xref: /openbmc/linux/mm/swap.c (revision fff74a93)
1 /*
2  *  linux/mm/swap.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6 
7 /*
8  * This file contains the default values for the operation of the
9  * Linux VM subsystem. Fine-tuning documentation can be found in
10  * Documentation/sysctl/vm.txt.
11  * Started 18.12.91
12  * Swap aging added 23.2.95, Stephen Tweedie.
13  * Buffermem limits added 12.3.98, Rik van Riel.
14  */
15 
16 #include <linux/mm.h>
17 #include <linux/sched.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/swap.h>
20 #include <linux/mman.h>
21 #include <linux/pagemap.h>
22 #include <linux/pagevec.h>
23 #include <linux/init.h>
24 #include <linux/export.h>
25 #include <linux/mm_inline.h>
26 #include <linux/percpu_counter.h>
27 #include <linux/percpu.h>
28 #include <linux/cpu.h>
29 #include <linux/notifier.h>
30 #include <linux/backing-dev.h>
31 #include <linux/memcontrol.h>
32 #include <linux/gfp.h>
33 #include <linux/uio.h>
34 
35 #include "internal.h"
36 
37 #define CREATE_TRACE_POINTS
38 #include <trace/events/pagemap.h>
39 
40 /* How many pages do we try to swap or page in/out together? */
41 int page_cluster;
42 
43 static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
44 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
45 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
46 
47 /*
48  * This path almost never happens for VM activity - pages are normally
49  * freed via pagevecs.  But it gets used by networking.
50  */
51 static void __page_cache_release(struct page *page)
52 {
53 	if (PageLRU(page)) {
54 		struct zone *zone = page_zone(page);
55 		struct lruvec *lruvec;
56 		unsigned long flags;
57 
58 		spin_lock_irqsave(&zone->lru_lock, flags);
59 		lruvec = mem_cgroup_page_lruvec(page, zone);
60 		VM_BUG_ON_PAGE(!PageLRU(page), page);
61 		__ClearPageLRU(page);
62 		del_page_from_lru_list(page, lruvec, page_off_lru(page));
63 		spin_unlock_irqrestore(&zone->lru_lock, flags);
64 	}
65 	mem_cgroup_uncharge(page);
66 }
67 
68 static void __put_single_page(struct page *page)
69 {
70 	__page_cache_release(page);
71 	free_hot_cold_page(page, false);
72 }
73 
74 static void __put_compound_page(struct page *page)
75 {
76 	compound_page_dtor *dtor;
77 
78 	__page_cache_release(page);
79 	dtor = get_compound_page_dtor(page);
80 	(*dtor)(page);
81 }
82 
83 /**
84  * Two special cases here: we could avoid taking compound_lock_irqsave
85  * and could skip the tail refcounting(in _mapcount).
86  *
87  * 1. Hugetlbfs page:
88  *
89  *    PageHeadHuge will remain true until the compound page
90  *    is released and enters the buddy allocator, and it could
91  *    not be split by __split_huge_page_refcount().
92  *
93  *    So if we see PageHeadHuge set, and we have the tail page pin,
94  *    then we could safely put head page.
95  *
96  * 2. Slab THP page:
97  *
98  *    PG_slab is cleared before the slab frees the head page, and
99  *    tail pin cannot be the last reference left on the head page,
100  *    because the slab code is free to reuse the compound page
101  *    after a kfree/kmem_cache_free without having to check if
102  *    there's any tail pin left.  In turn all tail pinsmust be always
103  *    released while the head is still pinned by the slab code
104  *    and so we know PG_slab will be still set too.
105  *
106  *    So if we see PageSlab set, and we have the tail page pin,
107  *    then we could safely put head page.
108  */
109 static __always_inline
110 void put_unrefcounted_compound_page(struct page *page_head, struct page *page)
111 {
112 	/*
113 	 * If @page is a THP tail, we must read the tail page
114 	 * flags after the head page flags. The
115 	 * __split_huge_page_refcount side enforces write memory barriers
116 	 * between clearing PageTail and before the head page
117 	 * can be freed and reallocated.
118 	 */
119 	smp_rmb();
120 	if (likely(PageTail(page))) {
121 		/*
122 		 * __split_huge_page_refcount cannot race
123 		 * here, see the comment above this function.
124 		 */
125 		VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
126 		VM_BUG_ON_PAGE(page_mapcount(page) != 0, page);
127 		if (put_page_testzero(page_head)) {
128 			/*
129 			 * If this is the tail of a slab THP page,
130 			 * the tail pin must not be the last reference
131 			 * held on the page, because the PG_slab cannot
132 			 * be cleared before all tail pins (which skips
133 			 * the _mapcount tail refcounting) have been
134 			 * released.
135 			 *
136 			 * If this is the tail of a hugetlbfs page,
137 			 * the tail pin may be the last reference on
138 			 * the page instead, because PageHeadHuge will
139 			 * not go away until the compound page enters
140 			 * the buddy allocator.
141 			 */
142 			VM_BUG_ON_PAGE(PageSlab(page_head), page_head);
143 			__put_compound_page(page_head);
144 		}
145 	} else
146 		/*
147 		 * __split_huge_page_refcount run before us,
148 		 * @page was a THP tail. The split @page_head
149 		 * has been freed and reallocated as slab or
150 		 * hugetlbfs page of smaller order (only
151 		 * possible if reallocated as slab on x86).
152 		 */
153 		if (put_page_testzero(page))
154 			__put_single_page(page);
155 }
156 
157 static __always_inline
158 void put_refcounted_compound_page(struct page *page_head, struct page *page)
159 {
160 	if (likely(page != page_head && get_page_unless_zero(page_head))) {
161 		unsigned long flags;
162 
163 		/*
164 		 * @page_head wasn't a dangling pointer but it may not
165 		 * be a head page anymore by the time we obtain the
166 		 * lock. That is ok as long as it can't be freed from
167 		 * under us.
168 		 */
169 		flags = compound_lock_irqsave(page_head);
170 		if (unlikely(!PageTail(page))) {
171 			/* __split_huge_page_refcount run before us */
172 			compound_unlock_irqrestore(page_head, flags);
173 			if (put_page_testzero(page_head)) {
174 				/*
175 				 * The @page_head may have been freed
176 				 * and reallocated as a compound page
177 				 * of smaller order and then freed
178 				 * again.  All we know is that it
179 				 * cannot have become: a THP page, a
180 				 * compound page of higher order, a
181 				 * tail page.  That is because we
182 				 * still hold the refcount of the
183 				 * split THP tail and page_head was
184 				 * the THP head before the split.
185 				 */
186 				if (PageHead(page_head))
187 					__put_compound_page(page_head);
188 				else
189 					__put_single_page(page_head);
190 			}
191 out_put_single:
192 			if (put_page_testzero(page))
193 				__put_single_page(page);
194 			return;
195 		}
196 		VM_BUG_ON_PAGE(page_head != page->first_page, page);
197 		/*
198 		 * We can release the refcount taken by
199 		 * get_page_unless_zero() now that
200 		 * __split_huge_page_refcount() is blocked on the
201 		 * compound_lock.
202 		 */
203 		if (put_page_testzero(page_head))
204 			VM_BUG_ON_PAGE(1, page_head);
205 		/* __split_huge_page_refcount will wait now */
206 		VM_BUG_ON_PAGE(page_mapcount(page) <= 0, page);
207 		atomic_dec(&page->_mapcount);
208 		VM_BUG_ON_PAGE(atomic_read(&page_head->_count) <= 0, page_head);
209 		VM_BUG_ON_PAGE(atomic_read(&page->_count) != 0, page);
210 		compound_unlock_irqrestore(page_head, flags);
211 
212 		if (put_page_testzero(page_head)) {
213 			if (PageHead(page_head))
214 				__put_compound_page(page_head);
215 			else
216 				__put_single_page(page_head);
217 		}
218 	} else {
219 		/* @page_head is a dangling pointer */
220 		VM_BUG_ON_PAGE(PageTail(page), page);
221 		goto out_put_single;
222 	}
223 }
224 
225 static void put_compound_page(struct page *page)
226 {
227 	struct page *page_head;
228 
229 	/*
230 	 * We see the PageCompound set and PageTail not set, so @page maybe:
231 	 *  1. hugetlbfs head page, or
232 	 *  2. THP head page.
233 	 */
234 	if (likely(!PageTail(page))) {
235 		if (put_page_testzero(page)) {
236 			/*
237 			 * By the time all refcounts have been released
238 			 * split_huge_page cannot run anymore from under us.
239 			 */
240 			if (PageHead(page))
241 				__put_compound_page(page);
242 			else
243 				__put_single_page(page);
244 		}
245 		return;
246 	}
247 
248 	/*
249 	 * We see the PageCompound set and PageTail set, so @page maybe:
250 	 *  1. a tail hugetlbfs page, or
251 	 *  2. a tail THP page, or
252 	 *  3. a split THP page.
253 	 *
254 	 *  Case 3 is possible, as we may race with
255 	 *  __split_huge_page_refcount tearing down a THP page.
256 	 */
257 	page_head = compound_head_by_tail(page);
258 	if (!__compound_tail_refcounted(page_head))
259 		put_unrefcounted_compound_page(page_head, page);
260 	else
261 		put_refcounted_compound_page(page_head, page);
262 }
263 
264 void put_page(struct page *page)
265 {
266 	if (unlikely(PageCompound(page)))
267 		put_compound_page(page);
268 	else if (put_page_testzero(page))
269 		__put_single_page(page);
270 }
271 EXPORT_SYMBOL(put_page);
272 
273 /*
274  * This function is exported but must not be called by anything other
275  * than get_page(). It implements the slow path of get_page().
276  */
277 bool __get_page_tail(struct page *page)
278 {
279 	/*
280 	 * This takes care of get_page() if run on a tail page
281 	 * returned by one of the get_user_pages/follow_page variants.
282 	 * get_user_pages/follow_page itself doesn't need the compound
283 	 * lock because it runs __get_page_tail_foll() under the
284 	 * proper PT lock that already serializes against
285 	 * split_huge_page().
286 	 */
287 	unsigned long flags;
288 	bool got;
289 	struct page *page_head = compound_head(page);
290 
291 	/* Ref to put_compound_page() comment. */
292 	if (!__compound_tail_refcounted(page_head)) {
293 		smp_rmb();
294 		if (likely(PageTail(page))) {
295 			/*
296 			 * This is a hugetlbfs page or a slab
297 			 * page. __split_huge_page_refcount
298 			 * cannot race here.
299 			 */
300 			VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
301 			__get_page_tail_foll(page, true);
302 			return true;
303 		} else {
304 			/*
305 			 * __split_huge_page_refcount run
306 			 * before us, "page" was a THP
307 			 * tail. The split page_head has been
308 			 * freed and reallocated as slab or
309 			 * hugetlbfs page of smaller order
310 			 * (only possible if reallocated as
311 			 * slab on x86).
312 			 */
313 			return false;
314 		}
315 	}
316 
317 	got = false;
318 	if (likely(page != page_head && get_page_unless_zero(page_head))) {
319 		/*
320 		 * page_head wasn't a dangling pointer but it
321 		 * may not be a head page anymore by the time
322 		 * we obtain the lock. That is ok as long as it
323 		 * can't be freed from under us.
324 		 */
325 		flags = compound_lock_irqsave(page_head);
326 		/* here __split_huge_page_refcount won't run anymore */
327 		if (likely(PageTail(page))) {
328 			__get_page_tail_foll(page, false);
329 			got = true;
330 		}
331 		compound_unlock_irqrestore(page_head, flags);
332 		if (unlikely(!got))
333 			put_page(page_head);
334 	}
335 	return got;
336 }
337 EXPORT_SYMBOL(__get_page_tail);
338 
339 /**
340  * put_pages_list() - release a list of pages
341  * @pages: list of pages threaded on page->lru
342  *
343  * Release a list of pages which are strung together on page.lru.  Currently
344  * used by read_cache_pages() and related error recovery code.
345  */
346 void put_pages_list(struct list_head *pages)
347 {
348 	while (!list_empty(pages)) {
349 		struct page *victim;
350 
351 		victim = list_entry(pages->prev, struct page, lru);
352 		list_del(&victim->lru);
353 		page_cache_release(victim);
354 	}
355 }
356 EXPORT_SYMBOL(put_pages_list);
357 
358 /*
359  * get_kernel_pages() - pin kernel pages in memory
360  * @kiov:	An array of struct kvec structures
361  * @nr_segs:	number of segments to pin
362  * @write:	pinning for read/write, currently ignored
363  * @pages:	array that receives pointers to the pages pinned.
364  *		Should be at least nr_segs long.
365  *
366  * Returns number of pages pinned. This may be fewer than the number
367  * requested. If nr_pages is 0 or negative, returns 0. If no pages
368  * were pinned, returns -errno. Each page returned must be released
369  * with a put_page() call when it is finished with.
370  */
371 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
372 		struct page **pages)
373 {
374 	int seg;
375 
376 	for (seg = 0; seg < nr_segs; seg++) {
377 		if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
378 			return seg;
379 
380 		pages[seg] = kmap_to_page(kiov[seg].iov_base);
381 		page_cache_get(pages[seg]);
382 	}
383 
384 	return seg;
385 }
386 EXPORT_SYMBOL_GPL(get_kernel_pages);
387 
388 /*
389  * get_kernel_page() - pin a kernel page in memory
390  * @start:	starting kernel address
391  * @write:	pinning for read/write, currently ignored
392  * @pages:	array that receives pointer to the page pinned.
393  *		Must be at least nr_segs long.
394  *
395  * Returns 1 if page is pinned. If the page was not pinned, returns
396  * -errno. The page returned must be released with a put_page() call
397  * when it is finished with.
398  */
399 int get_kernel_page(unsigned long start, int write, struct page **pages)
400 {
401 	const struct kvec kiov = {
402 		.iov_base = (void *)start,
403 		.iov_len = PAGE_SIZE
404 	};
405 
406 	return get_kernel_pages(&kiov, 1, write, pages);
407 }
408 EXPORT_SYMBOL_GPL(get_kernel_page);
409 
410 static void pagevec_lru_move_fn(struct pagevec *pvec,
411 	void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
412 	void *arg)
413 {
414 	int i;
415 	struct zone *zone = NULL;
416 	struct lruvec *lruvec;
417 	unsigned long flags = 0;
418 
419 	for (i = 0; i < pagevec_count(pvec); i++) {
420 		struct page *page = pvec->pages[i];
421 		struct zone *pagezone = page_zone(page);
422 
423 		if (pagezone != zone) {
424 			if (zone)
425 				spin_unlock_irqrestore(&zone->lru_lock, flags);
426 			zone = pagezone;
427 			spin_lock_irqsave(&zone->lru_lock, flags);
428 		}
429 
430 		lruvec = mem_cgroup_page_lruvec(page, zone);
431 		(*move_fn)(page, lruvec, arg);
432 	}
433 	if (zone)
434 		spin_unlock_irqrestore(&zone->lru_lock, flags);
435 	release_pages(pvec->pages, pvec->nr, pvec->cold);
436 	pagevec_reinit(pvec);
437 }
438 
439 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
440 				 void *arg)
441 {
442 	int *pgmoved = arg;
443 
444 	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
445 		enum lru_list lru = page_lru_base_type(page);
446 		list_move_tail(&page->lru, &lruvec->lists[lru]);
447 		(*pgmoved)++;
448 	}
449 }
450 
451 /*
452  * pagevec_move_tail() must be called with IRQ disabled.
453  * Otherwise this may cause nasty races.
454  */
455 static void pagevec_move_tail(struct pagevec *pvec)
456 {
457 	int pgmoved = 0;
458 
459 	pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
460 	__count_vm_events(PGROTATED, pgmoved);
461 }
462 
463 /*
464  * Writeback is about to end against a page which has been marked for immediate
465  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
466  * inactive list.
467  */
468 void rotate_reclaimable_page(struct page *page)
469 {
470 	if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
471 	    !PageUnevictable(page) && PageLRU(page)) {
472 		struct pagevec *pvec;
473 		unsigned long flags;
474 
475 		page_cache_get(page);
476 		local_irq_save(flags);
477 		pvec = this_cpu_ptr(&lru_rotate_pvecs);
478 		if (!pagevec_add(pvec, page))
479 			pagevec_move_tail(pvec);
480 		local_irq_restore(flags);
481 	}
482 }
483 
484 static void update_page_reclaim_stat(struct lruvec *lruvec,
485 				     int file, int rotated)
486 {
487 	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
488 
489 	reclaim_stat->recent_scanned[file]++;
490 	if (rotated)
491 		reclaim_stat->recent_rotated[file]++;
492 }
493 
494 static void __activate_page(struct page *page, struct lruvec *lruvec,
495 			    void *arg)
496 {
497 	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
498 		int file = page_is_file_cache(page);
499 		int lru = page_lru_base_type(page);
500 
501 		del_page_from_lru_list(page, lruvec, lru);
502 		SetPageActive(page);
503 		lru += LRU_ACTIVE;
504 		add_page_to_lru_list(page, lruvec, lru);
505 		trace_mm_lru_activate(page);
506 
507 		__count_vm_event(PGACTIVATE);
508 		update_page_reclaim_stat(lruvec, file, 1);
509 	}
510 }
511 
512 #ifdef CONFIG_SMP
513 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
514 
515 static void activate_page_drain(int cpu)
516 {
517 	struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
518 
519 	if (pagevec_count(pvec))
520 		pagevec_lru_move_fn(pvec, __activate_page, NULL);
521 }
522 
523 static bool need_activate_page_drain(int cpu)
524 {
525 	return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
526 }
527 
528 void activate_page(struct page *page)
529 {
530 	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
531 		struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
532 
533 		page_cache_get(page);
534 		if (!pagevec_add(pvec, page))
535 			pagevec_lru_move_fn(pvec, __activate_page, NULL);
536 		put_cpu_var(activate_page_pvecs);
537 	}
538 }
539 
540 #else
541 static inline void activate_page_drain(int cpu)
542 {
543 }
544 
545 static bool need_activate_page_drain(int cpu)
546 {
547 	return false;
548 }
549 
550 void activate_page(struct page *page)
551 {
552 	struct zone *zone = page_zone(page);
553 
554 	spin_lock_irq(&zone->lru_lock);
555 	__activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
556 	spin_unlock_irq(&zone->lru_lock);
557 }
558 #endif
559 
560 static void __lru_cache_activate_page(struct page *page)
561 {
562 	struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
563 	int i;
564 
565 	/*
566 	 * Search backwards on the optimistic assumption that the page being
567 	 * activated has just been added to this pagevec. Note that only
568 	 * the local pagevec is examined as a !PageLRU page could be in the
569 	 * process of being released, reclaimed, migrated or on a remote
570 	 * pagevec that is currently being drained. Furthermore, marking
571 	 * a remote pagevec's page PageActive potentially hits a race where
572 	 * a page is marked PageActive just after it is added to the inactive
573 	 * list causing accounting errors and BUG_ON checks to trigger.
574 	 */
575 	for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
576 		struct page *pagevec_page = pvec->pages[i];
577 
578 		if (pagevec_page == page) {
579 			SetPageActive(page);
580 			break;
581 		}
582 	}
583 
584 	put_cpu_var(lru_add_pvec);
585 }
586 
587 /*
588  * Mark a page as having seen activity.
589  *
590  * inactive,unreferenced	->	inactive,referenced
591  * inactive,referenced		->	active,unreferenced
592  * active,unreferenced		->	active,referenced
593  *
594  * When a newly allocated page is not yet visible, so safe for non-atomic ops,
595  * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
596  */
597 void mark_page_accessed(struct page *page)
598 {
599 	if (!PageActive(page) && !PageUnevictable(page) &&
600 			PageReferenced(page)) {
601 
602 		/*
603 		 * If the page is on the LRU, queue it for activation via
604 		 * activate_page_pvecs. Otherwise, assume the page is on a
605 		 * pagevec, mark it active and it'll be moved to the active
606 		 * LRU on the next drain.
607 		 */
608 		if (PageLRU(page))
609 			activate_page(page);
610 		else
611 			__lru_cache_activate_page(page);
612 		ClearPageReferenced(page);
613 		if (page_is_file_cache(page))
614 			workingset_activation(page);
615 	} else if (!PageReferenced(page)) {
616 		SetPageReferenced(page);
617 	}
618 }
619 EXPORT_SYMBOL(mark_page_accessed);
620 
621 static void __lru_cache_add(struct page *page)
622 {
623 	struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
624 
625 	page_cache_get(page);
626 	if (!pagevec_space(pvec))
627 		__pagevec_lru_add(pvec);
628 	pagevec_add(pvec, page);
629 	put_cpu_var(lru_add_pvec);
630 }
631 
632 /**
633  * lru_cache_add: add a page to the page lists
634  * @page: the page to add
635  */
636 void lru_cache_add_anon(struct page *page)
637 {
638 	if (PageActive(page))
639 		ClearPageActive(page);
640 	__lru_cache_add(page);
641 }
642 
643 void lru_cache_add_file(struct page *page)
644 {
645 	if (PageActive(page))
646 		ClearPageActive(page);
647 	__lru_cache_add(page);
648 }
649 EXPORT_SYMBOL(lru_cache_add_file);
650 
651 /**
652  * lru_cache_add - add a page to a page list
653  * @page: the page to be added to the LRU.
654  *
655  * Queue the page for addition to the LRU via pagevec. The decision on whether
656  * to add the page to the [in]active [file|anon] list is deferred until the
657  * pagevec is drained. This gives a chance for the caller of lru_cache_add()
658  * have the page added to the active list using mark_page_accessed().
659  */
660 void lru_cache_add(struct page *page)
661 {
662 	VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
663 	VM_BUG_ON_PAGE(PageLRU(page), page);
664 	__lru_cache_add(page);
665 }
666 
667 /**
668  * add_page_to_unevictable_list - add a page to the unevictable list
669  * @page:  the page to be added to the unevictable list
670  *
671  * Add page directly to its zone's unevictable list.  To avoid races with
672  * tasks that might be making the page evictable, through eg. munlock,
673  * munmap or exit, while it's not on the lru, we want to add the page
674  * while it's locked or otherwise "invisible" to other tasks.  This is
675  * difficult to do when using the pagevec cache, so bypass that.
676  */
677 void add_page_to_unevictable_list(struct page *page)
678 {
679 	struct zone *zone = page_zone(page);
680 	struct lruvec *lruvec;
681 
682 	spin_lock_irq(&zone->lru_lock);
683 	lruvec = mem_cgroup_page_lruvec(page, zone);
684 	ClearPageActive(page);
685 	SetPageUnevictable(page);
686 	SetPageLRU(page);
687 	add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
688 	spin_unlock_irq(&zone->lru_lock);
689 }
690 
691 /**
692  * lru_cache_add_active_or_unevictable
693  * @page:  the page to be added to LRU
694  * @vma:   vma in which page is mapped for determining reclaimability
695  *
696  * Place @page on the active or unevictable LRU list, depending on its
697  * evictability.  Note that if the page is not evictable, it goes
698  * directly back onto it's zone's unevictable list, it does NOT use a
699  * per cpu pagevec.
700  */
701 void lru_cache_add_active_or_unevictable(struct page *page,
702 					 struct vm_area_struct *vma)
703 {
704 	VM_BUG_ON_PAGE(PageLRU(page), page);
705 
706 	if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
707 		SetPageActive(page);
708 		lru_cache_add(page);
709 		return;
710 	}
711 
712 	if (!TestSetPageMlocked(page)) {
713 		/*
714 		 * We use the irq-unsafe __mod_zone_page_stat because this
715 		 * counter is not modified from interrupt context, and the pte
716 		 * lock is held(spinlock), which implies preemption disabled.
717 		 */
718 		__mod_zone_page_state(page_zone(page), NR_MLOCK,
719 				    hpage_nr_pages(page));
720 		count_vm_event(UNEVICTABLE_PGMLOCKED);
721 	}
722 	add_page_to_unevictable_list(page);
723 }
724 
725 /*
726  * If the page can not be invalidated, it is moved to the
727  * inactive list to speed up its reclaim.  It is moved to the
728  * head of the list, rather than the tail, to give the flusher
729  * threads some time to write it out, as this is much more
730  * effective than the single-page writeout from reclaim.
731  *
732  * If the page isn't page_mapped and dirty/writeback, the page
733  * could reclaim asap using PG_reclaim.
734  *
735  * 1. active, mapped page -> none
736  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
737  * 3. inactive, mapped page -> none
738  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
739  * 5. inactive, clean -> inactive, tail
740  * 6. Others -> none
741  *
742  * In 4, why it moves inactive's head, the VM expects the page would
743  * be write it out by flusher threads as this is much more effective
744  * than the single-page writeout from reclaim.
745  */
746 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
747 			      void *arg)
748 {
749 	int lru, file;
750 	bool active;
751 
752 	if (!PageLRU(page))
753 		return;
754 
755 	if (PageUnevictable(page))
756 		return;
757 
758 	/* Some processes are using the page */
759 	if (page_mapped(page))
760 		return;
761 
762 	active = PageActive(page);
763 	file = page_is_file_cache(page);
764 	lru = page_lru_base_type(page);
765 
766 	del_page_from_lru_list(page, lruvec, lru + active);
767 	ClearPageActive(page);
768 	ClearPageReferenced(page);
769 	add_page_to_lru_list(page, lruvec, lru);
770 
771 	if (PageWriteback(page) || PageDirty(page)) {
772 		/*
773 		 * PG_reclaim could be raced with end_page_writeback
774 		 * It can make readahead confusing.  But race window
775 		 * is _really_ small and  it's non-critical problem.
776 		 */
777 		SetPageReclaim(page);
778 	} else {
779 		/*
780 		 * The page's writeback ends up during pagevec
781 		 * We moves tha page into tail of inactive.
782 		 */
783 		list_move_tail(&page->lru, &lruvec->lists[lru]);
784 		__count_vm_event(PGROTATED);
785 	}
786 
787 	if (active)
788 		__count_vm_event(PGDEACTIVATE);
789 	update_page_reclaim_stat(lruvec, file, 0);
790 }
791 
792 /*
793  * Drain pages out of the cpu's pagevecs.
794  * Either "cpu" is the current CPU, and preemption has already been
795  * disabled; or "cpu" is being hot-unplugged, and is already dead.
796  */
797 void lru_add_drain_cpu(int cpu)
798 {
799 	struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
800 
801 	if (pagevec_count(pvec))
802 		__pagevec_lru_add(pvec);
803 
804 	pvec = &per_cpu(lru_rotate_pvecs, cpu);
805 	if (pagevec_count(pvec)) {
806 		unsigned long flags;
807 
808 		/* No harm done if a racing interrupt already did this */
809 		local_irq_save(flags);
810 		pagevec_move_tail(pvec);
811 		local_irq_restore(flags);
812 	}
813 
814 	pvec = &per_cpu(lru_deactivate_pvecs, cpu);
815 	if (pagevec_count(pvec))
816 		pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
817 
818 	activate_page_drain(cpu);
819 }
820 
821 /**
822  * deactivate_page - forcefully deactivate a page
823  * @page: page to deactivate
824  *
825  * This function hints the VM that @page is a good reclaim candidate,
826  * for example if its invalidation fails due to the page being dirty
827  * or under writeback.
828  */
829 void deactivate_page(struct page *page)
830 {
831 	/*
832 	 * In a workload with many unevictable page such as mprotect, unevictable
833 	 * page deactivation for accelerating reclaim is pointless.
834 	 */
835 	if (PageUnevictable(page))
836 		return;
837 
838 	if (likely(get_page_unless_zero(page))) {
839 		struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
840 
841 		if (!pagevec_add(pvec, page))
842 			pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
843 		put_cpu_var(lru_deactivate_pvecs);
844 	}
845 }
846 
847 void lru_add_drain(void)
848 {
849 	lru_add_drain_cpu(get_cpu());
850 	put_cpu();
851 }
852 
853 static void lru_add_drain_per_cpu(struct work_struct *dummy)
854 {
855 	lru_add_drain();
856 }
857 
858 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
859 
860 void lru_add_drain_all(void)
861 {
862 	static DEFINE_MUTEX(lock);
863 	static struct cpumask has_work;
864 	int cpu;
865 
866 	mutex_lock(&lock);
867 	get_online_cpus();
868 	cpumask_clear(&has_work);
869 
870 	for_each_online_cpu(cpu) {
871 		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
872 
873 		if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
874 		    pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
875 		    pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
876 		    need_activate_page_drain(cpu)) {
877 			INIT_WORK(work, lru_add_drain_per_cpu);
878 			schedule_work_on(cpu, work);
879 			cpumask_set_cpu(cpu, &has_work);
880 		}
881 	}
882 
883 	for_each_cpu(cpu, &has_work)
884 		flush_work(&per_cpu(lru_add_drain_work, cpu));
885 
886 	put_online_cpus();
887 	mutex_unlock(&lock);
888 }
889 
890 /*
891  * Batched page_cache_release().  Decrement the reference count on all the
892  * passed pages.  If it fell to zero then remove the page from the LRU and
893  * free it.
894  *
895  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
896  * for the remainder of the operation.
897  *
898  * The locking in this function is against shrink_inactive_list(): we recheck
899  * the page count inside the lock to see whether shrink_inactive_list()
900  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
901  * will free it.
902  */
903 void release_pages(struct page **pages, int nr, bool cold)
904 {
905 	int i;
906 	LIST_HEAD(pages_to_free);
907 	struct zone *zone = NULL;
908 	struct lruvec *lruvec;
909 	unsigned long uninitialized_var(flags);
910 
911 	for (i = 0; i < nr; i++) {
912 		struct page *page = pages[i];
913 
914 		if (unlikely(PageCompound(page))) {
915 			if (zone) {
916 				spin_unlock_irqrestore(&zone->lru_lock, flags);
917 				zone = NULL;
918 			}
919 			put_compound_page(page);
920 			continue;
921 		}
922 
923 		if (!put_page_testzero(page))
924 			continue;
925 
926 		if (PageLRU(page)) {
927 			struct zone *pagezone = page_zone(page);
928 
929 			if (pagezone != zone) {
930 				if (zone)
931 					spin_unlock_irqrestore(&zone->lru_lock,
932 									flags);
933 				zone = pagezone;
934 				spin_lock_irqsave(&zone->lru_lock, flags);
935 			}
936 
937 			lruvec = mem_cgroup_page_lruvec(page, zone);
938 			VM_BUG_ON_PAGE(!PageLRU(page), page);
939 			__ClearPageLRU(page);
940 			del_page_from_lru_list(page, lruvec, page_off_lru(page));
941 		}
942 
943 		/* Clear Active bit in case of parallel mark_page_accessed */
944 		__ClearPageActive(page);
945 
946 		list_add(&page->lru, &pages_to_free);
947 	}
948 	if (zone)
949 		spin_unlock_irqrestore(&zone->lru_lock, flags);
950 
951 	mem_cgroup_uncharge_list(&pages_to_free);
952 	free_hot_cold_page_list(&pages_to_free, cold);
953 }
954 EXPORT_SYMBOL(release_pages);
955 
956 /*
957  * The pages which we're about to release may be in the deferred lru-addition
958  * queues.  That would prevent them from really being freed right now.  That's
959  * OK from a correctness point of view but is inefficient - those pages may be
960  * cache-warm and we want to give them back to the page allocator ASAP.
961  *
962  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
963  * and __pagevec_lru_add_active() call release_pages() directly to avoid
964  * mutual recursion.
965  */
966 void __pagevec_release(struct pagevec *pvec)
967 {
968 	lru_add_drain();
969 	release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
970 	pagevec_reinit(pvec);
971 }
972 EXPORT_SYMBOL(__pagevec_release);
973 
974 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
975 /* used by __split_huge_page_refcount() */
976 void lru_add_page_tail(struct page *page, struct page *page_tail,
977 		       struct lruvec *lruvec, struct list_head *list)
978 {
979 	const int file = 0;
980 
981 	VM_BUG_ON_PAGE(!PageHead(page), page);
982 	VM_BUG_ON_PAGE(PageCompound(page_tail), page);
983 	VM_BUG_ON_PAGE(PageLRU(page_tail), page);
984 	VM_BUG_ON(NR_CPUS != 1 &&
985 		  !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
986 
987 	if (!list)
988 		SetPageLRU(page_tail);
989 
990 	if (likely(PageLRU(page)))
991 		list_add_tail(&page_tail->lru, &page->lru);
992 	else if (list) {
993 		/* page reclaim is reclaiming a huge page */
994 		get_page(page_tail);
995 		list_add_tail(&page_tail->lru, list);
996 	} else {
997 		struct list_head *list_head;
998 		/*
999 		 * Head page has not yet been counted, as an hpage,
1000 		 * so we must account for each subpage individually.
1001 		 *
1002 		 * Use the standard add function to put page_tail on the list,
1003 		 * but then correct its position so they all end up in order.
1004 		 */
1005 		add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
1006 		list_head = page_tail->lru.prev;
1007 		list_move_tail(&page_tail->lru, list_head);
1008 	}
1009 
1010 	if (!PageUnevictable(page))
1011 		update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
1012 }
1013 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1014 
1015 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
1016 				 void *arg)
1017 {
1018 	int file = page_is_file_cache(page);
1019 	int active = PageActive(page);
1020 	enum lru_list lru = page_lru(page);
1021 
1022 	VM_BUG_ON_PAGE(PageLRU(page), page);
1023 
1024 	SetPageLRU(page);
1025 	add_page_to_lru_list(page, lruvec, lru);
1026 	update_page_reclaim_stat(lruvec, file, active);
1027 	trace_mm_lru_insertion(page, lru);
1028 }
1029 
1030 /*
1031  * Add the passed pages to the LRU, then drop the caller's refcount
1032  * on them.  Reinitialises the caller's pagevec.
1033  */
1034 void __pagevec_lru_add(struct pagevec *pvec)
1035 {
1036 	pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1037 }
1038 EXPORT_SYMBOL(__pagevec_lru_add);
1039 
1040 /**
1041  * pagevec_lookup_entries - gang pagecache lookup
1042  * @pvec:	Where the resulting entries are placed
1043  * @mapping:	The address_space to search
1044  * @start:	The starting entry index
1045  * @nr_entries:	The maximum number of entries
1046  * @indices:	The cache indices corresponding to the entries in @pvec
1047  *
1048  * pagevec_lookup_entries() will search for and return a group of up
1049  * to @nr_entries pages and shadow entries in the mapping.  All
1050  * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1051  * reference against actual pages in @pvec.
1052  *
1053  * The search returns a group of mapping-contiguous entries with
1054  * ascending indexes.  There may be holes in the indices due to
1055  * not-present entries.
1056  *
1057  * pagevec_lookup_entries() returns the number of entries which were
1058  * found.
1059  */
1060 unsigned pagevec_lookup_entries(struct pagevec *pvec,
1061 				struct address_space *mapping,
1062 				pgoff_t start, unsigned nr_pages,
1063 				pgoff_t *indices)
1064 {
1065 	pvec->nr = find_get_entries(mapping, start, nr_pages,
1066 				    pvec->pages, indices);
1067 	return pagevec_count(pvec);
1068 }
1069 
1070 /**
1071  * pagevec_remove_exceptionals - pagevec exceptionals pruning
1072  * @pvec:	The pagevec to prune
1073  *
1074  * pagevec_lookup_entries() fills both pages and exceptional radix
1075  * tree entries into the pagevec.  This function prunes all
1076  * exceptionals from @pvec without leaving holes, so that it can be
1077  * passed on to page-only pagevec operations.
1078  */
1079 void pagevec_remove_exceptionals(struct pagevec *pvec)
1080 {
1081 	int i, j;
1082 
1083 	for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1084 		struct page *page = pvec->pages[i];
1085 		if (!radix_tree_exceptional_entry(page))
1086 			pvec->pages[j++] = page;
1087 	}
1088 	pvec->nr = j;
1089 }
1090 
1091 /**
1092  * pagevec_lookup - gang pagecache lookup
1093  * @pvec:	Where the resulting pages are placed
1094  * @mapping:	The address_space to search
1095  * @start:	The starting page index
1096  * @nr_pages:	The maximum number of pages
1097  *
1098  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
1099  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
1100  * reference against the pages in @pvec.
1101  *
1102  * The search returns a group of mapping-contiguous pages with ascending
1103  * indexes.  There may be holes in the indices due to not-present pages.
1104  *
1105  * pagevec_lookup() returns the number of pages which were found.
1106  */
1107 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
1108 		pgoff_t start, unsigned nr_pages)
1109 {
1110 	pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
1111 	return pagevec_count(pvec);
1112 }
1113 EXPORT_SYMBOL(pagevec_lookup);
1114 
1115 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
1116 		pgoff_t *index, int tag, unsigned nr_pages)
1117 {
1118 	pvec->nr = find_get_pages_tag(mapping, index, tag,
1119 					nr_pages, pvec->pages);
1120 	return pagevec_count(pvec);
1121 }
1122 EXPORT_SYMBOL(pagevec_lookup_tag);
1123 
1124 /*
1125  * Perform any setup for the swap system
1126  */
1127 void __init swap_setup(void)
1128 {
1129 	unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
1130 #ifdef CONFIG_SWAP
1131 	int i;
1132 
1133 	if (bdi_init(swapper_spaces[0].backing_dev_info))
1134 		panic("Failed to init swap bdi");
1135 	for (i = 0; i < MAX_SWAPFILES; i++) {
1136 		spin_lock_init(&swapper_spaces[i].tree_lock);
1137 		INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
1138 	}
1139 #endif
1140 
1141 	/* Use a smaller cluster for small-memory machines */
1142 	if (megs < 16)
1143 		page_cluster = 2;
1144 	else
1145 		page_cluster = 3;
1146 	/*
1147 	 * Right now other parts of the system means that we
1148 	 * _really_ don't want to cluster much more
1149 	 */
1150 }
1151