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