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 */
47 int page_cluster;
48 const int page_cluster_max = 31;
49
50 /* Protecting only lru_rotate.fbatch which requires disabling interrupts */
51 struct lru_rotate {
52 local_lock_t lock;
53 struct folio_batch fbatch;
54 };
55 static 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 */
63 struct 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 };
73 static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
74 .lock = INIT_LOCAL_LOCK(lock),
75 };
76
77 static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
78 unsigned long *flagsp)
79 {
80 if (folio_test_lru(folio)) {
81 folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
82 lruvec_del_folio(*lruvecp, folio);
83 __folio_clear_lru_flags(folio);
84 }
85
86 /*
87 * In rare cases, when truncation or holepunching raced with
88 * munlock after VM_LOCKED was cleared, Mlocked may still be
89 * found set here. This does not indicate a problem, unless
90 * "unevictable_pgs_cleared" appears worryingly large.
91 */
92 if (unlikely(folio_test_mlocked(folio))) {
93 long nr_pages = folio_nr_pages(folio);
94
95 __folio_clear_mlocked(folio);
96 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
97 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
98 }
99 }
100
101 /*
102 * This path almost never happens for VM activity - pages are normally freed
103 * in batches. But it gets used by networking - and for compound pages.
104 */
105 static void page_cache_release(struct folio *folio)
106 {
107 struct lruvec *lruvec = NULL;
108 unsigned long flags;
109
110 __page_cache_release(folio, &lruvec, &flags);
111 if (lruvec)
112 unlock_page_lruvec_irqrestore(lruvec, flags);
113 }
114
115 void __folio_put(struct folio *folio)
116 {
117 if (unlikely(folio_is_zone_device(folio))) {
118 free_zone_device_folio(folio);
119 return;
120 } else if (folio_test_hugetlb(folio)) {
121 free_huge_folio(folio);
122 return;
123 }
124
125 page_cache_release(folio);
126 folio_undo_large_rmappable(folio);
127 mem_cgroup_uncharge(folio);
128 free_unref_page(&folio->page, folio_order(folio));
129 }
130 EXPORT_SYMBOL(__folio_put);
131
132 /**
133 * put_pages_list() - release a list of pages
134 * @pages: list of pages threaded on page->lru
135 *
136 * Release a list of pages which are strung together on page.lru.
137 */
138 void put_pages_list(struct list_head *pages)
139 {
140 struct folio_batch fbatch;
141 struct folio *folio, *next;
142
143 folio_batch_init(&fbatch);
144 list_for_each_entry_safe(folio, next, pages, lru) {
145 if (!folio_put_testzero(folio))
146 continue;
147 if (folio_test_hugetlb(folio)) {
148 free_huge_folio(folio);
149 continue;
150 }
151 /* LRU flag must be clear because it's passed using the lru */
152 if (folio_batch_add(&fbatch, folio) > 0)
153 continue;
154 free_unref_folios(&fbatch);
155 }
156
157 if (fbatch.nr)
158 free_unref_folios(&fbatch);
159 INIT_LIST_HEAD(pages);
160 }
161 EXPORT_SYMBOL(put_pages_list);
162
163 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
164
165 static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
166 {
167 int was_unevictable = folio_test_clear_unevictable(folio);
168 long nr_pages = folio_nr_pages(folio);
169
170 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
171
172 /*
173 * Is an smp_mb__after_atomic() still required here, before
174 * folio_evictable() tests the mlocked flag, to rule out the possibility
175 * of stranding an evictable folio on an unevictable LRU? I think
176 * not, because __munlock_folio() only clears the mlocked flag
177 * while the LRU lock is held.
178 *
179 * (That is not true of __page_cache_release(), and not necessarily
180 * true of folios_put(): but those only clear the mlocked flag after
181 * folio_put_testzero() has excluded any other users of the folio.)
182 */
183 if (folio_evictable(folio)) {
184 if (was_unevictable)
185 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
186 } else {
187 folio_clear_active(folio);
188 folio_set_unevictable(folio);
189 /*
190 * folio->mlock_count = !!folio_test_mlocked(folio)?
191 * But that leaves __mlock_folio() in doubt whether another
192 * actor has already counted the mlock or not. Err on the
193 * safe side, underestimate, let page reclaim fix it, rather
194 * than leaving a page on the unevictable LRU indefinitely.
195 */
196 folio->mlock_count = 0;
197 if (!was_unevictable)
198 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
199 }
200
201 lruvec_add_folio(lruvec, folio);
202 trace_mm_lru_insertion(folio);
203 }
204
205 static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
206 {
207 int i;
208 struct lruvec *lruvec = NULL;
209 unsigned long flags = 0;
210
211 for (i = 0; i < folio_batch_count(fbatch); i++) {
212 struct folio *folio = fbatch->folios[i];
213
214 folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
215 move_fn(lruvec, folio);
216
217 folio_set_lru(folio);
218 }
219
220 if (lruvec)
221 unlock_page_lruvec_irqrestore(lruvec, flags);
222 folios_put(fbatch);
223 }
224
225 static void folio_batch_add_and_move(struct folio_batch *fbatch,
226 struct folio *folio, move_fn_t move_fn)
227 {
228 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
229 !lru_cache_disabled())
230 return;
231 folio_batch_move_lru(fbatch, move_fn);
232 }
233
234 static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
235 {
236 if (!folio_test_unevictable(folio)) {
237 lruvec_del_folio(lruvec, folio);
238 folio_clear_active(folio);
239 lruvec_add_folio_tail(lruvec, folio);
240 __count_vm_events(PGROTATED, folio_nr_pages(folio));
241 }
242 }
243
244 /*
245 * Writeback is about to end against a folio which has been marked for
246 * immediate reclaim. If it still appears to be reclaimable, move it
247 * to the tail of the inactive list.
248 *
249 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
250 */
251 void folio_rotate_reclaimable(struct folio *folio)
252 {
253 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
254 !folio_test_unevictable(folio)) {
255 struct folio_batch *fbatch;
256 unsigned long flags;
257
258 folio_get(folio);
259 if (!folio_test_clear_lru(folio)) {
260 folio_put(folio);
261 return;
262 }
263
264 local_lock_irqsave(&lru_rotate.lock, flags);
265 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
266 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
267 local_unlock_irqrestore(&lru_rotate.lock, flags);
268 }
269 }
270
271 void lru_note_cost(struct lruvec *lruvec, bool file,
272 unsigned int nr_io, unsigned int nr_rotated)
273 {
274 unsigned long cost;
275
276 /*
277 * Reflect the relative cost of incurring IO and spending CPU
278 * time on rotations. This doesn't attempt to make a precise
279 * comparison, it just says: if reloads are about comparable
280 * between the LRU lists, or rotations are overwhelmingly
281 * different between them, adjust scan balance for CPU work.
282 */
283 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
284
285 do {
286 unsigned long lrusize;
287
288 /*
289 * Hold lruvec->lru_lock is safe here, since
290 * 1) The pinned lruvec in reclaim, or
291 * 2) From a pre-LRU page during refault (which also holds the
292 * rcu lock, so would be safe even if the page was on the LRU
293 * and could move simultaneously to a new lruvec).
294 */
295 spin_lock_irq(&lruvec->lru_lock);
296 /* Record cost event */
297 if (file)
298 lruvec->file_cost += cost;
299 else
300 lruvec->anon_cost += cost;
301
302 /*
303 * Decay previous events
304 *
305 * Because workloads change over time (and to avoid
306 * overflow) we keep these statistics as a floating
307 * average, which ends up weighing recent refaults
308 * more than old ones.
309 */
310 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
311 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
312 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
313 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
314
315 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
316 lruvec->file_cost /= 2;
317 lruvec->anon_cost /= 2;
318 }
319 spin_unlock_irq(&lruvec->lru_lock);
320 } while ((lruvec = parent_lruvec(lruvec)));
321 }
322
323 void lru_note_cost_refault(struct folio *folio)
324 {
325 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
326 folio_nr_pages(folio), 0);
327 }
328
329 static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
330 {
331 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
332 long nr_pages = folio_nr_pages(folio);
333
334 lruvec_del_folio(lruvec, folio);
335 folio_set_active(folio);
336 lruvec_add_folio(lruvec, folio);
337 trace_mm_lru_activate(folio);
338
339 __count_vm_events(PGACTIVATE, nr_pages);
340 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
341 nr_pages);
342 }
343 }
344
345 #ifdef CONFIG_SMP
346 static void folio_activate_drain(int cpu)
347 {
348 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
349
350 if (folio_batch_count(fbatch))
351 folio_batch_move_lru(fbatch, folio_activate_fn);
352 }
353
354 void folio_activate(struct folio *folio)
355 {
356 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
357 struct folio_batch *fbatch;
358
359 folio_get(folio);
360 if (!folio_test_clear_lru(folio)) {
361 folio_put(folio);
362 return;
363 }
364
365 local_lock(&cpu_fbatches.lock);
366 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
367 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
368 local_unlock(&cpu_fbatches.lock);
369 }
370 }
371
372 #else
373 static inline void folio_activate_drain(int cpu)
374 {
375 }
376
377 void folio_activate(struct folio *folio)
378 {
379 struct lruvec *lruvec;
380
381 if (folio_test_clear_lru(folio)) {
382 lruvec = folio_lruvec_lock_irq(folio);
383 folio_activate_fn(lruvec, folio);
384 unlock_page_lruvec_irq(lruvec);
385 folio_set_lru(folio);
386 }
387 }
388 #endif
389
390 static void __lru_cache_activate_folio(struct folio *folio)
391 {
392 struct folio_batch *fbatch;
393 int i;
394
395 local_lock(&cpu_fbatches.lock);
396 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
397
398 /*
399 * Search backwards on the optimistic assumption that the folio being
400 * activated has just been added to this batch. Note that only
401 * the local batch is examined as a !LRU folio could be in the
402 * process of being released, reclaimed, migrated or on a remote
403 * batch that is currently being drained. Furthermore, marking
404 * a remote batch's folio active potentially hits a race where
405 * a folio is marked active just after it is added to the inactive
406 * list causing accounting errors and BUG_ON checks to trigger.
407 */
408 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
409 struct folio *batch_folio = fbatch->folios[i];
410
411 if (batch_folio == folio) {
412 folio_set_active(folio);
413 break;
414 }
415 }
416
417 local_unlock(&cpu_fbatches.lock);
418 }
419
420 #ifdef CONFIG_LRU_GEN
421 static void folio_inc_refs(struct folio *folio)
422 {
423 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
424
425 if (folio_test_unevictable(folio))
426 return;
427
428 if (!folio_test_referenced(folio)) {
429 folio_set_referenced(folio);
430 return;
431 }
432
433 if (!folio_test_workingset(folio)) {
434 folio_set_workingset(folio);
435 return;
436 }
437
438 /* see the comment on MAX_NR_TIERS */
439 do {
440 new_flags = old_flags & LRU_REFS_MASK;
441 if (new_flags == LRU_REFS_MASK)
442 break;
443
444 new_flags += BIT(LRU_REFS_PGOFF);
445 new_flags |= old_flags & ~LRU_REFS_MASK;
446 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
447 }
448 #else
449 static void folio_inc_refs(struct folio *folio)
450 {
451 }
452 #endif /* CONFIG_LRU_GEN */
453
454 /**
455 * folio_mark_accessed - Mark a folio as having seen activity.
456 * @folio: The folio to mark.
457 *
458 * This function will perform one of the following transitions:
459 *
460 * * inactive,unreferenced -> inactive,referenced
461 * * inactive,referenced -> active,unreferenced
462 * * active,unreferenced -> active,referenced
463 *
464 * When a newly allocated folio is not yet visible, so safe for non-atomic ops,
465 * __folio_set_referenced() may be substituted for folio_mark_accessed().
466 */
467 void folio_mark_accessed(struct folio *folio)
468 {
469 if (lru_gen_enabled()) {
470 folio_inc_refs(folio);
471 return;
472 }
473
474 if (!folio_test_referenced(folio)) {
475 folio_set_referenced(folio);
476 } else if (folio_test_unevictable(folio)) {
477 /*
478 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
479 * this list is never rotated or maintained, so marking an
480 * unevictable page accessed has no effect.
481 */
482 } else if (!folio_test_active(folio)) {
483 /*
484 * If the folio is on the LRU, queue it for activation via
485 * cpu_fbatches.activate. Otherwise, assume the folio is in a
486 * folio_batch, mark it active and it'll be moved to the active
487 * LRU on the next drain.
488 */
489 if (folio_test_lru(folio))
490 folio_activate(folio);
491 else
492 __lru_cache_activate_folio(folio);
493 folio_clear_referenced(folio);
494 workingset_activation(folio);
495 }
496 if (folio_test_idle(folio))
497 folio_clear_idle(folio);
498 }
499 EXPORT_SYMBOL(folio_mark_accessed);
500
501 /**
502 * folio_add_lru - Add a folio to an LRU list.
503 * @folio: The folio to be added to the LRU.
504 *
505 * Queue the folio for addition to the LRU. The decision on whether
506 * to add the page to the [in]active [file|anon] list is deferred until the
507 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
508 * have the folio added to the active list using folio_mark_accessed().
509 */
510 void folio_add_lru(struct folio *folio)
511 {
512 struct folio_batch *fbatch;
513
514 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
515 folio_test_unevictable(folio), folio);
516 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
517
518 /* see the comment in lru_gen_add_folio() */
519 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
520 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
521 folio_set_active(folio);
522
523 folio_get(folio);
524 local_lock(&cpu_fbatches.lock);
525 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
526 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
527 local_unlock(&cpu_fbatches.lock);
528 }
529 EXPORT_SYMBOL(folio_add_lru);
530
531 /**
532 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
533 * @folio: The folio to be added to the LRU.
534 * @vma: VMA in which the folio is mapped.
535 *
536 * If the VMA is mlocked, @folio is added to the unevictable list.
537 * Otherwise, it is treated the same way as folio_add_lru().
538 */
539 void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
540 {
541 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
542
543 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
544 mlock_new_folio(folio);
545 else
546 folio_add_lru(folio);
547 }
548
549 /*
550 * If the folio cannot be invalidated, it is moved to the
551 * inactive list to speed up its reclaim. It is moved to the
552 * head of the list, rather than the tail, to give the flusher
553 * threads some time to write it out, as this is much more
554 * effective than the single-page writeout from reclaim.
555 *
556 * If the folio isn't mapped and dirty/writeback, the folio
557 * could be reclaimed asap using the reclaim flag.
558 *
559 * 1. active, mapped folio -> none
560 * 2. active, dirty/writeback folio -> inactive, head, reclaim
561 * 3. inactive, mapped folio -> none
562 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
563 * 5. inactive, clean -> inactive, tail
564 * 6. Others -> none
565 *
566 * In 4, it moves to the head of the inactive list so the folio is
567 * written out by flusher threads as this is much more efficient
568 * than the single-page writeout from reclaim.
569 */
570 static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
571 {
572 bool active = folio_test_active(folio);
573 long nr_pages = folio_nr_pages(folio);
574
575 if (folio_test_unevictable(folio))
576 return;
577
578 /* Some processes are using the folio */
579 if (folio_mapped(folio))
580 return;
581
582 lruvec_del_folio(lruvec, folio);
583 folio_clear_active(folio);
584 folio_clear_referenced(folio);
585
586 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
587 /*
588 * Setting the reclaim flag could race with
589 * folio_end_writeback() and confuse readahead. But the
590 * race window is _really_ small and it's not a critical
591 * problem.
592 */
593 lruvec_add_folio(lruvec, folio);
594 folio_set_reclaim(folio);
595 } else {
596 /*
597 * The folio's writeback ended while it was in the batch.
598 * We move that folio to the tail of the inactive list.
599 */
600 lruvec_add_folio_tail(lruvec, folio);
601 __count_vm_events(PGROTATED, nr_pages);
602 }
603
604 if (active) {
605 __count_vm_events(PGDEACTIVATE, nr_pages);
606 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
607 nr_pages);
608 }
609 }
610
611 static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
612 {
613 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
614 long nr_pages = folio_nr_pages(folio);
615
616 lruvec_del_folio(lruvec, folio);
617 folio_clear_active(folio);
618 folio_clear_referenced(folio);
619 lruvec_add_folio(lruvec, folio);
620
621 __count_vm_events(PGDEACTIVATE, nr_pages);
622 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
623 nr_pages);
624 }
625 }
626
627 static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
628 {
629 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
630 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
631 long nr_pages = folio_nr_pages(folio);
632
633 lruvec_del_folio(lruvec, folio);
634 folio_clear_active(folio);
635 folio_clear_referenced(folio);
636 /*
637 * Lazyfree folios are clean anonymous folios. They have
638 * the swapbacked flag cleared, to distinguish them from normal
639 * anonymous folios
640 */
641 folio_clear_swapbacked(folio);
642 lruvec_add_folio(lruvec, folio);
643
644 __count_vm_events(PGLAZYFREE, nr_pages);
645 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
646 nr_pages);
647 }
648 }
649
650 /*
651 * Drain pages out of the cpu's folio_batch.
652 * Either "cpu" is the current CPU, and preemption has already been
653 * disabled; or "cpu" is being hot-unplugged, and is already dead.
654 */
655 void lru_add_drain_cpu(int cpu)
656 {
657 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
658 struct folio_batch *fbatch = &fbatches->lru_add;
659
660 if (folio_batch_count(fbatch))
661 folio_batch_move_lru(fbatch, lru_add_fn);
662
663 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
664 /* Disabling interrupts below acts as a compiler barrier. */
665 if (data_race(folio_batch_count(fbatch))) {
666 unsigned long flags;
667
668 /* No harm done if a racing interrupt already did this */
669 local_lock_irqsave(&lru_rotate.lock, flags);
670 folio_batch_move_lru(fbatch, lru_move_tail_fn);
671 local_unlock_irqrestore(&lru_rotate.lock, flags);
672 }
673
674 fbatch = &fbatches->lru_deactivate_file;
675 if (folio_batch_count(fbatch))
676 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
677
678 fbatch = &fbatches->lru_deactivate;
679 if (folio_batch_count(fbatch))
680 folio_batch_move_lru(fbatch, lru_deactivate_fn);
681
682 fbatch = &fbatches->lru_lazyfree;
683 if (folio_batch_count(fbatch))
684 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
685
686 folio_activate_drain(cpu);
687 }
688
689 /**
690 * deactivate_file_folio() - Deactivate a file folio.
691 * @folio: Folio to deactivate.
692 *
693 * This function hints to the VM that @folio is a good reclaim candidate,
694 * for example if its invalidation fails due to the folio being dirty
695 * or under writeback.
696 *
697 * Context: Caller holds a reference on the folio.
698 */
699 void deactivate_file_folio(struct folio *folio)
700 {
701 struct folio_batch *fbatch;
702
703 /* Deactivating an unevictable folio will not accelerate reclaim */
704 if (folio_test_unevictable(folio))
705 return;
706
707 folio_get(folio);
708 if (!folio_test_clear_lru(folio)) {
709 folio_put(folio);
710 return;
711 }
712
713 local_lock(&cpu_fbatches.lock);
714 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
715 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
716 local_unlock(&cpu_fbatches.lock);
717 }
718
719 /*
720 * folio_deactivate - deactivate a folio
721 * @folio: folio to deactivate
722 *
723 * folio_deactivate() moves @folio to the inactive list if @folio was on the
724 * active list and was not unevictable. This is done to accelerate the
725 * reclaim of @folio.
726 */
727 void folio_deactivate(struct folio *folio)
728 {
729 if (!folio_test_unevictable(folio) && (folio_test_active(folio) ||
730 lru_gen_enabled())) {
731 struct folio_batch *fbatch;
732
733 folio_get(folio);
734 if (!folio_test_clear_lru(folio)) {
735 folio_put(folio);
736 return;
737 }
738
739 local_lock(&cpu_fbatches.lock);
740 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
741 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
742 local_unlock(&cpu_fbatches.lock);
743 }
744 }
745
746 /**
747 * folio_mark_lazyfree - make an anon folio lazyfree
748 * @folio: folio to deactivate
749 *
750 * folio_mark_lazyfree() moves @folio to the inactive file list.
751 * This is done to accelerate the reclaim of @folio.
752 */
753 void folio_mark_lazyfree(struct folio *folio)
754 {
755 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
756 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
757 struct folio_batch *fbatch;
758
759 folio_get(folio);
760 if (!folio_test_clear_lru(folio)) {
761 folio_put(folio);
762 return;
763 }
764
765 local_lock(&cpu_fbatches.lock);
766 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
767 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
768 local_unlock(&cpu_fbatches.lock);
769 }
770 }
771
772 void lru_add_drain(void)
773 {
774 local_lock(&cpu_fbatches.lock);
775 lru_add_drain_cpu(smp_processor_id());
776 local_unlock(&cpu_fbatches.lock);
777 mlock_drain_local();
778 }
779
780 /*
781 * It's called from per-cpu workqueue context in SMP case so
782 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
783 * the same cpu. It shouldn't be a problem in !SMP case since
784 * the core is only one and the locks will disable preemption.
785 */
786 static void lru_add_and_bh_lrus_drain(void)
787 {
788 local_lock(&cpu_fbatches.lock);
789 lru_add_drain_cpu(smp_processor_id());
790 local_unlock(&cpu_fbatches.lock);
791 invalidate_bh_lrus_cpu();
792 mlock_drain_local();
793 }
794
795 void lru_add_drain_cpu_zone(struct zone *zone)
796 {
797 local_lock(&cpu_fbatches.lock);
798 lru_add_drain_cpu(smp_processor_id());
799 drain_local_pages(zone);
800 local_unlock(&cpu_fbatches.lock);
801 mlock_drain_local();
802 }
803
804 #ifdef CONFIG_SMP
805
806 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
807
808 static void lru_add_drain_per_cpu(struct work_struct *dummy)
809 {
810 lru_add_and_bh_lrus_drain();
811 }
812
813 static bool cpu_needs_drain(unsigned int cpu)
814 {
815 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
816
817 /* Check these in order of likelihood that they're not zero */
818 return folio_batch_count(&fbatches->lru_add) ||
819 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
820 folio_batch_count(&fbatches->lru_deactivate_file) ||
821 folio_batch_count(&fbatches->lru_deactivate) ||
822 folio_batch_count(&fbatches->lru_lazyfree) ||
823 folio_batch_count(&fbatches->activate) ||
824 need_mlock_drain(cpu) ||
825 has_bh_in_lru(cpu, NULL);
826 }
827
828 /*
829 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
830 * kworkers being shut down before our page_alloc_cpu_dead callback is
831 * executed on the offlined cpu.
832 * Calling this function with cpu hotplug locks held can actually lead
833 * to obscure indirect dependencies via WQ context.
834 */
835 static inline void __lru_add_drain_all(bool force_all_cpus)
836 {
837 /*
838 * lru_drain_gen - Global pages generation number
839 *
840 * (A) Definition: global lru_drain_gen = x implies that all generations
841 * 0 < n <= x are already *scheduled* for draining.
842 *
843 * This is an optimization for the highly-contended use case where a
844 * user space workload keeps constantly generating a flow of pages for
845 * each CPU.
846 */
847 static unsigned int lru_drain_gen;
848 static struct cpumask has_work;
849 static DEFINE_MUTEX(lock);
850 unsigned cpu, this_gen;
851
852 /*
853 * Make sure nobody triggers this path before mm_percpu_wq is fully
854 * initialized.
855 */
856 if (WARN_ON(!mm_percpu_wq))
857 return;
858
859 /*
860 * Guarantee folio_batch counter stores visible by this CPU
861 * are visible to other CPUs before loading the current drain
862 * generation.
863 */
864 smp_mb();
865
866 /*
867 * (B) Locally cache global LRU draining generation number
868 *
869 * The read barrier ensures that the counter is loaded before the mutex
870 * is taken. It pairs with smp_mb() inside the mutex critical section
871 * at (D).
872 */
873 this_gen = smp_load_acquire(&lru_drain_gen);
874
875 mutex_lock(&lock);
876
877 /*
878 * (C) Exit the draining operation if a newer generation, from another
879 * lru_add_drain_all(), was already scheduled for draining. Check (A).
880 */
881 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
882 goto done;
883
884 /*
885 * (D) Increment global generation number
886 *
887 * Pairs with smp_load_acquire() at (B), outside of the critical
888 * section. Use a full memory barrier to guarantee that the
889 * new global drain generation number is stored before loading
890 * folio_batch counters.
891 *
892 * This pairing must be done here, before the for_each_online_cpu loop
893 * below which drains the page vectors.
894 *
895 * Let x, y, and z represent some system CPU numbers, where x < y < z.
896 * Assume CPU #z is in the middle of the for_each_online_cpu loop
897 * below and has already reached CPU #y's per-cpu data. CPU #x comes
898 * along, adds some pages to its per-cpu vectors, then calls
899 * lru_add_drain_all().
900 *
901 * If the paired barrier is done at any later step, e.g. after the
902 * loop, CPU #x will just exit at (C) and miss flushing out all of its
903 * added pages.
904 */
905 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
906 smp_mb();
907
908 cpumask_clear(&has_work);
909 for_each_online_cpu(cpu) {
910 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
911
912 if (cpu_needs_drain(cpu)) {
913 INIT_WORK(work, lru_add_drain_per_cpu);
914 queue_work_on(cpu, mm_percpu_wq, work);
915 __cpumask_set_cpu(cpu, &has_work);
916 }
917 }
918
919 for_each_cpu(cpu, &has_work)
920 flush_work(&per_cpu(lru_add_drain_work, cpu));
921
922 done:
923 mutex_unlock(&lock);
924 }
925
926 void lru_add_drain_all(void)
927 {
928 __lru_add_drain_all(false);
929 }
930 #else
931 void lru_add_drain_all(void)
932 {
933 lru_add_drain();
934 }
935 #endif /* CONFIG_SMP */
936
937 atomic_t lru_disable_count = ATOMIC_INIT(0);
938
939 /*
940 * lru_cache_disable() needs to be called before we start compiling
941 * a list of pages to be migrated using isolate_lru_page().
942 * It drains pages on LRU cache and then disable on all cpus until
943 * lru_cache_enable is called.
944 *
945 * Must be paired with a call to lru_cache_enable().
946 */
947 void lru_cache_disable(void)
948 {
949 atomic_inc(&lru_disable_count);
950 /*
951 * Readers of lru_disable_count are protected by either disabling
952 * preemption or rcu_read_lock:
953 *
954 * preempt_disable, local_irq_disable [bh_lru_lock()]
955 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
956 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
957 *
958 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
959 * preempt_disable() regions of code. So any CPU which sees
960 * lru_disable_count = 0 will have exited the critical
961 * section when synchronize_rcu() returns.
962 */
963 synchronize_rcu_expedited();
964 #ifdef CONFIG_SMP
965 __lru_add_drain_all(true);
966 #else
967 lru_add_and_bh_lrus_drain();
968 #endif
969 }
970
971 /**
972 * folios_put_refs - Reduce the reference count on a batch of folios.
973 * @folios: The folios.
974 * @refs: The number of refs to subtract from each folio.
975 *
976 * Like folio_put(), but for a batch of folios. This is more efficient
977 * than writing the loop yourself as it will optimise the locks which need
978 * to be taken if the folios are freed. The folios batch is returned
979 * empty and ready to be reused for another batch; there is no need
980 * to reinitialise it. If @refs is NULL, we subtract one from each
981 * folio refcount.
982 *
983 * Context: May be called in process or interrupt context, but not in NMI
984 * context. May be called while holding a spinlock.
985 */
986 void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
987 {
988 int i, j;
989 struct lruvec *lruvec = NULL;
990 unsigned long flags = 0;
991
992 for (i = 0, j = 0; i < folios->nr; i++) {
993 struct folio *folio = folios->folios[i];
994 unsigned int nr_refs = refs ? refs[i] : 1;
995
996 if (is_huge_zero_folio(folio))
997 continue;
998
999 if (folio_is_zone_device(folio)) {
1000 if (lruvec) {
1001 unlock_page_lruvec_irqrestore(lruvec, flags);
1002 lruvec = NULL;
1003 }
1004 if (put_devmap_managed_folio_refs(folio, nr_refs))
1005 continue;
1006 if (folio_ref_sub_and_test(folio, nr_refs))
1007 free_zone_device_folio(folio);
1008 continue;
1009 }
1010
1011 if (!folio_ref_sub_and_test(folio, nr_refs))
1012 continue;
1013
1014 /* hugetlb has its own memcg */
1015 if (folio_test_hugetlb(folio)) {
1016 if (lruvec) {
1017 unlock_page_lruvec_irqrestore(lruvec, flags);
1018 lruvec = NULL;
1019 }
1020 free_huge_folio(folio);
1021 continue;
1022 }
1023 folio_undo_large_rmappable(folio);
1024 __page_cache_release(folio, &lruvec, &flags);
1025
1026 if (j != i)
1027 folios->folios[j] = folio;
1028 j++;
1029 }
1030 if (lruvec)
1031 unlock_page_lruvec_irqrestore(lruvec, flags);
1032 if (!j) {
1033 folio_batch_reinit(folios);
1034 return;
1035 }
1036
1037 folios->nr = j;
1038 mem_cgroup_uncharge_folios(folios);
1039 free_unref_folios(folios);
1040 }
1041 EXPORT_SYMBOL(folios_put_refs);
1042
1043 /**
1044 * release_pages - batched put_page()
1045 * @arg: array of pages to release
1046 * @nr: number of pages
1047 *
1048 * Decrement the reference count on all the pages in @arg. If it
1049 * fell to zero, remove the page from the LRU and free it.
1050 *
1051 * Note that the argument can be an array of pages, encoded pages,
1052 * or folio pointers. We ignore any encoded bits, and turn any of
1053 * them into just a folio that gets free'd.
1054 */
1055 void release_pages(release_pages_arg arg, int nr)
1056 {
1057 struct folio_batch fbatch;
1058 int refs[PAGEVEC_SIZE];
1059 struct encoded_page **encoded = arg.encoded_pages;
1060 int i;
1061
1062 folio_batch_init(&fbatch);
1063 for (i = 0; i < nr; i++) {
1064 /* Turn any of the argument types into a folio */
1065 struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
1066
1067 /* Is our next entry actually "nr_pages" -> "nr_refs" ? */
1068 refs[fbatch.nr] = 1;
1069 if (unlikely(encoded_page_flags(encoded[i]) &
1070 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
1071 refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
1072
1073 if (folio_batch_add(&fbatch, folio) > 0)
1074 continue;
1075 folios_put_refs(&fbatch, refs);
1076 }
1077
1078 if (fbatch.nr)
1079 folios_put_refs(&fbatch, refs);
1080 }
1081 EXPORT_SYMBOL(release_pages);
1082
1083 /*
1084 * The folios which we're about to release may be in the deferred lru-addition
1085 * queues. That would prevent them from really being freed right now. That's
1086 * OK from a correctness point of view but is inefficient - those folios may be
1087 * cache-warm and we want to give them back to the page allocator ASAP.
1088 *
1089 * So __folio_batch_release() will drain those queues here.
1090 * folio_batch_move_lru() calls folios_put() directly to avoid
1091 * mutual recursion.
1092 */
1093 void __folio_batch_release(struct folio_batch *fbatch)
1094 {
1095 if (!fbatch->percpu_pvec_drained) {
1096 lru_add_drain();
1097 fbatch->percpu_pvec_drained = true;
1098 }
1099 folios_put(fbatch);
1100 }
1101 EXPORT_SYMBOL(__folio_batch_release);
1102
1103 /**
1104 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1105 * @fbatch: The batch to prune
1106 *
1107 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1108 * entries. This function prunes all the non-folio entries from @fbatch
1109 * without leaving holes, so that it can be passed on to folio-only batch
1110 * operations.
1111 */
1112 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1113 {
1114 unsigned int i, j;
1115
1116 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1117 struct folio *folio = fbatch->folios[i];
1118 if (!xa_is_value(folio))
1119 fbatch->folios[j++] = folio;
1120 }
1121 fbatch->nr = j;
1122 }
1123
1124 /*
1125 * Perform any setup for the swap system
1126 */
1127 void __init swap_setup(void)
1128 {
1129 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1130
1131 /* Use a smaller cluster for small-memory machines */
1132 if (megs < 16)
1133 page_cluster = 2;
1134 else
1135 page_cluster = 3;
1136 /*
1137 * Right now other parts of the system means that we
1138 * _really_ don't want to cluster much more
1139 */
1140 }
1141
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.