1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
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.
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
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>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/pagemap.h>
46 /* How many pages do we try to swap or page in/out together? */
49 /* Protecting only lru_rotate.fbatch which requires disabling interrupts */
52 struct folio_batch fbatch
;
54 static DEFINE_PER_CPU(struct lru_rotate
, lru_rotate
) = {
55 .lock
= INIT_LOCAL_LOCK(lock
),
59 * The following folio batches are grouped together because they are protected
60 * by disabling preemption (and interrupts remain enabled).
64 struct folio_batch lru_add
;
65 struct folio_batch lru_deactivate_file
;
66 struct folio_batch lru_deactivate
;
67 struct folio_batch lru_lazyfree
;
69 struct folio_batch activate
;
72 static DEFINE_PER_CPU(struct cpu_fbatches
, cpu_fbatches
) = {
73 .lock
= INIT_LOCAL_LOCK(lock
),
77 * This path almost never happens for VM activity - pages are normally freed
78 * via pagevecs. But it gets used by networking - and for compound pages.
80 static void __page_cache_release(struct folio
*folio
)
82 if (folio_test_lru(folio
)) {
83 struct lruvec
*lruvec
;
86 lruvec
= folio_lruvec_lock_irqsave(folio
, &flags
);
87 lruvec_del_folio(lruvec
, folio
);
88 __folio_clear_lru_flags(folio
);
89 unlock_page_lruvec_irqrestore(lruvec
, flags
);
91 /* See comment on folio_test_mlocked in release_pages() */
92 if (unlikely(folio_test_mlocked(folio
))) {
93 long nr_pages
= folio_nr_pages(folio
);
95 __folio_clear_mlocked(folio
);
96 zone_stat_mod_folio(folio
, NR_MLOCK
, -nr_pages
);
97 count_vm_events(UNEVICTABLE_PGCLEARED
, nr_pages
);
101 static void __folio_put_small(struct folio
*folio
)
103 __page_cache_release(folio
);
104 mem_cgroup_uncharge(folio
);
105 free_unref_page(&folio
->page
, 0);
108 static void __folio_put_large(struct folio
*folio
)
111 * __page_cache_release() is supposed to be called for thp, not for
112 * hugetlb. This is because hugetlb page does never have PageLRU set
113 * (it's never listed to any LRU lists) and no memcg routines should
114 * be called for hugetlb (it has a separate hugetlb_cgroup.)
116 if (!folio_test_hugetlb(folio
))
117 __page_cache_release(folio
);
118 destroy_large_folio(folio
);
121 void __folio_put(struct folio
*folio
)
123 if (unlikely(folio_is_zone_device(folio
)))
124 free_zone_device_page(&folio
->page
);
125 else if (unlikely(folio_test_large(folio
)))
126 __folio_put_large(folio
);
128 __folio_put_small(folio
);
130 EXPORT_SYMBOL(__folio_put
);
133 * put_pages_list() - release a list of pages
134 * @pages: list of pages threaded on page->lru
136 * Release a list of pages which are strung together on page.lru.
138 void put_pages_list(struct list_head
*pages
)
140 struct folio
*folio
, *next
;
142 list_for_each_entry_safe(folio
, next
, pages
, lru
) {
143 if (!folio_put_testzero(folio
)) {
144 list_del(&folio
->lru
);
147 if (folio_test_large(folio
)) {
148 list_del(&folio
->lru
);
149 __folio_put_large(folio
);
152 /* LRU flag must be clear because it's passed using the lru */
155 free_unref_page_list(pages
);
156 INIT_LIST_HEAD(pages
);
158 EXPORT_SYMBOL(put_pages_list
);
161 * get_kernel_pages() - pin kernel pages in memory
162 * @kiov: An array of struct kvec structures
163 * @nr_segs: number of segments to pin
164 * @write: pinning for read/write, currently ignored
165 * @pages: array that receives pointers to the pages pinned.
166 * Should be at least nr_segs long.
168 * Returns number of pages pinned. This may be fewer than the number requested.
169 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0.
170 * Each page returned must be released with a put_page() call when it is
173 int get_kernel_pages(const struct kvec
*kiov
, int nr_segs
, int write
,
178 for (seg
= 0; seg
< nr_segs
; seg
++) {
179 if (WARN_ON(kiov
[seg
].iov_len
!= PAGE_SIZE
))
182 pages
[seg
] = kmap_to_page(kiov
[seg
].iov_base
);
183 get_page(pages
[seg
]);
188 EXPORT_SYMBOL_GPL(get_kernel_pages
);
190 typedef void (*move_fn_t
)(struct lruvec
*lruvec
, struct folio
*folio
);
192 static void lru_add_fn(struct lruvec
*lruvec
, struct folio
*folio
)
194 int was_unevictable
= folio_test_clear_unevictable(folio
);
195 long nr_pages
= folio_nr_pages(folio
);
197 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
200 * Is an smp_mb__after_atomic() still required here, before
201 * folio_evictable() tests the mlocked flag, to rule out the possibility
202 * of stranding an evictable folio on an unevictable LRU? I think
203 * not, because __munlock_page() only clears the mlocked flag
204 * while the LRU lock is held.
206 * (That is not true of __page_cache_release(), and not necessarily
207 * true of release_pages(): but those only clear the mlocked flag after
208 * folio_put_testzero() has excluded any other users of the folio.)
210 if (folio_evictable(folio
)) {
212 __count_vm_events(UNEVICTABLE_PGRESCUED
, nr_pages
);
214 folio_clear_active(folio
);
215 folio_set_unevictable(folio
);
217 * folio->mlock_count = !!folio_test_mlocked(folio)?
218 * But that leaves __mlock_page() in doubt whether another
219 * actor has already counted the mlock or not. Err on the
220 * safe side, underestimate, let page reclaim fix it, rather
221 * than leaving a page on the unevictable LRU indefinitely.
223 folio
->mlock_count
= 0;
224 if (!was_unevictable
)
225 __count_vm_events(UNEVICTABLE_PGCULLED
, nr_pages
);
228 lruvec_add_folio(lruvec
, folio
);
229 trace_mm_lru_insertion(folio
);
232 static void folio_batch_move_lru(struct folio_batch
*fbatch
, move_fn_t move_fn
)
235 struct lruvec
*lruvec
= NULL
;
236 unsigned long flags
= 0;
238 for (i
= 0; i
< folio_batch_count(fbatch
); i
++) {
239 struct folio
*folio
= fbatch
->folios
[i
];
241 /* block memcg migration while the folio moves between lru */
242 if (move_fn
!= lru_add_fn
&& !folio_test_clear_lru(folio
))
245 lruvec
= folio_lruvec_relock_irqsave(folio
, lruvec
, &flags
);
246 move_fn(lruvec
, folio
);
248 folio_set_lru(folio
);
252 unlock_page_lruvec_irqrestore(lruvec
, flags
);
253 folios_put(fbatch
->folios
, folio_batch_count(fbatch
));
254 folio_batch_init(fbatch
);
257 static void folio_batch_add_and_move(struct folio_batch
*fbatch
,
258 struct folio
*folio
, move_fn_t move_fn
)
260 if (folio_batch_add(fbatch
, folio
) && !folio_test_large(folio
) &&
261 !lru_cache_disabled())
263 folio_batch_move_lru(fbatch
, move_fn
);
266 static void lru_move_tail_fn(struct lruvec
*lruvec
, struct folio
*folio
)
268 if (!folio_test_unevictable(folio
)) {
269 lruvec_del_folio(lruvec
, folio
);
270 folio_clear_active(folio
);
271 lruvec_add_folio_tail(lruvec
, folio
);
272 __count_vm_events(PGROTATED
, folio_nr_pages(folio
));
277 * Writeback is about to end against a folio which has been marked for
278 * immediate reclaim. If it still appears to be reclaimable, move it
279 * to the tail of the inactive list.
281 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
283 void folio_rotate_reclaimable(struct folio
*folio
)
285 if (!folio_test_locked(folio
) && !folio_test_dirty(folio
) &&
286 !folio_test_unevictable(folio
) && folio_test_lru(folio
)) {
287 struct folio_batch
*fbatch
;
291 local_lock_irqsave(&lru_rotate
.lock
, flags
);
292 fbatch
= this_cpu_ptr(&lru_rotate
.fbatch
);
293 folio_batch_add_and_move(fbatch
, folio
, lru_move_tail_fn
);
294 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
298 void lru_note_cost(struct lruvec
*lruvec
, bool file
, unsigned int nr_pages
)
301 unsigned long lrusize
;
304 * Hold lruvec->lru_lock is safe here, since
305 * 1) The pinned lruvec in reclaim, or
306 * 2) From a pre-LRU page during refault (which also holds the
307 * rcu lock, so would be safe even if the page was on the LRU
308 * and could move simultaneously to a new lruvec).
310 spin_lock_irq(&lruvec
->lru_lock
);
311 /* Record cost event */
313 lruvec
->file_cost
+= nr_pages
;
315 lruvec
->anon_cost
+= nr_pages
;
318 * Decay previous events
320 * Because workloads change over time (and to avoid
321 * overflow) we keep these statistics as a floating
322 * average, which ends up weighing recent refaults
323 * more than old ones.
325 lrusize
= lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) +
326 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
) +
327 lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) +
328 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
);
330 if (lruvec
->file_cost
+ lruvec
->anon_cost
> lrusize
/ 4) {
331 lruvec
->file_cost
/= 2;
332 lruvec
->anon_cost
/= 2;
334 spin_unlock_irq(&lruvec
->lru_lock
);
335 } while ((lruvec
= parent_lruvec(lruvec
)));
338 void lru_note_cost_folio(struct folio
*folio
)
340 lru_note_cost(folio_lruvec(folio
), folio_is_file_lru(folio
),
341 folio_nr_pages(folio
));
344 static void folio_activate_fn(struct lruvec
*lruvec
, struct folio
*folio
)
346 if (!folio_test_active(folio
) && !folio_test_unevictable(folio
)) {
347 long nr_pages
= folio_nr_pages(folio
);
349 lruvec_del_folio(lruvec
, folio
);
350 folio_set_active(folio
);
351 lruvec_add_folio(lruvec
, folio
);
352 trace_mm_lru_activate(folio
);
354 __count_vm_events(PGACTIVATE
, nr_pages
);
355 __count_memcg_events(lruvec_memcg(lruvec
), PGACTIVATE
,
361 static void folio_activate_drain(int cpu
)
363 struct folio_batch
*fbatch
= &per_cpu(cpu_fbatches
.activate
, cpu
);
365 if (folio_batch_count(fbatch
))
366 folio_batch_move_lru(fbatch
, folio_activate_fn
);
369 static void folio_activate(struct folio
*folio
)
371 if (folio_test_lru(folio
) && !folio_test_active(folio
) &&
372 !folio_test_unevictable(folio
)) {
373 struct folio_batch
*fbatch
;
376 local_lock(&cpu_fbatches
.lock
);
377 fbatch
= this_cpu_ptr(&cpu_fbatches
.activate
);
378 folio_batch_add_and_move(fbatch
, folio
, folio_activate_fn
);
379 local_unlock(&cpu_fbatches
.lock
);
384 static inline void folio_activate_drain(int cpu
)
388 static void folio_activate(struct folio
*folio
)
390 struct lruvec
*lruvec
;
392 if (folio_test_clear_lru(folio
)) {
393 lruvec
= folio_lruvec_lock_irq(folio
);
394 folio_activate_fn(lruvec
, folio
);
395 unlock_page_lruvec_irq(lruvec
);
396 folio_set_lru(folio
);
401 static void __lru_cache_activate_folio(struct folio
*folio
)
403 struct folio_batch
*fbatch
;
406 local_lock(&cpu_fbatches
.lock
);
407 fbatch
= this_cpu_ptr(&cpu_fbatches
.lru_add
);
410 * Search backwards on the optimistic assumption that the folio being
411 * activated has just been added to this batch. Note that only
412 * the local batch is examined as a !LRU folio could be in the
413 * process of being released, reclaimed, migrated or on a remote
414 * batch that is currently being drained. Furthermore, marking
415 * a remote batch's folio active potentially hits a race where
416 * a folio is marked active just after it is added to the inactive
417 * list causing accounting errors and BUG_ON checks to trigger.
419 for (i
= folio_batch_count(fbatch
) - 1; i
>= 0; i
--) {
420 struct folio
*batch_folio
= fbatch
->folios
[i
];
422 if (batch_folio
== folio
) {
423 folio_set_active(folio
);
428 local_unlock(&cpu_fbatches
.lock
);
432 * Mark a page as having seen activity.
434 * inactive,unreferenced -> inactive,referenced
435 * inactive,referenced -> active,unreferenced
436 * active,unreferenced -> active,referenced
438 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
439 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
441 void folio_mark_accessed(struct folio
*folio
)
443 if (!folio_test_referenced(folio
)) {
444 folio_set_referenced(folio
);
445 } else if (folio_test_unevictable(folio
)) {
447 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
448 * this list is never rotated or maintained, so marking an
449 * unevictable page accessed has no effect.
451 } else if (!folio_test_active(folio
)) {
453 * If the folio is on the LRU, queue it for activation via
454 * cpu_fbatches.activate. Otherwise, assume the folio is in a
455 * folio_batch, mark it active and it'll be moved to the active
456 * LRU on the next drain.
458 if (folio_test_lru(folio
))
459 folio_activate(folio
);
461 __lru_cache_activate_folio(folio
);
462 folio_clear_referenced(folio
);
463 workingset_activation(folio
);
465 if (folio_test_idle(folio
))
466 folio_clear_idle(folio
);
468 EXPORT_SYMBOL(folio_mark_accessed
);
471 * folio_add_lru - Add a folio to an LRU list.
472 * @folio: The folio to be added to the LRU.
474 * Queue the folio for addition to the LRU. The decision on whether
475 * to add the page to the [in]active [file|anon] list is deferred until the
476 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
477 * have the folio added to the active list using folio_mark_accessed().
479 void folio_add_lru(struct folio
*folio
)
481 struct folio_batch
*fbatch
;
483 VM_BUG_ON_FOLIO(folio_test_active(folio
) &&
484 folio_test_unevictable(folio
), folio
);
485 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
488 local_lock(&cpu_fbatches
.lock
);
489 fbatch
= this_cpu_ptr(&cpu_fbatches
.lru_add
);
490 folio_batch_add_and_move(fbatch
, folio
, lru_add_fn
);
491 local_unlock(&cpu_fbatches
.lock
);
493 EXPORT_SYMBOL(folio_add_lru
);
496 * lru_cache_add_inactive_or_unevictable
497 * @page: the page to be added to LRU
498 * @vma: vma in which page is mapped for determining reclaimability
500 * Place @page on the inactive or unevictable LRU list, depending on its
503 void lru_cache_add_inactive_or_unevictable(struct page
*page
,
504 struct vm_area_struct
*vma
)
506 VM_BUG_ON_PAGE(PageLRU(page
), page
);
508 if (unlikely((vma
->vm_flags
& (VM_LOCKED
| VM_SPECIAL
)) == VM_LOCKED
))
509 mlock_new_page(page
);
515 * If the folio cannot be invalidated, it is moved to the
516 * inactive list to speed up its reclaim. It is moved to the
517 * head of the list, rather than the tail, to give the flusher
518 * threads some time to write it out, as this is much more
519 * effective than the single-page writeout from reclaim.
521 * If the folio isn't mapped and dirty/writeback, the folio
522 * could be reclaimed asap using the reclaim flag.
524 * 1. active, mapped folio -> none
525 * 2. active, dirty/writeback folio -> inactive, head, reclaim
526 * 3. inactive, mapped folio -> none
527 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
528 * 5. inactive, clean -> inactive, tail
531 * In 4, it moves to the head of the inactive list so the folio is
532 * written out by flusher threads as this is much more efficient
533 * than the single-page writeout from reclaim.
535 static void lru_deactivate_file_fn(struct lruvec
*lruvec
, struct folio
*folio
)
537 bool active
= folio_test_active(folio
);
538 long nr_pages
= folio_nr_pages(folio
);
540 if (folio_test_unevictable(folio
))
543 /* Some processes are using the folio */
544 if (folio_mapped(folio
))
547 lruvec_del_folio(lruvec
, folio
);
548 folio_clear_active(folio
);
549 folio_clear_referenced(folio
);
551 if (folio_test_writeback(folio
) || folio_test_dirty(folio
)) {
553 * Setting the reclaim flag could race with
554 * folio_end_writeback() and confuse readahead. But the
555 * race window is _really_ small and it's not a critical
558 lruvec_add_folio(lruvec
, folio
);
559 folio_set_reclaim(folio
);
562 * The folio's writeback ended while it was in the batch.
563 * We move that folio to the tail of the inactive list.
565 lruvec_add_folio_tail(lruvec
, folio
);
566 __count_vm_events(PGROTATED
, nr_pages
);
570 __count_vm_events(PGDEACTIVATE
, nr_pages
);
571 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
576 static void lru_deactivate_fn(struct lruvec
*lruvec
, struct folio
*folio
)
578 if (folio_test_active(folio
) && !folio_test_unevictable(folio
)) {
579 long nr_pages
= folio_nr_pages(folio
);
581 lruvec_del_folio(lruvec
, folio
);
582 folio_clear_active(folio
);
583 folio_clear_referenced(folio
);
584 lruvec_add_folio(lruvec
, folio
);
586 __count_vm_events(PGDEACTIVATE
, nr_pages
);
587 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
592 static void lru_lazyfree_fn(struct lruvec
*lruvec
, struct folio
*folio
)
594 if (folio_test_anon(folio
) && folio_test_swapbacked(folio
) &&
595 !folio_test_swapcache(folio
) && !folio_test_unevictable(folio
)) {
596 long nr_pages
= folio_nr_pages(folio
);
598 lruvec_del_folio(lruvec
, folio
);
599 folio_clear_active(folio
);
600 folio_clear_referenced(folio
);
602 * Lazyfree folios are clean anonymous folios. They have
603 * the swapbacked flag cleared, to distinguish them from normal
606 folio_clear_swapbacked(folio
);
607 lruvec_add_folio(lruvec
, folio
);
609 __count_vm_events(PGLAZYFREE
, nr_pages
);
610 __count_memcg_events(lruvec_memcg(lruvec
), PGLAZYFREE
,
616 * Drain pages out of the cpu's folio_batch.
617 * Either "cpu" is the current CPU, and preemption has already been
618 * disabled; or "cpu" is being hot-unplugged, and is already dead.
620 void lru_add_drain_cpu(int cpu
)
622 struct cpu_fbatches
*fbatches
= &per_cpu(cpu_fbatches
, cpu
);
623 struct folio_batch
*fbatch
= &fbatches
->lru_add
;
625 if (folio_batch_count(fbatch
))
626 folio_batch_move_lru(fbatch
, lru_add_fn
);
628 fbatch
= &per_cpu(lru_rotate
.fbatch
, cpu
);
629 /* Disabling interrupts below acts as a compiler barrier. */
630 if (data_race(folio_batch_count(fbatch
))) {
633 /* No harm done if a racing interrupt already did this */
634 local_lock_irqsave(&lru_rotate
.lock
, flags
);
635 folio_batch_move_lru(fbatch
, lru_move_tail_fn
);
636 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
639 fbatch
= &fbatches
->lru_deactivate_file
;
640 if (folio_batch_count(fbatch
))
641 folio_batch_move_lru(fbatch
, lru_deactivate_file_fn
);
643 fbatch
= &fbatches
->lru_deactivate
;
644 if (folio_batch_count(fbatch
))
645 folio_batch_move_lru(fbatch
, lru_deactivate_fn
);
647 fbatch
= &fbatches
->lru_lazyfree
;
648 if (folio_batch_count(fbatch
))
649 folio_batch_move_lru(fbatch
, lru_lazyfree_fn
);
651 folio_activate_drain(cpu
);
655 * deactivate_file_folio() - Deactivate a file folio.
656 * @folio: Folio to deactivate.
658 * This function hints to the VM that @folio is a good reclaim candidate,
659 * for example if its invalidation fails due to the folio being dirty
660 * or under writeback.
662 * Context: Caller holds a reference on the folio.
664 void deactivate_file_folio(struct folio
*folio
)
666 struct folio_batch
*fbatch
;
668 /* Deactivating an unevictable folio will not accelerate reclaim */
669 if (folio_test_unevictable(folio
))
673 local_lock(&cpu_fbatches
.lock
);
674 fbatch
= this_cpu_ptr(&cpu_fbatches
.lru_deactivate_file
);
675 folio_batch_add_and_move(fbatch
, folio
, lru_deactivate_file_fn
);
676 local_unlock(&cpu_fbatches
.lock
);
680 * deactivate_page - deactivate a page
681 * @page: page to deactivate
683 * deactivate_page() moves @page to the inactive list if @page was on the active
684 * list and was not an unevictable page. This is done to accelerate the reclaim
687 void deactivate_page(struct page
*page
)
689 struct folio
*folio
= page_folio(page
);
691 if (folio_test_lru(folio
) && folio_test_active(folio
) &&
692 !folio_test_unevictable(folio
)) {
693 struct folio_batch
*fbatch
;
696 local_lock(&cpu_fbatches
.lock
);
697 fbatch
= this_cpu_ptr(&cpu_fbatches
.lru_deactivate
);
698 folio_batch_add_and_move(fbatch
, folio
, lru_deactivate_fn
);
699 local_unlock(&cpu_fbatches
.lock
);
704 * mark_page_lazyfree - make an anon page lazyfree
705 * @page: page to deactivate
707 * mark_page_lazyfree() moves @page to the inactive file list.
708 * This is done to accelerate the reclaim of @page.
710 void mark_page_lazyfree(struct page
*page
)
712 struct folio
*folio
= page_folio(page
);
714 if (folio_test_lru(folio
) && folio_test_anon(folio
) &&
715 folio_test_swapbacked(folio
) && !folio_test_swapcache(folio
) &&
716 !folio_test_unevictable(folio
)) {
717 struct folio_batch
*fbatch
;
720 local_lock(&cpu_fbatches
.lock
);
721 fbatch
= this_cpu_ptr(&cpu_fbatches
.lru_lazyfree
);
722 folio_batch_add_and_move(fbatch
, folio
, lru_lazyfree_fn
);
723 local_unlock(&cpu_fbatches
.lock
);
727 void lru_add_drain(void)
729 local_lock(&cpu_fbatches
.lock
);
730 lru_add_drain_cpu(smp_processor_id());
731 local_unlock(&cpu_fbatches
.lock
);
732 mlock_page_drain_local();
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.
741 static void lru_add_and_bh_lrus_drain(void)
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_page_drain_local();
750 void lru_add_drain_cpu_zone(struct zone
*zone
)
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_page_drain_local();
761 static DEFINE_PER_CPU(struct work_struct
, lru_add_drain_work
);
763 static void lru_add_drain_per_cpu(struct work_struct
*dummy
)
765 lru_add_and_bh_lrus_drain();
768 static bool cpu_needs_drain(unsigned int cpu
)
770 struct cpu_fbatches
*fbatches
= &per_cpu(cpu_fbatches
, cpu
);
772 /* Check these in order of likelihood that they're not zero */
773 return folio_batch_count(&fbatches
->lru_add
) ||
774 data_race(folio_batch_count(&per_cpu(lru_rotate
.fbatch
, cpu
))) ||
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
->activate
) ||
779 need_mlock_page_drain(cpu
) ||
780 has_bh_in_lru(cpu
, NULL
);
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.
790 static inline void __lru_add_drain_all(bool force_all_cpus
)
793 * lru_drain_gen - Global pages generation number
795 * (A) Definition: global lru_drain_gen = x implies that all generations
796 * 0 < n <= x are already *scheduled* for draining.
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
802 static unsigned int lru_drain_gen
;
803 static struct cpumask has_work
;
804 static DEFINE_MUTEX(lock
);
805 unsigned cpu
, this_gen
;
808 * Make sure nobody triggers this path before mm_percpu_wq is fully
811 if (WARN_ON(!mm_percpu_wq
))
815 * Guarantee folio_batch counter stores visible by this CPU
816 * are visible to other CPUs before loading the current drain
822 * (B) Locally cache global LRU draining generation number
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
828 this_gen
= smp_load_acquire(&lru_drain_gen
);
833 * (C) Exit the draining operation if a newer generation, from another
834 * lru_add_drain_all(), was already scheduled for draining. Check (A).
836 if (unlikely(this_gen
!= lru_drain_gen
&& !force_all_cpus
))
840 * (D) Increment global generation number
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.
847 * This pairing must be done here, before the for_each_online_cpu loop
848 * below which drains the page vectors.
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().
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
860 WRITE_ONCE(lru_drain_gen
, lru_drain_gen
+ 1);
863 cpumask_clear(&has_work
);
864 for_each_online_cpu(cpu
) {
865 struct work_struct
*work
= &per_cpu(lru_add_drain_work
, cpu
);
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
);
874 for_each_cpu(cpu
, &has_work
)
875 flush_work(&per_cpu(lru_add_drain_work
, cpu
));
881 void lru_add_drain_all(void)
883 __lru_add_drain_all(false);
886 void lru_add_drain_all(void)
890 #endif /* CONFIG_SMP */
892 atomic_t lru_disable_count
= ATOMIC_INIT(0);
895 * lru_cache_disable() needs to be called before we start compiling
896 * a list of pages to be migrated using isolate_lru_page().
897 * It drains pages on LRU cache and then disable on all cpus until
898 * lru_cache_enable is called.
900 * Must be paired with a call to lru_cache_enable().
902 void lru_cache_disable(void)
904 atomic_inc(&lru_disable_count
);
906 * Readers of lru_disable_count are protected by either disabling
907 * preemption or rcu_read_lock:
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]
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.
918 synchronize_rcu_expedited();
920 __lru_add_drain_all(true);
922 lru_add_and_bh_lrus_drain();
927 * release_pages - batched put_page()
928 * @pages: array of pages to release
929 * @nr: number of pages
931 * Decrement the reference count on all the pages in @pages. If it
932 * fell to zero, remove the page from the LRU and free it.
934 void release_pages(struct page
**pages
, int nr
)
937 LIST_HEAD(pages_to_free
);
938 struct lruvec
*lruvec
= NULL
;
939 unsigned long flags
= 0;
940 unsigned int lock_batch
;
942 for (i
= 0; i
< nr
; i
++) {
943 struct folio
*folio
= page_folio(pages
[i
]);
946 * Make sure the IRQ-safe lock-holding time does not get
947 * excessive with a continuous string of pages from the
948 * same lruvec. The lock is held only if lruvec != NULL.
950 if (lruvec
&& ++lock_batch
== SWAP_CLUSTER_MAX
) {
951 unlock_page_lruvec_irqrestore(lruvec
, flags
);
955 if (is_huge_zero_page(&folio
->page
))
958 if (folio_is_zone_device(folio
)) {
960 unlock_page_lruvec_irqrestore(lruvec
, flags
);
963 if (put_devmap_managed_page(&folio
->page
))
965 if (folio_put_testzero(folio
))
966 free_zone_device_page(&folio
->page
);
970 if (!folio_put_testzero(folio
))
973 if (folio_test_large(folio
)) {
975 unlock_page_lruvec_irqrestore(lruvec
, flags
);
978 __folio_put_large(folio
);
982 if (folio_test_lru(folio
)) {
983 struct lruvec
*prev_lruvec
= lruvec
;
985 lruvec
= folio_lruvec_relock_irqsave(folio
, lruvec
,
987 if (prev_lruvec
!= lruvec
)
990 lruvec_del_folio(lruvec
, folio
);
991 __folio_clear_lru_flags(folio
);
995 * In rare cases, when truncation or holepunching raced with
996 * munlock after VM_LOCKED was cleared, Mlocked may still be
997 * found set here. This does not indicate a problem, unless
998 * "unevictable_pgs_cleared" appears worryingly large.
1000 if (unlikely(folio_test_mlocked(folio
))) {
1001 __folio_clear_mlocked(folio
);
1002 zone_stat_sub_folio(folio
, NR_MLOCK
);
1003 count_vm_event(UNEVICTABLE_PGCLEARED
);
1006 list_add(&folio
->lru
, &pages_to_free
);
1009 unlock_page_lruvec_irqrestore(lruvec
, flags
);
1011 mem_cgroup_uncharge_list(&pages_to_free
);
1012 free_unref_page_list(&pages_to_free
);
1014 EXPORT_SYMBOL(release_pages
);
1017 * The pages which we're about to release may be in the deferred lru-addition
1018 * queues. That would prevent them from really being freed right now. That's
1019 * OK from a correctness point of view but is inefficient - those pages may be
1020 * cache-warm and we want to give them back to the page allocator ASAP.
1022 * So __pagevec_release() will drain those queues here.
1023 * folio_batch_move_lru() calls folios_put() directly to avoid
1026 void __pagevec_release(struct pagevec
*pvec
)
1028 if (!pvec
->percpu_pvec_drained
) {
1030 pvec
->percpu_pvec_drained
= true;
1032 release_pages(pvec
->pages
, pagevec_count(pvec
));
1033 pagevec_reinit(pvec
);
1035 EXPORT_SYMBOL(__pagevec_release
);
1038 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1039 * @fbatch: The batch to prune
1041 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1042 * entries. This function prunes all the non-folio entries from @fbatch
1043 * without leaving holes, so that it can be passed on to folio-only batch
1046 void folio_batch_remove_exceptionals(struct folio_batch
*fbatch
)
1050 for (i
= 0, j
= 0; i
< folio_batch_count(fbatch
); i
++) {
1051 struct folio
*folio
= fbatch
->folios
[i
];
1052 if (!xa_is_value(folio
))
1053 fbatch
->folios
[j
++] = folio
;
1058 unsigned pagevec_lookup_range_tag(struct pagevec
*pvec
,
1059 struct address_space
*mapping
, pgoff_t
*index
, pgoff_t end
,
1062 pvec
->nr
= find_get_pages_range_tag(mapping
, index
, end
, tag
,
1063 PAGEVEC_SIZE
, pvec
->pages
);
1064 return pagevec_count(pvec
);
1066 EXPORT_SYMBOL(pagevec_lookup_range_tag
);
1069 * Perform any setup for the swap system
1071 void __init
swap_setup(void)
1073 unsigned long megs
= totalram_pages() >> (20 - PAGE_SHIFT
);
1075 /* Use a smaller cluster for small-memory machines */
1081 * Right now other parts of the system means that we
1082 * _really_ don't want to cluster much more