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 static const int page_cluster_max = 31;
49
50 struct cpu_fbatches {
51 /*
52 * The following folio batches are grouped together because they are protected
53 * by disabling preemption (and interrupts remain enabled).
54 */
55 local_lock_t lock;
56 struct folio_batch lru_add;
57 struct folio_batch lru_deactivate_file;
58 struct folio_batch lru_deactivate;
59 struct folio_batch lru_lazyfree;
60 #ifdef CONFIG_SMP
61 struct folio_batch lru_activate;
62 #endif
63 /* Protecting the following batches which require disabling interrupts */
64 local_lock_t lock_irq;
65 struct folio_batch lru_move_tail;
66 };
67
68 static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
69 .lock = INIT_LOCAL_LOCK(lock),
70 .lock_irq = INIT_LOCAL_LOCK(lock_irq),
71 };
72
__page_cache_release(struct folio * folio,struct lruvec ** lruvecp,unsigned long * flagsp)73 static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
74 unsigned long *flagsp)
75 {
76 if (folio_test_lru(folio)) {
77 folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
78 lruvec_del_folio(*lruvecp, folio);
79 __folio_clear_lru_flags(folio);
80 }
81 }
82
83 /*
84 * This path almost never happens for VM activity - pages are normally freed
85 * in batches. But it gets used by networking - and for compound pages.
86 */
page_cache_release(struct folio * folio)87 static void page_cache_release(struct folio *folio)
88 {
89 struct lruvec *lruvec = NULL;
90 unsigned long flags;
91
92 __page_cache_release(folio, &lruvec, &flags);
93 if (lruvec)
94 unlock_page_lruvec_irqrestore(lruvec, flags);
95 }
96
__folio_put(struct folio * folio)97 void __folio_put(struct folio *folio)
98 {
99 if (unlikely(folio_is_zone_device(folio))) {
100 free_zone_device_folio(folio);
101 return;
102 }
103
104 if (folio_test_hugetlb(folio)) {
105 free_huge_folio(folio);
106 return;
107 }
108
109 page_cache_release(folio);
110 folio_unqueue_deferred_split(folio);
111 mem_cgroup_uncharge(folio);
112 free_frozen_pages(&folio->page, folio_order(folio));
113 }
114 EXPORT_SYMBOL(__folio_put);
115
116 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
117
lru_add(struct lruvec * lruvec,struct folio * folio)118 static void lru_add(struct lruvec *lruvec, struct folio *folio)
119 {
120 int was_unevictable = folio_test_clear_unevictable(folio);
121 long nr_pages = folio_nr_pages(folio);
122
123 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
124
125 /*
126 * Is an smp_mb__after_atomic() still required here, before
127 * folio_evictable() tests the mlocked flag, to rule out the possibility
128 * of stranding an evictable folio on an unevictable LRU? I think
129 * not, because __munlock_folio() only clears the mlocked flag
130 * while the LRU lock is held.
131 *
132 * (That is not true of __page_cache_release(), and not necessarily
133 * true of folios_put(): but those only clear the mlocked flag after
134 * folio_put_testzero() has excluded any other users of the folio.)
135 */
136 if (folio_evictable(folio)) {
137 if (was_unevictable)
138 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
139 } else {
140 folio_clear_active(folio);
141 folio_set_unevictable(folio);
142 /*
143 * folio->mlock_count = !!folio_test_mlocked(folio)?
144 * But that leaves __mlock_folio() in doubt whether another
145 * actor has already counted the mlock or not. Err on the
146 * safe side, underestimate, let page reclaim fix it, rather
147 * than leaving a page on the unevictable LRU indefinitely.
148 */
149 folio->mlock_count = 0;
150 if (!was_unevictable)
151 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
152 }
153
154 lruvec_add_folio(lruvec, folio);
155 trace_mm_lru_insertion(folio);
156 }
157
folio_batch_move_lru(struct folio_batch * fbatch,move_fn_t move_fn)158 static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
159 {
160 int i;
161 struct lruvec *lruvec = NULL;
162 unsigned long flags = 0;
163
164 for (i = 0; i < folio_batch_count(fbatch); i++) {
165 struct folio *folio = fbatch->folios[i];
166
167 folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
168 move_fn(lruvec, folio);
169
170 folio_set_lru(folio);
171 }
172
173 if (lruvec)
174 unlock_page_lruvec_irqrestore(lruvec, flags);
175 folios_put(fbatch);
176 }
177
__folio_batch_add_and_move(struct folio_batch __percpu * fbatch,struct folio * folio,move_fn_t move_fn,bool on_lru,bool disable_irq)178 static void __folio_batch_add_and_move(struct folio_batch __percpu *fbatch,
179 struct folio *folio, move_fn_t move_fn,
180 bool on_lru, bool disable_irq)
181 {
182 unsigned long flags;
183
184 if (on_lru && !folio_test_clear_lru(folio))
185 return;
186
187 folio_get(folio);
188
189 if (disable_irq)
190 local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
191 else
192 local_lock(&cpu_fbatches.lock);
193
194 if (!folio_batch_add(this_cpu_ptr(fbatch), folio) || folio_test_large(folio) ||
195 lru_cache_disabled())
196 folio_batch_move_lru(this_cpu_ptr(fbatch), move_fn);
197
198 if (disable_irq)
199 local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
200 else
201 local_unlock(&cpu_fbatches.lock);
202 }
203
204 #define folio_batch_add_and_move(folio, op, on_lru) \
205 __folio_batch_add_and_move( \
206 &cpu_fbatches.op, \
207 folio, \
208 op, \
209 on_lru, \
210 offsetof(struct cpu_fbatches, op) >= offsetof(struct cpu_fbatches, lock_irq) \
211 )
212
lru_move_tail(struct lruvec * lruvec,struct folio * folio)213 static void lru_move_tail(struct lruvec *lruvec, struct folio *folio)
214 {
215 if (folio_test_unevictable(folio))
216 return;
217
218 lruvec_del_folio(lruvec, folio);
219 folio_clear_active(folio);
220 lruvec_add_folio_tail(lruvec, folio);
221 __count_vm_events(PGROTATED, folio_nr_pages(folio));
222 }
223
224 /*
225 * Writeback is about to end against a folio which has been marked for
226 * immediate reclaim. If it still appears to be reclaimable, move it
227 * to the tail of the inactive list.
228 *
229 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
230 */
folio_rotate_reclaimable(struct folio * folio)231 void folio_rotate_reclaimable(struct folio *folio)
232 {
233 if (folio_test_locked(folio) || folio_test_dirty(folio) ||
234 folio_test_unevictable(folio))
235 return;
236
237 folio_batch_add_and_move(folio, lru_move_tail, true);
238 }
239
lru_note_cost(struct lruvec * lruvec,bool file,unsigned int nr_io,unsigned int nr_rotated)240 void lru_note_cost(struct lruvec *lruvec, bool file,
241 unsigned int nr_io, unsigned int nr_rotated)
242 {
243 unsigned long cost;
244
245 /*
246 * Reflect the relative cost of incurring IO and spending CPU
247 * time on rotations. This doesn't attempt to make a precise
248 * comparison, it just says: if reloads are about comparable
249 * between the LRU lists, or rotations are overwhelmingly
250 * different between them, adjust scan balance for CPU work.
251 */
252 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
253
254 do {
255 unsigned long lrusize;
256
257 /*
258 * Hold lruvec->lru_lock is safe here, since
259 * 1) The pinned lruvec in reclaim, or
260 * 2) From a pre-LRU page during refault (which also holds the
261 * rcu lock, so would be safe even if the page was on the LRU
262 * and could move simultaneously to a new lruvec).
263 */
264 spin_lock_irq(&lruvec->lru_lock);
265 /* Record cost event */
266 if (file)
267 lruvec->file_cost += cost;
268 else
269 lruvec->anon_cost += cost;
270
271 /*
272 * Decay previous events
273 *
274 * Because workloads change over time (and to avoid
275 * overflow) we keep these statistics as a floating
276 * average, which ends up weighing recent refaults
277 * more than old ones.
278 */
279 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
280 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
281 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
282 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
283
284 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
285 lruvec->file_cost /= 2;
286 lruvec->anon_cost /= 2;
287 }
288 spin_unlock_irq(&lruvec->lru_lock);
289 } while ((lruvec = parent_lruvec(lruvec)));
290 }
291
lru_note_cost_refault(struct folio * folio)292 void lru_note_cost_refault(struct folio *folio)
293 {
294 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
295 folio_nr_pages(folio), 0);
296 }
297
lru_activate(struct lruvec * lruvec,struct folio * folio)298 static void lru_activate(struct lruvec *lruvec, struct folio *folio)
299 {
300 long nr_pages = folio_nr_pages(folio);
301
302 if (folio_test_active(folio) || folio_test_unevictable(folio))
303 return;
304
305
306 lruvec_del_folio(lruvec, folio);
307 folio_set_active(folio);
308 lruvec_add_folio(lruvec, folio);
309 trace_mm_lru_activate(folio);
310
311 __count_vm_events(PGACTIVATE, nr_pages);
312 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE, nr_pages);
313 }
314
315 #ifdef CONFIG_SMP
folio_activate_drain(int cpu)316 static void folio_activate_drain(int cpu)
317 {
318 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.lru_activate, cpu);
319
320 if (folio_batch_count(fbatch))
321 folio_batch_move_lru(fbatch, lru_activate);
322 }
323
folio_activate(struct folio * folio)324 void folio_activate(struct folio *folio)
325 {
326 if (folio_test_active(folio) || folio_test_unevictable(folio))
327 return;
328
329 folio_batch_add_and_move(folio, lru_activate, true);
330 }
331
332 #else
folio_activate_drain(int cpu)333 static inline void folio_activate_drain(int cpu)
334 {
335 }
336
folio_activate(struct folio * folio)337 void folio_activate(struct folio *folio)
338 {
339 struct lruvec *lruvec;
340
341 if (!folio_test_clear_lru(folio))
342 return;
343
344 lruvec = folio_lruvec_lock_irq(folio);
345 lru_activate(lruvec, folio);
346 unlock_page_lruvec_irq(lruvec);
347 folio_set_lru(folio);
348 }
349 #endif
350
__lru_cache_activate_folio(struct folio * folio)351 static void __lru_cache_activate_folio(struct folio *folio)
352 {
353 struct folio_batch *fbatch;
354 int i;
355
356 local_lock(&cpu_fbatches.lock);
357 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
358
359 /*
360 * Search backwards on the optimistic assumption that the folio being
361 * activated has just been added to this batch. Note that only
362 * the local batch is examined as a !LRU folio could be in the
363 * process of being released, reclaimed, migrated or on a remote
364 * batch that is currently being drained. Furthermore, marking
365 * a remote batch's folio active potentially hits a race where
366 * a folio is marked active just after it is added to the inactive
367 * list causing accounting errors and BUG_ON checks to trigger.
368 */
369 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
370 struct folio *batch_folio = fbatch->folios[i];
371
372 if (batch_folio == folio) {
373 folio_set_active(folio);
374 break;
375 }
376 }
377
378 local_unlock(&cpu_fbatches.lock);
379 }
380
381 #ifdef CONFIG_LRU_GEN
382
lru_gen_inc_refs(struct folio * folio)383 static void lru_gen_inc_refs(struct folio *folio)
384 {
385 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
386
387 if (folio_test_unevictable(folio))
388 return;
389
390 /* see the comment on LRU_REFS_FLAGS */
391 if (!folio_test_referenced(folio)) {
392 set_mask_bits(&folio->flags, LRU_REFS_MASK, BIT(PG_referenced));
393 return;
394 }
395
396 do {
397 if ((old_flags & LRU_REFS_MASK) == LRU_REFS_MASK) {
398 if (!folio_test_workingset(folio))
399 folio_set_workingset(folio);
400 return;
401 }
402
403 new_flags = old_flags + BIT(LRU_REFS_PGOFF);
404 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
405 }
406
lru_gen_clear_refs(struct folio * folio)407 static bool lru_gen_clear_refs(struct folio *folio)
408 {
409 struct lru_gen_folio *lrugen;
410 int gen = folio_lru_gen(folio);
411 int type = folio_is_file_lru(folio);
412
413 if (gen < 0)
414 return true;
415
416 set_mask_bits(&folio->flags, LRU_REFS_FLAGS | BIT(PG_workingset), 0);
417
418 lrugen = &folio_lruvec(folio)->lrugen;
419 /* whether can do without shuffling under the LRU lock */
420 return gen == lru_gen_from_seq(READ_ONCE(lrugen->min_seq[type]));
421 }
422
423 #else /* !CONFIG_LRU_GEN */
424
lru_gen_inc_refs(struct folio * folio)425 static void lru_gen_inc_refs(struct folio *folio)
426 {
427 }
428
lru_gen_clear_refs(struct folio * folio)429 static bool lru_gen_clear_refs(struct folio *folio)
430 {
431 return false;
432 }
433
434 #endif /* CONFIG_LRU_GEN */
435
436 /**
437 * folio_mark_accessed - Mark a folio as having seen activity.
438 * @folio: The folio to mark.
439 *
440 * This function will perform one of the following transitions:
441 *
442 * * inactive,unreferenced -> inactive,referenced
443 * * inactive,referenced -> active,unreferenced
444 * * active,unreferenced -> active,referenced
445 *
446 * When a newly allocated folio is not yet visible, so safe for non-atomic ops,
447 * __folio_set_referenced() may be substituted for folio_mark_accessed().
448 */
folio_mark_accessed(struct folio * folio)449 void folio_mark_accessed(struct folio *folio)
450 {
451 if (folio_test_dropbehind(folio))
452 return;
453 if (lru_gen_enabled()) {
454 lru_gen_inc_refs(folio);
455 return;
456 }
457
458 if (!folio_test_referenced(folio)) {
459 folio_set_referenced(folio);
460 } else if (folio_test_unevictable(folio)) {
461 /*
462 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
463 * this list is never rotated or maintained, so marking an
464 * unevictable page accessed has no effect.
465 */
466 } else if (!folio_test_active(folio)) {
467 /*
468 * If the folio is on the LRU, queue it for activation via
469 * cpu_fbatches.lru_activate. Otherwise, assume the folio is in a
470 * folio_batch, mark it active and it'll be moved to the active
471 * LRU on the next drain.
472 */
473 if (folio_test_lru(folio))
474 folio_activate(folio);
475 else
476 __lru_cache_activate_folio(folio);
477 folio_clear_referenced(folio);
478 workingset_activation(folio);
479 }
480 if (folio_test_idle(folio))
481 folio_clear_idle(folio);
482 }
483 EXPORT_SYMBOL(folio_mark_accessed);
484
485 /**
486 * folio_add_lru - Add a folio to an LRU list.
487 * @folio: The folio to be added to the LRU.
488 *
489 * Queue the folio for addition to the LRU. The decision on whether
490 * to add the page to the [in]active [file|anon] list is deferred until the
491 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
492 * have the folio added to the active list using folio_mark_accessed().
493 */
folio_add_lru(struct folio * folio)494 void folio_add_lru(struct folio *folio)
495 {
496 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
497 folio_test_unevictable(folio), folio);
498 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
499
500 /* see the comment in lru_gen_folio_seq() */
501 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
502 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
503 folio_set_active(folio);
504
505 folio_batch_add_and_move(folio, lru_add, false);
506 }
507 EXPORT_SYMBOL(folio_add_lru);
508
509 /**
510 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
511 * @folio: The folio to be added to the LRU.
512 * @vma: VMA in which the folio is mapped.
513 *
514 * If the VMA is mlocked, @folio is added to the unevictable list.
515 * Otherwise, it is treated the same way as folio_add_lru().
516 */
folio_add_lru_vma(struct folio * folio,struct vm_area_struct * vma)517 void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
518 {
519 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
520
521 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
522 mlock_new_folio(folio);
523 else
524 folio_add_lru(folio);
525 }
526
527 /*
528 * If the folio cannot be invalidated, it is moved to the
529 * inactive list to speed up its reclaim. It is moved to the
530 * head of the list, rather than the tail, to give the flusher
531 * threads some time to write it out, as this is much more
532 * effective than the single-page writeout from reclaim.
533 *
534 * If the folio isn't mapped and dirty/writeback, the folio
535 * could be reclaimed asap using the reclaim flag.
536 *
537 * 1. active, mapped folio -> none
538 * 2. active, dirty/writeback folio -> inactive, head, reclaim
539 * 3. inactive, mapped folio -> none
540 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
541 * 5. inactive, clean -> inactive, tail
542 * 6. Others -> none
543 *
544 * In 4, it moves to the head of the inactive list so the folio is
545 * written out by flusher threads as this is much more efficient
546 * than the single-page writeout from reclaim.
547 */
lru_deactivate_file(struct lruvec * lruvec,struct folio * folio)548 static void lru_deactivate_file(struct lruvec *lruvec, struct folio *folio)
549 {
550 bool active = folio_test_active(folio) || lru_gen_enabled();
551 long nr_pages = folio_nr_pages(folio);
552
553 if (folio_test_unevictable(folio))
554 return;
555
556 /* Some processes are using the folio */
557 if (folio_mapped(folio))
558 return;
559
560 lruvec_del_folio(lruvec, folio);
561 folio_clear_active(folio);
562 folio_clear_referenced(folio);
563
564 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
565 /*
566 * Setting the reclaim flag could race with
567 * folio_end_writeback() and confuse readahead. But the
568 * race window is _really_ small and it's not a critical
569 * problem.
570 */
571 lruvec_add_folio(lruvec, folio);
572 folio_set_reclaim(folio);
573 } else {
574 /*
575 * The folio's writeback ended while it was in the batch.
576 * We move that folio to the tail of the inactive list.
577 */
578 lruvec_add_folio_tail(lruvec, folio);
579 __count_vm_events(PGROTATED, nr_pages);
580 }
581
582 if (active) {
583 __count_vm_events(PGDEACTIVATE, nr_pages);
584 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
585 nr_pages);
586 }
587 }
588
lru_deactivate(struct lruvec * lruvec,struct folio * folio)589 static void lru_deactivate(struct lruvec *lruvec, struct folio *folio)
590 {
591 long nr_pages = folio_nr_pages(folio);
592
593 if (folio_test_unevictable(folio) || !(folio_test_active(folio) || lru_gen_enabled()))
594 return;
595
596 lruvec_del_folio(lruvec, folio);
597 folio_clear_active(folio);
598 folio_clear_referenced(folio);
599 lruvec_add_folio(lruvec, folio);
600
601 __count_vm_events(PGDEACTIVATE, nr_pages);
602 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_pages);
603 }
604
lru_lazyfree(struct lruvec * lruvec,struct folio * folio)605 static void lru_lazyfree(struct lruvec *lruvec, struct folio *folio)
606 {
607 long nr_pages = folio_nr_pages(folio);
608
609 if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
610 folio_test_swapcache(folio) || folio_test_unevictable(folio))
611 return;
612
613 lruvec_del_folio(lruvec, folio);
614 folio_clear_active(folio);
615 if (lru_gen_enabled())
616 lru_gen_clear_refs(folio);
617 else
618 folio_clear_referenced(folio);
619 /*
620 * Lazyfree folios are clean anonymous folios. They have
621 * the swapbacked flag cleared, to distinguish them from normal
622 * anonymous folios
623 */
624 folio_clear_swapbacked(folio);
625 lruvec_add_folio(lruvec, folio);
626
627 __count_vm_events(PGLAZYFREE, nr_pages);
628 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE, nr_pages);
629 }
630
631 /*
632 * Drain pages out of the cpu's folio_batch.
633 * Either "cpu" is the current CPU, and preemption has already been
634 * disabled; or "cpu" is being hot-unplugged, and is already dead.
635 */
lru_add_drain_cpu(int cpu)636 void lru_add_drain_cpu(int cpu)
637 {
638 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
639 struct folio_batch *fbatch = &fbatches->lru_add;
640
641 if (folio_batch_count(fbatch))
642 folio_batch_move_lru(fbatch, lru_add);
643
644 fbatch = &fbatches->lru_move_tail;
645 /* Disabling interrupts below acts as a compiler barrier. */
646 if (data_race(folio_batch_count(fbatch))) {
647 unsigned long flags;
648
649 /* No harm done if a racing interrupt already did this */
650 local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
651 folio_batch_move_lru(fbatch, lru_move_tail);
652 local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
653 }
654
655 fbatch = &fbatches->lru_deactivate_file;
656 if (folio_batch_count(fbatch))
657 folio_batch_move_lru(fbatch, lru_deactivate_file);
658
659 fbatch = &fbatches->lru_deactivate;
660 if (folio_batch_count(fbatch))
661 folio_batch_move_lru(fbatch, lru_deactivate);
662
663 fbatch = &fbatches->lru_lazyfree;
664 if (folio_batch_count(fbatch))
665 folio_batch_move_lru(fbatch, lru_lazyfree);
666
667 folio_activate_drain(cpu);
668 }
669
670 /**
671 * deactivate_file_folio() - Deactivate a file folio.
672 * @folio: Folio to deactivate.
673 *
674 * This function hints to the VM that @folio is a good reclaim candidate,
675 * for example if its invalidation fails due to the folio being dirty
676 * or under writeback.
677 *
678 * Context: Caller holds a reference on the folio.
679 */
deactivate_file_folio(struct folio * folio)680 void deactivate_file_folio(struct folio *folio)
681 {
682 /* Deactivating an unevictable folio will not accelerate reclaim */
683 if (folio_test_unevictable(folio))
684 return;
685
686 if (lru_gen_enabled() && lru_gen_clear_refs(folio))
687 return;
688
689 folio_batch_add_and_move(folio, lru_deactivate_file, true);
690 }
691
692 /*
693 * folio_deactivate - deactivate a folio
694 * @folio: folio to deactivate
695 *
696 * folio_deactivate() moves @folio to the inactive list if @folio was on the
697 * active list and was not unevictable. This is done to accelerate the
698 * reclaim of @folio.
699 */
folio_deactivate(struct folio * folio)700 void folio_deactivate(struct folio *folio)
701 {
702 if (folio_test_unevictable(folio))
703 return;
704
705 if (lru_gen_enabled() ? lru_gen_clear_refs(folio) : !folio_test_active(folio))
706 return;
707
708 folio_batch_add_and_move(folio, lru_deactivate, true);
709 }
710
711 /**
712 * folio_mark_lazyfree - make an anon folio lazyfree
713 * @folio: folio to deactivate
714 *
715 * folio_mark_lazyfree() moves @folio to the inactive file list.
716 * This is done to accelerate the reclaim of @folio.
717 */
folio_mark_lazyfree(struct folio * folio)718 void folio_mark_lazyfree(struct folio *folio)
719 {
720 if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
721 folio_test_swapcache(folio) || folio_test_unevictable(folio))
722 return;
723
724 folio_batch_add_and_move(folio, lru_lazyfree, true);
725 }
726
lru_add_drain(void)727 void lru_add_drain(void)
728 {
729 local_lock(&cpu_fbatches.lock);
730 lru_add_drain_cpu(smp_processor_id());
731 local_unlock(&cpu_fbatches.lock);
732 mlock_drain_local();
733 }
734
735 /*
736 * It's called from per-cpu workqueue context in SMP case so
737 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
738 * the same cpu. It shouldn't be a problem in !SMP case since
739 * the core is only one and the locks will disable preemption.
740 */
lru_add_and_bh_lrus_drain(void)741 static void lru_add_and_bh_lrus_drain(void)
742 {
743 local_lock(&cpu_fbatches.lock);
744 lru_add_drain_cpu(smp_processor_id());
745 local_unlock(&cpu_fbatches.lock);
746 invalidate_bh_lrus_cpu();
747 mlock_drain_local();
748 }
749
lru_add_drain_cpu_zone(struct zone * zone)750 void lru_add_drain_cpu_zone(struct zone *zone)
751 {
752 local_lock(&cpu_fbatches.lock);
753 lru_add_drain_cpu(smp_processor_id());
754 drain_local_pages(zone);
755 local_unlock(&cpu_fbatches.lock);
756 mlock_drain_local();
757 }
758
759 #ifdef CONFIG_SMP
760
761 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
762
lru_add_drain_per_cpu(struct work_struct * dummy)763 static void lru_add_drain_per_cpu(struct work_struct *dummy)
764 {
765 lru_add_and_bh_lrus_drain();
766 }
767
cpu_needs_drain(unsigned int cpu)768 static bool cpu_needs_drain(unsigned int cpu)
769 {
770 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
771
772 /* Check these in order of likelihood that they're not zero */
773 return folio_batch_count(&fbatches->lru_add) ||
774 folio_batch_count(&fbatches->lru_move_tail) ||
775 folio_batch_count(&fbatches->lru_deactivate_file) ||
776 folio_batch_count(&fbatches->lru_deactivate) ||
777 folio_batch_count(&fbatches->lru_lazyfree) ||
778 folio_batch_count(&fbatches->lru_activate) ||
779 need_mlock_drain(cpu) ||
780 has_bh_in_lru(cpu, NULL);
781 }
782
783 /*
784 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
785 * kworkers being shut down before our page_alloc_cpu_dead callback is
786 * executed on the offlined cpu.
787 * Calling this function with cpu hotplug locks held can actually lead
788 * to obscure indirect dependencies via WQ context.
789 */
__lru_add_drain_all(bool force_all_cpus)790 static inline void __lru_add_drain_all(bool force_all_cpus)
791 {
792 /*
793 * lru_drain_gen - Global pages generation number
794 *
795 * (A) Definition: global lru_drain_gen = x implies that all generations
796 * 0 < n <= x are already *scheduled* for draining.
797 *
798 * This is an optimization for the highly-contended use case where a
799 * user space workload keeps constantly generating a flow of pages for
800 * each CPU.
801 */
802 static unsigned int lru_drain_gen;
803 static struct cpumask has_work;
804 static DEFINE_MUTEX(lock);
805 unsigned cpu, this_gen;
806
807 /*
808 * Make sure nobody triggers this path before mm_percpu_wq is fully
809 * initialized.
810 */
811 if (WARN_ON(!mm_percpu_wq))
812 return;
813
814 /*
815 * Guarantee folio_batch counter stores visible by this CPU
816 * are visible to other CPUs before loading the current drain
817 * generation.
818 */
819 smp_mb();
820
821 /*
822 * (B) Locally cache global LRU draining generation number
823 *
824 * The read barrier ensures that the counter is loaded before the mutex
825 * is taken. It pairs with smp_mb() inside the mutex critical section
826 * at (D).
827 */
828 this_gen = smp_load_acquire(&lru_drain_gen);
829
830 mutex_lock(&lock);
831
832 /*
833 * (C) Exit the draining operation if a newer generation, from another
834 * lru_add_drain_all(), was already scheduled for draining. Check (A).
835 */
836 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
837 goto done;
838
839 /*
840 * (D) Increment global generation number
841 *
842 * Pairs with smp_load_acquire() at (B), outside of the critical
843 * section. Use a full memory barrier to guarantee that the
844 * new global drain generation number is stored before loading
845 * folio_batch counters.
846 *
847 * This pairing must be done here, before the for_each_online_cpu loop
848 * below which drains the page vectors.
849 *
850 * Let x, y, and z represent some system CPU numbers, where x < y < z.
851 * Assume CPU #z is in the middle of the for_each_online_cpu loop
852 * below and has already reached CPU #y's per-cpu data. CPU #x comes
853 * along, adds some pages to its per-cpu vectors, then calls
854 * lru_add_drain_all().
855 *
856 * If the paired barrier is done at any later step, e.g. after the
857 * loop, CPU #x will just exit at (C) and miss flushing out all of its
858 * added pages.
859 */
860 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
861 smp_mb();
862
863 cpumask_clear(&has_work);
864 for_each_online_cpu(cpu) {
865 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
866
867 if (cpu_needs_drain(cpu)) {
868 INIT_WORK(work, lru_add_drain_per_cpu);
869 queue_work_on(cpu, mm_percpu_wq, work);
870 __cpumask_set_cpu(cpu, &has_work);
871 }
872 }
873
874 for_each_cpu(cpu, &has_work)
875 flush_work(&per_cpu(lru_add_drain_work, cpu));
876
877 done:
878 mutex_unlock(&lock);
879 }
880
lru_add_drain_all(void)881 void lru_add_drain_all(void)
882 {
883 __lru_add_drain_all(false);
884 }
885 #else
lru_add_drain_all(void)886 void lru_add_drain_all(void)
887 {
888 lru_add_drain();
889 }
890 #endif /* CONFIG_SMP */
891
892 atomic_t lru_disable_count = ATOMIC_INIT(0);
893
894 /*
895 * lru_cache_disable() needs to be called before we start compiling
896 * a list of folios to be migrated using folio_isolate_lru().
897 * It drains folios on LRU cache and then disable on all cpus until
898 * lru_cache_enable is called.
899 *
900 * Must be paired with a call to lru_cache_enable().
901 */
lru_cache_disable(void)902 void lru_cache_disable(void)
903 {
904 atomic_inc(&lru_disable_count);
905 /*
906 * Readers of lru_disable_count are protected by either disabling
907 * preemption or rcu_read_lock:
908 *
909 * preempt_disable, local_irq_disable [bh_lru_lock()]
910 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
911 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
912 *
913 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
914 * preempt_disable() regions of code. So any CPU which sees
915 * lru_disable_count = 0 will have exited the critical
916 * section when synchronize_rcu() returns.
917 */
918 synchronize_rcu_expedited();
919 #ifdef CONFIG_SMP
920 __lru_add_drain_all(true);
921 #else
922 lru_add_and_bh_lrus_drain();
923 #endif
924 }
925
926 /**
927 * folios_put_refs - Reduce the reference count on a batch of folios.
928 * @folios: The folios.
929 * @refs: The number of refs to subtract from each folio.
930 *
931 * Like folio_put(), but for a batch of folios. This is more efficient
932 * than writing the loop yourself as it will optimise the locks which need
933 * to be taken if the folios are freed. The folios batch is returned
934 * empty and ready to be reused for another batch; there is no need
935 * to reinitialise it. If @refs is NULL, we subtract one from each
936 * folio refcount.
937 *
938 * Context: May be called in process or interrupt context, but not in NMI
939 * context. May be called while holding a spinlock.
940 */
folios_put_refs(struct folio_batch * folios,unsigned int * refs)941 void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
942 {
943 int i, j;
944 struct lruvec *lruvec = NULL;
945 unsigned long flags = 0;
946
947 for (i = 0, j = 0; i < folios->nr; i++) {
948 struct folio *folio = folios->folios[i];
949 unsigned int nr_refs = refs ? refs[i] : 1;
950
951 if (is_huge_zero_folio(folio))
952 continue;
953
954 if (folio_is_zone_device(folio)) {
955 if (lruvec) {
956 unlock_page_lruvec_irqrestore(lruvec, flags);
957 lruvec = NULL;
958 }
959 if (put_devmap_managed_folio_refs(folio, nr_refs))
960 continue;
961 if (folio_ref_sub_and_test(folio, nr_refs))
962 free_zone_device_folio(folio);
963 continue;
964 }
965
966 if (!folio_ref_sub_and_test(folio, nr_refs))
967 continue;
968
969 /* hugetlb has its own memcg */
970 if (folio_test_hugetlb(folio)) {
971 if (lruvec) {
972 unlock_page_lruvec_irqrestore(lruvec, flags);
973 lruvec = NULL;
974 }
975 free_huge_folio(folio);
976 continue;
977 }
978 folio_unqueue_deferred_split(folio);
979 __page_cache_release(folio, &lruvec, &flags);
980
981 if (j != i)
982 folios->folios[j] = folio;
983 j++;
984 }
985 if (lruvec)
986 unlock_page_lruvec_irqrestore(lruvec, flags);
987 if (!j) {
988 folio_batch_reinit(folios);
989 return;
990 }
991
992 folios->nr = j;
993 mem_cgroup_uncharge_folios(folios);
994 free_unref_folios(folios);
995 }
996 EXPORT_SYMBOL(folios_put_refs);
997
998 /**
999 * release_pages - batched put_page()
1000 * @arg: array of pages to release
1001 * @nr: number of pages
1002 *
1003 * Decrement the reference count on all the pages in @arg. If it
1004 * fell to zero, remove the page from the LRU and free it.
1005 *
1006 * Note that the argument can be an array of pages, encoded pages,
1007 * or folio pointers. We ignore any encoded bits, and turn any of
1008 * them into just a folio that gets free'd.
1009 */
release_pages(release_pages_arg arg,int nr)1010 void release_pages(release_pages_arg arg, int nr)
1011 {
1012 struct folio_batch fbatch;
1013 int refs[PAGEVEC_SIZE];
1014 struct encoded_page **encoded = arg.encoded_pages;
1015 int i;
1016
1017 folio_batch_init(&fbatch);
1018 for (i = 0; i < nr; i++) {
1019 /* Turn any of the argument types into a folio */
1020 struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
1021
1022 /* Is our next entry actually "nr_pages" -> "nr_refs" ? */
1023 refs[fbatch.nr] = 1;
1024 if (unlikely(encoded_page_flags(encoded[i]) &
1025 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
1026 refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
1027
1028 if (folio_batch_add(&fbatch, folio) > 0)
1029 continue;
1030 folios_put_refs(&fbatch, refs);
1031 }
1032
1033 if (fbatch.nr)
1034 folios_put_refs(&fbatch, refs);
1035 }
1036 EXPORT_SYMBOL(release_pages);
1037
1038 /*
1039 * The folios which we're about to release may be in the deferred lru-addition
1040 * queues. That would prevent them from really being freed right now. That's
1041 * OK from a correctness point of view but is inefficient - those folios may be
1042 * cache-warm and we want to give them back to the page allocator ASAP.
1043 *
1044 * So __folio_batch_release() will drain those queues here.
1045 * folio_batch_move_lru() calls folios_put() directly to avoid
1046 * mutual recursion.
1047 */
__folio_batch_release(struct folio_batch * fbatch)1048 void __folio_batch_release(struct folio_batch *fbatch)
1049 {
1050 if (!fbatch->percpu_pvec_drained) {
1051 lru_add_drain();
1052 fbatch->percpu_pvec_drained = true;
1053 }
1054 folios_put(fbatch);
1055 }
1056 EXPORT_SYMBOL(__folio_batch_release);
1057
1058 /**
1059 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1060 * @fbatch: The batch to prune
1061 *
1062 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1063 * entries. This function prunes all the non-folio entries from @fbatch
1064 * without leaving holes, so that it can be passed on to folio-only batch
1065 * operations.
1066 */
folio_batch_remove_exceptionals(struct folio_batch * fbatch)1067 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1068 {
1069 unsigned int i, j;
1070
1071 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1072 struct folio *folio = fbatch->folios[i];
1073 if (!xa_is_value(folio))
1074 fbatch->folios[j++] = folio;
1075 }
1076 fbatch->nr = j;
1077 }
1078
1079 static const struct ctl_table swap_sysctl_table[] = {
1080 {
1081 .procname = "page-cluster",
1082 .data = &page_cluster,
1083 .maxlen = sizeof(int),
1084 .mode = 0644,
1085 .proc_handler = proc_dointvec_minmax,
1086 .extra1 = SYSCTL_ZERO,
1087 .extra2 = (void *)&page_cluster_max,
1088 }
1089 };
1090
1091 /*
1092 * Perform any setup for the swap system
1093 */
swap_setup(void)1094 void __init swap_setup(void)
1095 {
1096 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1097
1098 /* Use a smaller cluster for small-memory machines */
1099 if (megs < 16)
1100 page_cluster = 2;
1101 else
1102 page_cluster = 3;
1103 /*
1104 * Right now other parts of the system means that we
1105 * _really_ don't want to cluster much more
1106 */
1107
1108 register_sysctl_init("vm", swap_sysctl_table);
1109 }
1110