• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <linux/hugetlb.h>
35 #include <linux/page_idle.h>
36 
37 #include "internal.h"
38 
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/pagemap.h>
41 
42 /* How many pages do we try to swap or page in/out together? */
43 int page_cluster;
44 
45 static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
46 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
47 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_file_pvecs);
48 
49 /*
50  * This path almost never happens for VM activity - pages are normally
51  * freed via pagevecs.  But it gets used by networking.
52  */
__page_cache_release(struct page * page)53 static void __page_cache_release(struct page *page)
54 {
55 	if (PageLRU(page)) {
56 		struct zone *zone = page_zone(page);
57 		struct lruvec *lruvec;
58 		unsigned long flags;
59 
60 		spin_lock_irqsave(&zone->lru_lock, flags);
61 		lruvec = mem_cgroup_page_lruvec(page, zone);
62 		VM_BUG_ON_PAGE(!PageLRU(page), page);
63 		__ClearPageLRU(page);
64 		del_page_from_lru_list(page, lruvec, page_off_lru(page));
65 		spin_unlock_irqrestore(&zone->lru_lock, flags);
66 	}
67 	mem_cgroup_uncharge(page);
68 }
69 
__put_single_page(struct page * page)70 static void __put_single_page(struct page *page)
71 {
72 	__page_cache_release(page);
73 	free_hot_cold_page(page, false);
74 }
75 
__put_compound_page(struct page * page)76 static void __put_compound_page(struct page *page)
77 {
78 	compound_page_dtor *dtor;
79 
80 	/*
81 	 * __page_cache_release() is supposed to be called for thp, not for
82 	 * hugetlb. This is because hugetlb page does never have PageLRU set
83 	 * (it's never listed to any LRU lists) and no memcg routines should
84 	 * be called for hugetlb (it has a separate hugetlb_cgroup.)
85 	 */
86 	if (!PageHuge(page))
87 		__page_cache_release(page);
88 	dtor = get_compound_page_dtor(page);
89 	(*dtor)(page);
90 }
91 
92 /**
93  * Two special cases here: we could avoid taking compound_lock_irqsave
94  * and could skip the tail refcounting(in _mapcount).
95  *
96  * 1. Hugetlbfs page:
97  *
98  *    PageHeadHuge will remain true until the compound page
99  *    is released and enters the buddy allocator, and it could
100  *    not be split by __split_huge_page_refcount().
101  *
102  *    So if we see PageHeadHuge set, and we have the tail page pin,
103  *    then we could safely put head page.
104  *
105  * 2. Slab THP page:
106  *
107  *    PG_slab is cleared before the slab frees the head page, and
108  *    tail pin cannot be the last reference left on the head page,
109  *    because the slab code is free to reuse the compound page
110  *    after a kfree/kmem_cache_free without having to check if
111  *    there's any tail pin left.  In turn all tail pinsmust be always
112  *    released while the head is still pinned by the slab code
113  *    and so we know PG_slab will be still set too.
114  *
115  *    So if we see PageSlab set, and we have the tail page pin,
116  *    then we could safely put head page.
117  */
118 static __always_inline
put_unrefcounted_compound_page(struct page * page_head,struct page * page)119 void put_unrefcounted_compound_page(struct page *page_head, struct page *page)
120 {
121 	/*
122 	 * If @page is a THP tail, we must read the tail page
123 	 * flags after the head page flags. The
124 	 * __split_huge_page_refcount side enforces write memory barriers
125 	 * between clearing PageTail and before the head page
126 	 * can be freed and reallocated.
127 	 */
128 	smp_rmb();
129 	if (likely(PageTail(page))) {
130 		/*
131 		 * __split_huge_page_refcount cannot race
132 		 * here, see the comment above this function.
133 		 */
134 		VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
135 		if (put_page_testzero(page_head)) {
136 			/*
137 			 * If this is the tail of a slab THP page,
138 			 * the tail pin must not be the last reference
139 			 * held on the page, because the PG_slab cannot
140 			 * be cleared before all tail pins (which skips
141 			 * the _mapcount tail refcounting) have been
142 			 * released.
143 			 *
144 			 * If this is the tail of a hugetlbfs page,
145 			 * the tail pin may be the last reference on
146 			 * the page instead, because PageHeadHuge will
147 			 * not go away until the compound page enters
148 			 * the buddy allocator.
149 			 */
150 			VM_BUG_ON_PAGE(PageSlab(page_head), page_head);
151 			__put_compound_page(page_head);
152 		}
153 	} else
154 		/*
155 		 * __split_huge_page_refcount run before us,
156 		 * @page was a THP tail. The split @page_head
157 		 * has been freed and reallocated as slab or
158 		 * hugetlbfs page of smaller order (only
159 		 * possible if reallocated as slab on x86).
160 		 */
161 		if (put_page_testzero(page))
162 			__put_single_page(page);
163 }
164 
165 static __always_inline
put_refcounted_compound_page(struct page * page_head,struct page * page)166 void put_refcounted_compound_page(struct page *page_head, struct page *page)
167 {
168 	if (likely(page != page_head && get_page_unless_zero(page_head))) {
169 		unsigned long flags;
170 
171 		/*
172 		 * @page_head wasn't a dangling pointer but it may not
173 		 * be a head page anymore by the time we obtain the
174 		 * lock. That is ok as long as it can't be freed from
175 		 * under us.
176 		 */
177 		flags = compound_lock_irqsave(page_head);
178 		if (unlikely(!PageTail(page))) {
179 			/* __split_huge_page_refcount run before us */
180 			compound_unlock_irqrestore(page_head, flags);
181 			if (put_page_testzero(page_head)) {
182 				/*
183 				 * The @page_head may have been freed
184 				 * and reallocated as a compound page
185 				 * of smaller order and then freed
186 				 * again.  All we know is that it
187 				 * cannot have become: a THP page, a
188 				 * compound page of higher order, a
189 				 * tail page.  That is because we
190 				 * still hold the refcount of the
191 				 * split THP tail and page_head was
192 				 * the THP head before the split.
193 				 */
194 				if (PageHead(page_head))
195 					__put_compound_page(page_head);
196 				else
197 					__put_single_page(page_head);
198 			}
199 out_put_single:
200 			if (put_page_testzero(page))
201 				__put_single_page(page);
202 			return;
203 		}
204 		VM_BUG_ON_PAGE(page_head != compound_head(page), page);
205 		/*
206 		 * We can release the refcount taken by
207 		 * get_page_unless_zero() now that
208 		 * __split_huge_page_refcount() is blocked on the
209 		 * compound_lock.
210 		 */
211 		if (put_page_testzero(page_head))
212 			VM_BUG_ON_PAGE(1, page_head);
213 		/* __split_huge_page_refcount will wait now */
214 		VM_BUG_ON_PAGE(page_mapcount(page) <= 0, page);
215 		atomic_dec(&page->_mapcount);
216 		VM_BUG_ON_PAGE(atomic_read(&page_head->_count) <= 0, page_head);
217 		VM_BUG_ON_PAGE(atomic_read(&page->_count) != 0, page);
218 		compound_unlock_irqrestore(page_head, flags);
219 
220 		if (put_page_testzero(page_head)) {
221 			if (PageHead(page_head))
222 				__put_compound_page(page_head);
223 			else
224 				__put_single_page(page_head);
225 		}
226 	} else {
227 		/* @page_head is a dangling pointer */
228 		VM_BUG_ON_PAGE(PageTail(page), page);
229 		goto out_put_single;
230 	}
231 }
232 
put_compound_page(struct page * page)233 static void put_compound_page(struct page *page)
234 {
235 	struct page *page_head;
236 
237 	/*
238 	 * We see the PageCompound set and PageTail not set, so @page maybe:
239 	 *  1. hugetlbfs head page, or
240 	 *  2. THP head page.
241 	 */
242 	if (likely(!PageTail(page))) {
243 		if (put_page_testzero(page)) {
244 			/*
245 			 * By the time all refcounts have been released
246 			 * split_huge_page cannot run anymore from under us.
247 			 */
248 			if (PageHead(page))
249 				__put_compound_page(page);
250 			else
251 				__put_single_page(page);
252 		}
253 		return;
254 	}
255 
256 	/*
257 	 * We see the PageCompound set and PageTail set, so @page maybe:
258 	 *  1. a tail hugetlbfs page, or
259 	 *  2. a tail THP page, or
260 	 *  3. a split THP page.
261 	 *
262 	 *  Case 3 is possible, as we may race with
263 	 *  __split_huge_page_refcount tearing down a THP page.
264 	 */
265 	page_head = compound_head(page);
266 	if (!__compound_tail_refcounted(page_head))
267 		put_unrefcounted_compound_page(page_head, page);
268 	else
269 		put_refcounted_compound_page(page_head, page);
270 }
271 
put_page(struct page * page)272 void put_page(struct page *page)
273 {
274 	if (unlikely(PageCompound(page)))
275 		put_compound_page(page);
276 	else if (put_page_testzero(page))
277 		__put_single_page(page);
278 }
279 EXPORT_SYMBOL(put_page);
280 
281 /*
282  * This function is exported but must not be called by anything other
283  * than get_page(). It implements the slow path of get_page().
284  */
__get_page_tail(struct page * page)285 bool __get_page_tail(struct page *page)
286 {
287 	/*
288 	 * This takes care of get_page() if run on a tail page
289 	 * returned by one of the get_user_pages/follow_page variants.
290 	 * get_user_pages/follow_page itself doesn't need the compound
291 	 * lock because it runs __get_page_tail_foll() under the
292 	 * proper PT lock that already serializes against
293 	 * split_huge_page().
294 	 */
295 	unsigned long flags;
296 	bool got;
297 	struct page *page_head = compound_head(page);
298 
299 	/* Ref to put_compound_page() comment. */
300 	if (!__compound_tail_refcounted(page_head)) {
301 		smp_rmb();
302 		if (likely(PageTail(page))) {
303 			/*
304 			 * This is a hugetlbfs page or a slab
305 			 * page. __split_huge_page_refcount
306 			 * cannot race here.
307 			 */
308 			VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
309 			__get_page_tail_foll(page, true);
310 			return true;
311 		} else {
312 			/*
313 			 * __split_huge_page_refcount run
314 			 * before us, "page" was a THP
315 			 * tail. The split page_head has been
316 			 * freed and reallocated as slab or
317 			 * hugetlbfs page of smaller order
318 			 * (only possible if reallocated as
319 			 * slab on x86).
320 			 */
321 			return false;
322 		}
323 	}
324 
325 	got = false;
326 	if (likely(page != page_head && get_page_unless_zero(page_head))) {
327 		/*
328 		 * page_head wasn't a dangling pointer but it
329 		 * may not be a head page anymore by the time
330 		 * we obtain the lock. That is ok as long as it
331 		 * can't be freed from under us.
332 		 */
333 		flags = compound_lock_irqsave(page_head);
334 		/* here __split_huge_page_refcount won't run anymore */
335 		if (likely(PageTail(page))) {
336 			__get_page_tail_foll(page, false);
337 			got = true;
338 		}
339 		compound_unlock_irqrestore(page_head, flags);
340 		if (unlikely(!got))
341 			put_page(page_head);
342 	}
343 	return got;
344 }
345 EXPORT_SYMBOL(__get_page_tail);
346 
347 /**
348  * put_pages_list() - release a list of pages
349  * @pages: list of pages threaded on page->lru
350  *
351  * Release a list of pages which are strung together on page.lru.  Currently
352  * used by read_cache_pages() and related error recovery code.
353  */
put_pages_list(struct list_head * pages)354 void put_pages_list(struct list_head *pages)
355 {
356 	while (!list_empty(pages)) {
357 		struct page *victim;
358 
359 		victim = list_entry(pages->prev, struct page, lru);
360 		list_del(&victim->lru);
361 		page_cache_release(victim);
362 	}
363 }
364 EXPORT_SYMBOL(put_pages_list);
365 
366 /*
367  * get_kernel_pages() - pin kernel pages in memory
368  * @kiov:	An array of struct kvec structures
369  * @nr_segs:	number of segments to pin
370  * @write:	pinning for read/write, currently ignored
371  * @pages:	array that receives pointers to the pages pinned.
372  *		Should be at least nr_segs long.
373  *
374  * Returns number of pages pinned. This may be fewer than the number
375  * requested. If nr_pages is 0 or negative, returns 0. If no pages
376  * were pinned, returns -errno. Each page returned must be released
377  * with a put_page() call when it is finished with.
378  */
get_kernel_pages(const struct kvec * kiov,int nr_segs,int write,struct page ** pages)379 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
380 		struct page **pages)
381 {
382 	int seg;
383 
384 	for (seg = 0; seg < nr_segs; seg++) {
385 		if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
386 			return seg;
387 
388 		pages[seg] = kmap_to_page(kiov[seg].iov_base);
389 		page_cache_get(pages[seg]);
390 	}
391 
392 	return seg;
393 }
394 EXPORT_SYMBOL_GPL(get_kernel_pages);
395 
396 /*
397  * get_kernel_page() - pin a kernel page in memory
398  * @start:	starting kernel address
399  * @write:	pinning for read/write, currently ignored
400  * @pages:	array that receives pointer to the page pinned.
401  *		Must be at least nr_segs long.
402  *
403  * Returns 1 if page is pinned. If the page was not pinned, returns
404  * -errno. The page returned must be released with a put_page() call
405  * when it is finished with.
406  */
get_kernel_page(unsigned long start,int write,struct page ** pages)407 int get_kernel_page(unsigned long start, int write, struct page **pages)
408 {
409 	const struct kvec kiov = {
410 		.iov_base = (void *)start,
411 		.iov_len = PAGE_SIZE
412 	};
413 
414 	return get_kernel_pages(&kiov, 1, write, pages);
415 }
416 EXPORT_SYMBOL_GPL(get_kernel_page);
417 
pagevec_lru_move_fn(struct pagevec * pvec,void (* move_fn)(struct page * page,struct lruvec * lruvec,void * arg),void * arg)418 static void pagevec_lru_move_fn(struct pagevec *pvec,
419 	void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
420 	void *arg)
421 {
422 	int i;
423 	struct zone *zone = NULL;
424 	struct lruvec *lruvec;
425 	unsigned long flags = 0;
426 
427 	for (i = 0; i < pagevec_count(pvec); i++) {
428 		struct page *page = pvec->pages[i];
429 		struct zone *pagezone = page_zone(page);
430 
431 		if (pagezone != zone) {
432 			if (zone)
433 				spin_unlock_irqrestore(&zone->lru_lock, flags);
434 			zone = pagezone;
435 			spin_lock_irqsave(&zone->lru_lock, flags);
436 		}
437 
438 		lruvec = mem_cgroup_page_lruvec(page, zone);
439 		(*move_fn)(page, lruvec, arg);
440 	}
441 	if (zone)
442 		spin_unlock_irqrestore(&zone->lru_lock, flags);
443 	release_pages(pvec->pages, pvec->nr, pvec->cold);
444 	pagevec_reinit(pvec);
445 }
446 
pagevec_move_tail_fn(struct page * page,struct lruvec * lruvec,void * arg)447 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
448 				 void *arg)
449 {
450 	int *pgmoved = arg;
451 
452 	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
453 		enum lru_list lru = page_lru_base_type(page);
454 		list_move_tail(&page->lru, &lruvec->lists[lru]);
455 		(*pgmoved)++;
456 	}
457 }
458 
459 /*
460  * pagevec_move_tail() must be called with IRQ disabled.
461  * Otherwise this may cause nasty races.
462  */
pagevec_move_tail(struct pagevec * pvec)463 static void pagevec_move_tail(struct pagevec *pvec)
464 {
465 	int pgmoved = 0;
466 
467 	pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
468 	__count_vm_events(PGROTATED, pgmoved);
469 }
470 
471 /*
472  * Writeback is about to end against a page which has been marked for immediate
473  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
474  * inactive list.
475  */
rotate_reclaimable_page(struct page * page)476 void rotate_reclaimable_page(struct page *page)
477 {
478 	if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
479 	    !PageUnevictable(page) && PageLRU(page)) {
480 		struct pagevec *pvec;
481 		unsigned long flags;
482 
483 		page_cache_get(page);
484 		local_irq_save(flags);
485 		pvec = this_cpu_ptr(&lru_rotate_pvecs);
486 		if (!pagevec_add(pvec, page))
487 			pagevec_move_tail(pvec);
488 		local_irq_restore(flags);
489 	}
490 }
491 
update_page_reclaim_stat(struct lruvec * lruvec,int file,int rotated)492 static void update_page_reclaim_stat(struct lruvec *lruvec,
493 				     int file, int rotated)
494 {
495 	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
496 
497 	reclaim_stat->recent_scanned[file]++;
498 	if (rotated)
499 		reclaim_stat->recent_rotated[file]++;
500 }
501 
__activate_page(struct page * page,struct lruvec * lruvec,void * arg)502 static void __activate_page(struct page *page, struct lruvec *lruvec,
503 			    void *arg)
504 {
505 	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
506 		int file = page_is_file_cache(page);
507 		int lru = page_lru_base_type(page);
508 
509 		del_page_from_lru_list(page, lruvec, lru);
510 		SetPageActive(page);
511 		lru += LRU_ACTIVE;
512 		add_page_to_lru_list(page, lruvec, lru);
513 		trace_mm_lru_activate(page);
514 
515 		__count_vm_event(PGACTIVATE);
516 		update_page_reclaim_stat(lruvec, file, 1);
517 	}
518 }
519 
520 #ifdef CONFIG_SMP
521 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
522 
activate_page_drain(int cpu)523 static void activate_page_drain(int cpu)
524 {
525 	struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
526 
527 	if (pagevec_count(pvec))
528 		pagevec_lru_move_fn(pvec, __activate_page, NULL);
529 }
530 
need_activate_page_drain(int cpu)531 static bool need_activate_page_drain(int cpu)
532 {
533 	return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
534 }
535 
activate_page(struct page * page)536 void activate_page(struct page *page)
537 {
538 	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
539 		struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
540 
541 		page_cache_get(page);
542 		if (!pagevec_add(pvec, page))
543 			pagevec_lru_move_fn(pvec, __activate_page, NULL);
544 		put_cpu_var(activate_page_pvecs);
545 	}
546 }
547 
548 #else
activate_page_drain(int cpu)549 static inline void activate_page_drain(int cpu)
550 {
551 }
552 
need_activate_page_drain(int cpu)553 static bool need_activate_page_drain(int cpu)
554 {
555 	return false;
556 }
557 
activate_page(struct page * page)558 void activate_page(struct page *page)
559 {
560 	struct zone *zone = page_zone(page);
561 
562 	spin_lock_irq(&zone->lru_lock);
563 	__activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
564 	spin_unlock_irq(&zone->lru_lock);
565 }
566 #endif
567 
__lru_cache_activate_page(struct page * page)568 static void __lru_cache_activate_page(struct page *page)
569 {
570 	struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
571 	int i;
572 
573 	/*
574 	 * Search backwards on the optimistic assumption that the page being
575 	 * activated has just been added to this pagevec. Note that only
576 	 * the local pagevec is examined as a !PageLRU page could be in the
577 	 * process of being released, reclaimed, migrated or on a remote
578 	 * pagevec that is currently being drained. Furthermore, marking
579 	 * a remote pagevec's page PageActive potentially hits a race where
580 	 * a page is marked PageActive just after it is added to the inactive
581 	 * list causing accounting errors and BUG_ON checks to trigger.
582 	 */
583 	for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
584 		struct page *pagevec_page = pvec->pages[i];
585 
586 		if (pagevec_page == page) {
587 			SetPageActive(page);
588 			break;
589 		}
590 	}
591 
592 	put_cpu_var(lru_add_pvec);
593 }
594 
595 /*
596  * Mark a page as having seen activity.
597  *
598  * inactive,unreferenced	->	inactive,referenced
599  * inactive,referenced		->	active,unreferenced
600  * active,unreferenced		->	active,referenced
601  *
602  * When a newly allocated page is not yet visible, so safe for non-atomic ops,
603  * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
604  */
mark_page_accessed(struct page * page)605 void mark_page_accessed(struct page *page)
606 {
607 	if (!PageActive(page) && !PageUnevictable(page) &&
608 			PageReferenced(page)) {
609 
610 		/*
611 		 * If the page is on the LRU, queue it for activation via
612 		 * activate_page_pvecs. Otherwise, assume the page is on a
613 		 * pagevec, mark it active and it'll be moved to the active
614 		 * LRU on the next drain.
615 		 */
616 		if (PageLRU(page))
617 			activate_page(page);
618 		else
619 			__lru_cache_activate_page(page);
620 		ClearPageReferenced(page);
621 		if (page_is_file_cache(page))
622 			workingset_activation(page);
623 	} else if (!PageReferenced(page)) {
624 		SetPageReferenced(page);
625 	}
626 	if (page_is_idle(page))
627 		clear_page_idle(page);
628 }
629 EXPORT_SYMBOL(mark_page_accessed);
630 
__lru_cache_add(struct page * page)631 static void __lru_cache_add(struct page *page)
632 {
633 	struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
634 
635 	page_cache_get(page);
636 	if (!pagevec_space(pvec))
637 		__pagevec_lru_add(pvec);
638 	pagevec_add(pvec, page);
639 	put_cpu_var(lru_add_pvec);
640 }
641 
642 /**
643  * lru_cache_add: add a page to the page lists
644  * @page: the page to add
645  */
lru_cache_add_anon(struct page * page)646 void lru_cache_add_anon(struct page *page)
647 {
648 	if (PageActive(page))
649 		ClearPageActive(page);
650 	__lru_cache_add(page);
651 }
652 
lru_cache_add_file(struct page * page)653 void lru_cache_add_file(struct page *page)
654 {
655 	if (PageActive(page))
656 		ClearPageActive(page);
657 	__lru_cache_add(page);
658 }
659 EXPORT_SYMBOL(lru_cache_add_file);
660 
661 /**
662  * lru_cache_add - add a page to a page list
663  * @page: the page to be added to the LRU.
664  *
665  * Queue the page for addition to the LRU via pagevec. The decision on whether
666  * to add the page to the [in]active [file|anon] list is deferred until the
667  * pagevec is drained. This gives a chance for the caller of lru_cache_add()
668  * have the page added to the active list using mark_page_accessed().
669  */
lru_cache_add(struct page * page)670 void lru_cache_add(struct page *page)
671 {
672 	VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
673 	VM_BUG_ON_PAGE(PageLRU(page), page);
674 	__lru_cache_add(page);
675 }
676 
677 /**
678  * add_page_to_unevictable_list - add a page to the unevictable list
679  * @page:  the page to be added to the unevictable list
680  *
681  * Add page directly to its zone's unevictable list.  To avoid races with
682  * tasks that might be making the page evictable, through eg. munlock,
683  * munmap or exit, while it's not on the lru, we want to add the page
684  * while it's locked or otherwise "invisible" to other tasks.  This is
685  * difficult to do when using the pagevec cache, so bypass that.
686  */
add_page_to_unevictable_list(struct page * page)687 void add_page_to_unevictable_list(struct page *page)
688 {
689 	struct zone *zone = page_zone(page);
690 	struct lruvec *lruvec;
691 
692 	spin_lock_irq(&zone->lru_lock);
693 	lruvec = mem_cgroup_page_lruvec(page, zone);
694 	ClearPageActive(page);
695 	SetPageUnevictable(page);
696 	SetPageLRU(page);
697 	add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
698 	spin_unlock_irq(&zone->lru_lock);
699 }
700 
701 /**
702  * lru_cache_add_active_or_unevictable
703  * @page:  the page to be added to LRU
704  * @vma:   vma in which page is mapped for determining reclaimability
705  *
706  * Place @page on the active or unevictable LRU list, depending on its
707  * evictability.  Note that if the page is not evictable, it goes
708  * directly back onto it's zone's unevictable list, it does NOT use a
709  * per cpu pagevec.
710  */
lru_cache_add_active_or_unevictable(struct page * page,struct vm_area_struct * vma)711 void lru_cache_add_active_or_unevictable(struct page *page,
712 					 struct vm_area_struct *vma)
713 {
714 	VM_BUG_ON_PAGE(PageLRU(page), page);
715 
716 	if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
717 		SetPageActive(page);
718 		lru_cache_add(page);
719 		return;
720 	}
721 
722 	if (!TestSetPageMlocked(page)) {
723 		/*
724 		 * We use the irq-unsafe __mod_zone_page_stat because this
725 		 * counter is not modified from interrupt context, and the pte
726 		 * lock is held(spinlock), which implies preemption disabled.
727 		 */
728 		__mod_zone_page_state(page_zone(page), NR_MLOCK,
729 				    hpage_nr_pages(page));
730 		count_vm_event(UNEVICTABLE_PGMLOCKED);
731 	}
732 	add_page_to_unevictable_list(page);
733 }
734 
735 /*
736  * If the page can not be invalidated, it is moved to the
737  * inactive list to speed up its reclaim.  It is moved to the
738  * head of the list, rather than the tail, to give the flusher
739  * threads some time to write it out, as this is much more
740  * effective than the single-page writeout from reclaim.
741  *
742  * If the page isn't page_mapped and dirty/writeback, the page
743  * could reclaim asap using PG_reclaim.
744  *
745  * 1. active, mapped page -> none
746  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
747  * 3. inactive, mapped page -> none
748  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
749  * 5. inactive, clean -> inactive, tail
750  * 6. Others -> none
751  *
752  * In 4, why it moves inactive's head, the VM expects the page would
753  * be write it out by flusher threads as this is much more effective
754  * than the single-page writeout from reclaim.
755  */
lru_deactivate_file_fn(struct page * page,struct lruvec * lruvec,void * arg)756 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
757 			      void *arg)
758 {
759 	int lru, file;
760 	bool active;
761 
762 	if (!PageLRU(page))
763 		return;
764 
765 	if (PageUnevictable(page))
766 		return;
767 
768 	/* Some processes are using the page */
769 	if (page_mapped(page))
770 		return;
771 
772 	active = PageActive(page);
773 	file = page_is_file_cache(page);
774 	lru = page_lru_base_type(page);
775 
776 	del_page_from_lru_list(page, lruvec, lru + active);
777 	ClearPageActive(page);
778 	ClearPageReferenced(page);
779 	add_page_to_lru_list(page, lruvec, lru);
780 
781 	if (PageWriteback(page) || PageDirty(page)) {
782 		/*
783 		 * PG_reclaim could be raced with end_page_writeback
784 		 * It can make readahead confusing.  But race window
785 		 * is _really_ small and  it's non-critical problem.
786 		 */
787 		SetPageReclaim(page);
788 	} else {
789 		/*
790 		 * The page's writeback ends up during pagevec
791 		 * We moves tha page into tail of inactive.
792 		 */
793 		list_move_tail(&page->lru, &lruvec->lists[lru]);
794 		__count_vm_event(PGROTATED);
795 	}
796 
797 	if (active)
798 		__count_vm_event(PGDEACTIVATE);
799 	update_page_reclaim_stat(lruvec, file, 0);
800 }
801 
802 /*
803  * Drain pages out of the cpu's pagevecs.
804  * Either "cpu" is the current CPU, and preemption has already been
805  * disabled; or "cpu" is being hot-unplugged, and is already dead.
806  */
lru_add_drain_cpu(int cpu)807 void lru_add_drain_cpu(int cpu)
808 {
809 	struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
810 
811 	if (pagevec_count(pvec))
812 		__pagevec_lru_add(pvec);
813 
814 	pvec = &per_cpu(lru_rotate_pvecs, cpu);
815 	if (pagevec_count(pvec)) {
816 		unsigned long flags;
817 
818 		/* No harm done if a racing interrupt already did this */
819 		local_irq_save(flags);
820 		pagevec_move_tail(pvec);
821 		local_irq_restore(flags);
822 	}
823 
824 	pvec = &per_cpu(lru_deactivate_file_pvecs, cpu);
825 	if (pagevec_count(pvec))
826 		pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
827 
828 	activate_page_drain(cpu);
829 }
830 
831 /**
832  * deactivate_file_page - forcefully deactivate a file page
833  * @page: page to deactivate
834  *
835  * This function hints the VM that @page is a good reclaim candidate,
836  * for example if its invalidation fails due to the page being dirty
837  * or under writeback.
838  */
deactivate_file_page(struct page * page)839 void deactivate_file_page(struct page *page)
840 {
841 	/*
842 	 * In a workload with many unevictable page such as mprotect,
843 	 * unevictable page deactivation for accelerating reclaim is pointless.
844 	 */
845 	if (PageUnevictable(page))
846 		return;
847 
848 	if (likely(get_page_unless_zero(page))) {
849 		struct pagevec *pvec = &get_cpu_var(lru_deactivate_file_pvecs);
850 
851 		if (!pagevec_add(pvec, page))
852 			pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
853 		put_cpu_var(lru_deactivate_file_pvecs);
854 	}
855 }
856 
lru_add_drain(void)857 void lru_add_drain(void)
858 {
859 	lru_add_drain_cpu(get_cpu());
860 	put_cpu();
861 }
862 
lru_add_drain_per_cpu(struct work_struct * dummy)863 static void lru_add_drain_per_cpu(struct work_struct *dummy)
864 {
865 	lru_add_drain();
866 }
867 
868 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
869 
lru_add_drain_all(void)870 void lru_add_drain_all(void)
871 {
872 	static DEFINE_MUTEX(lock);
873 	static struct cpumask has_work;
874 	int cpu;
875 
876 	mutex_lock(&lock);
877 	get_online_cpus();
878 	cpumask_clear(&has_work);
879 
880 	for_each_online_cpu(cpu) {
881 		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
882 
883 		if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
884 		    pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
885 		    pagevec_count(&per_cpu(lru_deactivate_file_pvecs, cpu)) ||
886 		    need_activate_page_drain(cpu)) {
887 			INIT_WORK(work, lru_add_drain_per_cpu);
888 			schedule_work_on(cpu, work);
889 			cpumask_set_cpu(cpu, &has_work);
890 		}
891 	}
892 
893 	for_each_cpu(cpu, &has_work)
894 		flush_work(&per_cpu(lru_add_drain_work, cpu));
895 
896 	put_online_cpus();
897 	mutex_unlock(&lock);
898 }
899 
900 /**
901  * release_pages - batched page_cache_release()
902  * @pages: array of pages to release
903  * @nr: number of pages
904  * @cold: whether the pages are cache cold
905  *
906  * Decrement the reference count on all the pages in @pages.  If it
907  * fell to zero, remove the page from the LRU and free it.
908  */
release_pages(struct page ** pages,int nr,bool cold)909 void release_pages(struct page **pages, int nr, bool cold)
910 {
911 	int i;
912 	LIST_HEAD(pages_to_free);
913 	struct zone *zone = NULL;
914 	struct lruvec *lruvec;
915 	unsigned long uninitialized_var(flags);
916 	unsigned int uninitialized_var(lock_batch);
917 
918 	for (i = 0; i < nr; i++) {
919 		struct page *page = pages[i];
920 
921 		if (unlikely(PageCompound(page))) {
922 			if (zone) {
923 				spin_unlock_irqrestore(&zone->lru_lock, flags);
924 				zone = NULL;
925 			}
926 			put_compound_page(page);
927 			continue;
928 		}
929 
930 		/*
931 		 * Make sure the IRQ-safe lock-holding time does not get
932 		 * excessive with a continuous string of pages from the
933 		 * same zone. The lock is held only if zone != NULL.
934 		 */
935 		if (zone && ++lock_batch == SWAP_CLUSTER_MAX) {
936 			spin_unlock_irqrestore(&zone->lru_lock, flags);
937 			zone = NULL;
938 		}
939 
940 		if (!put_page_testzero(page))
941 			continue;
942 
943 		if (PageLRU(page)) {
944 			struct zone *pagezone = page_zone(page);
945 
946 			if (pagezone != zone) {
947 				if (zone)
948 					spin_unlock_irqrestore(&zone->lru_lock,
949 									flags);
950 				lock_batch = 0;
951 				zone = pagezone;
952 				spin_lock_irqsave(&zone->lru_lock, flags);
953 			}
954 
955 			lruvec = mem_cgroup_page_lruvec(page, zone);
956 			VM_BUG_ON_PAGE(!PageLRU(page), page);
957 			__ClearPageLRU(page);
958 			del_page_from_lru_list(page, lruvec, page_off_lru(page));
959 		}
960 
961 		/* Clear Active bit in case of parallel mark_page_accessed */
962 		__ClearPageActive(page);
963 
964 		list_add(&page->lru, &pages_to_free);
965 	}
966 	if (zone)
967 		spin_unlock_irqrestore(&zone->lru_lock, flags);
968 
969 	mem_cgroup_uncharge_list(&pages_to_free);
970 	free_hot_cold_page_list(&pages_to_free, cold);
971 }
972 EXPORT_SYMBOL(release_pages);
973 
974 /*
975  * The pages which we're about to release may be in the deferred lru-addition
976  * queues.  That would prevent them from really being freed right now.  That's
977  * OK from a correctness point of view but is inefficient - those pages may be
978  * cache-warm and we want to give them back to the page allocator ASAP.
979  *
980  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
981  * and __pagevec_lru_add_active() call release_pages() directly to avoid
982  * mutual recursion.
983  */
__pagevec_release(struct pagevec * pvec)984 void __pagevec_release(struct pagevec *pvec)
985 {
986 	lru_add_drain();
987 	release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
988 	pagevec_reinit(pvec);
989 }
990 EXPORT_SYMBOL(__pagevec_release);
991 
992 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
993 /* used by __split_huge_page_refcount() */
lru_add_page_tail(struct page * page,struct page * page_tail,struct lruvec * lruvec,struct list_head * list)994 void lru_add_page_tail(struct page *page, struct page *page_tail,
995 		       struct lruvec *lruvec, struct list_head *list)
996 {
997 	const int file = 0;
998 
999 	VM_BUG_ON_PAGE(!PageHead(page), page);
1000 	VM_BUG_ON_PAGE(PageCompound(page_tail), page);
1001 	VM_BUG_ON_PAGE(PageLRU(page_tail), page);
1002 	VM_BUG_ON(NR_CPUS != 1 &&
1003 		  !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
1004 
1005 	if (!list)
1006 		SetPageLRU(page_tail);
1007 
1008 	if (likely(PageLRU(page)))
1009 		list_add_tail(&page_tail->lru, &page->lru);
1010 	else if (list) {
1011 		/* page reclaim is reclaiming a huge page */
1012 		get_page(page_tail);
1013 		list_add_tail(&page_tail->lru, list);
1014 	} else {
1015 		struct list_head *list_head;
1016 		/*
1017 		 * Head page has not yet been counted, as an hpage,
1018 		 * so we must account for each subpage individually.
1019 		 *
1020 		 * Use the standard add function to put page_tail on the list,
1021 		 * but then correct its position so they all end up in order.
1022 		 */
1023 		add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
1024 		list_head = page_tail->lru.prev;
1025 		list_move_tail(&page_tail->lru, list_head);
1026 	}
1027 
1028 	if (!PageUnevictable(page))
1029 		update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
1030 }
1031 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1032 
__pagevec_lru_add_fn(struct page * page,struct lruvec * lruvec,void * arg)1033 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
1034 				 void *arg)
1035 {
1036 	int file = page_is_file_cache(page);
1037 	int active = PageActive(page);
1038 	enum lru_list lru = page_lru(page);
1039 
1040 	VM_BUG_ON_PAGE(PageLRU(page), page);
1041 
1042 	SetPageLRU(page);
1043 	add_page_to_lru_list(page, lruvec, lru);
1044 	update_page_reclaim_stat(lruvec, file, active);
1045 	trace_mm_lru_insertion(page, lru);
1046 }
1047 
1048 /*
1049  * Add the passed pages to the LRU, then drop the caller's refcount
1050  * on them.  Reinitialises the caller's pagevec.
1051  */
__pagevec_lru_add(struct pagevec * pvec)1052 void __pagevec_lru_add(struct pagevec *pvec)
1053 {
1054 	pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1055 }
1056 EXPORT_SYMBOL(__pagevec_lru_add);
1057 
1058 /**
1059  * pagevec_lookup_entries - gang pagecache lookup
1060  * @pvec:	Where the resulting entries are placed
1061  * @mapping:	The address_space to search
1062  * @start:	The starting entry index
1063  * @nr_entries:	The maximum number of entries
1064  * @indices:	The cache indices corresponding to the entries in @pvec
1065  *
1066  * pagevec_lookup_entries() will search for and return a group of up
1067  * to @nr_entries pages and shadow entries in the mapping.  All
1068  * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1069  * reference against actual pages in @pvec.
1070  *
1071  * The search returns a group of mapping-contiguous entries with
1072  * ascending indexes.  There may be holes in the indices due to
1073  * not-present entries.
1074  *
1075  * pagevec_lookup_entries() returns the number of entries which were
1076  * found.
1077  */
pagevec_lookup_entries(struct pagevec * pvec,struct address_space * mapping,pgoff_t start,unsigned nr_pages,pgoff_t * indices)1078 unsigned pagevec_lookup_entries(struct pagevec *pvec,
1079 				struct address_space *mapping,
1080 				pgoff_t start, unsigned nr_pages,
1081 				pgoff_t *indices)
1082 {
1083 	pvec->nr = find_get_entries(mapping, start, nr_pages,
1084 				    pvec->pages, indices);
1085 	return pagevec_count(pvec);
1086 }
1087 
1088 /**
1089  * pagevec_remove_exceptionals - pagevec exceptionals pruning
1090  * @pvec:	The pagevec to prune
1091  *
1092  * pagevec_lookup_entries() fills both pages and exceptional radix
1093  * tree entries into the pagevec.  This function prunes all
1094  * exceptionals from @pvec without leaving holes, so that it can be
1095  * passed on to page-only pagevec operations.
1096  */
pagevec_remove_exceptionals(struct pagevec * pvec)1097 void pagevec_remove_exceptionals(struct pagevec *pvec)
1098 {
1099 	int i, j;
1100 
1101 	for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1102 		struct page *page = pvec->pages[i];
1103 		if (!radix_tree_exceptional_entry(page))
1104 			pvec->pages[j++] = page;
1105 	}
1106 	pvec->nr = j;
1107 }
1108 
1109 /**
1110  * pagevec_lookup - gang pagecache lookup
1111  * @pvec:	Where the resulting pages are placed
1112  * @mapping:	The address_space to search
1113  * @start:	The starting page index
1114  * @nr_pages:	The maximum number of pages
1115  *
1116  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
1117  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
1118  * reference against the pages in @pvec.
1119  *
1120  * The search returns a group of mapping-contiguous pages with ascending
1121  * indexes.  There may be holes in the indices due to not-present pages.
1122  *
1123  * pagevec_lookup() returns the number of pages which were found.
1124  */
pagevec_lookup(struct pagevec * pvec,struct address_space * mapping,pgoff_t start,unsigned nr_pages)1125 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
1126 		pgoff_t start, unsigned nr_pages)
1127 {
1128 	pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
1129 	return pagevec_count(pvec);
1130 }
1131 EXPORT_SYMBOL(pagevec_lookup);
1132 
pagevec_lookup_tag(struct pagevec * pvec,struct address_space * mapping,pgoff_t * index,int tag,unsigned nr_pages)1133 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
1134 		pgoff_t *index, int tag, unsigned nr_pages)
1135 {
1136 	pvec->nr = find_get_pages_tag(mapping, index, tag,
1137 					nr_pages, pvec->pages);
1138 	return pagevec_count(pvec);
1139 }
1140 EXPORT_SYMBOL(pagevec_lookup_tag);
1141 
1142 /*
1143  * Perform any setup for the swap system
1144  */
swap_setup(void)1145 void __init swap_setup(void)
1146 {
1147 	unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
1148 #ifdef CONFIG_SWAP
1149 	int i;
1150 
1151 	for (i = 0; i < MAX_SWAPFILES; i++)
1152 		spin_lock_init(&swapper_spaces[i].tree_lock);
1153 #endif
1154 
1155 	/* Use a smaller cluster for small-memory machines */
1156 	if (megs < 16)
1157 		page_cluster = 2;
1158 	else
1159 		page_cluster = 3;
1160 	/*
1161 	 * Right now other parts of the system means that we
1162 	 * _really_ don't want to cluster much more
1163 	 */
1164 }
1165