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? */
47 int page_cluster;
48 
49 /* Protecting only lru_rotate.pvec which requires disabling interrupts */
50 struct lru_rotate {
51 	local_lock_t lock;
52 	struct pagevec pvec;
53 };
54 static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
55 	.lock = INIT_LOCAL_LOCK(lock),
56 };
57 
58 /*
59  * The following struct pagevec are grouped together because they are protected
60  * by disabling preemption (and interrupts remain enabled).
61  */
62 struct lru_pvecs {
63 	local_lock_t lock;
64 	struct pagevec lru_add;
65 	struct pagevec lru_deactivate_file;
66 	struct pagevec lru_deactivate;
67 	struct pagevec lru_lazyfree;
68 #ifdef CONFIG_SMP
69 	struct pagevec activate_page;
70 #endif
71 };
72 static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
73 	.lock = INIT_LOCAL_LOCK(lock),
74 };
75 
76 /*
77  * This path almost never happens for VM activity - pages are normally freed
78  * via pagevecs.  But it gets used by networking - and for compound pages.
79  */
__page_cache_release(struct page * page)80 static void __page_cache_release(struct page *page)
81 {
82 	if (PageLRU(page)) {
83 		struct folio *folio = page_folio(page);
84 		struct lruvec *lruvec;
85 		unsigned long flags;
86 
87 		lruvec = folio_lruvec_lock_irqsave(folio, &flags);
88 		del_page_from_lru_list(page, lruvec);
89 		__clear_page_lru_flags(page);
90 		unlock_page_lruvec_irqrestore(lruvec, flags);
91 	}
92 	/* See comment on PageMlocked in release_pages() */
93 	if (unlikely(PageMlocked(page))) {
94 		int nr_pages = thp_nr_pages(page);
95 
96 		__ClearPageMlocked(page);
97 		mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
98 		count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
99 	}
100 }
101 
__put_single_page(struct page * page)102 static void __put_single_page(struct page *page)
103 {
104 	__page_cache_release(page);
105 	mem_cgroup_uncharge(page_folio(page));
106 	free_unref_page(page, 0);
107 }
108 
__put_compound_page(struct page * page)109 static void __put_compound_page(struct page *page)
110 {
111 	/*
112 	 * __page_cache_release() is supposed to be called for thp, not for
113 	 * hugetlb. This is because hugetlb page does never have PageLRU set
114 	 * (it's never listed to any LRU lists) and no memcg routines should
115 	 * be called for hugetlb (it has a separate hugetlb_cgroup.)
116 	 */
117 	if (!PageHuge(page))
118 		__page_cache_release(page);
119 	destroy_compound_page(page);
120 }
121 
__put_page(struct page * page)122 void __put_page(struct page *page)
123 {
124 	if (unlikely(is_zone_device_page(page)))
125 		free_zone_device_page(page);
126 	else if (unlikely(PageCompound(page)))
127 		__put_compound_page(page);
128 	else
129 		__put_single_page(page);
130 }
131 EXPORT_SYMBOL(__put_page);
132 
133 /**
134  * put_pages_list() - release a list of pages
135  * @pages: list of pages threaded on page->lru
136  *
137  * Release a list of pages which are strung together on page.lru.
138  */
put_pages_list(struct list_head * pages)139 void put_pages_list(struct list_head *pages)
140 {
141 	struct page *page, *next;
142 
143 	list_for_each_entry_safe(page, next, pages, lru) {
144 		if (!put_page_testzero(page)) {
145 			list_del(&page->lru);
146 			continue;
147 		}
148 		if (PageHead(page)) {
149 			list_del(&page->lru);
150 			__put_compound_page(page);
151 			continue;
152 		}
153 		/* Cannot be PageLRU because it's passed to us using the lru */
154 	}
155 
156 	free_unref_page_list(pages);
157 	INIT_LIST_HEAD(pages);
158 }
159 EXPORT_SYMBOL(put_pages_list);
160 
161 /*
162  * get_kernel_pages() - pin kernel pages in memory
163  * @kiov:	An array of struct kvec structures
164  * @nr_segs:	number of segments to pin
165  * @write:	pinning for read/write, currently ignored
166  * @pages:	array that receives pointers to the pages pinned.
167  *		Should be at least nr_segs long.
168  *
169  * Returns number of pages pinned. This may be fewer than the number requested.
170  * If nr_segs is 0 or negative, returns 0.  If no pages were pinned, returns 0.
171  * Each page returned must be released with a put_page() call when it is
172  * finished with.
173  */
get_kernel_pages(const struct kvec * kiov,int nr_segs,int write,struct page ** pages)174 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
175 		struct page **pages)
176 {
177 	int seg;
178 
179 	for (seg = 0; seg < nr_segs; seg++) {
180 		if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
181 			return seg;
182 
183 		pages[seg] = kmap_to_page(kiov[seg].iov_base);
184 		get_page(pages[seg]);
185 	}
186 
187 	return seg;
188 }
189 EXPORT_SYMBOL_GPL(get_kernel_pages);
190 
pagevec_lru_move_fn(struct pagevec * pvec,void (* move_fn)(struct page * page,struct lruvec * lruvec))191 static void pagevec_lru_move_fn(struct pagevec *pvec,
192 	void (*move_fn)(struct page *page, struct lruvec *lruvec))
193 {
194 	int i;
195 	struct lruvec *lruvec = NULL;
196 	unsigned long flags = 0;
197 
198 	for (i = 0; i < pagevec_count(pvec); i++) {
199 		struct page *page = pvec->pages[i];
200 		struct folio *folio = page_folio(page);
201 
202 		/* block memcg migration during page moving between lru */
203 		if (!TestClearPageLRU(page))
204 			continue;
205 
206 		lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
207 		(*move_fn)(page, lruvec);
208 
209 		SetPageLRU(page);
210 	}
211 	if (lruvec)
212 		unlock_page_lruvec_irqrestore(lruvec, flags);
213 	release_pages(pvec->pages, pvec->nr);
214 	pagevec_reinit(pvec);
215 }
216 
pagevec_move_tail_fn(struct page * page,struct lruvec * lruvec)217 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
218 {
219 	struct folio *folio = page_folio(page);
220 
221 	if (!folio_test_unevictable(folio)) {
222 		lruvec_del_folio(lruvec, folio);
223 		folio_clear_active(folio);
224 		lruvec_add_folio_tail(lruvec, folio);
225 		__count_vm_events(PGROTATED, folio_nr_pages(folio));
226 	}
227 }
228 
229 /* return true if pagevec needs to drain */
pagevec_add_and_need_flush(struct pagevec * pvec,struct page * page)230 static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
231 {
232 	bool ret = false;
233 
234 	if (!pagevec_add(pvec, page) || PageCompound(page) ||
235 			lru_cache_disabled())
236 		ret = true;
237 
238 	return ret;
239 }
240 
241 /*
242  * Writeback is about to end against a folio which has been marked for
243  * immediate reclaim.  If it still appears to be reclaimable, move it
244  * to the tail of the inactive list.
245  *
246  * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
247  */
folio_rotate_reclaimable(struct folio * folio)248 void folio_rotate_reclaimable(struct folio *folio)
249 {
250 	if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
251 	    !folio_test_unevictable(folio) && folio_test_lru(folio)) {
252 		struct pagevec *pvec;
253 		unsigned long flags;
254 
255 		folio_get(folio);
256 		local_lock_irqsave(&lru_rotate.lock, flags);
257 		pvec = this_cpu_ptr(&lru_rotate.pvec);
258 		if (pagevec_add_and_need_flush(pvec, &folio->page))
259 			pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
260 		local_unlock_irqrestore(&lru_rotate.lock, flags);
261 	}
262 }
263 
lru_note_cost(struct lruvec * lruvec,bool file,unsigned int nr_pages)264 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
265 {
266 	do {
267 		unsigned long lrusize;
268 
269 		/*
270 		 * Hold lruvec->lru_lock is safe here, since
271 		 * 1) The pinned lruvec in reclaim, or
272 		 * 2) From a pre-LRU page during refault (which also holds the
273 		 *    rcu lock, so would be safe even if the page was on the LRU
274 		 *    and could move simultaneously to a new lruvec).
275 		 */
276 		spin_lock_irq(&lruvec->lru_lock);
277 		/* Record cost event */
278 		if (file)
279 			lruvec->file_cost += nr_pages;
280 		else
281 			lruvec->anon_cost += nr_pages;
282 
283 		/*
284 		 * Decay previous events
285 		 *
286 		 * Because workloads change over time (and to avoid
287 		 * overflow) we keep these statistics as a floating
288 		 * average, which ends up weighing recent refaults
289 		 * more than old ones.
290 		 */
291 		lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
292 			  lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
293 			  lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
294 			  lruvec_page_state(lruvec, NR_ACTIVE_FILE);
295 
296 		if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
297 			lruvec->file_cost /= 2;
298 			lruvec->anon_cost /= 2;
299 		}
300 		spin_unlock_irq(&lruvec->lru_lock);
301 	} while ((lruvec = parent_lruvec(lruvec)));
302 }
303 
lru_note_cost_folio(struct folio * folio)304 void lru_note_cost_folio(struct folio *folio)
305 {
306 	lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
307 			folio_nr_pages(folio));
308 }
309 
__folio_activate(struct folio * folio,struct lruvec * lruvec)310 static void __folio_activate(struct folio *folio, struct lruvec *lruvec)
311 {
312 	if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
313 		long nr_pages = folio_nr_pages(folio);
314 
315 		lruvec_del_folio(lruvec, folio);
316 		folio_set_active(folio);
317 		lruvec_add_folio(lruvec, folio);
318 		trace_mm_lru_activate(folio);
319 
320 		__count_vm_events(PGACTIVATE, nr_pages);
321 		__count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
322 				     nr_pages);
323 	}
324 }
325 
326 #ifdef CONFIG_SMP
__activate_page(struct page * page,struct lruvec * lruvec)327 static void __activate_page(struct page *page, struct lruvec *lruvec)
328 {
329 	return __folio_activate(page_folio(page), lruvec);
330 }
331 
activate_page_drain(int cpu)332 static void activate_page_drain(int cpu)
333 {
334 	struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
335 
336 	if (pagevec_count(pvec))
337 		pagevec_lru_move_fn(pvec, __activate_page);
338 }
339 
need_activate_page_drain(int cpu)340 static bool need_activate_page_drain(int cpu)
341 {
342 	return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
343 }
344 
folio_activate(struct folio * folio)345 static void folio_activate(struct folio *folio)
346 {
347 	if (folio_test_lru(folio) && !folio_test_active(folio) &&
348 	    !folio_test_unevictable(folio)) {
349 		struct pagevec *pvec;
350 
351 		folio_get(folio);
352 		local_lock(&lru_pvecs.lock);
353 		pvec = this_cpu_ptr(&lru_pvecs.activate_page);
354 		if (pagevec_add_and_need_flush(pvec, &folio->page))
355 			pagevec_lru_move_fn(pvec, __activate_page);
356 		local_unlock(&lru_pvecs.lock);
357 	}
358 }
359 
360 #else
activate_page_drain(int cpu)361 static inline void activate_page_drain(int cpu)
362 {
363 }
364 
folio_activate(struct folio * folio)365 static void folio_activate(struct folio *folio)
366 {
367 	struct lruvec *lruvec;
368 
369 	if (folio_test_clear_lru(folio)) {
370 		lruvec = folio_lruvec_lock_irq(folio);
371 		__folio_activate(folio, lruvec);
372 		unlock_page_lruvec_irq(lruvec);
373 		folio_set_lru(folio);
374 	}
375 }
376 #endif
377 
__lru_cache_activate_folio(struct folio * folio)378 static void __lru_cache_activate_folio(struct folio *folio)
379 {
380 	struct pagevec *pvec;
381 	int i;
382 
383 	local_lock(&lru_pvecs.lock);
384 	pvec = this_cpu_ptr(&lru_pvecs.lru_add);
385 
386 	/*
387 	 * Search backwards on the optimistic assumption that the page being
388 	 * activated has just been added to this pagevec. Note that only
389 	 * the local pagevec is examined as a !PageLRU page could be in the
390 	 * process of being released, reclaimed, migrated or on a remote
391 	 * pagevec that is currently being drained. Furthermore, marking
392 	 * a remote pagevec's page PageActive potentially hits a race where
393 	 * a page is marked PageActive just after it is added to the inactive
394 	 * list causing accounting errors and BUG_ON checks to trigger.
395 	 */
396 	for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
397 		struct page *pagevec_page = pvec->pages[i];
398 
399 		if (pagevec_page == &folio->page) {
400 			folio_set_active(folio);
401 			break;
402 		}
403 	}
404 
405 	local_unlock(&lru_pvecs.lock);
406 }
407 
408 /*
409  * Mark a page as having seen activity.
410  *
411  * inactive,unreferenced	->	inactive,referenced
412  * inactive,referenced		->	active,unreferenced
413  * active,unreferenced		->	active,referenced
414  *
415  * When a newly allocated page is not yet visible, so safe for non-atomic ops,
416  * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
417  */
folio_mark_accessed(struct folio * folio)418 void folio_mark_accessed(struct folio *folio)
419 {
420 	if (!folio_test_referenced(folio)) {
421 		folio_set_referenced(folio);
422 	} else if (folio_test_unevictable(folio)) {
423 		/*
424 		 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
425 		 * this list is never rotated or maintained, so marking an
426 		 * unevictable page accessed has no effect.
427 		 */
428 	} else if (!folio_test_active(folio)) {
429 		/*
430 		 * If the page is on the LRU, queue it for activation via
431 		 * lru_pvecs.activate_page. Otherwise, assume the page is on a
432 		 * pagevec, mark it active and it'll be moved to the active
433 		 * LRU on the next drain.
434 		 */
435 		if (folio_test_lru(folio))
436 			folio_activate(folio);
437 		else
438 			__lru_cache_activate_folio(folio);
439 		folio_clear_referenced(folio);
440 		workingset_activation(folio);
441 	}
442 	if (folio_test_idle(folio))
443 		folio_clear_idle(folio);
444 }
445 EXPORT_SYMBOL(folio_mark_accessed);
446 
447 /**
448  * folio_add_lru - Add a folio to an LRU list.
449  * @folio: The folio to be added to the LRU.
450  *
451  * Queue the folio for addition to the LRU. The decision on whether
452  * to add the page to the [in]active [file|anon] list is deferred until the
453  * pagevec is drained. This gives a chance for the caller of folio_add_lru()
454  * have the folio added to the active list using folio_mark_accessed().
455  */
folio_add_lru(struct folio * folio)456 void folio_add_lru(struct folio *folio)
457 {
458 	struct pagevec *pvec;
459 
460 	VM_BUG_ON_FOLIO(folio_test_active(folio) && folio_test_unevictable(folio), folio);
461 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
462 
463 	folio_get(folio);
464 	local_lock(&lru_pvecs.lock);
465 	pvec = this_cpu_ptr(&lru_pvecs.lru_add);
466 	if (pagevec_add_and_need_flush(pvec, &folio->page))
467 		__pagevec_lru_add(pvec);
468 	local_unlock(&lru_pvecs.lock);
469 }
470 EXPORT_SYMBOL(folio_add_lru);
471 
472 /**
473  * lru_cache_add_inactive_or_unevictable
474  * @page:  the page to be added to LRU
475  * @vma:   vma in which page is mapped for determining reclaimability
476  *
477  * Place @page on the inactive or unevictable LRU list, depending on its
478  * evictability.
479  */
lru_cache_add_inactive_or_unevictable(struct page * page,struct vm_area_struct * vma)480 void lru_cache_add_inactive_or_unevictable(struct page *page,
481 					 struct vm_area_struct *vma)
482 {
483 	VM_BUG_ON_PAGE(PageLRU(page), page);
484 
485 	if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
486 		mlock_new_page(page);
487 	else
488 		lru_cache_add(page);
489 }
490 
491 /*
492  * If the page can not be invalidated, it is moved to the
493  * inactive list to speed up its reclaim.  It is moved to the
494  * head of the list, rather than the tail, to give the flusher
495  * threads some time to write it out, as this is much more
496  * effective than the single-page writeout from reclaim.
497  *
498  * If the page isn't page_mapped and dirty/writeback, the page
499  * could reclaim asap using PG_reclaim.
500  *
501  * 1. active, mapped page -> none
502  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
503  * 3. inactive, mapped page -> none
504  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
505  * 5. inactive, clean -> inactive, tail
506  * 6. Others -> none
507  *
508  * In 4, why it moves inactive's head, the VM expects the page would
509  * be write it out by flusher threads as this is much more effective
510  * than the single-page writeout from reclaim.
511  */
lru_deactivate_file_fn(struct page * page,struct lruvec * lruvec)512 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
513 {
514 	bool active = PageActive(page);
515 	int nr_pages = thp_nr_pages(page);
516 
517 	if (PageUnevictable(page))
518 		return;
519 
520 	/* Some processes are using the page */
521 	if (page_mapped(page))
522 		return;
523 
524 	del_page_from_lru_list(page, lruvec);
525 	ClearPageActive(page);
526 	ClearPageReferenced(page);
527 
528 	if (PageWriteback(page) || PageDirty(page)) {
529 		/*
530 		 * PG_reclaim could be raced with end_page_writeback
531 		 * It can make readahead confusing.  But race window
532 		 * is _really_ small and  it's non-critical problem.
533 		 */
534 		add_page_to_lru_list(page, lruvec);
535 		SetPageReclaim(page);
536 	} else {
537 		/*
538 		 * The page's writeback ends up during pagevec
539 		 * We move that page into tail of inactive.
540 		 */
541 		add_page_to_lru_list_tail(page, lruvec);
542 		__count_vm_events(PGROTATED, nr_pages);
543 	}
544 
545 	if (active) {
546 		__count_vm_events(PGDEACTIVATE, nr_pages);
547 		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
548 				     nr_pages);
549 	}
550 }
551 
lru_deactivate_fn(struct page * page,struct lruvec * lruvec)552 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
553 {
554 	if (PageActive(page) && !PageUnevictable(page)) {
555 		int nr_pages = thp_nr_pages(page);
556 
557 		del_page_from_lru_list(page, lruvec);
558 		ClearPageActive(page);
559 		ClearPageReferenced(page);
560 		add_page_to_lru_list(page, lruvec);
561 
562 		__count_vm_events(PGDEACTIVATE, nr_pages);
563 		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
564 				     nr_pages);
565 	}
566 }
567 
lru_lazyfree_fn(struct page * page,struct lruvec * lruvec)568 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
569 {
570 	if (PageAnon(page) && PageSwapBacked(page) &&
571 	    !PageSwapCache(page) && !PageUnevictable(page)) {
572 		int nr_pages = thp_nr_pages(page);
573 
574 		del_page_from_lru_list(page, lruvec);
575 		ClearPageActive(page);
576 		ClearPageReferenced(page);
577 		/*
578 		 * Lazyfree pages are clean anonymous pages.  They have
579 		 * PG_swapbacked flag cleared, to distinguish them from normal
580 		 * anonymous pages
581 		 */
582 		ClearPageSwapBacked(page);
583 		add_page_to_lru_list(page, lruvec);
584 
585 		__count_vm_events(PGLAZYFREE, nr_pages);
586 		__count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
587 				     nr_pages);
588 	}
589 }
590 
591 /*
592  * Drain pages out of the cpu's pagevecs.
593  * Either "cpu" is the current CPU, and preemption has already been
594  * disabled; or "cpu" is being hot-unplugged, and is already dead.
595  */
lru_add_drain_cpu(int cpu)596 void lru_add_drain_cpu(int cpu)
597 {
598 	struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
599 
600 	if (pagevec_count(pvec))
601 		__pagevec_lru_add(pvec);
602 
603 	pvec = &per_cpu(lru_rotate.pvec, cpu);
604 	/* Disabling interrupts below acts as a compiler barrier. */
605 	if (data_race(pagevec_count(pvec))) {
606 		unsigned long flags;
607 
608 		/* No harm done if a racing interrupt already did this */
609 		local_lock_irqsave(&lru_rotate.lock, flags);
610 		pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
611 		local_unlock_irqrestore(&lru_rotate.lock, flags);
612 	}
613 
614 	pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
615 	if (pagevec_count(pvec))
616 		pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
617 
618 	pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
619 	if (pagevec_count(pvec))
620 		pagevec_lru_move_fn(pvec, lru_deactivate_fn);
621 
622 	pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
623 	if (pagevec_count(pvec))
624 		pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
625 
626 	activate_page_drain(cpu);
627 }
628 
629 /**
630  * deactivate_file_folio() - Forcefully deactivate a file folio.
631  * @folio: Folio to deactivate.
632  *
633  * This function hints to the VM that @folio is a good reclaim candidate,
634  * for example if its invalidation fails due to the folio being dirty
635  * or under writeback.
636  *
637  * Context: Caller holds a reference on the page.
638  */
deactivate_file_folio(struct folio * folio)639 void deactivate_file_folio(struct folio *folio)
640 {
641 	struct pagevec *pvec;
642 
643 	/*
644 	 * In a workload with many unevictable pages such as mprotect,
645 	 * unevictable folio deactivation for accelerating reclaim is pointless.
646 	 */
647 	if (folio_test_unevictable(folio))
648 		return;
649 
650 	folio_get(folio);
651 	local_lock(&lru_pvecs.lock);
652 	pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
653 
654 	if (pagevec_add_and_need_flush(pvec, &folio->page))
655 		pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
656 	local_unlock(&lru_pvecs.lock);
657 }
658 
659 /*
660  * deactivate_page - deactivate a page
661  * @page: page to deactivate
662  *
663  * deactivate_page() moves @page to the inactive list if @page was on the active
664  * list and was not an unevictable page.  This is done to accelerate the reclaim
665  * of @page.
666  */
deactivate_page(struct page * page)667 void deactivate_page(struct page *page)
668 {
669 	if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
670 		struct pagevec *pvec;
671 
672 		local_lock(&lru_pvecs.lock);
673 		pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
674 		get_page(page);
675 		if (pagevec_add_and_need_flush(pvec, page))
676 			pagevec_lru_move_fn(pvec, lru_deactivate_fn);
677 		local_unlock(&lru_pvecs.lock);
678 	}
679 }
680 
681 /**
682  * mark_page_lazyfree - make an anon page lazyfree
683  * @page: page to deactivate
684  *
685  * mark_page_lazyfree() moves @page to the inactive file list.
686  * This is done to accelerate the reclaim of @page.
687  */
mark_page_lazyfree(struct page * page)688 void mark_page_lazyfree(struct page *page)
689 {
690 	if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
691 	    !PageSwapCache(page) && !PageUnevictable(page)) {
692 		struct pagevec *pvec;
693 
694 		local_lock(&lru_pvecs.lock);
695 		pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
696 		get_page(page);
697 		if (pagevec_add_and_need_flush(pvec, page))
698 			pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
699 		local_unlock(&lru_pvecs.lock);
700 	}
701 }
702 
lru_add_drain(void)703 void lru_add_drain(void)
704 {
705 	local_lock(&lru_pvecs.lock);
706 	lru_add_drain_cpu(smp_processor_id());
707 	local_unlock(&lru_pvecs.lock);
708 	mlock_page_drain_local();
709 }
710 
711 /*
712  * It's called from per-cpu workqueue context in SMP case so
713  * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
714  * the same cpu. It shouldn't be a problem in !SMP case since
715  * the core is only one and the locks will disable preemption.
716  */
lru_add_and_bh_lrus_drain(void)717 static void lru_add_and_bh_lrus_drain(void)
718 {
719 	local_lock(&lru_pvecs.lock);
720 	lru_add_drain_cpu(smp_processor_id());
721 	local_unlock(&lru_pvecs.lock);
722 	invalidate_bh_lrus_cpu();
723 	mlock_page_drain_local();
724 }
725 
lru_add_drain_cpu_zone(struct zone * zone)726 void lru_add_drain_cpu_zone(struct zone *zone)
727 {
728 	local_lock(&lru_pvecs.lock);
729 	lru_add_drain_cpu(smp_processor_id());
730 	drain_local_pages(zone);
731 	local_unlock(&lru_pvecs.lock);
732 	mlock_page_drain_local();
733 }
734 
735 #ifdef CONFIG_SMP
736 
737 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
738 
lru_add_drain_per_cpu(struct work_struct * dummy)739 static void lru_add_drain_per_cpu(struct work_struct *dummy)
740 {
741 	lru_add_and_bh_lrus_drain();
742 }
743 
744 /*
745  * Doesn't need any cpu hotplug locking because we do rely on per-cpu
746  * kworkers being shut down before our page_alloc_cpu_dead callback is
747  * executed on the offlined cpu.
748  * Calling this function with cpu hotplug locks held can actually lead
749  * to obscure indirect dependencies via WQ context.
750  */
__lru_add_drain_all(bool force_all_cpus)751 static inline void __lru_add_drain_all(bool force_all_cpus)
752 {
753 	/*
754 	 * lru_drain_gen - Global pages generation number
755 	 *
756 	 * (A) Definition: global lru_drain_gen = x implies that all generations
757 	 *     0 < n <= x are already *scheduled* for draining.
758 	 *
759 	 * This is an optimization for the highly-contended use case where a
760 	 * user space workload keeps constantly generating a flow of pages for
761 	 * each CPU.
762 	 */
763 	static unsigned int lru_drain_gen;
764 	static struct cpumask has_work;
765 	static DEFINE_MUTEX(lock);
766 	unsigned cpu, this_gen;
767 
768 	/*
769 	 * Make sure nobody triggers this path before mm_percpu_wq is fully
770 	 * initialized.
771 	 */
772 	if (WARN_ON(!mm_percpu_wq))
773 		return;
774 
775 	/*
776 	 * Guarantee pagevec counter stores visible by this CPU are visible to
777 	 * other CPUs before loading the current drain generation.
778 	 */
779 	smp_mb();
780 
781 	/*
782 	 * (B) Locally cache global LRU draining generation number
783 	 *
784 	 * The read barrier ensures that the counter is loaded before the mutex
785 	 * is taken. It pairs with smp_mb() inside the mutex critical section
786 	 * at (D).
787 	 */
788 	this_gen = smp_load_acquire(&lru_drain_gen);
789 
790 	mutex_lock(&lock);
791 
792 	/*
793 	 * (C) Exit the draining operation if a newer generation, from another
794 	 * lru_add_drain_all(), was already scheduled for draining. Check (A).
795 	 */
796 	if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
797 		goto done;
798 
799 	/*
800 	 * (D) Increment global generation number
801 	 *
802 	 * Pairs with smp_load_acquire() at (B), outside of the critical
803 	 * section. Use a full memory barrier to guarantee that the new global
804 	 * drain generation number is stored before loading pagevec counters.
805 	 *
806 	 * This pairing must be done here, before the for_each_online_cpu loop
807 	 * below which drains the page vectors.
808 	 *
809 	 * Let x, y, and z represent some system CPU numbers, where x < y < z.
810 	 * Assume CPU #z is in the middle of the for_each_online_cpu loop
811 	 * below and has already reached CPU #y's per-cpu data. CPU #x comes
812 	 * along, adds some pages to its per-cpu vectors, then calls
813 	 * lru_add_drain_all().
814 	 *
815 	 * If the paired barrier is done at any later step, e.g. after the
816 	 * loop, CPU #x will just exit at (C) and miss flushing out all of its
817 	 * added pages.
818 	 */
819 	WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
820 	smp_mb();
821 
822 	cpumask_clear(&has_work);
823 	for_each_online_cpu(cpu) {
824 		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
825 
826 		if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
827 		    data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
828 		    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
829 		    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
830 		    pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
831 		    need_activate_page_drain(cpu) ||
832 		    need_mlock_page_drain(cpu) ||
833 		    has_bh_in_lru(cpu, NULL)) {
834 			INIT_WORK(work, lru_add_drain_per_cpu);
835 			queue_work_on(cpu, mm_percpu_wq, work);
836 			__cpumask_set_cpu(cpu, &has_work);
837 		}
838 	}
839 
840 	for_each_cpu(cpu, &has_work)
841 		flush_work(&per_cpu(lru_add_drain_work, cpu));
842 
843 done:
844 	mutex_unlock(&lock);
845 }
846 
lru_add_drain_all(void)847 void lru_add_drain_all(void)
848 {
849 	__lru_add_drain_all(false);
850 }
851 #else
lru_add_drain_all(void)852 void lru_add_drain_all(void)
853 {
854 	lru_add_drain();
855 }
856 #endif /* CONFIG_SMP */
857 
858 atomic_t lru_disable_count = ATOMIC_INIT(0);
859 
860 /*
861  * lru_cache_disable() needs to be called before we start compiling
862  * a list of pages to be migrated using isolate_lru_page().
863  * It drains pages on LRU cache and then disable on all cpus until
864  * lru_cache_enable is called.
865  *
866  * Must be paired with a call to lru_cache_enable().
867  */
lru_cache_disable(void)868 void lru_cache_disable(void)
869 {
870 	atomic_inc(&lru_disable_count);
871 	/*
872 	 * Readers of lru_disable_count are protected by either disabling
873 	 * preemption or rcu_read_lock:
874 	 *
875 	 * preempt_disable, local_irq_disable  [bh_lru_lock()]
876 	 * rcu_read_lock		       [rt_spin_lock CONFIG_PREEMPT_RT]
877 	 * preempt_disable		       [local_lock !CONFIG_PREEMPT_RT]
878 	 *
879 	 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
880 	 * preempt_disable() regions of code. So any CPU which sees
881 	 * lru_disable_count = 0 will have exited the critical
882 	 * section when synchronize_rcu() returns.
883 	 */
884 	synchronize_rcu_expedited();
885 #ifdef CONFIG_SMP
886 	__lru_add_drain_all(true);
887 #else
888 	lru_add_and_bh_lrus_drain();
889 #endif
890 }
891 
892 /**
893  * release_pages - batched put_page()
894  * @pages: array of pages to release
895  * @nr: number of pages
896  *
897  * Decrement the reference count on all the pages in @pages.  If it
898  * fell to zero, remove the page from the LRU and free it.
899  */
release_pages(struct page ** pages,int nr)900 void release_pages(struct page **pages, int nr)
901 {
902 	int i;
903 	LIST_HEAD(pages_to_free);
904 	struct lruvec *lruvec = NULL;
905 	unsigned long flags = 0;
906 	unsigned int lock_batch;
907 
908 	for (i = 0; i < nr; i++) {
909 		struct page *page = pages[i];
910 		struct folio *folio = page_folio(page);
911 
912 		/*
913 		 * Make sure the IRQ-safe lock-holding time does not get
914 		 * excessive with a continuous string of pages from the
915 		 * same lruvec. The lock is held only if lruvec != NULL.
916 		 */
917 		if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
918 			unlock_page_lruvec_irqrestore(lruvec, flags);
919 			lruvec = NULL;
920 		}
921 
922 		page = &folio->page;
923 		if (is_huge_zero_page(page))
924 			continue;
925 
926 		if (is_zone_device_page(page)) {
927 			if (lruvec) {
928 				unlock_page_lruvec_irqrestore(lruvec, flags);
929 				lruvec = NULL;
930 			}
931 			if (put_devmap_managed_page(page))
932 				continue;
933 			if (put_page_testzero(page))
934 				free_zone_device_page(page);
935 			continue;
936 		}
937 
938 		if (!put_page_testzero(page))
939 			continue;
940 
941 		if (PageCompound(page)) {
942 			if (lruvec) {
943 				unlock_page_lruvec_irqrestore(lruvec, flags);
944 				lruvec = NULL;
945 			}
946 			__put_compound_page(page);
947 			continue;
948 		}
949 
950 		if (PageLRU(page)) {
951 			struct lruvec *prev_lruvec = lruvec;
952 
953 			lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
954 									&flags);
955 			if (prev_lruvec != lruvec)
956 				lock_batch = 0;
957 
958 			del_page_from_lru_list(page, lruvec);
959 			__clear_page_lru_flags(page);
960 		}
961 
962 		/*
963 		 * In rare cases, when truncation or holepunching raced with
964 		 * munlock after VM_LOCKED was cleared, Mlocked may still be
965 		 * found set here.  This does not indicate a problem, unless
966 		 * "unevictable_pgs_cleared" appears worryingly large.
967 		 */
968 		if (unlikely(PageMlocked(page))) {
969 			__ClearPageMlocked(page);
970 			dec_zone_page_state(page, NR_MLOCK);
971 			count_vm_event(UNEVICTABLE_PGCLEARED);
972 		}
973 
974 		list_add(&page->lru, &pages_to_free);
975 	}
976 	if (lruvec)
977 		unlock_page_lruvec_irqrestore(lruvec, flags);
978 
979 	mem_cgroup_uncharge_list(&pages_to_free);
980 	free_unref_page_list(&pages_to_free);
981 }
982 EXPORT_SYMBOL(release_pages);
983 
984 /*
985  * The pages which we're about to release may be in the deferred lru-addition
986  * queues.  That would prevent them from really being freed right now.  That's
987  * OK from a correctness point of view but is inefficient - those pages may be
988  * cache-warm and we want to give them back to the page allocator ASAP.
989  *
990  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
991  * and __pagevec_lru_add_active() call release_pages() directly to avoid
992  * mutual recursion.
993  */
__pagevec_release(struct pagevec * pvec)994 void __pagevec_release(struct pagevec *pvec)
995 {
996 	if (!pvec->percpu_pvec_drained) {
997 		lru_add_drain();
998 		pvec->percpu_pvec_drained = true;
999 	}
1000 	release_pages(pvec->pages, pagevec_count(pvec));
1001 	pagevec_reinit(pvec);
1002 }
1003 EXPORT_SYMBOL(__pagevec_release);
1004 
__pagevec_lru_add_fn(struct folio * folio,struct lruvec * lruvec)1005 static void __pagevec_lru_add_fn(struct folio *folio, struct lruvec *lruvec)
1006 {
1007 	int was_unevictable = folio_test_clear_unevictable(folio);
1008 	long nr_pages = folio_nr_pages(folio);
1009 
1010 	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1011 
1012 	folio_set_lru(folio);
1013 	/*
1014 	 * Is an smp_mb__after_atomic() still required here, before
1015 	 * folio_evictable() tests PageMlocked, to rule out the possibility
1016 	 * of stranding an evictable folio on an unevictable LRU?  I think
1017 	 * not, because __munlock_page() only clears PageMlocked while the LRU
1018 	 * lock is held.
1019 	 *
1020 	 * (That is not true of __page_cache_release(), and not necessarily
1021 	 * true of release_pages(): but those only clear PageMlocked after
1022 	 * put_page_testzero() has excluded any other users of the page.)
1023 	 */
1024 	if (folio_evictable(folio)) {
1025 		if (was_unevictable)
1026 			__count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1027 	} else {
1028 		folio_clear_active(folio);
1029 		folio_set_unevictable(folio);
1030 		/*
1031 		 * folio->mlock_count = !!folio_test_mlocked(folio)?
1032 		 * But that leaves __mlock_page() in doubt whether another
1033 		 * actor has already counted the mlock or not.  Err on the
1034 		 * safe side, underestimate, let page reclaim fix it, rather
1035 		 * than leaving a page on the unevictable LRU indefinitely.
1036 		 */
1037 		folio->mlock_count = 0;
1038 		if (!was_unevictable)
1039 			__count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1040 	}
1041 
1042 	lruvec_add_folio(lruvec, folio);
1043 	trace_mm_lru_insertion(folio);
1044 }
1045 
1046 /*
1047  * Add the passed pages to the LRU, then drop the caller's refcount
1048  * on them.  Reinitialises the caller's pagevec.
1049  */
__pagevec_lru_add(struct pagevec * pvec)1050 void __pagevec_lru_add(struct pagevec *pvec)
1051 {
1052 	int i;
1053 	struct lruvec *lruvec = NULL;
1054 	unsigned long flags = 0;
1055 
1056 	for (i = 0; i < pagevec_count(pvec); i++) {
1057 		struct folio *folio = page_folio(pvec->pages[i]);
1058 
1059 		lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
1060 		__pagevec_lru_add_fn(folio, lruvec);
1061 	}
1062 	if (lruvec)
1063 		unlock_page_lruvec_irqrestore(lruvec, flags);
1064 	release_pages(pvec->pages, pvec->nr);
1065 	pagevec_reinit(pvec);
1066 }
1067 
1068 /**
1069  * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1070  * @fbatch: The batch to prune
1071  *
1072  * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1073  * entries.  This function prunes all the non-folio entries from @fbatch
1074  * without leaving holes, so that it can be passed on to folio-only batch
1075  * operations.
1076  */
folio_batch_remove_exceptionals(struct folio_batch * fbatch)1077 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1078 {
1079 	unsigned int i, j;
1080 
1081 	for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1082 		struct folio *folio = fbatch->folios[i];
1083 		if (!xa_is_value(folio))
1084 			fbatch->folios[j++] = folio;
1085 	}
1086 	fbatch->nr = j;
1087 }
1088 
1089 /**
1090  * pagevec_lookup_range - gang pagecache lookup
1091  * @pvec:	Where the resulting pages are placed
1092  * @mapping:	The address_space to search
1093  * @start:	The starting page index
1094  * @end:	The final page index
1095  *
1096  * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1097  * pages in the mapping starting from index @start and upto index @end
1098  * (inclusive).  The pages are placed in @pvec.  pagevec_lookup() takes a
1099  * reference against the pages in @pvec.
1100  *
1101  * The search returns a group of mapping-contiguous pages with ascending
1102  * indexes.  There may be holes in the indices due to not-present pages. We
1103  * also update @start to index the next page for the traversal.
1104  *
1105  * pagevec_lookup_range() returns the number of pages which were found. If this
1106  * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1107  * reached.
1108  */
pagevec_lookup_range(struct pagevec * pvec,struct address_space * mapping,pgoff_t * start,pgoff_t end)1109 unsigned pagevec_lookup_range(struct pagevec *pvec,
1110 		struct address_space *mapping, pgoff_t *start, pgoff_t end)
1111 {
1112 	pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1113 					pvec->pages);
1114 	return pagevec_count(pvec);
1115 }
1116 EXPORT_SYMBOL(pagevec_lookup_range);
1117 
pagevec_lookup_range_tag(struct pagevec * pvec,struct address_space * mapping,pgoff_t * index,pgoff_t end,xa_mark_t tag)1118 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1119 		struct address_space *mapping, pgoff_t *index, pgoff_t end,
1120 		xa_mark_t tag)
1121 {
1122 	pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1123 					PAGEVEC_SIZE, pvec->pages);
1124 	return pagevec_count(pvec);
1125 }
1126 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1127 
1128 /*
1129  * Perform any setup for the swap system
1130  */
swap_setup(void)1131 void __init swap_setup(void)
1132 {
1133 	unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1134 
1135 	/* Use a smaller cluster for small-memory machines */
1136 	if (megs < 16)
1137 		page_cluster = 2;
1138 	else
1139 		page_cluster = 3;
1140 	/*
1141 	 * Right now other parts of the system means that we
1142 	 * _really_ don't want to cluster much more
1143 	 */
1144 }
1145