blob: 21d15f99805c96517c21927817512d749017c244 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 opereation 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/module.h>
25#include <linux/mm_inline.h>
26#include <linux/buffer_head.h> /* for try_to_release_page() */
27#include <linux/module.h>
28#include <linux/percpu_counter.h>
29#include <linux/percpu.h>
30#include <linux/cpu.h>
31#include <linux/notifier.h>
32#include <linux/init.h>
33
34/* How many pages do we try to swap or page in/out together? */
35int page_cluster;
36
37#ifdef CONFIG_HUGETLB_PAGE
38
39void put_page(struct page *page)
40{
41 if (unlikely(PageCompound(page))) {
42 page = (struct page *)page->private;
43 if (put_page_testzero(page)) {
44 void (*dtor)(struct page *page);
45
46 dtor = (void (*)(struct page *))page[1].mapping;
47 (*dtor)(page);
48 }
49 return;
50 }
Nick Pigginb5810032005-10-29 18:16:12 -070051 if (put_page_testzero(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 __page_cache_release(page);
53}
54EXPORT_SYMBOL(put_page);
55#endif
56
57/*
58 * Writeback is about to end against a page which has been marked for immediate
59 * reclaim. If it still appears to be reclaimable, move it to the tail of the
60 * inactive list. The page still has PageWriteback set, which will pin it.
61 *
62 * We don't expect many pages to come through here, so don't bother batching
63 * things up.
64 *
65 * To avoid placing the page at the tail of the LRU while PG_writeback is still
66 * set, this function will clear PG_writeback before performing the page
67 * motion. Do that inside the lru lock because once PG_writeback is cleared
68 * we may not touch the page.
69 *
70 * Returns zero if it cleared PG_writeback.
71 */
72int rotate_reclaimable_page(struct page *page)
73{
74 struct zone *zone;
75 unsigned long flags;
76
77 if (PageLocked(page))
78 return 1;
79 if (PageDirty(page))
80 return 1;
81 if (PageActive(page))
82 return 1;
83 if (!PageLRU(page))
84 return 1;
85
86 zone = page_zone(page);
87 spin_lock_irqsave(&zone->lru_lock, flags);
88 if (PageLRU(page) && !PageActive(page)) {
89 list_del(&page->lru);
90 list_add_tail(&page->lru, &zone->inactive_list);
91 inc_page_state(pgrotated);
92 }
93 if (!test_clear_page_writeback(page))
94 BUG();
95 spin_unlock_irqrestore(&zone->lru_lock, flags);
96 return 0;
97}
98
99/*
100 * FIXME: speed this up?
101 */
102void fastcall activate_page(struct page *page)
103{
104 struct zone *zone = page_zone(page);
105
106 spin_lock_irq(&zone->lru_lock);
107 if (PageLRU(page) && !PageActive(page)) {
108 del_page_from_inactive_list(zone, page);
109 SetPageActive(page);
110 add_page_to_active_list(zone, page);
111 inc_page_state(pgactivate);
112 }
113 spin_unlock_irq(&zone->lru_lock);
114}
115
116/*
117 * Mark a page as having seen activity.
118 *
119 * inactive,unreferenced -> inactive,referenced
120 * inactive,referenced -> active,unreferenced
121 * active,unreferenced -> active,referenced
122 */
123void fastcall mark_page_accessed(struct page *page)
124{
125 if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
126 activate_page(page);
127 ClearPageReferenced(page);
128 } else if (!PageReferenced(page)) {
129 SetPageReferenced(page);
130 }
131}
132
133EXPORT_SYMBOL(mark_page_accessed);
134
135/**
136 * lru_cache_add: add a page to the page lists
137 * @page: the page to add
138 */
139static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
140static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
141
142void fastcall lru_cache_add(struct page *page)
143{
144 struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
145
146 page_cache_get(page);
147 if (!pagevec_add(pvec, page))
148 __pagevec_lru_add(pvec);
149 put_cpu_var(lru_add_pvecs);
150}
151
152void fastcall lru_cache_add_active(struct page *page)
153{
154 struct pagevec *pvec = &get_cpu_var(lru_add_active_pvecs);
155
156 page_cache_get(page);
157 if (!pagevec_add(pvec, page))
158 __pagevec_lru_add_active(pvec);
159 put_cpu_var(lru_add_active_pvecs);
160}
161
162void lru_add_drain(void)
163{
164 struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
165
166 if (pagevec_count(pvec))
167 __pagevec_lru_add(pvec);
168 pvec = &__get_cpu_var(lru_add_active_pvecs);
169 if (pagevec_count(pvec))
170 __pagevec_lru_add_active(pvec);
171 put_cpu_var(lru_add_pvecs);
172}
173
174/*
175 * This path almost never happens for VM activity - pages are normally
176 * freed via pagevecs. But it gets used by networking.
177 */
178void fastcall __page_cache_release(struct page *page)
179{
180 unsigned long flags;
181 struct zone *zone = page_zone(page);
182
183 spin_lock_irqsave(&zone->lru_lock, flags);
184 if (TestClearPageLRU(page))
185 del_page_from_lru(zone, page);
186 if (page_count(page) != 0)
187 page = NULL;
188 spin_unlock_irqrestore(&zone->lru_lock, flags);
189 if (page)
190 free_hot_page(page);
191}
192
193EXPORT_SYMBOL(__page_cache_release);
194
195/*
196 * Batched page_cache_release(). Decrement the reference count on all the
197 * passed pages. If it fell to zero then remove the page from the LRU and
198 * free it.
199 *
200 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
201 * for the remainder of the operation.
202 *
203 * The locking in this function is against shrink_cache(): we recheck the
204 * page count inside the lock to see whether shrink_cache grabbed the page
205 * via the LRU. If it did, give up: shrink_cache will free it.
206 */
207void release_pages(struct page **pages, int nr, int cold)
208{
209 int i;
210 struct pagevec pages_to_free;
211 struct zone *zone = NULL;
212
213 pagevec_init(&pages_to_free, cold);
214 for (i = 0; i < nr; i++) {
215 struct page *page = pages[i];
216 struct zone *pagezone;
217
Nick Pigginb5810032005-10-29 18:16:12 -0700218 if (!put_page_testzero(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 continue;
220
221 pagezone = page_zone(page);
222 if (pagezone != zone) {
223 if (zone)
224 spin_unlock_irq(&zone->lru_lock);
225 zone = pagezone;
226 spin_lock_irq(&zone->lru_lock);
227 }
228 if (TestClearPageLRU(page))
229 del_page_from_lru(zone, page);
230 if (page_count(page) == 0) {
231 if (!pagevec_add(&pages_to_free, page)) {
232 spin_unlock_irq(&zone->lru_lock);
233 __pagevec_free(&pages_to_free);
234 pagevec_reinit(&pages_to_free);
235 zone = NULL; /* No lock is held */
236 }
237 }
238 }
239 if (zone)
240 spin_unlock_irq(&zone->lru_lock);
241
242 pagevec_free(&pages_to_free);
243}
244
245/*
246 * The pages which we're about to release may be in the deferred lru-addition
247 * queues. That would prevent them from really being freed right now. That's
248 * OK from a correctness point of view but is inefficient - those pages may be
249 * cache-warm and we want to give them back to the page allocator ASAP.
250 *
251 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
252 * and __pagevec_lru_add_active() call release_pages() directly to avoid
253 * mutual recursion.
254 */
255void __pagevec_release(struct pagevec *pvec)
256{
257 lru_add_drain();
258 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
259 pagevec_reinit(pvec);
260}
261
262/*
263 * pagevec_release() for pages which are known to not be on the LRU
264 *
265 * This function reinitialises the caller's pagevec.
266 */
267void __pagevec_release_nonlru(struct pagevec *pvec)
268{
269 int i;
270 struct pagevec pages_to_free;
271
272 pagevec_init(&pages_to_free, pvec->cold);
273 pages_to_free.cold = pvec->cold;
274 for (i = 0; i < pagevec_count(pvec); i++) {
275 struct page *page = pvec->pages[i];
276
277 BUG_ON(PageLRU(page));
278 if (put_page_testzero(page))
279 pagevec_add(&pages_to_free, page);
280 }
281 pagevec_free(&pages_to_free);
282 pagevec_reinit(pvec);
283}
284
285/*
286 * Add the passed pages to the LRU, then drop the caller's refcount
287 * on them. Reinitialises the caller's pagevec.
288 */
289void __pagevec_lru_add(struct pagevec *pvec)
290{
291 int i;
292 struct zone *zone = NULL;
293
294 for (i = 0; i < pagevec_count(pvec); i++) {
295 struct page *page = pvec->pages[i];
296 struct zone *pagezone = page_zone(page);
297
298 if (pagezone != zone) {
299 if (zone)
300 spin_unlock_irq(&zone->lru_lock);
301 zone = pagezone;
302 spin_lock_irq(&zone->lru_lock);
303 }
304 if (TestSetPageLRU(page))
305 BUG();
306 add_page_to_inactive_list(zone, page);
307 }
308 if (zone)
309 spin_unlock_irq(&zone->lru_lock);
310 release_pages(pvec->pages, pvec->nr, pvec->cold);
311 pagevec_reinit(pvec);
312}
313
314EXPORT_SYMBOL(__pagevec_lru_add);
315
316void __pagevec_lru_add_active(struct pagevec *pvec)
317{
318 int i;
319 struct zone *zone = NULL;
320
321 for (i = 0; i < pagevec_count(pvec); i++) {
322 struct page *page = pvec->pages[i];
323 struct zone *pagezone = page_zone(page);
324
325 if (pagezone != zone) {
326 if (zone)
327 spin_unlock_irq(&zone->lru_lock);
328 zone = pagezone;
329 spin_lock_irq(&zone->lru_lock);
330 }
331 if (TestSetPageLRU(page))
332 BUG();
333 if (TestSetPageActive(page))
334 BUG();
335 add_page_to_active_list(zone, page);
336 }
337 if (zone)
338 spin_unlock_irq(&zone->lru_lock);
339 release_pages(pvec->pages, pvec->nr, pvec->cold);
340 pagevec_reinit(pvec);
341}
342
343/*
344 * Try to drop buffers from the pages in a pagevec
345 */
346void pagevec_strip(struct pagevec *pvec)
347{
348 int i;
349
350 for (i = 0; i < pagevec_count(pvec); i++) {
351 struct page *page = pvec->pages[i];
352
353 if (PagePrivate(page) && !TestSetPageLocked(page)) {
354 try_to_release_page(page, 0);
355 unlock_page(page);
356 }
357 }
358}
359
360/**
361 * pagevec_lookup - gang pagecache lookup
362 * @pvec: Where the resulting pages are placed
363 * @mapping: The address_space to search
364 * @start: The starting page index
365 * @nr_pages: The maximum number of pages
366 *
367 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
368 * in the mapping. The pages are placed in @pvec. pagevec_lookup() takes a
369 * reference against the pages in @pvec.
370 *
371 * The search returns a group of mapping-contiguous pages with ascending
372 * indexes. There may be holes in the indices due to not-present pages.
373 *
374 * pagevec_lookup() returns the number of pages which were found.
375 */
376unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
377 pgoff_t start, unsigned nr_pages)
378{
379 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
380 return pagevec_count(pvec);
381}
382
383unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
384 pgoff_t *index, int tag, unsigned nr_pages)
385{
386 pvec->nr = find_get_pages_tag(mapping, index, tag,
387 nr_pages, pvec->pages);
388 return pagevec_count(pvec);
389}
390
391
392#ifdef CONFIG_SMP
393/*
394 * We tolerate a little inaccuracy to avoid ping-ponging the counter between
395 * CPUs
396 */
397#define ACCT_THRESHOLD max(16, NR_CPUS * 2)
398
399static DEFINE_PER_CPU(long, committed_space) = 0;
400
401void vm_acct_memory(long pages)
402{
403 long *local;
404
405 preempt_disable();
406 local = &__get_cpu_var(committed_space);
407 *local += pages;
408 if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
409 atomic_add(*local, &vm_committed_space);
410 *local = 0;
411 }
412 preempt_enable();
413}
414EXPORT_SYMBOL(vm_acct_memory);
415
416#ifdef CONFIG_HOTPLUG_CPU
417static void lru_drain_cache(unsigned int cpu)
418{
419 struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
420
421 /* CPU is dead, so no locking needed. */
422 if (pagevec_count(pvec))
423 __pagevec_lru_add(pvec);
424 pvec = &per_cpu(lru_add_active_pvecs, cpu);
425 if (pagevec_count(pvec))
426 __pagevec_lru_add_active(pvec);
427}
428
429/* Drop the CPU's cached committed space back into the central pool. */
430static int cpu_swap_callback(struct notifier_block *nfb,
431 unsigned long action,
432 void *hcpu)
433{
434 long *committed;
435
436 committed = &per_cpu(committed_space, (long)hcpu);
437 if (action == CPU_DEAD) {
438 atomic_add(*committed, &vm_committed_space);
439 *committed = 0;
440 lru_drain_cache((long)hcpu);
441 }
442 return NOTIFY_OK;
443}
444#endif /* CONFIG_HOTPLUG_CPU */
445#endif /* CONFIG_SMP */
446
447#ifdef CONFIG_SMP
448void percpu_counter_mod(struct percpu_counter *fbc, long amount)
449{
450 long count;
451 long *pcount;
452 int cpu = get_cpu();
453
454 pcount = per_cpu_ptr(fbc->counters, cpu);
455 count = *pcount + amount;
456 if (count >= FBC_BATCH || count <= -FBC_BATCH) {
457 spin_lock(&fbc->lock);
458 fbc->count += count;
459 spin_unlock(&fbc->lock);
460 count = 0;
461 }
462 *pcount = count;
463 put_cpu();
464}
465EXPORT_SYMBOL(percpu_counter_mod);
466#endif
467
468/*
469 * Perform any setup for the swap system
470 */
471void __init swap_setup(void)
472{
473 unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
474
475 /* Use a smaller cluster for small-memory machines */
476 if (megs < 16)
477 page_cluster = 2;
478 else
479 page_cluster = 3;
480 /*
481 * Right now other parts of the system means that we
482 * _really_ don't want to cluster much more
483 */
484 hotcpu_notifier(cpu_swap_callback, 0);
485}