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 * via pagevecs. 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
161/*
162 * get_kernel_pages() - pin kernel pages in memory
163 * @kiov: An array of struct kvec structures
164 * @nr_segs: number of segments to pin
165 * @write: pinning for read/write, currently ignored
166 * @pages: array that receives pointers to the pages pinned.
167 * Should be at least nr_segs long.
168 *
169 * Returns number of pages pinned. This may be fewer than the number requested.
170 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0.
171 * Each page returned must be released with a put_page() call when it is
172 * finished with.
173 */
174int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
175 struct page **pages)
176{
177 int seg;
178
179 for (seg = 0; seg < nr_segs; seg++) {
180 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
181 return seg;
182
183 pages[seg] = kmap_to_page(kiov[seg].iov_base);
184 get_page(pages[seg]);
185 }
186
187 return seg;
188}
189EXPORT_SYMBOL_GPL(get_kernel_pages);
190
191typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
192
193static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
194{
195 int was_unevictable = folio_test_clear_unevictable(folio);
196 long nr_pages = folio_nr_pages(folio);
197
198 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
199
200 /*
201 * Is an smp_mb__after_atomic() still required here, before
202 * folio_evictable() tests the mlocked flag, to rule out the possibility
203 * of stranding an evictable folio on an unevictable LRU? I think
204 * not, because __munlock_page() only clears the mlocked flag
205 * while the LRU lock is held.
206 *
207 * (That is not true of __page_cache_release(), and not necessarily
208 * true of release_pages(): but those only clear the mlocked flag after
209 * folio_put_testzero() has excluded any other users of the folio.)
210 */
211 if (folio_evictable(folio)) {
212 if (was_unevictable)
213 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
214 } else {
215 folio_clear_active(folio);
216 folio_set_unevictable(folio);
217 /*
218 * folio->mlock_count = !!folio_test_mlocked(folio)?
219 * But that leaves __mlock_page() in doubt whether another
220 * actor has already counted the mlock or not. Err on the
221 * safe side, underestimate, let page reclaim fix it, rather
222 * than leaving a page on the unevictable LRU indefinitely.
223 */
224 folio->mlock_count = 0;
225 if (!was_unevictable)
226 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
227 }
228
229 lruvec_add_folio(lruvec, folio);
230 trace_mm_lru_insertion(folio);
231}
232
233static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
234{
235 int i;
236 struct lruvec *lruvec = NULL;
237 unsigned long flags = 0;
238
239 for (i = 0; i < folio_batch_count(fbatch); i++) {
240 struct folio *folio = fbatch->folios[i];
241
242 /* block memcg migration while the folio moves between lru */
243 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
244 continue;
245
246 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
247 move_fn(lruvec, folio);
248
249 folio_set_lru(folio);
250 }
251
252 if (lruvec)
253 unlock_page_lruvec_irqrestore(lruvec, flags);
254 folios_put(fbatch->folios, folio_batch_count(fbatch));
255 folio_batch_init(fbatch);
256}
257
258static void folio_batch_add_and_move(struct folio_batch *fbatch,
259 struct folio *folio, move_fn_t move_fn)
260{
261 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
262 !lru_cache_disabled())
263 return;
264 folio_batch_move_lru(fbatch, move_fn);
265}
266
267static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
268{
269 if (!folio_test_unevictable(folio)) {
270 lruvec_del_folio(lruvec, folio);
271 folio_clear_active(folio);
272 lruvec_add_folio_tail(lruvec, folio);
273 __count_vm_events(PGROTATED, folio_nr_pages(folio));
274 }
275}
276
277/*
278 * Writeback is about to end against a folio which has been marked for
279 * immediate reclaim. If it still appears to be reclaimable, move it
280 * to the tail of the inactive list.
281 *
282 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
283 */
284void folio_rotate_reclaimable(struct folio *folio)
285{
286 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
287 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
288 struct folio_batch *fbatch;
289 unsigned long flags;
290
291 folio_get(folio);
292 local_lock_irqsave(&lru_rotate.lock, flags);
293 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
294 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
295 local_unlock_irqrestore(&lru_rotate.lock, flags);
296 }
297}
298
299void lru_note_cost(struct lruvec *lruvec, bool file,
300 unsigned int nr_io, unsigned int nr_rotated)
301{
302 unsigned long cost;
303
304 /*
305 * Reflect the relative cost of incurring IO and spending CPU
306 * time on rotations. This doesn't attempt to make a precise
307 * comparison, it just says: if reloads are about comparable
308 * between the LRU lists, or rotations are overwhelmingly
309 * different between them, adjust scan balance for CPU work.
310 */
311 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
312
313 do {
314 unsigned long lrusize;
315
316 /*
317 * Hold lruvec->lru_lock is safe here, since
318 * 1) The pinned lruvec in reclaim, or
319 * 2) From a pre-LRU page during refault (which also holds the
320 * rcu lock, so would be safe even if the page was on the LRU
321 * and could move simultaneously to a new lruvec).
322 */
323 spin_lock_irq(&lruvec->lru_lock);
324 /* Record cost event */
325 if (file)
326 lruvec->file_cost += cost;
327 else
328 lruvec->anon_cost += cost;
329
330 /*
331 * Decay previous events
332 *
333 * Because workloads change over time (and to avoid
334 * overflow) we keep these statistics as a floating
335 * average, which ends up weighing recent refaults
336 * more than old ones.
337 */
338 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
339 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
340 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
341 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
342
343 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
344 lruvec->file_cost /= 2;
345 lruvec->anon_cost /= 2;
346 }
347 spin_unlock_irq(&lruvec->lru_lock);
348 } while ((lruvec = parent_lruvec(lruvec)));
349}
350
351void lru_note_cost_refault(struct folio *folio)
352{
353 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
354 folio_nr_pages(folio), 0);
355}
356
357static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
358{
359 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
360 long nr_pages = folio_nr_pages(folio);
361
362 lruvec_del_folio(lruvec, folio);
363 folio_set_active(folio);
364 lruvec_add_folio(lruvec, folio);
365 trace_mm_lru_activate(folio);
366
367 __count_vm_events(PGACTIVATE, nr_pages);
368 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
369 nr_pages);
370 }
371}
372
373#ifdef CONFIG_SMP
374static void folio_activate_drain(int cpu)
375{
376 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
377
378 if (folio_batch_count(fbatch))
379 folio_batch_move_lru(fbatch, folio_activate_fn);
380}
381
382void folio_activate(struct folio *folio)
383{
384 if (folio_test_lru(folio) && !folio_test_active(folio) &&
385 !folio_test_unevictable(folio)) {
386 struct folio_batch *fbatch;
387
388 folio_get(folio);
389 local_lock(&cpu_fbatches.lock);
390 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
391 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
392 local_unlock(&cpu_fbatches.lock);
393 }
394}
395
396#else
397static inline void folio_activate_drain(int cpu)
398{
399}
400
401void folio_activate(struct folio *folio)
402{
403 struct lruvec *lruvec;
404
405 if (folio_test_clear_lru(folio)) {
406 lruvec = folio_lruvec_lock_irq(folio);
407 folio_activate_fn(lruvec, folio);
408 unlock_page_lruvec_irq(lruvec);
409 folio_set_lru(folio);
410 }
411}
412#endif
413
414static void __lru_cache_activate_folio(struct folio *folio)
415{
416 struct folio_batch *fbatch;
417 int i;
418
419 local_lock(&cpu_fbatches.lock);
420 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
421
422 /*
423 * Search backwards on the optimistic assumption that the folio being
424 * activated has just been added to this batch. Note that only
425 * the local batch is examined as a !LRU folio could be in the
426 * process of being released, reclaimed, migrated or on a remote
427 * batch that is currently being drained. Furthermore, marking
428 * a remote batch's folio active potentially hits a race where
429 * a folio is marked active just after it is added to the inactive
430 * list causing accounting errors and BUG_ON checks to trigger.
431 */
432 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
433 struct folio *batch_folio = fbatch->folios[i];
434
435 if (batch_folio == folio) {
436 folio_set_active(folio);
437 break;
438 }
439 }
440
441 local_unlock(&cpu_fbatches.lock);
442}
443
444#ifdef CONFIG_LRU_GEN
445static void folio_inc_refs(struct folio *folio)
446{
447 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
448
449 if (folio_test_unevictable(folio))
450 return;
451
452 if (!folio_test_referenced(folio)) {
453 folio_set_referenced(folio);
454 return;
455 }
456
457 if (!folio_test_workingset(folio)) {
458 folio_set_workingset(folio);
459 return;
460 }
461
462 /* see the comment on MAX_NR_TIERS */
463 do {
464 new_flags = old_flags & LRU_REFS_MASK;
465 if (new_flags == LRU_REFS_MASK)
466 break;
467
468 new_flags += BIT(LRU_REFS_PGOFF);
469 new_flags |= old_flags & ~LRU_REFS_MASK;
470 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
471}
472#else
473static void folio_inc_refs(struct folio *folio)
474{
475}
476#endif /* CONFIG_LRU_GEN */
477
478/*
479 * Mark a page as having seen activity.
480 *
481 * inactive,unreferenced -> inactive,referenced
482 * inactive,referenced -> active,unreferenced
483 * active,unreferenced -> active,referenced
484 *
485 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
486 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
487 */
488void folio_mark_accessed(struct folio *folio)
489{
490 if (lru_gen_enabled()) {
491 folio_inc_refs(folio);
492 return;
493 }
494
495 if (!folio_test_referenced(folio)) {
496 folio_set_referenced(folio);
497 } else if (folio_test_unevictable(folio)) {
498 /*
499 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
500 * this list is never rotated or maintained, so marking an
501 * unevictable page accessed has no effect.
502 */
503 } else if (!folio_test_active(folio)) {
504 /*
505 * If the folio is on the LRU, queue it for activation via
506 * cpu_fbatches.activate. Otherwise, assume the folio is in a
507 * folio_batch, mark it active and it'll be moved to the active
508 * LRU on the next drain.
509 */
510 if (folio_test_lru(folio))
511 folio_activate(folio);
512 else
513 __lru_cache_activate_folio(folio);
514 folio_clear_referenced(folio);
515 workingset_activation(folio);
516 }
517 if (folio_test_idle(folio))
518 folio_clear_idle(folio);
519}
520EXPORT_SYMBOL(folio_mark_accessed);
521
522/**
523 * folio_add_lru - Add a folio to an LRU list.
524 * @folio: The folio to be added to the LRU.
525 *
526 * Queue the folio for addition to the LRU. The decision on whether
527 * to add the page to the [in]active [file|anon] list is deferred until the
528 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
529 * have the folio added to the active list using folio_mark_accessed().
530 */
531void folio_add_lru(struct folio *folio)
532{
533 struct folio_batch *fbatch;
534
535 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
536 folio_test_unevictable(folio), folio);
537 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
538
539 /* see the comment in lru_gen_add_folio() */
540 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
541 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
542 folio_set_active(folio);
543
544 folio_get(folio);
545 local_lock(&cpu_fbatches.lock);
546 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
547 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
548 local_unlock(&cpu_fbatches.lock);
549}
550EXPORT_SYMBOL(folio_add_lru);
551
552/**
553 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
554 * @folio: The folio to be added to the LRU.
555 * @vma: VMA in which the folio is mapped.
556 *
557 * If the VMA is mlocked, @folio is added to the unevictable list.
558 * Otherwise, it is treated the same way as folio_add_lru().
559 */
560void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
561{
562 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
563
564 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
565 mlock_new_page(&folio->page);
566 else
567 folio_add_lru(folio);
568}
569
570/*
571 * If the folio cannot be invalidated, it is moved to the
572 * inactive list to speed up its reclaim. It is moved to the
573 * head of the list, rather than the tail, to give the flusher
574 * threads some time to write it out, as this is much more
575 * effective than the single-page writeout from reclaim.
576 *
577 * If the folio isn't mapped and dirty/writeback, the folio
578 * could be reclaimed asap using the reclaim flag.
579 *
580 * 1. active, mapped folio -> none
581 * 2. active, dirty/writeback folio -> inactive, head, reclaim
582 * 3. inactive, mapped folio -> none
583 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
584 * 5. inactive, clean -> inactive, tail
585 * 6. Others -> none
586 *
587 * In 4, it moves to the head of the inactive list so the folio is
588 * written out by flusher threads as this is much more efficient
589 * than the single-page writeout from reclaim.
590 */
591static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
592{
593 bool active = folio_test_active(folio);
594 long nr_pages = folio_nr_pages(folio);
595
596 if (folio_test_unevictable(folio))
597 return;
598
599 /* Some processes are using the folio */
600 if (folio_mapped(folio))
601 return;
602
603 lruvec_del_folio(lruvec, folio);
604 folio_clear_active(folio);
605 folio_clear_referenced(folio);
606
607 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
608 /*
609 * Setting the reclaim flag could race with
610 * folio_end_writeback() and confuse readahead. But the
611 * race window is _really_ small and it's not a critical
612 * problem.
613 */
614 lruvec_add_folio(lruvec, folio);
615 folio_set_reclaim(folio);
616 } else {
617 /*
618 * The folio's writeback ended while it was in the batch.
619 * We move that folio to the tail of the inactive list.
620 */
621 lruvec_add_folio_tail(lruvec, folio);
622 __count_vm_events(PGROTATED, nr_pages);
623 }
624
625 if (active) {
626 __count_vm_events(PGDEACTIVATE, nr_pages);
627 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
628 nr_pages);
629 }
630}
631
632static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
633{
634 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
635 long nr_pages = folio_nr_pages(folio);
636
637 lruvec_del_folio(lruvec, folio);
638 folio_clear_active(folio);
639 folio_clear_referenced(folio);
640 lruvec_add_folio(lruvec, folio);
641
642 __count_vm_events(PGDEACTIVATE, nr_pages);
643 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
644 nr_pages);
645 }
646}
647
648static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
649{
650 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
651 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
652 long nr_pages = folio_nr_pages(folio);
653
654 lruvec_del_folio(lruvec, folio);
655 folio_clear_active(folio);
656 folio_clear_referenced(folio);
657 /*
658 * Lazyfree folios are clean anonymous folios. They have
659 * the swapbacked flag cleared, to distinguish them from normal
660 * anonymous folios
661 */
662 folio_clear_swapbacked(folio);
663 lruvec_add_folio(lruvec, folio);
664
665 __count_vm_events(PGLAZYFREE, nr_pages);
666 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
667 nr_pages);
668 }
669}
670
671/*
672 * Drain pages out of the cpu's folio_batch.
673 * Either "cpu" is the current CPU, and preemption has already been
674 * disabled; or "cpu" is being hot-unplugged, and is already dead.
675 */
676void lru_add_drain_cpu(int cpu)
677{
678 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
679 struct folio_batch *fbatch = &fbatches->lru_add;
680
681 if (folio_batch_count(fbatch))
682 folio_batch_move_lru(fbatch, lru_add_fn);
683
684 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
685 /* Disabling interrupts below acts as a compiler barrier. */
686 if (data_race(folio_batch_count(fbatch))) {
687 unsigned long flags;
688
689 /* No harm done if a racing interrupt already did this */
690 local_lock_irqsave(&lru_rotate.lock, flags);
691 folio_batch_move_lru(fbatch, lru_move_tail_fn);
692 local_unlock_irqrestore(&lru_rotate.lock, flags);
693 }
694
695 fbatch = &fbatches->lru_deactivate_file;
696 if (folio_batch_count(fbatch))
697 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
698
699 fbatch = &fbatches->lru_deactivate;
700 if (folio_batch_count(fbatch))
701 folio_batch_move_lru(fbatch, lru_deactivate_fn);
702
703 fbatch = &fbatches->lru_lazyfree;
704 if (folio_batch_count(fbatch))
705 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
706
707 folio_activate_drain(cpu);
708}
709
710/**
711 * deactivate_file_folio() - Deactivate a file folio.
712 * @folio: Folio to deactivate.
713 *
714 * This function hints to the VM that @folio is a good reclaim candidate,
715 * for example if its invalidation fails due to the folio being dirty
716 * or under writeback.
717 *
718 * Context: Caller holds a reference on the folio.
719 */
720void deactivate_file_folio(struct folio *folio)
721{
722 struct folio_batch *fbatch;
723
724 /* Deactivating an unevictable folio will not accelerate reclaim */
725 if (folio_test_unevictable(folio))
726 return;
727
728 folio_get(folio);
729 local_lock(&cpu_fbatches.lock);
730 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
731 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
732 local_unlock(&cpu_fbatches.lock);
733}
734
735/*
736 * deactivate_page - deactivate a page
737 * @page: page to deactivate
738 *
739 * deactivate_page() moves @page to the inactive list if @page was on the active
740 * list and was not an unevictable page. This is done to accelerate the reclaim
741 * of @page.
742 */
743void deactivate_page(struct page *page)
744{
745 struct folio *folio = page_folio(page);
746
747 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
748 (folio_test_active(folio) || lru_gen_enabled())) {
749 struct folio_batch *fbatch;
750
751 folio_get(folio);
752 local_lock(&cpu_fbatches.lock);
753 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
754 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
755 local_unlock(&cpu_fbatches.lock);
756 }
757}
758
759/**
760 * mark_page_lazyfree - make an anon page lazyfree
761 * @page: page to deactivate
762 *
763 * mark_page_lazyfree() moves @page to the inactive file list.
764 * This is done to accelerate the reclaim of @page.
765 */
766void mark_page_lazyfree(struct page *page)
767{
768 struct folio *folio = page_folio(page);
769
770 if (folio_test_lru(folio) && folio_test_anon(folio) &&
771 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
772 !folio_test_unevictable(folio)) {
773 struct folio_batch *fbatch;
774
775 folio_get(folio);
776 local_lock(&cpu_fbatches.lock);
777 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
778 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
779 local_unlock(&cpu_fbatches.lock);
780 }
781}
782
783void lru_add_drain(void)
784{
785 local_lock(&cpu_fbatches.lock);
786 lru_add_drain_cpu(smp_processor_id());
787 local_unlock(&cpu_fbatches.lock);
788 mlock_page_drain_local();
789}
790
791/*
792 * It's called from per-cpu workqueue context in SMP case so
793 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
794 * the same cpu. It shouldn't be a problem in !SMP case since
795 * the core is only one and the locks will disable preemption.
796 */
797static void lru_add_and_bh_lrus_drain(void)
798{
799 local_lock(&cpu_fbatches.lock);
800 lru_add_drain_cpu(smp_processor_id());
801 local_unlock(&cpu_fbatches.lock);
802 invalidate_bh_lrus_cpu();
803 mlock_page_drain_local();
804}
805
806void lru_add_drain_cpu_zone(struct zone *zone)
807{
808 local_lock(&cpu_fbatches.lock);
809 lru_add_drain_cpu(smp_processor_id());
810 drain_local_pages(zone);
811 local_unlock(&cpu_fbatches.lock);
812 mlock_page_drain_local();
813}
814
815#ifdef CONFIG_SMP
816
817static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
818
819static void lru_add_drain_per_cpu(struct work_struct *dummy)
820{
821 lru_add_and_bh_lrus_drain();
822}
823
824static bool cpu_needs_drain(unsigned int cpu)
825{
826 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
827
828 /* Check these in order of likelihood that they're not zero */
829 return folio_batch_count(&fbatches->lru_add) ||
830 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
831 folio_batch_count(&fbatches->lru_deactivate_file) ||
832 folio_batch_count(&fbatches->lru_deactivate) ||
833 folio_batch_count(&fbatches->lru_lazyfree) ||
834 folio_batch_count(&fbatches->activate) ||
835 need_mlock_page_drain(cpu) ||
836 has_bh_in_lru(cpu, NULL);
837}
838
839/*
840 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
841 * kworkers being shut down before our page_alloc_cpu_dead callback is
842 * executed on the offlined cpu.
843 * Calling this function with cpu hotplug locks held can actually lead
844 * to obscure indirect dependencies via WQ context.
845 */
846static inline void __lru_add_drain_all(bool force_all_cpus)
847{
848 /*
849 * lru_drain_gen - Global pages generation number
850 *
851 * (A) Definition: global lru_drain_gen = x implies that all generations
852 * 0 < n <= x are already *scheduled* for draining.
853 *
854 * This is an optimization for the highly-contended use case where a
855 * user space workload keeps constantly generating a flow of pages for
856 * each CPU.
857 */
858 static unsigned int lru_drain_gen;
859 static struct cpumask has_work;
860 static DEFINE_MUTEX(lock);
861 unsigned cpu, this_gen;
862
863 /*
864 * Make sure nobody triggers this path before mm_percpu_wq is fully
865 * initialized.
866 */
867 if (WARN_ON(!mm_percpu_wq))
868 return;
869
870 /*
871 * Guarantee folio_batch counter stores visible by this CPU
872 * are visible to other CPUs before loading the current drain
873 * generation.
874 */
875 smp_mb();
876
877 /*
878 * (B) Locally cache global LRU draining generation number
879 *
880 * The read barrier ensures that the counter is loaded before the mutex
881 * is taken. It pairs with smp_mb() inside the mutex critical section
882 * at (D).
883 */
884 this_gen = smp_load_acquire(&lru_drain_gen);
885
886 mutex_lock(&lock);
887
888 /*
889 * (C) Exit the draining operation if a newer generation, from another
890 * lru_add_drain_all(), was already scheduled for draining. Check (A).
891 */
892 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
893 goto done;
894
895 /*
896 * (D) Increment global generation number
897 *
898 * Pairs with smp_load_acquire() at (B), outside of the critical
899 * section. Use a full memory barrier to guarantee that the
900 * new global drain generation number is stored before loading
901 * folio_batch counters.
902 *
903 * This pairing must be done here, before the for_each_online_cpu loop
904 * below which drains the page vectors.
905 *
906 * Let x, y, and z represent some system CPU numbers, where x < y < z.
907 * Assume CPU #z is in the middle of the for_each_online_cpu loop
908 * below and has already reached CPU #y's per-cpu data. CPU #x comes
909 * along, adds some pages to its per-cpu vectors, then calls
910 * lru_add_drain_all().
911 *
912 * If the paired barrier is done at any later step, e.g. after the
913 * loop, CPU #x will just exit at (C) and miss flushing out all of its
914 * added pages.
915 */
916 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
917 smp_mb();
918
919 cpumask_clear(&has_work);
920 for_each_online_cpu(cpu) {
921 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
922
923 if (cpu_needs_drain(cpu)) {
924 INIT_WORK(work, lru_add_drain_per_cpu);
925 queue_work_on(cpu, mm_percpu_wq, work);
926 __cpumask_set_cpu(cpu, &has_work);
927 }
928 }
929
930 for_each_cpu(cpu, &has_work)
931 flush_work(&per_cpu(lru_add_drain_work, cpu));
932
933done:
934 mutex_unlock(&lock);
935}
936
937void lru_add_drain_all(void)
938{
939 __lru_add_drain_all(false);
940}
941#else
942void lru_add_drain_all(void)
943{
944 lru_add_drain();
945}
946#endif /* CONFIG_SMP */
947
948atomic_t lru_disable_count = ATOMIC_INIT(0);
949
950/*
951 * lru_cache_disable() needs to be called before we start compiling
952 * a list of pages to be migrated using isolate_lru_page().
953 * It drains pages on LRU cache and then disable on all cpus until
954 * lru_cache_enable is called.
955 *
956 * Must be paired with a call to lru_cache_enable().
957 */
958void lru_cache_disable(void)
959{
960 atomic_inc(&lru_disable_count);
961 /*
962 * Readers of lru_disable_count are protected by either disabling
963 * preemption or rcu_read_lock:
964 *
965 * preempt_disable, local_irq_disable [bh_lru_lock()]
966 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
967 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
968 *
969 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
970 * preempt_disable() regions of code. So any CPU which sees
971 * lru_disable_count = 0 will have exited the critical
972 * section when synchronize_rcu() returns.
973 */
974 synchronize_rcu_expedited();
975#ifdef CONFIG_SMP
976 __lru_add_drain_all(true);
977#else
978 lru_add_and_bh_lrus_drain();
979#endif
980}
981
982/**
983 * release_pages - batched put_page()
984 * @arg: array of pages to release
985 * @nr: number of pages
986 *
987 * Decrement the reference count on all the pages in @arg. If it
988 * fell to zero, remove the page from the LRU and free it.
989 *
990 * Note that the argument can be an array of pages, encoded pages,
991 * or folio pointers. We ignore any encoded bits, and turn any of
992 * them into just a folio that gets free'd.
993 */
994void release_pages(release_pages_arg arg, int nr)
995{
996 int i;
997 struct encoded_page **encoded = arg.encoded_pages;
998 LIST_HEAD(pages_to_free);
999 struct lruvec *lruvec = NULL;
1000 unsigned long flags = 0;
1001 unsigned int lock_batch;
1002
1003 for (i = 0; i < nr; i++) {
1004 struct folio *folio;
1005
1006 /* Turn any of the argument types into a folio */
1007 folio = page_folio(encoded_page_ptr(encoded[i]));
1008
1009 /*
1010 * Make sure the IRQ-safe lock-holding time does not get
1011 * excessive with a continuous string of pages from the
1012 * same lruvec. The lock is held only if lruvec != NULL.
1013 */
1014 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
1015 unlock_page_lruvec_irqrestore(lruvec, flags);
1016 lruvec = NULL;
1017 }
1018
1019 if (is_huge_zero_page(&folio->page))
1020 continue;
1021
1022 if (folio_is_zone_device(folio)) {
1023 if (lruvec) {
1024 unlock_page_lruvec_irqrestore(lruvec, flags);
1025 lruvec = NULL;
1026 }
1027 if (put_devmap_managed_page(&folio->page))
1028 continue;
1029 if (folio_put_testzero(folio))
1030 free_zone_device_page(&folio->page);
1031 continue;
1032 }
1033
1034 if (!folio_put_testzero(folio))
1035 continue;
1036
1037 if (folio_test_large(folio)) {
1038 if (lruvec) {
1039 unlock_page_lruvec_irqrestore(lruvec, flags);
1040 lruvec = NULL;
1041 }
1042 __folio_put_large(folio);
1043 continue;
1044 }
1045
1046 if (folio_test_lru(folio)) {
1047 struct lruvec *prev_lruvec = lruvec;
1048
1049 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1050 &flags);
1051 if (prev_lruvec != lruvec)
1052 lock_batch = 0;
1053
1054 lruvec_del_folio(lruvec, folio);
1055 __folio_clear_lru_flags(folio);
1056 }
1057
1058 /*
1059 * In rare cases, when truncation or holepunching raced with
1060 * munlock after VM_LOCKED was cleared, Mlocked may still be
1061 * found set here. This does not indicate a problem, unless
1062 * "unevictable_pgs_cleared" appears worryingly large.
1063 */
1064 if (unlikely(folio_test_mlocked(folio))) {
1065 __folio_clear_mlocked(folio);
1066 zone_stat_sub_folio(folio, NR_MLOCK);
1067 count_vm_event(UNEVICTABLE_PGCLEARED);
1068 }
1069
1070 list_add(&folio->lru, &pages_to_free);
1071 }
1072 if (lruvec)
1073 unlock_page_lruvec_irqrestore(lruvec, flags);
1074
1075 mem_cgroup_uncharge_list(&pages_to_free);
1076 free_unref_page_list(&pages_to_free);
1077}
1078EXPORT_SYMBOL(release_pages);
1079
1080/*
1081 * The pages which we're about to release may be in the deferred lru-addition
1082 * queues. That would prevent them from really being freed right now. That's
1083 * OK from a correctness point of view but is inefficient - those pages may be
1084 * cache-warm and we want to give them back to the page allocator ASAP.
1085 *
1086 * So __pagevec_release() will drain those queues here.
1087 * folio_batch_move_lru() calls folios_put() directly to avoid
1088 * mutual recursion.
1089 */
1090void __pagevec_release(struct pagevec *pvec)
1091{
1092 if (!pvec->percpu_pvec_drained) {
1093 lru_add_drain();
1094 pvec->percpu_pvec_drained = true;
1095 }
1096 release_pages(pvec->pages, pagevec_count(pvec));
1097 pagevec_reinit(pvec);
1098}
1099EXPORT_SYMBOL(__pagevec_release);
1100
1101/**
1102 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1103 * @fbatch: The batch to prune
1104 *
1105 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1106 * entries. This function prunes all the non-folio entries from @fbatch
1107 * without leaving holes, so that it can be passed on to folio-only batch
1108 * operations.
1109 */
1110void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1111{
1112 unsigned int i, j;
1113
1114 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1115 struct folio *folio = fbatch->folios[i];
1116 if (!xa_is_value(folio))
1117 fbatch->folios[j++] = folio;
1118 }
1119 fbatch->nr = j;
1120}
1121
1122unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1123 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1124 xa_mark_t tag)
1125{
1126 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1127 PAGEVEC_SIZE, pvec->pages);
1128 return pagevec_count(pvec);
1129}
1130EXPORT_SYMBOL(pagevec_lookup_range_tag);
1131
1132/*
1133 * Perform any setup for the swap system
1134 */
1135void __init swap_setup(void)
1136{
1137 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1138
1139 /* Use a smaller cluster for small-memory machines */
1140 if (megs < 16)
1141 page_cluster = 2;
1142 else
1143 page_cluster = 3;
1144 /*
1145 * Right now other parts of the system means that we
1146 * _really_ don't want to cluster much more
1147 */
1148}
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? */
47int page_cluster;
48
49/* Protecting only lru_rotate.pvec which requires disabling interrupts */
50struct lru_rotate {
51 local_lock_t lock;
52 struct pagevec pvec;
53};
54static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
55 .lock = INIT_LOCAL_LOCK(lock),
56};
57
58/*
59 * The following struct pagevec are grouped together because they are protected
60 * by disabling preemption (and interrupts remain enabled).
61 */
62struct lru_pvecs {
63 local_lock_t lock;
64 struct pagevec lru_add;
65 struct pagevec lru_deactivate_file;
66 struct pagevec lru_deactivate;
67 struct pagevec lru_lazyfree;
68#ifdef CONFIG_SMP
69 struct pagevec activate_page;
70#endif
71};
72static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
73 .lock = INIT_LOCAL_LOCK(lock),
74};
75
76/*
77 * This path almost never happens for VM activity - pages are normally
78 * freed via pagevecs. But it gets used by networking.
79 */
80static void __page_cache_release(struct page *page)
81{
82 if (PageLRU(page)) {
83 struct lruvec *lruvec;
84 unsigned long flags;
85
86 lruvec = lock_page_lruvec_irqsave(page, &flags);
87 del_page_from_lru_list(page, lruvec);
88 __clear_page_lru_flags(page);
89 unlock_page_lruvec_irqrestore(lruvec, flags);
90 }
91 __ClearPageWaiters(page);
92}
93
94static void __put_single_page(struct page *page)
95{
96 __page_cache_release(page);
97 mem_cgroup_uncharge(page);
98 free_unref_page(page, 0);
99}
100
101static void __put_compound_page(struct page *page)
102{
103 /*
104 * __page_cache_release() is supposed to be called for thp, not for
105 * hugetlb. This is because hugetlb page does never have PageLRU set
106 * (it's never listed to any LRU lists) and no memcg routines should
107 * be called for hugetlb (it has a separate hugetlb_cgroup.)
108 */
109 if (!PageHuge(page))
110 __page_cache_release(page);
111 destroy_compound_page(page);
112}
113
114void __put_page(struct page *page)
115{
116 if (is_zone_device_page(page)) {
117 put_dev_pagemap(page->pgmap);
118
119 /*
120 * The page belongs to the device that created pgmap. Do
121 * not return it to page allocator.
122 */
123 return;
124 }
125
126 if (unlikely(PageCompound(page)))
127 __put_compound_page(page);
128 else
129 __put_single_page(page);
130}
131EXPORT_SYMBOL(__put_page);
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. Currently
138 * used by read_cache_pages() and related error recovery code.
139 */
140void put_pages_list(struct list_head *pages)
141{
142 while (!list_empty(pages)) {
143 struct page *victim;
144
145 victim = lru_to_page(pages);
146 list_del(&victim->lru);
147 put_page(victim);
148 }
149}
150EXPORT_SYMBOL(put_pages_list);
151
152/*
153 * get_kernel_pages() - pin kernel pages in memory
154 * @kiov: An array of struct kvec structures
155 * @nr_segs: number of segments to pin
156 * @write: pinning for read/write, currently ignored
157 * @pages: array that receives pointers to the pages pinned.
158 * Should be at least nr_segs long.
159 *
160 * Returns number of pages pinned. This may be fewer than the number
161 * requested. If nr_pages is 0 or negative, returns 0. If no pages
162 * were pinned, returns -errno. Each page returned must be released
163 * with a put_page() call when it is finished with.
164 */
165int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
166 struct page **pages)
167{
168 int seg;
169
170 for (seg = 0; seg < nr_segs; seg++) {
171 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
172 return seg;
173
174 pages[seg] = kmap_to_page(kiov[seg].iov_base);
175 get_page(pages[seg]);
176 }
177
178 return seg;
179}
180EXPORT_SYMBOL_GPL(get_kernel_pages);
181
182/*
183 * get_kernel_page() - pin a kernel page in memory
184 * @start: starting kernel address
185 * @write: pinning for read/write, currently ignored
186 * @pages: array that receives pointer to the page pinned.
187 * Must be at least nr_segs long.
188 *
189 * Returns 1 if page is pinned. If the page was not pinned, returns
190 * -errno. The page returned must be released with a put_page() call
191 * when it is finished with.
192 */
193int get_kernel_page(unsigned long start, int write, struct page **pages)
194{
195 const struct kvec kiov = {
196 .iov_base = (void *)start,
197 .iov_len = PAGE_SIZE
198 };
199
200 return get_kernel_pages(&kiov, 1, write, pages);
201}
202EXPORT_SYMBOL_GPL(get_kernel_page);
203
204static void pagevec_lru_move_fn(struct pagevec *pvec,
205 void (*move_fn)(struct page *page, struct lruvec *lruvec))
206{
207 int i;
208 struct lruvec *lruvec = NULL;
209 unsigned long flags = 0;
210
211 for (i = 0; i < pagevec_count(pvec); i++) {
212 struct page *page = pvec->pages[i];
213
214 /* block memcg migration during page moving between lru */
215 if (!TestClearPageLRU(page))
216 continue;
217
218 lruvec = relock_page_lruvec_irqsave(page, lruvec, &flags);
219 (*move_fn)(page, lruvec);
220
221 SetPageLRU(page);
222 }
223 if (lruvec)
224 unlock_page_lruvec_irqrestore(lruvec, flags);
225 release_pages(pvec->pages, pvec->nr);
226 pagevec_reinit(pvec);
227}
228
229static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
230{
231 if (!PageUnevictable(page)) {
232 del_page_from_lru_list(page, lruvec);
233 ClearPageActive(page);
234 add_page_to_lru_list_tail(page, lruvec);
235 __count_vm_events(PGROTATED, thp_nr_pages(page));
236 }
237}
238
239/* return true if pagevec needs to drain */
240static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
241{
242 bool ret = false;
243
244 if (!pagevec_add(pvec, page) || PageCompound(page) ||
245 lru_cache_disabled())
246 ret = true;
247
248 return ret;
249}
250
251/*
252 * Writeback is about to end against a page which has been marked for immediate
253 * reclaim. If it still appears to be reclaimable, move it to the tail of the
254 * inactive list.
255 *
256 * rotate_reclaimable_page() must disable IRQs, to prevent nasty races.
257 */
258void rotate_reclaimable_page(struct page *page)
259{
260 if (!PageLocked(page) && !PageDirty(page) &&
261 !PageUnevictable(page) && PageLRU(page)) {
262 struct pagevec *pvec;
263 unsigned long flags;
264
265 get_page(page);
266 local_lock_irqsave(&lru_rotate.lock, flags);
267 pvec = this_cpu_ptr(&lru_rotate.pvec);
268 if (pagevec_add_and_need_flush(pvec, page))
269 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
270 local_unlock_irqrestore(&lru_rotate.lock, flags);
271 }
272}
273
274void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
275{
276 do {
277 unsigned long lrusize;
278
279 /*
280 * Hold lruvec->lru_lock is safe here, since
281 * 1) The pinned lruvec in reclaim, or
282 * 2) From a pre-LRU page during refault (which also holds the
283 * rcu lock, so would be safe even if the page was on the LRU
284 * and could move simultaneously to a new lruvec).
285 */
286 spin_lock_irq(&lruvec->lru_lock);
287 /* Record cost event */
288 if (file)
289 lruvec->file_cost += nr_pages;
290 else
291 lruvec->anon_cost += nr_pages;
292
293 /*
294 * Decay previous events
295 *
296 * Because workloads change over time (and to avoid
297 * overflow) we keep these statistics as a floating
298 * average, which ends up weighing recent refaults
299 * more than old ones.
300 */
301 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
302 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
303 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
304 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
305
306 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
307 lruvec->file_cost /= 2;
308 lruvec->anon_cost /= 2;
309 }
310 spin_unlock_irq(&lruvec->lru_lock);
311 } while ((lruvec = parent_lruvec(lruvec)));
312}
313
314void lru_note_cost_page(struct page *page)
315{
316 lru_note_cost(mem_cgroup_page_lruvec(page),
317 page_is_file_lru(page), thp_nr_pages(page));
318}
319
320static void __activate_page(struct page *page, struct lruvec *lruvec)
321{
322 if (!PageActive(page) && !PageUnevictable(page)) {
323 int nr_pages = thp_nr_pages(page);
324
325 del_page_from_lru_list(page, lruvec);
326 SetPageActive(page);
327 add_page_to_lru_list(page, lruvec);
328 trace_mm_lru_activate(page);
329
330 __count_vm_events(PGACTIVATE, nr_pages);
331 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
332 nr_pages);
333 }
334}
335
336#ifdef CONFIG_SMP
337static void activate_page_drain(int cpu)
338{
339 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
340
341 if (pagevec_count(pvec))
342 pagevec_lru_move_fn(pvec, __activate_page);
343}
344
345static bool need_activate_page_drain(int cpu)
346{
347 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
348}
349
350static void activate_page(struct page *page)
351{
352 page = compound_head(page);
353 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
354 struct pagevec *pvec;
355
356 local_lock(&lru_pvecs.lock);
357 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
358 get_page(page);
359 if (pagevec_add_and_need_flush(pvec, page))
360 pagevec_lru_move_fn(pvec, __activate_page);
361 local_unlock(&lru_pvecs.lock);
362 }
363}
364
365#else
366static inline void activate_page_drain(int cpu)
367{
368}
369
370static void activate_page(struct page *page)
371{
372 struct lruvec *lruvec;
373
374 page = compound_head(page);
375 if (TestClearPageLRU(page)) {
376 lruvec = lock_page_lruvec_irq(page);
377 __activate_page(page, lruvec);
378 unlock_page_lruvec_irq(lruvec);
379 SetPageLRU(page);
380 }
381}
382#endif
383
384static void __lru_cache_activate_page(struct page *page)
385{
386 struct pagevec *pvec;
387 int i;
388
389 local_lock(&lru_pvecs.lock);
390 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
391
392 /*
393 * Search backwards on the optimistic assumption that the page being
394 * activated has just been added to this pagevec. Note that only
395 * the local pagevec is examined as a !PageLRU page could be in the
396 * process of being released, reclaimed, migrated or on a remote
397 * pagevec that is currently being drained. Furthermore, marking
398 * a remote pagevec's page PageActive potentially hits a race where
399 * a page is marked PageActive just after it is added to the inactive
400 * list causing accounting errors and BUG_ON checks to trigger.
401 */
402 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
403 struct page *pagevec_page = pvec->pages[i];
404
405 if (pagevec_page == page) {
406 SetPageActive(page);
407 break;
408 }
409 }
410
411 local_unlock(&lru_pvecs.lock);
412}
413
414/*
415 * Mark a page as having seen activity.
416 *
417 * inactive,unreferenced -> inactive,referenced
418 * inactive,referenced -> active,unreferenced
419 * active,unreferenced -> active,referenced
420 *
421 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
422 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
423 */
424void mark_page_accessed(struct page *page)
425{
426 page = compound_head(page);
427
428 if (!PageReferenced(page)) {
429 SetPageReferenced(page);
430 } else if (PageUnevictable(page)) {
431 /*
432 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
433 * this list is never rotated or maintained, so marking an
434 * evictable page accessed has no effect.
435 */
436 } else if (!PageActive(page)) {
437 /*
438 * If the page is on the LRU, queue it for activation via
439 * lru_pvecs.activate_page. Otherwise, assume the page is on a
440 * pagevec, mark it active and it'll be moved to the active
441 * LRU on the next drain.
442 */
443 if (PageLRU(page))
444 activate_page(page);
445 else
446 __lru_cache_activate_page(page);
447 ClearPageReferenced(page);
448 workingset_activation(page);
449 }
450 if (page_is_idle(page))
451 clear_page_idle(page);
452}
453EXPORT_SYMBOL(mark_page_accessed);
454
455/**
456 * lru_cache_add - add a page to a page list
457 * @page: the page to be added to the LRU.
458 *
459 * Queue the page for addition to the LRU via pagevec. The decision on whether
460 * to add the page to the [in]active [file|anon] list is deferred until the
461 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
462 * have the page added to the active list using mark_page_accessed().
463 */
464void lru_cache_add(struct page *page)
465{
466 struct pagevec *pvec;
467
468 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
469 VM_BUG_ON_PAGE(PageLRU(page), page);
470
471 get_page(page);
472 local_lock(&lru_pvecs.lock);
473 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
474 if (pagevec_add_and_need_flush(pvec, page))
475 __pagevec_lru_add(pvec);
476 local_unlock(&lru_pvecs.lock);
477}
478EXPORT_SYMBOL(lru_cache_add);
479
480/**
481 * lru_cache_add_inactive_or_unevictable
482 * @page: the page to be added to LRU
483 * @vma: vma in which page is mapped for determining reclaimability
484 *
485 * Place @page on the inactive or unevictable LRU list, depending on its
486 * evictability.
487 */
488void lru_cache_add_inactive_or_unevictable(struct page *page,
489 struct vm_area_struct *vma)
490{
491 bool unevictable;
492
493 VM_BUG_ON_PAGE(PageLRU(page), page);
494
495 unevictable = (vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED;
496 if (unlikely(unevictable) && !TestSetPageMlocked(page)) {
497 int nr_pages = thp_nr_pages(page);
498 /*
499 * We use the irq-unsafe __mod_zone_page_state because this
500 * counter is not modified from interrupt context, and the pte
501 * lock is held(spinlock), which implies preemption disabled.
502 */
503 __mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
504 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
505 }
506 lru_cache_add(page);
507}
508
509/*
510 * If the page can not be invalidated, it is moved to the
511 * inactive list to speed up its reclaim. It is moved to the
512 * head of the list, rather than the tail, to give the flusher
513 * threads some time to write it out, as this is much more
514 * effective than the single-page writeout from reclaim.
515 *
516 * If the page isn't page_mapped and dirty/writeback, the page
517 * could reclaim asap using PG_reclaim.
518 *
519 * 1. active, mapped page -> none
520 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
521 * 3. inactive, mapped page -> none
522 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
523 * 5. inactive, clean -> inactive, tail
524 * 6. Others -> none
525 *
526 * In 4, why it moves inactive's head, the VM expects the page would
527 * be write it out by flusher threads as this is much more effective
528 * than the single-page writeout from reclaim.
529 */
530static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
531{
532 bool active = PageActive(page);
533 int nr_pages = thp_nr_pages(page);
534
535 if (PageUnevictable(page))
536 return;
537
538 /* Some processes are using the page */
539 if (page_mapped(page))
540 return;
541
542 del_page_from_lru_list(page, lruvec);
543 ClearPageActive(page);
544 ClearPageReferenced(page);
545
546 if (PageWriteback(page) || PageDirty(page)) {
547 /*
548 * PG_reclaim could be raced with end_page_writeback
549 * It can make readahead confusing. But race window
550 * is _really_ small and it's non-critical problem.
551 */
552 add_page_to_lru_list(page, lruvec);
553 SetPageReclaim(page);
554 } else {
555 /*
556 * The page's writeback ends up during pagevec
557 * We move that page into tail of inactive.
558 */
559 add_page_to_lru_list_tail(page, lruvec);
560 __count_vm_events(PGROTATED, nr_pages);
561 }
562
563 if (active) {
564 __count_vm_events(PGDEACTIVATE, nr_pages);
565 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
566 nr_pages);
567 }
568}
569
570static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
571{
572 if (PageActive(page) && !PageUnevictable(page)) {
573 int nr_pages = thp_nr_pages(page);
574
575 del_page_from_lru_list(page, lruvec);
576 ClearPageActive(page);
577 ClearPageReferenced(page);
578 add_page_to_lru_list(page, lruvec);
579
580 __count_vm_events(PGDEACTIVATE, nr_pages);
581 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
582 nr_pages);
583 }
584}
585
586static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
587{
588 if (PageAnon(page) && PageSwapBacked(page) &&
589 !PageSwapCache(page) && !PageUnevictable(page)) {
590 int nr_pages = thp_nr_pages(page);
591
592 del_page_from_lru_list(page, lruvec);
593 ClearPageActive(page);
594 ClearPageReferenced(page);
595 /*
596 * Lazyfree pages are clean anonymous pages. They have
597 * PG_swapbacked flag cleared, to distinguish them from normal
598 * anonymous pages
599 */
600 ClearPageSwapBacked(page);
601 add_page_to_lru_list(page, lruvec);
602
603 __count_vm_events(PGLAZYFREE, nr_pages);
604 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
605 nr_pages);
606 }
607}
608
609/*
610 * Drain pages out of the cpu's pagevecs.
611 * Either "cpu" is the current CPU, and preemption has already been
612 * disabled; or "cpu" is being hot-unplugged, and is already dead.
613 */
614void lru_add_drain_cpu(int cpu)
615{
616 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
617
618 if (pagevec_count(pvec))
619 __pagevec_lru_add(pvec);
620
621 pvec = &per_cpu(lru_rotate.pvec, cpu);
622 /* Disabling interrupts below acts as a compiler barrier. */
623 if (data_race(pagevec_count(pvec))) {
624 unsigned long flags;
625
626 /* No harm done if a racing interrupt already did this */
627 local_lock_irqsave(&lru_rotate.lock, flags);
628 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
629 local_unlock_irqrestore(&lru_rotate.lock, flags);
630 }
631
632 pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
633 if (pagevec_count(pvec))
634 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
635
636 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
637 if (pagevec_count(pvec))
638 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
639
640 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
641 if (pagevec_count(pvec))
642 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
643
644 activate_page_drain(cpu);
645 invalidate_bh_lrus_cpu(cpu);
646}
647
648/**
649 * deactivate_file_page - forcefully deactivate a file page
650 * @page: page to deactivate
651 *
652 * This function hints the VM that @page is a good reclaim candidate,
653 * for example if its invalidation fails due to the page being dirty
654 * or under writeback.
655 */
656void deactivate_file_page(struct page *page)
657{
658 /*
659 * In a workload with many unevictable page such as mprotect,
660 * unevictable page deactivation for accelerating reclaim is pointless.
661 */
662 if (PageUnevictable(page))
663 return;
664
665 if (likely(get_page_unless_zero(page))) {
666 struct pagevec *pvec;
667
668 local_lock(&lru_pvecs.lock);
669 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
670
671 if (pagevec_add_and_need_flush(pvec, page))
672 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
673 local_unlock(&lru_pvecs.lock);
674 }
675}
676
677/*
678 * deactivate_page - deactivate a page
679 * @page: page to deactivate
680 *
681 * deactivate_page() moves @page to the inactive list if @page was on the active
682 * list and was not an unevictable page. This is done to accelerate the reclaim
683 * of @page.
684 */
685void deactivate_page(struct page *page)
686{
687 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
688 struct pagevec *pvec;
689
690 local_lock(&lru_pvecs.lock);
691 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
692 get_page(page);
693 if (pagevec_add_and_need_flush(pvec, page))
694 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
695 local_unlock(&lru_pvecs.lock);
696 }
697}
698
699/**
700 * mark_page_lazyfree - make an anon page lazyfree
701 * @page: page to deactivate
702 *
703 * mark_page_lazyfree() moves @page to the inactive file list.
704 * This is done to accelerate the reclaim of @page.
705 */
706void mark_page_lazyfree(struct page *page)
707{
708 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
709 !PageSwapCache(page) && !PageUnevictable(page)) {
710 struct pagevec *pvec;
711
712 local_lock(&lru_pvecs.lock);
713 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
714 get_page(page);
715 if (pagevec_add_and_need_flush(pvec, page))
716 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
717 local_unlock(&lru_pvecs.lock);
718 }
719}
720
721void lru_add_drain(void)
722{
723 local_lock(&lru_pvecs.lock);
724 lru_add_drain_cpu(smp_processor_id());
725 local_unlock(&lru_pvecs.lock);
726}
727
728void lru_add_drain_cpu_zone(struct zone *zone)
729{
730 local_lock(&lru_pvecs.lock);
731 lru_add_drain_cpu(smp_processor_id());
732 drain_local_pages(zone);
733 local_unlock(&lru_pvecs.lock);
734}
735
736#ifdef CONFIG_SMP
737
738static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
739
740static void lru_add_drain_per_cpu(struct work_struct *dummy)
741{
742 lru_add_drain();
743}
744
745/*
746 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
747 * kworkers being shut down before our page_alloc_cpu_dead callback is
748 * executed on the offlined cpu.
749 * Calling this function with cpu hotplug locks held can actually lead
750 * to obscure indirect dependencies via WQ context.
751 */
752inline void __lru_add_drain_all(bool force_all_cpus)
753{
754 /*
755 * lru_drain_gen - Global pages generation number
756 *
757 * (A) Definition: global lru_drain_gen = x implies that all generations
758 * 0 < n <= x are already *scheduled* for draining.
759 *
760 * This is an optimization for the highly-contended use case where a
761 * user space workload keeps constantly generating a flow of pages for
762 * each CPU.
763 */
764 static unsigned int lru_drain_gen;
765 static struct cpumask has_work;
766 static DEFINE_MUTEX(lock);
767 unsigned cpu, this_gen;
768
769 /*
770 * Make sure nobody triggers this path before mm_percpu_wq is fully
771 * initialized.
772 */
773 if (WARN_ON(!mm_percpu_wq))
774 return;
775
776 /*
777 * Guarantee pagevec counter stores visible by this CPU are visible to
778 * other CPUs before loading the current drain generation.
779 */
780 smp_mb();
781
782 /*
783 * (B) Locally cache global LRU draining generation number
784 *
785 * The read barrier ensures that the counter is loaded before the mutex
786 * is taken. It pairs with smp_mb() inside the mutex critical section
787 * at (D).
788 */
789 this_gen = smp_load_acquire(&lru_drain_gen);
790
791 mutex_lock(&lock);
792
793 /*
794 * (C) Exit the draining operation if a newer generation, from another
795 * lru_add_drain_all(), was already scheduled for draining. Check (A).
796 */
797 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
798 goto done;
799
800 /*
801 * (D) Increment global generation number
802 *
803 * Pairs with smp_load_acquire() at (B), outside of the critical
804 * section. Use a full memory barrier to guarantee that the new global
805 * drain generation number is stored before loading pagevec counters.
806 *
807 * This pairing must be done here, before the for_each_online_cpu loop
808 * below which drains the page vectors.
809 *
810 * Let x, y, and z represent some system CPU numbers, where x < y < z.
811 * Assume CPU #z is in the middle of the for_each_online_cpu loop
812 * below and has already reached CPU #y's per-cpu data. CPU #x comes
813 * along, adds some pages to its per-cpu vectors, then calls
814 * lru_add_drain_all().
815 *
816 * If the paired barrier is done at any later step, e.g. after the
817 * loop, CPU #x will just exit at (C) and miss flushing out all of its
818 * added pages.
819 */
820 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
821 smp_mb();
822
823 cpumask_clear(&has_work);
824 for_each_online_cpu(cpu) {
825 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
826
827 if (force_all_cpus ||
828 pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
829 data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
830 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
831 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
832 pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
833 need_activate_page_drain(cpu) ||
834 has_bh_in_lru(cpu, NULL)) {
835 INIT_WORK(work, lru_add_drain_per_cpu);
836 queue_work_on(cpu, mm_percpu_wq, work);
837 __cpumask_set_cpu(cpu, &has_work);
838 }
839 }
840
841 for_each_cpu(cpu, &has_work)
842 flush_work(&per_cpu(lru_add_drain_work, cpu));
843
844done:
845 mutex_unlock(&lock);
846}
847
848void lru_add_drain_all(void)
849{
850 __lru_add_drain_all(false);
851}
852#else
853void lru_add_drain_all(void)
854{
855 lru_add_drain();
856}
857#endif /* CONFIG_SMP */
858
859atomic_t lru_disable_count = ATOMIC_INIT(0);
860
861/*
862 * lru_cache_disable() needs to be called before we start compiling
863 * a list of pages to be migrated using isolate_lru_page().
864 * It drains pages on LRU cache and then disable on all cpus until
865 * lru_cache_enable is called.
866 *
867 * Must be paired with a call to lru_cache_enable().
868 */
869void lru_cache_disable(void)
870{
871 atomic_inc(&lru_disable_count);
872#ifdef CONFIG_SMP
873 /*
874 * lru_add_drain_all in the force mode will schedule draining on
875 * all online CPUs so any calls of lru_cache_disabled wrapped by
876 * local_lock or preemption disabled would be ordered by that.
877 * The atomic operation doesn't need to have stronger ordering
878 * requirements because that is enforeced by the scheduling
879 * guarantees.
880 */
881 __lru_add_drain_all(true);
882#else
883 lru_add_drain();
884#endif
885}
886
887/**
888 * release_pages - batched put_page()
889 * @pages: array of pages to release
890 * @nr: number of pages
891 *
892 * Decrement the reference count on all the pages in @pages. If it
893 * fell to zero, remove the page from the LRU and free it.
894 */
895void release_pages(struct page **pages, int nr)
896{
897 int i;
898 LIST_HEAD(pages_to_free);
899 struct lruvec *lruvec = NULL;
900 unsigned long flags;
901 unsigned int lock_batch;
902
903 for (i = 0; i < nr; i++) {
904 struct page *page = pages[i];
905
906 /*
907 * Make sure the IRQ-safe lock-holding time does not get
908 * excessive with a continuous string of pages from the
909 * same lruvec. The lock is held only if lruvec != NULL.
910 */
911 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
912 unlock_page_lruvec_irqrestore(lruvec, flags);
913 lruvec = NULL;
914 }
915
916 page = compound_head(page);
917 if (is_huge_zero_page(page))
918 continue;
919
920 if (is_zone_device_page(page)) {
921 if (lruvec) {
922 unlock_page_lruvec_irqrestore(lruvec, flags);
923 lruvec = NULL;
924 }
925 /*
926 * ZONE_DEVICE pages that return 'false' from
927 * page_is_devmap_managed() do not require special
928 * processing, and instead, expect a call to
929 * put_page_testzero().
930 */
931 if (page_is_devmap_managed(page)) {
932 put_devmap_managed_page(page);
933 continue;
934 }
935 if (put_page_testzero(page))
936 put_dev_pagemap(page->pgmap);
937 continue;
938 }
939
940 if (!put_page_testzero(page))
941 continue;
942
943 if (PageCompound(page)) {
944 if (lruvec) {
945 unlock_page_lruvec_irqrestore(lruvec, flags);
946 lruvec = NULL;
947 }
948 __put_compound_page(page);
949 continue;
950 }
951
952 if (PageLRU(page)) {
953 struct lruvec *prev_lruvec = lruvec;
954
955 lruvec = relock_page_lruvec_irqsave(page, lruvec,
956 &flags);
957 if (prev_lruvec != lruvec)
958 lock_batch = 0;
959
960 del_page_from_lru_list(page, lruvec);
961 __clear_page_lru_flags(page);
962 }
963
964 __ClearPageWaiters(page);
965
966 list_add(&page->lru, &pages_to_free);
967 }
968 if (lruvec)
969 unlock_page_lruvec_irqrestore(lruvec, flags);
970
971 mem_cgroup_uncharge_list(&pages_to_free);
972 free_unref_page_list(&pages_to_free);
973}
974EXPORT_SYMBOL(release_pages);
975
976/*
977 * The pages which we're about to release may be in the deferred lru-addition
978 * queues. That would prevent them from really being freed right now. That's
979 * OK from a correctness point of view but is inefficient - those pages may be
980 * cache-warm and we want to give them back to the page allocator ASAP.
981 *
982 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
983 * and __pagevec_lru_add_active() call release_pages() directly to avoid
984 * mutual recursion.
985 */
986void __pagevec_release(struct pagevec *pvec)
987{
988 if (!pvec->percpu_pvec_drained) {
989 lru_add_drain();
990 pvec->percpu_pvec_drained = true;
991 }
992 release_pages(pvec->pages, pagevec_count(pvec));
993 pagevec_reinit(pvec);
994}
995EXPORT_SYMBOL(__pagevec_release);
996
997static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec)
998{
999 int was_unevictable = TestClearPageUnevictable(page);
1000 int nr_pages = thp_nr_pages(page);
1001
1002 VM_BUG_ON_PAGE(PageLRU(page), page);
1003
1004 /*
1005 * Page becomes evictable in two ways:
1006 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1007 * 2) Before acquiring LRU lock to put the page to correct LRU and then
1008 * a) do PageLRU check with lock [check_move_unevictable_pages]
1009 * b) do PageLRU check before lock [clear_page_mlock]
1010 *
1011 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1012 * following strict ordering:
1013 *
1014 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1015 *
1016 * SetPageLRU() TestClearPageMlocked()
1017 * smp_mb() // explicit ordering // above provides strict
1018 * // ordering
1019 * PageMlocked() PageLRU()
1020 *
1021 *
1022 * if '#1' does not observe setting of PG_lru by '#0' and fails
1023 * isolation, the explicit barrier will make sure that page_evictable
1024 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1025 * can be reordered after PageMlocked check and can make '#1' to fail
1026 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1027 * looking at the same page) and the evictable page will be stranded
1028 * in an unevictable LRU.
1029 */
1030 SetPageLRU(page);
1031 smp_mb__after_atomic();
1032
1033 if (page_evictable(page)) {
1034 if (was_unevictable)
1035 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1036 } else {
1037 ClearPageActive(page);
1038 SetPageUnevictable(page);
1039 if (!was_unevictable)
1040 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1041 }
1042
1043 add_page_to_lru_list(page, lruvec);
1044 trace_mm_lru_insertion(page);
1045}
1046
1047/*
1048 * Add the passed pages to the LRU, then drop the caller's refcount
1049 * on them. Reinitialises the caller's pagevec.
1050 */
1051void __pagevec_lru_add(struct pagevec *pvec)
1052{
1053 int i;
1054 struct lruvec *lruvec = NULL;
1055 unsigned long flags = 0;
1056
1057 for (i = 0; i < pagevec_count(pvec); i++) {
1058 struct page *page = pvec->pages[i];
1059
1060 lruvec = relock_page_lruvec_irqsave(page, lruvec, &flags);
1061 __pagevec_lru_add_fn(page, lruvec);
1062 }
1063 if (lruvec)
1064 unlock_page_lruvec_irqrestore(lruvec, flags);
1065 release_pages(pvec->pages, pvec->nr);
1066 pagevec_reinit(pvec);
1067}
1068
1069/**
1070 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1071 * @pvec: The pagevec to prune
1072 *
1073 * find_get_entries() fills both pages and XArray value entries (aka
1074 * exceptional entries) into the pagevec. This function prunes all
1075 * exceptionals from @pvec without leaving holes, so that it can be
1076 * passed on to page-only pagevec operations.
1077 */
1078void pagevec_remove_exceptionals(struct pagevec *pvec)
1079{
1080 int i, j;
1081
1082 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1083 struct page *page = pvec->pages[i];
1084 if (!xa_is_value(page))
1085 pvec->pages[j++] = page;
1086 }
1087 pvec->nr = j;
1088}
1089
1090/**
1091 * pagevec_lookup_range - gang pagecache lookup
1092 * @pvec: Where the resulting pages are placed
1093 * @mapping: The address_space to search
1094 * @start: The starting page index
1095 * @end: The final page index
1096 *
1097 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1098 * pages in the mapping starting from index @start and upto index @end
1099 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1100 * reference against the pages in @pvec.
1101 *
1102 * The search returns a group of mapping-contiguous pages with ascending
1103 * indexes. There may be holes in the indices due to not-present pages. We
1104 * also update @start to index the next page for the traversal.
1105 *
1106 * pagevec_lookup_range() returns the number of pages which were found. If this
1107 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1108 * reached.
1109 */
1110unsigned pagevec_lookup_range(struct pagevec *pvec,
1111 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1112{
1113 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1114 pvec->pages);
1115 return pagevec_count(pvec);
1116}
1117EXPORT_SYMBOL(pagevec_lookup_range);
1118
1119unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1120 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1121 xa_mark_t tag)
1122{
1123 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1124 PAGEVEC_SIZE, pvec->pages);
1125 return pagevec_count(pvec);
1126}
1127EXPORT_SYMBOL(pagevec_lookup_range_tag);
1128
1129/*
1130 * Perform any setup for the swap system
1131 */
1132void __init swap_setup(void)
1133{
1134 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1135
1136 /* Use a smaller cluster for small-memory machines */
1137 if (megs < 16)
1138 page_cluster = 2;
1139 else
1140 page_cluster = 3;
1141 /*
1142 * Right now other parts of the system means that we
1143 * _really_ don't want to cluster much more
1144 */
1145}
1146
1147#ifdef CONFIG_DEV_PAGEMAP_OPS
1148void put_devmap_managed_page(struct page *page)
1149{
1150 int count;
1151
1152 if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1153 return;
1154
1155 count = page_ref_dec_return(page);
1156
1157 /*
1158 * devmap page refcounts are 1-based, rather than 0-based: if
1159 * refcount is 1, then the page is free and the refcount is
1160 * stable because nobody holds a reference on the page.
1161 */
1162 if (count == 1)
1163 free_devmap_managed_page(page);
1164 else if (!count)
1165 __put_page(page);
1166}
1167EXPORT_SYMBOL(put_devmap_managed_page);
1168#endif