blob: 5db32fdfaf392ae8a8ecea01aaef8589f55d161e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
22#include <linux/file.h>
23#include <linux/writeback.h>
24#include <linux/blkdev.h>
25#include <linux/buffer_head.h> /* for try_to_release_page(),
26 buffer_heads_over_limit */
27#include <linux/mm_inline.h>
28#include <linux/pagevec.h>
29#include <linux/backing-dev.h>
30#include <linux/rmap.h>
31#include <linux/topology.h>
32#include <linux/cpu.h>
33#include <linux/cpuset.h>
34#include <linux/notifier.h>
35#include <linux/rwsem.h>
36
37#include <asm/tlbflush.h>
38#include <asm/div64.h>
39
40#include <linux/swapops.h>
41
42/* possible outcome of pageout() */
43typedef enum {
44 /* failed to write page out, page is locked */
45 PAGE_KEEP,
46 /* move page to the active list, page is locked */
47 PAGE_ACTIVATE,
48 /* page has been sent to the disk successfully, page is unlocked */
49 PAGE_SUCCESS,
50 /* page is clean and locked */
51 PAGE_CLEAN,
52} pageout_t;
53
54struct scan_control {
55 /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
56 unsigned long nr_to_scan;
57
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned;
60
61 /* Incremented by the number of pages reclaimed */
62 unsigned long nr_reclaimed;
63
64 unsigned long nr_mapped; /* From page_state */
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 /* Ask shrink_caches, or shrink_zone to scan at this priority */
67 unsigned int priority;
68
69 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040070 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070071
72 int may_writepage;
73
Christoph Lameterf1fd1062006-01-18 17:42:30 -080074 /* Can pages be swapped as part of reclaim? */
75 int may_swap;
76
Linus Torvalds1da177e2005-04-16 15:20:36 -070077 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
78 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
79 * In this context, it doesn't matter that we scan the
80 * whole list at once. */
81 int swap_cluster_max;
82};
83
84/*
85 * The list of shrinker callbacks used by to apply pressure to
86 * ageable caches.
87 */
88struct shrinker {
89 shrinker_t shrinker;
90 struct list_head list;
91 int seeks; /* seeks to recreate an obj */
92 long nr; /* objs pending delete */
93};
94
95#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
96
97#ifdef ARCH_HAS_PREFETCH
98#define prefetch_prev_lru_page(_page, _base, _field) \
99 do { \
100 if ((_page)->lru.prev != _base) { \
101 struct page *prev; \
102 \
103 prev = lru_to_page(&(_page->lru)); \
104 prefetch(&prev->_field); \
105 } \
106 } while (0)
107#else
108#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
109#endif
110
111#ifdef ARCH_HAS_PREFETCHW
112#define prefetchw_prev_lru_page(_page, _base, _field) \
113 do { \
114 if ((_page)->lru.prev != _base) { \
115 struct page *prev; \
116 \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetchw(&prev->_field); \
119 } \
120 } while (0)
121#else
122#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
123#endif
124
125/*
126 * From 0 .. 100. Higher means more swappy.
127 */
128int vm_swappiness = 60;
129static long total_memory;
130
131static LIST_HEAD(shrinker_list);
132static DECLARE_RWSEM(shrinker_rwsem);
133
134/*
135 * Add a shrinker callback to be called from the vm
136 */
137struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
138{
139 struct shrinker *shrinker;
140
141 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
142 if (shrinker) {
143 shrinker->shrinker = theshrinker;
144 shrinker->seeks = seeks;
145 shrinker->nr = 0;
146 down_write(&shrinker_rwsem);
147 list_add_tail(&shrinker->list, &shrinker_list);
148 up_write(&shrinker_rwsem);
149 }
150 return shrinker;
151}
152EXPORT_SYMBOL(set_shrinker);
153
154/*
155 * Remove one
156 */
157void remove_shrinker(struct shrinker *shrinker)
158{
159 down_write(&shrinker_rwsem);
160 list_del(&shrinker->list);
161 up_write(&shrinker_rwsem);
162 kfree(shrinker);
163}
164EXPORT_SYMBOL(remove_shrinker);
165
166#define SHRINK_BATCH 128
167/*
168 * Call the shrink functions to age shrinkable caches
169 *
170 * Here we assume it costs one seek to replace a lru page and that it also
171 * takes a seek to recreate a cache object. With this in mind we age equal
172 * percentages of the lru and ageable caches. This should balance the seeks
173 * generated by these structures.
174 *
175 * If the vm encounted mapped pages on the LRU it increase the pressure on
176 * slab to avoid swapping.
177 *
178 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
179 *
180 * `lru_pages' represents the number of on-LRU pages in all the zones which
181 * are eligible for the caller's allocation attempt. It is used for balancing
182 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700183 *
184 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 */
Andrew Morton9d0243b2006-01-08 01:00:39 -0800186int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187{
188 struct shrinker *shrinker;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700189 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190
191 if (scanned == 0)
192 scanned = SWAP_CLUSTER_MAX;
193
194 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700195 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196
197 list_for_each_entry(shrinker, &shrinker_list, list) {
198 unsigned long long delta;
199 unsigned long total_scan;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800200 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201
202 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800203 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 do_div(delta, lru_pages + 1);
205 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800206 if (shrinker->nr < 0) {
207 printk(KERN_ERR "%s: nr=%ld\n",
208 __FUNCTION__, shrinker->nr);
209 shrinker->nr = max_pass;
210 }
211
212 /*
213 * Avoid risking looping forever due to too large nr value:
214 * never try to free more than twice the estimate number of
215 * freeable entries.
216 */
217 if (shrinker->nr > max_pass * 2)
218 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219
220 total_scan = shrinker->nr;
221 shrinker->nr = 0;
222
223 while (total_scan >= SHRINK_BATCH) {
224 long this_scan = SHRINK_BATCH;
225 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700226 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700228 nr_before = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
230 if (shrink_ret == -1)
231 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700232 if (shrink_ret < nr_before)
233 ret += nr_before - shrink_ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 mod_page_state(slabs_scanned, this_scan);
235 total_scan -= this_scan;
236
237 cond_resched();
238 }
239
240 shrinker->nr += total_scan;
241 }
242 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700243 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244}
245
246/* Called without lock on whether page is mapped, so answer is unstable */
247static inline int page_mapping_inuse(struct page *page)
248{
249 struct address_space *mapping;
250
251 /* Page is in somebody's page tables. */
252 if (page_mapped(page))
253 return 1;
254
255 /* Be more reluctant to reclaim swapcache than pagecache */
256 if (PageSwapCache(page))
257 return 1;
258
259 mapping = page_mapping(page);
260 if (!mapping)
261 return 0;
262
263 /* File is mmap'd by somebody? */
264 return mapping_mapped(mapping);
265}
266
267static inline int is_page_cache_freeable(struct page *page)
268{
269 return page_count(page) - !!PagePrivate(page) == 2;
270}
271
272static int may_write_to_queue(struct backing_dev_info *bdi)
273{
Christoph Lameter930d9152006-01-08 01:00:47 -0800274 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 return 1;
276 if (!bdi_write_congested(bdi))
277 return 1;
278 if (bdi == current->backing_dev_info)
279 return 1;
280 return 0;
281}
282
283/*
284 * We detected a synchronous write error writing a page out. Probably
285 * -ENOSPC. We need to propagate that into the address_space for a subsequent
286 * fsync(), msync() or close().
287 *
288 * The tricky part is that after writepage we cannot touch the mapping: nothing
289 * prevents it from being freed up. But we have a ref on the page and once
290 * that page is locked, the mapping is pinned.
291 *
292 * We're allowed to run sleeping lock_page() here because we know the caller has
293 * __GFP_FS.
294 */
295static void handle_write_error(struct address_space *mapping,
296 struct page *page, int error)
297{
298 lock_page(page);
299 if (page_mapping(page) == mapping) {
300 if (error == -ENOSPC)
301 set_bit(AS_ENOSPC, &mapping->flags);
302 else
303 set_bit(AS_EIO, &mapping->flags);
304 }
305 unlock_page(page);
306}
307
308/*
309 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
310 */
311static pageout_t pageout(struct page *page, struct address_space *mapping)
312{
313 /*
314 * If the page is dirty, only perform writeback if that write
315 * will be non-blocking. To prevent this allocation from being
316 * stalled by pagecache activity. But note that there may be
317 * stalls if we need to run get_block(). We could test
318 * PagePrivate for that.
319 *
320 * If this process is currently in generic_file_write() against
321 * this page's queue, we can perform writeback even if that
322 * will block.
323 *
324 * If the page is swapcache, write it back even if that would
325 * block, for some throttling. This happens by accident, because
326 * swap_backing_dev_info is bust: it doesn't reflect the
327 * congestion state of the swapdevs. Easy to fix, if needed.
328 * See swapfile.c:page_queue_congested().
329 */
330 if (!is_page_cache_freeable(page))
331 return PAGE_KEEP;
332 if (!mapping) {
333 /*
334 * Some data journaling orphaned pages can have
335 * page->mapping == NULL while being dirty with clean buffers.
336 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700337 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 if (try_to_free_buffers(page)) {
339 ClearPageDirty(page);
340 printk("%s: orphaned page\n", __FUNCTION__);
341 return PAGE_CLEAN;
342 }
343 }
344 return PAGE_KEEP;
345 }
346 if (mapping->a_ops->writepage == NULL)
347 return PAGE_ACTIVATE;
348 if (!may_write_to_queue(mapping->backing_dev_info))
349 return PAGE_KEEP;
350
351 if (clear_page_dirty_for_io(page)) {
352 int res;
353 struct writeback_control wbc = {
354 .sync_mode = WB_SYNC_NONE,
355 .nr_to_write = SWAP_CLUSTER_MAX,
356 .nonblocking = 1,
357 .for_reclaim = 1,
358 };
359
360 SetPageReclaim(page);
361 res = mapping->a_ops->writepage(page, &wbc);
362 if (res < 0)
363 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800364 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 ClearPageReclaim(page);
366 return PAGE_ACTIVATE;
367 }
368 if (!PageWriteback(page)) {
369 /* synchronous write or broken a_ops? */
370 ClearPageReclaim(page);
371 }
372
373 return PAGE_SUCCESS;
374 }
375
376 return PAGE_CLEAN;
377}
378
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800379static int remove_mapping(struct address_space *mapping, struct page *page)
380{
381 if (!mapping)
382 return 0; /* truncate got there first */
383
384 write_lock_irq(&mapping->tree_lock);
385
386 /*
387 * The non-racy check for busy page. It is critical to check
388 * PageDirty _after_ making sure that the page is freeable and
389 * not in use by anybody. (pagecache + us == 2)
390 */
391 if (unlikely(page_count(page) != 2))
392 goto cannot_free;
393 smp_rmb();
394 if (unlikely(PageDirty(page)))
395 goto cannot_free;
396
397 if (PageSwapCache(page)) {
398 swp_entry_t swap = { .val = page_private(page) };
399 __delete_from_swap_cache(page);
400 write_unlock_irq(&mapping->tree_lock);
401 swap_free(swap);
402 __put_page(page); /* The pagecache ref */
403 return 1;
404 }
405
406 __remove_from_page_cache(page);
407 write_unlock_irq(&mapping->tree_lock);
408 __put_page(page);
409 return 1;
410
411cannot_free:
412 write_unlock_irq(&mapping->tree_lock);
413 return 0;
414}
415
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416/*
417 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
418 */
419static int shrink_list(struct list_head *page_list, struct scan_control *sc)
420{
421 LIST_HEAD(ret_pages);
422 struct pagevec freed_pvec;
423 int pgactivate = 0;
424 int reclaimed = 0;
425
426 cond_resched();
427
428 pagevec_init(&freed_pvec, 1);
429 while (!list_empty(page_list)) {
430 struct address_space *mapping;
431 struct page *page;
432 int may_enter_fs;
433 int referenced;
434
435 cond_resched();
436
437 page = lru_to_page(page_list);
438 list_del(&page->lru);
439
440 if (TestSetPageLocked(page))
441 goto keep;
442
443 BUG_ON(PageActive(page));
444
445 sc->nr_scanned++;
446 /* Double the slab pressure for mapped and swapcache pages */
447 if (page_mapped(page) || PageSwapCache(page))
448 sc->nr_scanned++;
449
450 if (PageWriteback(page))
451 goto keep_locked;
452
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800453 referenced = page_referenced(page, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 /* In active use or really unfreeable? Activate it. */
455 if (referenced && page_mapping_inuse(page))
456 goto activate_locked;
457
458#ifdef CONFIG_SWAP
459 /*
460 * Anonymous process memory has backing store?
461 * Try to allocate it some swap space here.
462 */
Lee Schermerhornc3400102005-10-29 18:15:51 -0700463 if (PageAnon(page) && !PageSwapCache(page)) {
Christoph Lameterf1fd1062006-01-18 17:42:30 -0800464 if (!sc->may_swap)
465 goto keep_locked;
Christoph Lameter1480a542006-01-08 01:00:53 -0800466 if (!add_to_swap(page, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 goto activate_locked;
468 }
469#endif /* CONFIG_SWAP */
470
471 mapping = page_mapping(page);
472 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
473 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
474
475 /*
476 * The page is mapped into the page tables of one or more
477 * processes. Try to unmap it here.
478 */
479 if (page_mapped(page) && mapping) {
Christoph Lameteraa3f18b2006-02-01 03:05:32 -0800480 /*
481 * No unmapping if we do not swap
482 */
483 if (!sc->may_swap)
484 goto keep_locked;
485
Christoph Lametera48d07a2006-02-01 03:05:38 -0800486 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 case SWAP_FAIL:
488 goto activate_locked;
489 case SWAP_AGAIN:
490 goto keep_locked;
491 case SWAP_SUCCESS:
492 ; /* try to free the page below */
493 }
494 }
495
496 if (PageDirty(page)) {
497 if (referenced)
498 goto keep_locked;
499 if (!may_enter_fs)
500 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800501 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 goto keep_locked;
503
504 /* Page is dirty, try to write it out here */
505 switch(pageout(page, mapping)) {
506 case PAGE_KEEP:
507 goto keep_locked;
508 case PAGE_ACTIVATE:
509 goto activate_locked;
510 case PAGE_SUCCESS:
511 if (PageWriteback(page) || PageDirty(page))
512 goto keep;
513 /*
514 * A synchronous write - probably a ramdisk. Go
515 * ahead and try to reclaim the page.
516 */
517 if (TestSetPageLocked(page))
518 goto keep;
519 if (PageDirty(page) || PageWriteback(page))
520 goto keep_locked;
521 mapping = page_mapping(page);
522 case PAGE_CLEAN:
523 ; /* try to free the page below */
524 }
525 }
526
527 /*
528 * If the page has buffers, try to free the buffer mappings
529 * associated with this page. If we succeed we try to free
530 * the page as well.
531 *
532 * We do this even if the page is PageDirty().
533 * try_to_release_page() does not perform I/O, but it is
534 * possible for a page to have PageDirty set, but it is actually
535 * clean (all its buffers are clean). This happens if the
536 * buffers were written out directly, with submit_bh(). ext3
537 * will do this, as well as the blockdev mapping.
538 * try_to_release_page() will discover that cleanness and will
539 * drop the buffers and mark the page clean - it can be freed.
540 *
541 * Rarely, pages can have buffers and no ->mapping. These are
542 * the pages which were not successfully invalidated in
543 * truncate_complete_page(). We try to drop those buffers here
544 * and if that worked, and the page is no longer mapped into
545 * process address space (page_count == 1) it can be freed.
546 * Otherwise, leave the page on the LRU so it is swappable.
547 */
548 if (PagePrivate(page)) {
549 if (!try_to_release_page(page, sc->gfp_mask))
550 goto activate_locked;
551 if (!mapping && page_count(page) == 1)
552 goto free_it;
553 }
554
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800555 if (!remove_mapping(mapping, page))
556 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
558free_it:
559 unlock_page(page);
560 reclaimed++;
561 if (!pagevec_add(&freed_pvec, page))
562 __pagevec_release_nonlru(&freed_pvec);
563 continue;
564
565activate_locked:
566 SetPageActive(page);
567 pgactivate++;
568keep_locked:
569 unlock_page(page);
570keep:
571 list_add(&page->lru, &ret_pages);
572 BUG_ON(PageLRU(page));
573 }
574 list_splice(&ret_pages, page_list);
575 if (pagevec_count(&freed_pvec))
576 __pagevec_release_nonlru(&freed_pvec);
577 mod_page_state(pgactivate, pgactivate);
578 sc->nr_reclaimed += reclaimed;
579 return reclaimed;
580}
581
Christoph Lameter7cbe34c2006-01-08 01:00:49 -0800582#ifdef CONFIG_MIGRATION
Christoph Lameter8419c312006-01-08 01:00:52 -0800583static inline void move_to_lru(struct page *page)
584{
585 list_del(&page->lru);
586 if (PageActive(page)) {
587 /*
588 * lru_cache_add_active checks that
589 * the PG_active bit is off.
590 */
591 ClearPageActive(page);
592 lru_cache_add_active(page);
593 } else {
594 lru_cache_add(page);
595 }
596 put_page(page);
597}
598
599/*
Nick Piggin053837f2006-01-18 17:42:27 -0800600 * Add isolated pages on the list back to the LRU.
Christoph Lameter8419c312006-01-08 01:00:52 -0800601 *
602 * returns the number of pages put back.
603 */
604int putback_lru_pages(struct list_head *l)
605{
606 struct page *page;
607 struct page *page2;
608 int count = 0;
609
610 list_for_each_entry_safe(page, page2, l, lru) {
611 move_to_lru(page);
612 count++;
613 }
614 return count;
615}
616
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617/*
Christoph Lametere965f962006-02-01 03:05:41 -0800618 * Non migratable page
619 */
620int fail_migrate_page(struct page *newpage, struct page *page)
621{
622 return -EIO;
623}
624EXPORT_SYMBOL(fail_migrate_page);
625
626/*
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800627 * swapout a single page
628 * page is locked upon entry, unlocked on exit
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800629 */
630static int swap_page(struct page *page)
631{
632 struct address_space *mapping = page_mapping(page);
633
634 if (page_mapped(page) && mapping)
Christoph Lameter418aade2006-02-10 01:51:15 -0800635 if (try_to_unmap(page, 1) != SWAP_SUCCESS)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800636 goto unlock_retry;
637
638 if (PageDirty(page)) {
639 /* Page is dirty, try to write it out here */
640 switch(pageout(page, mapping)) {
641 case PAGE_KEEP:
642 case PAGE_ACTIVATE:
643 goto unlock_retry;
644
645 case PAGE_SUCCESS:
646 goto retry;
647
648 case PAGE_CLEAN:
649 ; /* try to free the page below */
650 }
651 }
652
653 if (PagePrivate(page)) {
654 if (!try_to_release_page(page, GFP_KERNEL) ||
655 (!mapping && page_count(page) == 1))
656 goto unlock_retry;
657 }
658
659 if (remove_mapping(mapping, page)) {
660 /* Success */
661 unlock_page(page);
662 return 0;
663 }
664
665unlock_retry:
666 unlock_page(page);
667
668retry:
Christoph Lameterd0d96322006-01-08 01:00:55 -0800669 return -EAGAIN;
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800670}
Christoph Lametere965f962006-02-01 03:05:41 -0800671EXPORT_SYMBOL(swap_page);
Christoph Lametera48d07a2006-02-01 03:05:38 -0800672
673/*
674 * Page migration was first developed in the context of the memory hotplug
675 * project. The main authors of the migration code are:
676 *
677 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
678 * Hirokazu Takahashi <taka@valinux.co.jp>
679 * Dave Hansen <haveblue@us.ibm.com>
680 * Christoph Lameter <clameter@sgi.com>
681 */
682
683/*
684 * Remove references for a page and establish the new page with the correct
685 * basic settings to be able to stop accesses to the page.
686 */
Christoph Lametere965f962006-02-01 03:05:41 -0800687int migrate_page_remove_references(struct page *newpage,
Christoph Lametera48d07a2006-02-01 03:05:38 -0800688 struct page *page, int nr_refs)
689{
690 struct address_space *mapping = page_mapping(page);
691 struct page **radix_pointer;
692
693 /*
694 * Avoid doing any of the following work if the page count
695 * indicates that the page is in use or truncate has removed
696 * the page.
697 */
698 if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
699 return 1;
700
701 /*
702 * Establish swap ptes for anonymous pages or destroy pte
703 * maps for files.
704 *
705 * In order to reestablish file backed mappings the fault handlers
706 * will take the radix tree_lock which may then be used to stop
707 * processses from accessing this page until the new page is ready.
708 *
709 * A process accessing via a swap pte (an anonymous page) will take a
710 * page_lock on the old page which will block the process until the
711 * migration attempt is complete. At that time the PageSwapCache bit
712 * will be examined. If the page was migrated then the PageSwapCache
713 * bit will be clear and the operation to retrieve the page will be
714 * retried which will find the new page in the radix tree. Then a new
715 * direct mapping may be generated based on the radix tree contents.
716 *
717 * If the page was not migrated then the PageSwapCache bit
718 * is still set and the operation may continue.
719 */
720 try_to_unmap(page, 1);
721
722 /*
723 * Give up if we were unable to remove all mappings.
724 */
725 if (page_mapcount(page))
726 return 1;
727
728 write_lock_irq(&mapping->tree_lock);
729
730 radix_pointer = (struct page **)radix_tree_lookup_slot(
731 &mapping->page_tree,
732 page_index(page));
733
734 if (!page_mapping(page) || page_count(page) != nr_refs ||
735 *radix_pointer != page) {
736 write_unlock_irq(&mapping->tree_lock);
737 return 1;
738 }
739
740 /*
741 * Now we know that no one else is looking at the page.
742 *
743 * Certain minimal information about a page must be available
744 * in order for other subsystems to properly handle the page if they
745 * find it through the radix tree update before we are finished
746 * copying the page.
747 */
748 get_page(newpage);
749 newpage->index = page->index;
750 newpage->mapping = page->mapping;
751 if (PageSwapCache(page)) {
752 SetPageSwapCache(newpage);
753 set_page_private(newpage, page_private(page));
754 }
755
756 *radix_pointer = newpage;
757 __put_page(page);
758 write_unlock_irq(&mapping->tree_lock);
759
760 return 0;
761}
Christoph Lametere965f962006-02-01 03:05:41 -0800762EXPORT_SYMBOL(migrate_page_remove_references);
Christoph Lametera48d07a2006-02-01 03:05:38 -0800763
764/*
765 * Copy the page to its new location
766 */
767void migrate_page_copy(struct page *newpage, struct page *page)
768{
769 copy_highpage(newpage, page);
770
771 if (PageError(page))
772 SetPageError(newpage);
773 if (PageReferenced(page))
774 SetPageReferenced(newpage);
775 if (PageUptodate(page))
776 SetPageUptodate(newpage);
777 if (PageActive(page))
778 SetPageActive(newpage);
779 if (PageChecked(page))
780 SetPageChecked(newpage);
781 if (PageMappedToDisk(page))
782 SetPageMappedToDisk(newpage);
783
784 if (PageDirty(page)) {
785 clear_page_dirty_for_io(page);
786 set_page_dirty(newpage);
787 }
788
789 ClearPageSwapCache(page);
790 ClearPageActive(page);
791 ClearPagePrivate(page);
792 set_page_private(page, 0);
793 page->mapping = NULL;
794
795 /*
796 * If any waiters have accumulated on the new page then
797 * wake them up.
798 */
799 if (PageWriteback(newpage))
800 end_page_writeback(newpage);
801}
Christoph Lametere965f962006-02-01 03:05:41 -0800802EXPORT_SYMBOL(migrate_page_copy);
Christoph Lametera48d07a2006-02-01 03:05:38 -0800803
804/*
805 * Common logic to directly migrate a single page suitable for
806 * pages that do not use PagePrivate.
807 *
808 * Pages are locked upon entry and exit.
809 */
810int migrate_page(struct page *newpage, struct page *page)
811{
812 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
813
814 if (migrate_page_remove_references(newpage, page, 2))
815 return -EAGAIN;
816
817 migrate_page_copy(newpage, page);
818
Christoph Lametera3351e52006-02-01 03:05:39 -0800819 /*
820 * Remove auxiliary swap entries and replace
821 * them with real ptes.
822 *
823 * Note that a real pte entry will allow processes that are not
824 * waiting on the page lock to use the new page via the page tables
825 * before the new page is unlocked.
826 */
827 remove_from_swap(newpage);
Christoph Lametera48d07a2006-02-01 03:05:38 -0800828 return 0;
829}
Christoph Lametere965f962006-02-01 03:05:41 -0800830EXPORT_SYMBOL(migrate_page);
Christoph Lametera48d07a2006-02-01 03:05:38 -0800831
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800832/*
833 * migrate_pages
834 *
835 * Two lists are passed to this function. The first list
836 * contains the pages isolated from the LRU to be migrated.
837 * The second list contains new pages that the pages isolated
838 * can be moved to. If the second list is NULL then all
839 * pages are swapped out.
840 *
841 * The function returns after 10 attempts or if no pages
Christoph Lameter418aade2006-02-10 01:51:15 -0800842 * are movable anymore because to has become empty
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800843 * or no retryable pages exist anymore.
844 *
Christoph Lameterd0d96322006-01-08 01:00:55 -0800845 * Return: Number of pages not migrated when "to" ran empty.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800846 */
Christoph Lameterd4984712006-01-08 01:00:55 -0800847int migrate_pages(struct list_head *from, struct list_head *to,
848 struct list_head *moved, struct list_head *failed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800849{
850 int retry;
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800851 int nr_failed = 0;
852 int pass = 0;
853 struct page *page;
854 struct page *page2;
855 int swapwrite = current->flags & PF_SWAPWRITE;
Christoph Lameterd0d96322006-01-08 01:00:55 -0800856 int rc;
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800857
858 if (!swapwrite)
859 current->flags |= PF_SWAPWRITE;
860
861redo:
862 retry = 0;
863
Christoph Lameterd4984712006-01-08 01:00:55 -0800864 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800865 struct page *newpage = NULL;
866 struct address_space *mapping;
867
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800868 cond_resched();
869
Christoph Lameterd0d96322006-01-08 01:00:55 -0800870 rc = 0;
871 if (page_count(page) == 1)
Christoph Lameteree274972006-01-08 01:00:54 -0800872 /* page was freed from under us. So we are done. */
Christoph Lameterd0d96322006-01-08 01:00:55 -0800873 goto next;
874
Christoph Lametera48d07a2006-02-01 03:05:38 -0800875 if (to && list_empty(to))
876 break;
877
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800878 /*
879 * Skip locked pages during the first two passes to give the
Christoph Lameter7cbe34c2006-01-08 01:00:49 -0800880 * functions holding the lock time to release the page. Later we
881 * use lock_page() to have a higher chance of acquiring the
882 * lock.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800883 */
Christoph Lameterd0d96322006-01-08 01:00:55 -0800884 rc = -EAGAIN;
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800885 if (pass > 2)
886 lock_page(page);
887 else
888 if (TestSetPageLocked(page))
Christoph Lameterd0d96322006-01-08 01:00:55 -0800889 goto next;
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800890
891 /*
892 * Only wait on writeback if we have already done a pass where
893 * we we may have triggered writeouts for lots of pages.
894 */
Christoph Lameter7cbe34c2006-01-08 01:00:49 -0800895 if (pass > 0) {
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800896 wait_on_page_writeback(page);
Christoph Lameter7cbe34c2006-01-08 01:00:49 -0800897 } else {
Christoph Lameterd0d96322006-01-08 01:00:55 -0800898 if (PageWriteback(page))
899 goto unlock_page;
Christoph Lameter7cbe34c2006-01-08 01:00:49 -0800900 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800901
Christoph Lameterd0d96322006-01-08 01:00:55 -0800902 /*
903 * Anonymous pages must have swap cache references otherwise
904 * the information contained in the page maps cannot be
905 * preserved.
906 */
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800907 if (PageAnon(page) && !PageSwapCache(page)) {
Christoph Lameter1480a542006-01-08 01:00:53 -0800908 if (!add_to_swap(page, GFP_KERNEL)) {
Christoph Lameterd0d96322006-01-08 01:00:55 -0800909 rc = -ENOMEM;
910 goto unlock_page;
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800911 }
912 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800913
Christoph Lametera48d07a2006-02-01 03:05:38 -0800914 if (!to) {
915 rc = swap_page(page);
916 goto next;
917 }
918
919 newpage = lru_to_page(to);
920 lock_page(newpage);
921
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800922 /*
Christoph Lametera48d07a2006-02-01 03:05:38 -0800923 * Pages are properly locked and writeback is complete.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800924 * Try to migrate the page.
925 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800926 mapping = page_mapping(page);
927 if (!mapping)
928 goto unlock_both;
929
Christoph Lametere965f962006-02-01 03:05:41 -0800930 if (mapping->a_ops->migratepage) {
Christoph Lameter418aade2006-02-10 01:51:15 -0800931 /*
932 * Most pages have a mapping and most filesystems
933 * should provide a migration function. Anonymous
934 * pages are part of swap space which also has its
935 * own migration function. This is the most common
936 * path for page migration.
937 */
Christoph Lametere965f962006-02-01 03:05:41 -0800938 rc = mapping->a_ops->migratepage(newpage, page);
939 goto unlock_both;
940 }
941
Christoph Lametera48d07a2006-02-01 03:05:38 -0800942 /*
Christoph Lameter418aade2006-02-10 01:51:15 -0800943 * Default handling if a filesystem does not provide
944 * a migration function. We can only migrate clean
945 * pages so try to write out any dirty pages first.
Christoph Lametera48d07a2006-02-01 03:05:38 -0800946 */
947 if (PageDirty(page)) {
948 switch (pageout(page, mapping)) {
949 case PAGE_KEEP:
950 case PAGE_ACTIVATE:
951 goto unlock_both;
952
953 case PAGE_SUCCESS:
954 unlock_page(newpage);
955 goto next;
956
957 case PAGE_CLEAN:
958 ; /* try to migrate the page below */
959 }
960 }
Christoph Lameter418aade2006-02-10 01:51:15 -0800961
Christoph Lametera48d07a2006-02-01 03:05:38 -0800962 /*
Christoph Lameter418aade2006-02-10 01:51:15 -0800963 * Buffers are managed in a filesystem specific way.
964 * We must have no buffers or drop them.
Christoph Lametera48d07a2006-02-01 03:05:38 -0800965 */
966 if (!page_has_buffers(page) ||
967 try_to_release_page(page, GFP_KERNEL)) {
968 rc = migrate_page(newpage, page);
969 goto unlock_both;
970 }
971
972 /*
973 * On early passes with mapped pages simply
974 * retry. There may be a lock held for some
975 * buffers that may go away. Later
976 * swap them out.
977 */
978 if (pass > 4) {
Christoph Lameter418aade2006-02-10 01:51:15 -0800979 /*
980 * Persistently unable to drop buffers..... As a
981 * measure of last resort we fall back to
982 * swap_page().
983 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800984 unlock_page(newpage);
985 newpage = NULL;
986 rc = swap_page(page);
987 goto next;
988 }
989
990unlock_both:
991 unlock_page(newpage);
Christoph Lameterd0d96322006-01-08 01:00:55 -0800992
993unlock_page:
994 unlock_page(page);
995
996next:
997 if (rc == -EAGAIN) {
998 retry++;
999 } else if (rc) {
1000 /* Permanent failure */
1001 list_move(&page->lru, failed);
1002 nr_failed++;
1003 } else {
Christoph Lametera48d07a2006-02-01 03:05:38 -08001004 if (newpage) {
1005 /* Successful migration. Return page to LRU */
1006 move_to_lru(newpage);
1007 }
Christoph Lameterd4984712006-01-08 01:00:55 -08001008 list_move(&page->lru, moved);
Christoph Lameterd4984712006-01-08 01:00:55 -08001009 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001010 }
1011 if (retry && pass++ < 10)
1012 goto redo;
1013
1014 if (!swapwrite)
1015 current->flags &= ~PF_SWAPWRITE;
1016
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001017 return nr_failed + retry;
1018}
Christoph Lameter8419c312006-01-08 01:00:52 -08001019
Christoph Lameter8419c312006-01-08 01:00:52 -08001020/*
1021 * Isolate one page from the LRU lists and put it on the
Nick Piggin053837f2006-01-18 17:42:27 -08001022 * indicated list with elevated refcount.
Christoph Lameter8419c312006-01-08 01:00:52 -08001023 *
1024 * Result:
1025 * 0 = page not on LRU list
1026 * 1 = page removed from LRU list and added to the specified list.
Christoph Lameter8419c312006-01-08 01:00:52 -08001027 */
1028int isolate_lru_page(struct page *page)
1029{
Nick Piggin053837f2006-01-18 17:42:27 -08001030 int ret = 0;
Christoph Lameter8419c312006-01-08 01:00:52 -08001031
Nick Piggin053837f2006-01-18 17:42:27 -08001032 if (PageLRU(page)) {
1033 struct zone *zone = page_zone(page);
1034 spin_lock_irq(&zone->lru_lock);
1035 if (TestClearPageLRU(page)) {
1036 ret = 1;
1037 get_page(page);
1038 if (PageActive(page))
1039 del_page_from_active_list(zone, page);
1040 else
1041 del_page_from_inactive_list(zone, page);
1042 }
1043 spin_unlock_irq(&zone->lru_lock);
Christoph Lameter8419c312006-01-08 01:00:52 -08001044 }
Nick Piggin053837f2006-01-18 17:42:27 -08001045
1046 return ret;
Christoph Lameter8419c312006-01-08 01:00:52 -08001047}
Christoph Lameter7cbe34c2006-01-08 01:00:49 -08001048#endif
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001049
1050/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 * zone->lru_lock is heavily contended. Some of the functions that
1052 * shrink the lists perform better by taking out a batch of pages
1053 * and working on them outside the LRU lock.
1054 *
1055 * For pagecache intensive workloads, this function is the hottest
1056 * spot in the kernel (apart from copy_*_user functions).
1057 *
1058 * Appropriate locks must be held before calling this function.
1059 *
1060 * @nr_to_scan: The number of pages to look through on the list.
1061 * @src: The LRU list to pull pages off.
1062 * @dst: The temp list to put pages on to.
1063 * @scanned: The number of pages that were scanned.
1064 *
1065 * returns how many pages were moved onto *@dst.
1066 */
1067static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
1068 struct list_head *dst, int *scanned)
1069{
1070 int nr_taken = 0;
1071 struct page *page;
1072 int scan = 0;
1073
1074 while (scan++ < nr_to_scan && !list_empty(src)) {
1075 page = lru_to_page(src);
1076 prefetchw_prev_lru_page(page, src, flags);
1077
Nick Piggin053837f2006-01-18 17:42:27 -08001078 if (!TestClearPageLRU(page))
Christoph Lameter21eac812006-01-08 01:00:45 -08001079 BUG();
Nick Piggin053837f2006-01-18 17:42:27 -08001080 list_del(&page->lru);
1081 if (get_page_testone(page)) {
1082 /*
1083 * It is being freed elsewhere
1084 */
1085 __put_page(page);
1086 SetPageLRU(page);
1087 list_add(&page->lru, src);
1088 continue;
1089 } else {
1090 list_add(&page->lru, dst);
1091 nr_taken++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 }
1093 }
1094
1095 *scanned = scan;
1096 return nr_taken;
1097}
1098
1099/*
1100 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
1101 */
1102static void shrink_cache(struct zone *zone, struct scan_control *sc)
1103{
1104 LIST_HEAD(page_list);
1105 struct pagevec pvec;
1106 int max_scan = sc->nr_to_scan;
1107
1108 pagevec_init(&pvec, 1);
1109
1110 lru_add_drain();
1111 spin_lock_irq(&zone->lru_lock);
1112 while (max_scan > 0) {
1113 struct page *page;
1114 int nr_taken;
1115 int nr_scan;
1116 int nr_freed;
1117
1118 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
1119 &zone->inactive_list,
1120 &page_list, &nr_scan);
1121 zone->nr_inactive -= nr_taken;
1122 zone->pages_scanned += nr_scan;
1123 spin_unlock_irq(&zone->lru_lock);
1124
1125 if (nr_taken == 0)
1126 goto done;
1127
1128 max_scan -= nr_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 nr_freed = shrink_list(&page_list, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Nick Piggina74609f2006-01-06 00:11:20 -08001131 local_irq_disable();
1132 if (current_is_kswapd()) {
1133 __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
1134 __mod_page_state(kswapd_steal, nr_freed);
1135 } else
1136 __mod_page_state_zone(zone, pgscan_direct, nr_scan);
1137 __mod_page_state_zone(zone, pgsteal, nr_freed);
1138
1139 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 /*
1141 * Put back any unfreeable pages.
1142 */
1143 while (!list_empty(&page_list)) {
1144 page = lru_to_page(&page_list);
1145 if (TestSetPageLRU(page))
1146 BUG();
1147 list_del(&page->lru);
1148 if (PageActive(page))
1149 add_page_to_active_list(zone, page);
1150 else
1151 add_page_to_inactive_list(zone, page);
1152 if (!pagevec_add(&pvec, page)) {
1153 spin_unlock_irq(&zone->lru_lock);
1154 __pagevec_release(&pvec);
1155 spin_lock_irq(&zone->lru_lock);
1156 }
1157 }
1158 }
1159 spin_unlock_irq(&zone->lru_lock);
1160done:
1161 pagevec_release(&pvec);
1162}
1163
1164/*
1165 * This moves pages from the active list to the inactive list.
1166 *
1167 * We move them the other way if the page is referenced by one or more
1168 * processes, from rmap.
1169 *
1170 * If the pages are mostly unmapped, the processing is fast and it is
1171 * appropriate to hold zone->lru_lock across the whole operation. But if
1172 * the pages are mapped, the processing is slow (page_referenced()) so we
1173 * should drop zone->lru_lock around each page. It's impossible to balance
1174 * this, so instead we remove the pages from the LRU while processing them.
1175 * It is safe to rely on PG_active against the non-LRU pages in here because
1176 * nobody will play with that bit on a non-LRU page.
1177 *
1178 * The downside is that we have to touch page->_count against each page.
1179 * But we had to alter page->flags anyway.
1180 */
1181static void
1182refill_inactive_zone(struct zone *zone, struct scan_control *sc)
1183{
1184 int pgmoved;
1185 int pgdeactivate = 0;
1186 int pgscanned;
1187 int nr_pages = sc->nr_to_scan;
1188 LIST_HEAD(l_hold); /* The pages which were snipped off */
1189 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
1190 LIST_HEAD(l_active); /* Pages to go onto the active_list */
1191 struct page *page;
1192 struct pagevec pvec;
1193 int reclaim_mapped = 0;
1194 long mapped_ratio;
1195 long distress;
1196 long swap_tendency;
1197
1198 lru_add_drain();
1199 spin_lock_irq(&zone->lru_lock);
1200 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
1201 &l_hold, &pgscanned);
1202 zone->pages_scanned += pgscanned;
1203 zone->nr_active -= pgmoved;
1204 spin_unlock_irq(&zone->lru_lock);
1205
1206 /*
1207 * `distress' is a measure of how much trouble we're having reclaiming
1208 * pages. 0 -> no problems. 100 -> great trouble.
1209 */
1210 distress = 100 >> zone->prev_priority;
1211
1212 /*
1213 * The point of this algorithm is to decide when to start reclaiming
1214 * mapped memory instead of just pagecache. Work out how much memory
1215 * is mapped.
1216 */
1217 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
1218
1219 /*
1220 * Now decide how much we really want to unmap some pages. The mapped
1221 * ratio is downgraded - just because there's a lot of mapped memory
1222 * doesn't necessarily mean that page reclaim isn't succeeding.
1223 *
1224 * The distress ratio is important - we don't want to start going oom.
1225 *
1226 * A 100% value of vm_swappiness overrides this algorithm altogether.
1227 */
1228 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
1229
1230 /*
1231 * Now use this metric to decide whether to start moving mapped memory
1232 * onto the inactive list.
1233 */
1234 if (swap_tendency >= 100)
1235 reclaim_mapped = 1;
1236
1237 while (!list_empty(&l_hold)) {
1238 cond_resched();
1239 page = lru_to_page(&l_hold);
1240 list_del(&page->lru);
1241 if (page_mapped(page)) {
1242 if (!reclaim_mapped ||
1243 (total_swap_pages == 0 && PageAnon(page)) ||
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001244 page_referenced(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 list_add(&page->lru, &l_active);
1246 continue;
1247 }
1248 }
1249 list_add(&page->lru, &l_inactive);
1250 }
1251
1252 pagevec_init(&pvec, 1);
1253 pgmoved = 0;
1254 spin_lock_irq(&zone->lru_lock);
1255 while (!list_empty(&l_inactive)) {
1256 page = lru_to_page(&l_inactive);
1257 prefetchw_prev_lru_page(page, &l_inactive, flags);
1258 if (TestSetPageLRU(page))
1259 BUG();
1260 if (!TestClearPageActive(page))
1261 BUG();
1262 list_move(&page->lru, &zone->inactive_list);
1263 pgmoved++;
1264 if (!pagevec_add(&pvec, page)) {
1265 zone->nr_inactive += pgmoved;
1266 spin_unlock_irq(&zone->lru_lock);
1267 pgdeactivate += pgmoved;
1268 pgmoved = 0;
1269 if (buffer_heads_over_limit)
1270 pagevec_strip(&pvec);
1271 __pagevec_release(&pvec);
1272 spin_lock_irq(&zone->lru_lock);
1273 }
1274 }
1275 zone->nr_inactive += pgmoved;
1276 pgdeactivate += pgmoved;
1277 if (buffer_heads_over_limit) {
1278 spin_unlock_irq(&zone->lru_lock);
1279 pagevec_strip(&pvec);
1280 spin_lock_irq(&zone->lru_lock);
1281 }
1282
1283 pgmoved = 0;
1284 while (!list_empty(&l_active)) {
1285 page = lru_to_page(&l_active);
1286 prefetchw_prev_lru_page(page, &l_active, flags);
1287 if (TestSetPageLRU(page))
1288 BUG();
1289 BUG_ON(!PageActive(page));
1290 list_move(&page->lru, &zone->active_list);
1291 pgmoved++;
1292 if (!pagevec_add(&pvec, page)) {
1293 zone->nr_active += pgmoved;
1294 pgmoved = 0;
1295 spin_unlock_irq(&zone->lru_lock);
1296 __pagevec_release(&pvec);
1297 spin_lock_irq(&zone->lru_lock);
1298 }
1299 }
1300 zone->nr_active += pgmoved;
Nick Piggina74609f2006-01-06 00:11:20 -08001301 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302
Nick Piggina74609f2006-01-06 00:11:20 -08001303 __mod_page_state_zone(zone, pgrefill, pgscanned);
1304 __mod_page_state(pgdeactivate, pgdeactivate);
1305 local_irq_enable();
1306
1307 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308}
1309
1310/*
1311 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1312 */
1313static void
1314shrink_zone(struct zone *zone, struct scan_control *sc)
1315{
1316 unsigned long nr_active;
1317 unsigned long nr_inactive;
1318
Martin Hicks53e9a612005-09-03 15:54:51 -07001319 atomic_inc(&zone->reclaim_in_progress);
1320
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 /*
1322 * Add one to `nr_to_scan' just to make sure that the kernel will
1323 * slowly sift through the active list.
1324 */
1325 zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1326 nr_active = zone->nr_scan_active;
1327 if (nr_active >= sc->swap_cluster_max)
1328 zone->nr_scan_active = 0;
1329 else
1330 nr_active = 0;
1331
1332 zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1333 nr_inactive = zone->nr_scan_inactive;
1334 if (nr_inactive >= sc->swap_cluster_max)
1335 zone->nr_scan_inactive = 0;
1336 else
1337 nr_inactive = 0;
1338
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 while (nr_active || nr_inactive) {
1340 if (nr_active) {
1341 sc->nr_to_scan = min(nr_active,
1342 (unsigned long)sc->swap_cluster_max);
1343 nr_active -= sc->nr_to_scan;
1344 refill_inactive_zone(zone, sc);
1345 }
1346
1347 if (nr_inactive) {
1348 sc->nr_to_scan = min(nr_inactive,
1349 (unsigned long)sc->swap_cluster_max);
1350 nr_inactive -= sc->nr_to_scan;
1351 shrink_cache(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 }
1353 }
1354
1355 throttle_vm_writeout();
Martin Hicks53e9a612005-09-03 15:54:51 -07001356
1357 atomic_dec(&zone->reclaim_in_progress);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358}
1359
1360/*
1361 * This is the direct reclaim path, for page-allocating processes. We only
1362 * try to reclaim pages from zones which will satisfy the caller's allocation
1363 * request.
1364 *
1365 * We reclaim from a zone even if that zone is over pages_high. Because:
1366 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1367 * allocation or
1368 * b) The zones may be over pages_high but they must go *over* pages_high to
1369 * satisfy the `incremental min' zone defense algorithm.
1370 *
1371 * Returns the number of reclaimed pages.
1372 *
1373 * If a zone is deemed to be full of pinned pages then just give it a light
1374 * scan then give up on it.
1375 */
1376static void
1377shrink_caches(struct zone **zones, struct scan_control *sc)
1378{
1379 int i;
1380
1381 for (i = 0; zones[i] != NULL; i++) {
1382 struct zone *zone = zones[i];
1383
Con Kolivasf3fe6512006-01-06 00:11:15 -08001384 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 continue;
1386
Paul Jackson9bf22292005-09-06 15:18:12 -07001387 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 continue;
1389
1390 zone->temp_priority = sc->priority;
1391 if (zone->prev_priority > sc->priority)
1392 zone->prev_priority = sc->priority;
1393
1394 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1395 continue; /* Let kswapd poll it */
1396
1397 shrink_zone(zone, sc);
1398 }
1399}
1400
1401/*
1402 * This is the main entry point to direct page reclaim.
1403 *
1404 * If a full scan of the inactive list fails to free enough memory then we
1405 * are "out of memory" and something needs to be killed.
1406 *
1407 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1408 * high - the zone may be full of dirty or under-writeback pages, which this
1409 * caller can't do much about. We kick pdflush and take explicit naps in the
1410 * hope that some of these pages can be written. But if the allocating task
1411 * holds filesystem locks which prevent writeout this might not work, and the
1412 * allocation attempt will fail.
1413 */
Al Viro6daa0e22005-10-21 03:18:50 -04001414int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415{
1416 int priority;
1417 int ret = 0;
1418 int total_scanned = 0, total_reclaimed = 0;
1419 struct reclaim_state *reclaim_state = current->reclaim_state;
1420 struct scan_control sc;
1421 unsigned long lru_pages = 0;
1422 int i;
1423
1424 sc.gfp_mask = gfp_mask;
Christoph Lameter52a83632006-02-01 03:05:28 -08001425 sc.may_writepage = !laptop_mode;
Christoph Lameterf1fd1062006-01-18 17:42:30 -08001426 sc.may_swap = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427
1428 inc_page_state(allocstall);
1429
1430 for (i = 0; zones[i] != NULL; i++) {
1431 struct zone *zone = zones[i];
1432
Paul Jackson9bf22292005-09-06 15:18:12 -07001433 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434 continue;
1435
1436 zone->temp_priority = DEF_PRIORITY;
1437 lru_pages += zone->nr_active + zone->nr_inactive;
1438 }
1439
1440 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1441 sc.nr_mapped = read_page_state(nr_mapped);
1442 sc.nr_scanned = 0;
1443 sc.nr_reclaimed = 0;
1444 sc.priority = priority;
1445 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001446 if (!priority)
1447 disable_swap_token();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 shrink_caches(zones, &sc);
1449 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1450 if (reclaim_state) {
1451 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1452 reclaim_state->reclaimed_slab = 0;
1453 }
1454 total_scanned += sc.nr_scanned;
1455 total_reclaimed += sc.nr_reclaimed;
1456 if (total_reclaimed >= sc.swap_cluster_max) {
1457 ret = 1;
1458 goto out;
1459 }
1460
1461 /*
1462 * Try to write back as many pages as we just scanned. This
1463 * tends to cause slow streaming writers to write data to the
1464 * disk smoothly, at the dirtying rate, which is nice. But
1465 * that's undesirable in laptop mode, where we *want* lumpy
1466 * writeout. So in laptop mode, write out the whole world.
1467 */
1468 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001469 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 sc.may_writepage = 1;
1471 }
1472
1473 /* Take a nap, wait for some writeback to complete */
1474 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1475 blk_congestion_wait(WRITE, HZ/10);
1476 }
1477out:
1478 for (i = 0; zones[i] != 0; i++) {
1479 struct zone *zone = zones[i];
1480
Paul Jackson9bf22292005-09-06 15:18:12 -07001481 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 continue;
1483
1484 zone->prev_priority = zone->temp_priority;
1485 }
1486 return ret;
1487}
1488
1489/*
1490 * For kswapd, balance_pgdat() will work across all this node's zones until
1491 * they are all at pages_high.
1492 *
1493 * If `nr_pages' is non-zero then it is the number of pages which are to be
1494 * reclaimed, regardless of the zone occupancies. This is a software suspend
1495 * special.
1496 *
1497 * Returns the number of pages which were actually freed.
1498 *
1499 * There is special handling here for zones which are full of pinned pages.
1500 * This can happen if the pages are all mlocked, or if they are all used by
1501 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1502 * What we do is to detect the case where all pages in the zone have been
1503 * scanned twice and there has been zero successful reclaim. Mark the zone as
1504 * dead and from now on, only perform a short scan. Basically we're polling
1505 * the zone for when the problem goes away.
1506 *
1507 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1508 * zones which have free_pages > pages_high, but once a zone is found to have
1509 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1510 * of the number of free pages in the lower zones. This interoperates with
1511 * the page allocator fallback scheme to ensure that aging of pages is balanced
1512 * across the zones.
1513 */
1514static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1515{
1516 int to_free = nr_pages;
1517 int all_zones_ok;
1518 int priority;
1519 int i;
1520 int total_scanned, total_reclaimed;
1521 struct reclaim_state *reclaim_state = current->reclaim_state;
1522 struct scan_control sc;
1523
1524loop_again:
1525 total_scanned = 0;
1526 total_reclaimed = 0;
1527 sc.gfp_mask = GFP_KERNEL;
Christoph Lameter52a83632006-02-01 03:05:28 -08001528 sc.may_writepage = !laptop_mode;
Christoph Lameterf1fd1062006-01-18 17:42:30 -08001529 sc.may_swap = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 sc.nr_mapped = read_page_state(nr_mapped);
1531
1532 inc_page_state(pageoutrun);
1533
1534 for (i = 0; i < pgdat->nr_zones; i++) {
1535 struct zone *zone = pgdat->node_zones + i;
1536
1537 zone->temp_priority = DEF_PRIORITY;
1538 }
1539
1540 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1541 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1542 unsigned long lru_pages = 0;
1543
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001544 /* The swap token gets in the way of swapout... */
1545 if (!priority)
1546 disable_swap_token();
1547
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 all_zones_ok = 1;
1549
1550 if (nr_pages == 0) {
1551 /*
1552 * Scan in the highmem->dma direction for the highest
1553 * zone which needs scanning
1554 */
1555 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1556 struct zone *zone = pgdat->node_zones + i;
1557
Con Kolivasf3fe6512006-01-06 00:11:15 -08001558 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 continue;
1560
1561 if (zone->all_unreclaimable &&
1562 priority != DEF_PRIORITY)
1563 continue;
1564
1565 if (!zone_watermark_ok(zone, order,
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001566 zone->pages_high, 0, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 end_zone = i;
1568 goto scan;
1569 }
1570 }
1571 goto out;
1572 } else {
1573 end_zone = pgdat->nr_zones - 1;
1574 }
1575scan:
1576 for (i = 0; i <= end_zone; i++) {
1577 struct zone *zone = pgdat->node_zones + i;
1578
1579 lru_pages += zone->nr_active + zone->nr_inactive;
1580 }
1581
1582 /*
1583 * Now scan the zone in the dma->highmem direction, stopping
1584 * at the last zone which needs scanning.
1585 *
1586 * We do this because the page allocator works in the opposite
1587 * direction. This prevents the page allocator from allocating
1588 * pages behind kswapd's direction of progress, which would
1589 * cause too much scanning of the lower zones.
1590 */
1591 for (i = 0; i <= end_zone; i++) {
1592 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001593 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594
Con Kolivasf3fe6512006-01-06 00:11:15 -08001595 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 continue;
1597
1598 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1599 continue;
1600
1601 if (nr_pages == 0) { /* Not software suspend */
1602 if (!zone_watermark_ok(zone, order,
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001603 zone->pages_high, end_zone, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 all_zones_ok = 0;
1605 }
1606 zone->temp_priority = priority;
1607 if (zone->prev_priority > priority)
1608 zone->prev_priority = priority;
1609 sc.nr_scanned = 0;
1610 sc.nr_reclaimed = 0;
1611 sc.priority = priority;
1612 sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001613 atomic_inc(&zone->reclaim_in_progress);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 shrink_zone(zone, &sc);
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001615 atomic_dec(&zone->reclaim_in_progress);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001617 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1618 lru_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1620 total_reclaimed += sc.nr_reclaimed;
1621 total_scanned += sc.nr_scanned;
1622 if (zone->all_unreclaimable)
1623 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001624 if (nr_slab == 0 && zone->pages_scanned >=
1625 (zone->nr_active + zone->nr_inactive) * 4)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 zone->all_unreclaimable = 1;
1627 /*
1628 * If we've done a decent amount of scanning and
1629 * the reclaim ratio is low, start doing writepage
1630 * even in laptop mode
1631 */
1632 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1633 total_scanned > total_reclaimed+total_reclaimed/2)
1634 sc.may_writepage = 1;
1635 }
1636 if (nr_pages && to_free > total_reclaimed)
1637 continue; /* swsusp: need to do more work */
1638 if (all_zones_ok)
1639 break; /* kswapd: all done */
1640 /*
1641 * OK, kswapd is getting into trouble. Take a nap, then take
1642 * another pass across the zones.
1643 */
1644 if (total_scanned && priority < DEF_PRIORITY - 2)
1645 blk_congestion_wait(WRITE, HZ/10);
1646
1647 /*
1648 * We do this so kswapd doesn't build up large priorities for
1649 * example when it is freeing in parallel with allocators. It
1650 * matches the direct reclaim path behaviour in terms of impact
1651 * on zone->*_priority.
1652 */
1653 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1654 break;
1655 }
1656out:
1657 for (i = 0; i < pgdat->nr_zones; i++) {
1658 struct zone *zone = pgdat->node_zones + i;
1659
1660 zone->prev_priority = zone->temp_priority;
1661 }
1662 if (!all_zones_ok) {
1663 cond_resched();
1664 goto loop_again;
1665 }
1666
1667 return total_reclaimed;
1668}
1669
1670/*
1671 * The background pageout daemon, started as a kernel thread
1672 * from the init process.
1673 *
1674 * This basically trickles out pages so that we have _some_
1675 * free memory available even if there is no other activity
1676 * that frees anything up. This is needed for things like routing
1677 * etc, where we otherwise might have all activity going on in
1678 * asynchronous contexts that cannot page things out.
1679 *
1680 * If there are applications that are active memory-allocators
1681 * (most normal use), this basically shouldn't matter.
1682 */
1683static int kswapd(void *p)
1684{
1685 unsigned long order;
1686 pg_data_t *pgdat = (pg_data_t*)p;
1687 struct task_struct *tsk = current;
1688 DEFINE_WAIT(wait);
1689 struct reclaim_state reclaim_state = {
1690 .reclaimed_slab = 0,
1691 };
1692 cpumask_t cpumask;
1693
1694 daemonize("kswapd%d", pgdat->node_id);
1695 cpumask = node_to_cpumask(pgdat->node_id);
1696 if (!cpus_empty(cpumask))
1697 set_cpus_allowed(tsk, cpumask);
1698 current->reclaim_state = &reclaim_state;
1699
1700 /*
1701 * Tell the memory management that we're a "memory allocator",
1702 * and that if we need more memory we should get access to it
1703 * regardless (see "__alloc_pages()"). "kswapd" should
1704 * never get caught in the normal page freeing logic.
1705 *
1706 * (Kswapd normally doesn't need memory anyway, but sometimes
1707 * you need a small amount of memory in order to be able to
1708 * page out something else, and this flag essentially protects
1709 * us from recursively trying to free more memory as we're
1710 * trying to free the first piece of memory in the first place).
1711 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001712 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
1714 order = 0;
1715 for ( ; ; ) {
1716 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001717
1718 try_to_freeze();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719
1720 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1721 new_order = pgdat->kswapd_max_order;
1722 pgdat->kswapd_max_order = 0;
1723 if (order < new_order) {
1724 /*
1725 * Don't sleep if someone wants a larger 'order'
1726 * allocation
1727 */
1728 order = new_order;
1729 } else {
1730 schedule();
1731 order = pgdat->kswapd_max_order;
1732 }
1733 finish_wait(&pgdat->kswapd_wait, &wait);
1734
1735 balance_pgdat(pgdat, 0, order);
1736 }
1737 return 0;
1738}
1739
1740/*
1741 * A zone is low on free memory, so wake its kswapd task to service it.
1742 */
1743void wakeup_kswapd(struct zone *zone, int order)
1744{
1745 pg_data_t *pgdat;
1746
Con Kolivasf3fe6512006-01-06 00:11:15 -08001747 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 return;
1749
1750 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001751 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 return;
1753 if (pgdat->kswapd_max_order < order)
1754 pgdat->kswapd_max_order = order;
Paul Jackson9bf22292005-09-06 15:18:12 -07001755 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001757 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001759 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760}
1761
1762#ifdef CONFIG_PM
1763/*
1764 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1765 * pages.
1766 */
1767int shrink_all_memory(int nr_pages)
1768{
1769 pg_data_t *pgdat;
1770 int nr_to_free = nr_pages;
1771 int ret = 0;
1772 struct reclaim_state reclaim_state = {
1773 .reclaimed_slab = 0,
1774 };
1775
1776 current->reclaim_state = &reclaim_state;
1777 for_each_pgdat(pgdat) {
1778 int freed;
1779 freed = balance_pgdat(pgdat, nr_to_free, 0);
1780 ret += freed;
1781 nr_to_free -= freed;
1782 if (nr_to_free <= 0)
1783 break;
1784 }
1785 current->reclaim_state = NULL;
1786 return ret;
1787}
1788#endif
1789
1790#ifdef CONFIG_HOTPLUG_CPU
1791/* It's optimal to keep kswapds on the same CPUs as their memory, but
1792 not required for correctness. So if the last cpu in a node goes
1793 away, we get changed to run anywhere: as the first one comes back,
1794 restore their cpu bindings. */
1795static int __devinit cpu_callback(struct notifier_block *nfb,
1796 unsigned long action,
1797 void *hcpu)
1798{
1799 pg_data_t *pgdat;
1800 cpumask_t mask;
1801
1802 if (action == CPU_ONLINE) {
1803 for_each_pgdat(pgdat) {
1804 mask = node_to_cpumask(pgdat->node_id);
1805 if (any_online_cpu(mask) != NR_CPUS)
1806 /* One of our CPUs online: restore mask */
1807 set_cpus_allowed(pgdat->kswapd, mask);
1808 }
1809 }
1810 return NOTIFY_OK;
1811}
1812#endif /* CONFIG_HOTPLUG_CPU */
1813
1814static int __init kswapd_init(void)
1815{
1816 pg_data_t *pgdat;
1817 swap_setup();
1818 for_each_pgdat(pgdat)
1819 pgdat->kswapd
1820 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1821 total_memory = nr_free_pagecache_pages();
1822 hotcpu_notifier(cpu_callback, 0);
1823 return 0;
1824}
1825
1826module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001827
1828#ifdef CONFIG_NUMA
1829/*
1830 * Zone reclaim mode
1831 *
1832 * If non-zero call zone_reclaim when the number of free pages falls below
1833 * the watermarks.
1834 *
1835 * In the future we may add flags to the mode. However, the page allocator
1836 * should only have to check that zone_reclaim_mode != 0 before calling
1837 * zone_reclaim().
1838 */
1839int zone_reclaim_mode __read_mostly;
1840
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001841#define RECLAIM_OFF 0
1842#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1843#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1844#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001845#define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001846
Christoph Lameter9eeff232006-01-18 17:42:31 -08001847/*
1848 * Mininum time between zone reclaim scans
1849 */
Christoph Lameter2a11ff02006-02-01 03:05:33 -08001850int zone_reclaim_interval __read_mostly = 30*HZ;
Christoph Lametera92f7122006-02-01 03:05:32 -08001851
1852/*
1853 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1854 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1855 * a zone.
1856 */
1857#define ZONE_RECLAIM_PRIORITY 4
1858
Christoph Lameter9eeff232006-01-18 17:42:31 -08001859/*
1860 * Try to free up some pages from this zone through reclaim.
1861 */
1862int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1863{
Christoph Lameter89288622006-02-01 03:05:25 -08001864 int nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001865 struct task_struct *p = current;
1866 struct reclaim_state reclaim_state;
Christoph Lameter89288622006-02-01 03:05:25 -08001867 struct scan_control sc;
Christoph Lameter42c722d2006-02-01 03:05:26 -08001868 cpumask_t mask;
1869 int node_id;
Christoph Lameter89288622006-02-01 03:05:25 -08001870
1871 if (time_before(jiffies,
Christoph Lameter2a11ff02006-02-01 03:05:33 -08001872 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
Christoph Lameter89288622006-02-01 03:05:25 -08001873 return 0;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001874
1875 if (!(gfp_mask & __GFP_WAIT) ||
Christoph Lameter9eeff232006-01-18 17:42:31 -08001876 zone->all_unreclaimable ||
1877 atomic_read(&zone->reclaim_in_progress) > 0)
1878 return 0;
1879
Christoph Lameter42c722d2006-02-01 03:05:26 -08001880 node_id = zone->zone_pgdat->node_id;
1881 mask = node_to_cpumask(node_id);
1882 if (!cpus_empty(mask) && node_id != numa_node_id())
1883 return 0;
1884
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001885 sc.may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE);
1886 sc.may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP);
Christoph Lameter89288622006-02-01 03:05:25 -08001887 sc.nr_scanned = 0;
1888 sc.nr_reclaimed = 0;
Christoph Lametera92f7122006-02-01 03:05:32 -08001889 sc.priority = ZONE_RECLAIM_PRIORITY + 1;
Christoph Lameter89288622006-02-01 03:05:25 -08001890 sc.nr_mapped = read_page_state(nr_mapped);
1891 sc.gfp_mask = gfp_mask;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001892
1893 disable_swap_token();
1894
Christoph Lameter89288622006-02-01 03:05:25 -08001895 nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001896 if (nr_pages > SWAP_CLUSTER_MAX)
1897 sc.swap_cluster_max = nr_pages;
1898 else
1899 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1900
1901 cond_resched();
1902 p->flags |= PF_MEMALLOC;
1903 reclaim_state.reclaimed_slab = 0;
1904 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08001905
Christoph Lametera92f7122006-02-01 03:05:32 -08001906 /*
1907 * Free memory by calling shrink zone with increasing priorities
1908 * until we have enough memory freed.
1909 */
1910 do {
1911 sc.priority--;
1912 shrink_zone(zone, &sc);
1913
1914 } while (sc.nr_reclaimed < nr_pages && sc.priority > 0);
Christoph Lameterc84db232006-02-01 03:05:29 -08001915
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001916 if (sc.nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1917 /*
1918 * shrink_slab does not currently allow us to determine
1919 * how many pages were freed in the zone. So we just
1920 * shake the slab and then go offnode for a single allocation.
1921 *
1922 * shrink_slab will free memory on all zones and may take
1923 * a long time.
1924 */
1925 shrink_slab(sc.nr_scanned, gfp_mask, order);
1926 sc.nr_reclaimed = 1; /* Avoid getting the off node timeout */
1927 }
1928
Christoph Lameter9eeff232006-01-18 17:42:31 -08001929 p->reclaim_state = NULL;
1930 current->flags &= ~PF_MEMALLOC;
1931
1932 if (sc.nr_reclaimed == 0)
1933 zone->last_unsuccessful_zone_reclaim = jiffies;
1934
Christoph Lameterc84db232006-02-01 03:05:29 -08001935 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001936}
1937#endif
1938