page));
487 count_vm_event(UNEVICTABLE_PGMLOCKED);
488 }
489 add_page_to_unevictable_list(page);
490}
491
492/*
493 * If the page can not be invalidated, it is moved to the
494 * inactive list to speed up its reclaim. It is moved to the
495 * head of the list, rather than the tail, to give the flusher
496 * threads some time to write it out, as this is much more
497 * effective than the single-page writeout from reclaim.
498 *
499 * If the page isn't page_mapped and dirty/writeback, the page
500 * could reclaim asap using PG_reclaim.
501 *
502 * 1. active, mapped page -> none
503 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
504 * 3. inactive, mapped page -> none
505 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
506 * 5. inactive, clean -> inactive, tail
507 * 6. Others -> none
508 *
509 * In 4, why it moves inactive's head, the VM expects the page would
510 * be write it out by flusher threads as this is much more effective
511 * than the single-page writeout from reclaim.
512 */
513static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
514 void *arg)
515{
516 int lru, file;
517 bool active;
518
519 if (!PageLRU(page))
520 return;
521
522 if (PageUnevictable(page))
523 return;
524
525 /* Some processes are using the page */
526 if (page_mapped(page))
527 return;
528
529 active = PageActive(page);
530 file = page_is_file_cache(page);
531 lru = page_lru_base_type(page);
532
533 del_page_from_lru_list(page, lruvec, lru + active);
534 ClearPageActive(page);
535 ClearPageReferenced(page);
536 add_page_to_lru_list(page, lruvec, lru);
537
538 if (PageWriteback(page) || PageDirty(page)) {
539 /*
540 * PG_reclaim could be raced with end_page_writeback
541 * It can make readahead confusing. But race window
542 * is _really_ small and it's non-critical problem.
543 */
544 SetPageReclaim(page);
545 } else {
546 /*
547 * The page's writeback ends up during pagevec
548 * We moves tha page into tail of inactive.
549 */
550 list_move_tail(&page->lru, &lruvec->lists[lru]);
551 __count_vm_event(PGROTATED);
552 }
553
554 if (active)
555 __count_vm_event(PGDEACTIVATE);
556 update_page_reclaim_stat(lruvec, file, 0);
557}
558
559
560static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
561 void *arg)
562{
563 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
564 int file = page_is_file_cache(page);
565 int lru = page_lru_base_type(page);
566
567 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
568 ClearPageActive(page);
569 ClearPageReferenced(page);
570 add_page_to_lru_list(page, lruvec, lru);
571
572 __count_vm_event(PGDEACTIVATE);
573 update_page_reclaim_stat(lruvec, file, 0);
574 }
575}
576
577/*
578 * Drain pages out of the cpu's pagevecs.
579 * Either "cpu" is the current CPU, and preemption has already been
580 * disabled; or "cpu" is being hot-unplugged, and is already dead.
581 */
582void lru_add_drain_cpu(int cpu)
583{
584 struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
585
586 if (pagevec_count(pvec))
587 __pagevec_lru_add(pvec);
588
589 pvec = &per_cpu(lru_rotate_pvecs, cpu);
590 if (pagevec_count(pvec)) {
591 unsigned long flags;
592
593 /* No harm done if a racing interrupt already did this */
594 local_irq_save(flags);
595 pagevec_move_tail(pvec);
596 local_irq_restore(flags);
597 }
598
599 pvec = &per_cpu(lru_deactivate_file_pvecs, cpu);
600 if (pagevec_count(pvec))
601 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
602
603 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
604 if (pagevec_count(pvec))
605 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
606
607 activate_page_drain(cpu);
608}
609
610/**
611 * deactivate_file_page - forcefully deactivate a file page
612 * @page: page to deactivate
613 *
614 * This function hints the VM that @page is a good reclaim candidate,
615 * for example if its invalidation fails due to the page being dirty
616 * or under writeback.
617 */
618void deactivate_file_page(struct page *page)
619{
620 /*
621 * In a workload with many unevictable page such as mprotect,
622 * unevictable page deactivation for accelerating reclaim is pointless.
623 */
624 if (PageUnevictable(page))
625 return;
626
627 if (likely(get_page_unless_zero(page))) {
628 struct pagevec *pvec = &get_cpu_var(lru_deactivate_file_pvecs);
629
630 if (!pagevec_add(pvec, page))
631 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
632 put_cpu_var(lru_deactivate_file_pvecs);
633 }
634}
635
636/**
637 * deactivate_page - deactivate a page
638 * @page: page to deactivate
639 *
640 * deactivate_page() moves @page to the inactive list if @page was on the active
641 * list and was not an unevictable page. This is done to accelerate the reclaim
642 * of @page.
643 */
644void deactivate_page(struct page *page)
645{
646 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
647 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
648
649 get_page(page);
650 if (!pagevec_add(pvec, page))
651 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
652 put_cpu_var(lru_deactivate_pvecs);
653 }
654}
655
656void lru_add_drain(void)
657{
658 lru_add_drain_cpu(get_cpu());
659 put_cpu();
660}
661
662static void lru_add_drain_per_cpu(struct work_struct *dummy)
663{
664 lru_add_drain();
665}
666
667static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
668
669void lru_add_drain_all(void)
670{
671 static DEFINE_MUTEX(lock);
672 static struct cpumask has_work;
673 int cpu;
674
675 mutex_lock(&lock);
676 get_online_cpus();
677 cpumask_clear(&has_work);
678
679 for_each_online_cpu(cpu) {
680 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
681
682 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
683 pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
684 pagevec_count(&per_cpu(lru_deactivate_file_pvecs, cpu)) ||
685 pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
686 need_activate_page_drain(cpu)) {
687 INIT_WORK(work, lru_add_drain_per_cpu);
688 schedule_work_on(cpu, work);
689 cpumask_set_cpu(cpu, &has_work);
690 }
691 }
692
693 for_each_cpu(cpu, &has_work)
694 flush_work(&per_cpu(lru_add_drain_work, cpu));
695
696 put_online_cpus();
697 mutex_unlock(&lock);
698}
699
700/**
701 * release_pages - batched put_page()
702 * @pages: array of pages to release
703 * @nr: number of pages
704 * @cold: whether the pages are cache cold
705 *
706 * Decrement the reference count on all the pages in @pages. If it
707 * fell to zero, remove the page from the LRU and free it.
708 */
709void release_pages(struct page **pages, int nr, bool cold)
710{
711 int i;
712 LIST_HEAD(pages_to_free);
713 struct zone *zone = NULL;
714 struct lruvec *lruvec;
715 unsigned long uninitialized_var(flags);
716 unsigned int uninitialized_var(lock_batch);
717
718 for (i = 0; i < nr; i++) {
719 struct page *page = pages[i];
720
721 /*
722 * Make sure the IRQ-safe lock-holding time does not get
723 * excessive with a continuous string of pages from the
724 * same zone. The lock is held only if zone != NULL.
725 */
726 if (zone && ++lock_batch == SWAP_CLUSTER_MAX) {
727 spin_unlock_irqrestore(&zone->lru_lock, flags);
728 zone = NULL;
729 }
730
731 if (is_huge_zero_page(page)) {
732 put_huge_zero_page();
733 continue;
734 }
735
736 page = compound_head(page);
737 if (!put_page_testzero(page))
738 continue;
739
740 if (PageCompound(page)) {
741 if (zone) {
742 spin_unlock_irqrestore(&zone->lru_lock, flags);
743 zone = NULL;
744 }
745 __put_compound_page(page);
746 continue;
747 }
748
749 if (PageLRU(page)) {
750 struct zone *pagezone = page_zone(page);
751
752 if (pagezone != zone) {
753 if (zone)
754 spin_unlock_irqrestore(&zone->lru_lock,
755 flags);
756 lock_batch = 0;
757 zone = pagezone;
758 spin_lock_irqsave(&zone->lru_lock, flags);
759 }
760
761 lruvec = mem_cgroup_page_lruvec(page, zone);
762 VM_BUG_ON_PAGE(!PageLRU(page), page);
763 __ClearPageLRU(page);
764 del_page_from_lru_list(page, lruvec, page_off_lru(page));
765 }
766
767 /* Clear Active bit in case of parallel mark_page_accessed */
768 __ClearPageActive(page);
769
770 list_add(&page->lru, &pages_to_free);
771 }
772 if (zone)
773 spin_unlock_irqrestore(&zone->lru_lock, flags);
774
775 mem_cgroup_uncharge_list(&pages_to_free);
776 free_hot_cold_page_list(&pages_to_free, cold);
777}
778EXPORT_SYMBOL(release_pages);
779
780/*
781 * The pages which we're about to release may be in the deferred lru-addition
782 * queues. That would prevent them from really being freed right now. That's
783 * OK from a correctness point of view but is inefficient - those pages may be
784 * cache-warm and we want to give them back to the page allocator ASAP.
785 *
786 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
787 * and __pagevec_lru_add_active() call release_pages() directly to avoid
788 * mutual recursion.
789 */
790void __pagevec_release(struct pagevec *pvec)
791{
792 lru_add_drain();
793 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
794 pagevec_reinit(pvec);
795}
796EXPORT_SYMBOL(__pagevec_release);
797
798#ifdef CONFIG_TRANSPARENT_HUGEPAGE
799/* used by __split_huge_page_refcount() */
800void lru_add_page_tail(struct page *page, struct page *page_tail,
801 struct lruvec *lruvec, struct list_head *list)
802{
803 const int file = 0;
804
805 VM_BUG_ON_PAGE(!PageHead(page), page);
806 VM_BUG_ON_PAGE(PageCompound(page_tail), page);
807 VM_BUG_ON_PAGE(PageLRU(page_tail), page);
808 VM_BUG_ON(NR_CPUS != 1 &&
809 !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
810
811 if (!list)
812 SetPageLRU(page_tail);
813
814 if (likely(PageLRU(page)))
815 list_add_tail(&page_tail->lru, &page->lru);
816 else if (list) {
817 /* page reclaim is reclaiming a huge page */
818 get_page(page_tail);
819 list_add_tail(&page_tail->lru, list);
820 } else {
821 struct list_head *list_head;
822 /*
823 * Head page has not yet been counted, as an hpage,
824 * so we must account for each subpage individually.
825 *
826 * Use the standard add function to put page_tail on the list,
827 * but then correct its position so they all end up in order.
828 */
829 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
830 list_head = page_tail->lru.prev;
831 list_move_tail(&page_tail->lru, list_head);
832 }
833
834 if (!PageUnevictable(page))
835 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
836}
837#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
838
839static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
840 void *arg)
841{
842 int file = page_is_file_cache(page);
843 int active = PageActive(page);
844 enum lru_list lru = page_lru(page);
845
846 VM_BUG_ON_PAGE(PageLRU(page), page);
847
848 SetPageLRU(page);
849 add_page_to_lru_list(page, lruvec, lru);
850 update_page_reclaim_stat(lruvec, file, active);
851 trace_mm_lru_insertion(page, lru);
852}
853
854/*
855 * Add the passed pages to the LRU, then drop the caller's refcount
856 * on them. Reinitialises the caller's pagevec.
857 */
858void __pagevec_lru_add(struct pagevec *pvec)
859{
860 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
861}
862EXPORT_SYMBOL(__pagevec_lru_add);
863
864/**
865 * pagevec_lookup_entries - gang pagecache lookup
866 * @pvec: Where the resulting entries are placed
867 * @mapping: The address_space to search
868 * @start: The starting entry index
869 * @nr_entries: The maximum number of entries
870 * @indices: The cache indices corresponding to the entries in @pvec
871 *
872 * pagevec_lookup_entries() will search for and return a group of up
873 * to @nr_entries pages and shadow entries in the mapping. All
874 * entries are placed in @pvec. pagevec_lookup_entries() takes a
875 * reference against actual pages in @pvec.
876 *
877 * The search returns a group of mapping-contiguous entries with
878 * ascending indexes. There may be holes in the indices due to
879 * not-present entries.
880 *
881 * pagevec_lookup_entries() returns the number of entries which were
882 * found.
883 */
884unsigned pagevec_lookup_entries(struct pagevec *pvec,
885 struct address_space *mapping,
886 pgoff_t start, unsigned nr_pages,
887 pgoff_t *indices)
888{
889 pvec->nr = find_get_entries(mapping, start, nr_pages,
890 pvec->pages, indices);
891 return pagevec_count(pvec);
892}
893
894/**
895 * pagevec_remove_exceptionals - pagevec exceptionals pruning
896 * @pvec: The pagevec to prune
897 *
898 * pagevec_lookup_entries() fills both pages and exceptional radix
899 * tree entries into the pagevec. This function prunes all
900 * exceptionals from @pvec without leaving holes, so that it can be
901 * passed on to page-only pagevec operations.
902 */
903void pagevec_remove_exceptionals(struct pagevec *pvec)
904{
905 int i, j;
906
907 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
908 struct page *page = pvec->pages[i];
909 if (!radix_tree_exceptional_entry(page))
910 pvec->pages[j++] = page;
911 }
912 pvec->nr = j;
913}
914
915/**
916 * pagevec_lookup - gang pagecache lookup
917 * @pvec: Where the resulting pages are placed
918 * @mapping: The address_space to search
919 * @start: The starting page index
920 * @nr_pages: The maximum number of pages
921 *
922 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
923 * in the mapping. The pages are placed in @pvec. pagevec_lookup() takes a
924 * reference against the pages in @pvec.
925 *
926 * The search returns a group of mapping-contiguous pages with ascending
927 * indexes. There may be holes in the indices due to not-present pages.
928 *
929 * pagevec_lookup() returns the number of pages which were found.
930 */
931unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
932 pgoff_t start, unsigned nr_pages)
933{
934 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
935 return pagevec_count(pvec);
936}
937EXPORT_SYMBOL(pagevec_lookup);
938
939unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
940 pgoff_t *index, int tag, unsigned nr_pages)
941{
942 pvec->nr = find_get_pages_tag(mapping, index, tag,
943 nr_pages, pvec->pages);
944 return pagevec_count(pvec);
945}
946EXPORT_SYMBOL(pagevec_lookup_tag);
947
948/*
949 * Perform any setup for the swap system
950 */
951void __init swap_setup(void)
952{
953 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
954#ifdef CONFIG_SWAP
955 int i;
956
957 for (i = 0; i < MAX_SWAPFILES; i++)
958 spin_lock_init(&swapper_spaces[i].tree_lock);
959#endif
960
961 /* Use a smaller cluster for small-memory machines */
962 if (megs < 16)
963 page_cluster = 2;
964 else
965 page_cluster = 3;
966 /*
967 * Right now other parts of the system means that we
968 * _really_ don't want to cluster much more
969 */
970}