blob: b0cd81c32de6ca0f601bb1be3c12764beb37a8af [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
66 /* How many pages shrink_cache() should reclaim */
67 int nr_to_reclaim;
68
69 /* Ask shrink_caches, or shrink_zone to scan at this priority */
70 unsigned int priority;
71
72 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040073 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75 int may_writepage;
76
Martin Hicksbfbb38f2005-06-21 17:14:40 -070077 /* Can pages be swapped as part of reclaim? */
78 int may_swap;
79
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
81 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
82 * In this context, it doesn't matter that we scan the
83 * whole list at once. */
84 int swap_cluster_max;
85};
86
87/*
88 * The list of shrinker callbacks used by to apply pressure to
89 * ageable caches.
90 */
91struct shrinker {
92 shrinker_t shrinker;
93 struct list_head list;
94 int seeks; /* seeks to recreate an obj */
95 long nr; /* objs pending delete */
96};
97
98#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99
100#ifdef ARCH_HAS_PREFETCH
101#define prefetch_prev_lru_page(_page, _base, _field) \
102 do { \
103 if ((_page)->lru.prev != _base) { \
104 struct page *prev; \
105 \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
108 } \
109 } while (0)
110#else
111#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
112#endif
113
114#ifdef ARCH_HAS_PREFETCHW
115#define prefetchw_prev_lru_page(_page, _base, _field) \
116 do { \
117 if ((_page)->lru.prev != _base) { \
118 struct page *prev; \
119 \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
122 } \
123 } while (0)
124#else
125#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
126#endif
127
128/*
129 * From 0 .. 100. Higher means more swappy.
130 */
131int vm_swappiness = 60;
132static long total_memory;
133
134static LIST_HEAD(shrinker_list);
135static DECLARE_RWSEM(shrinker_rwsem);
136
137/*
138 * Add a shrinker callback to be called from the vm
139 */
140struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
141{
142 struct shrinker *shrinker;
143
144 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
145 if (shrinker) {
146 shrinker->shrinker = theshrinker;
147 shrinker->seeks = seeks;
148 shrinker->nr = 0;
149 down_write(&shrinker_rwsem);
150 list_add_tail(&shrinker->list, &shrinker_list);
151 up_write(&shrinker_rwsem);
152 }
153 return shrinker;
154}
155EXPORT_SYMBOL(set_shrinker);
156
157/*
158 * Remove one
159 */
160void remove_shrinker(struct shrinker *shrinker)
161{
162 down_write(&shrinker_rwsem);
163 list_del(&shrinker->list);
164 up_write(&shrinker_rwsem);
165 kfree(shrinker);
166}
167EXPORT_SYMBOL(remove_shrinker);
168
169#define SHRINK_BATCH 128
170/*
171 * Call the shrink functions to age shrinkable caches
172 *
173 * Here we assume it costs one seek to replace a lru page and that it also
174 * takes a seek to recreate a cache object. With this in mind we age equal
175 * percentages of the lru and ageable caches. This should balance the seeks
176 * generated by these structures.
177 *
178 * If the vm encounted mapped pages on the LRU it increase the pressure on
179 * slab to avoid swapping.
180 *
181 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
182 *
183 * `lru_pages' represents the number of on-LRU pages in all the zones which
184 * are eligible for the caller's allocation attempt. It is used for balancing
185 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700186 *
187 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 */
Al Viro6daa0e22005-10-21 03:18:50 -0400189static int shrink_slab(unsigned long scanned, gfp_t gfp_mask,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 unsigned long lru_pages)
191{
192 struct shrinker *shrinker;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700193 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194
195 if (scanned == 0)
196 scanned = SWAP_CLUSTER_MAX;
197
198 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700199 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
201 list_for_each_entry(shrinker, &shrinker_list, list) {
202 unsigned long long delta;
203 unsigned long total_scan;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800204 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
206 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800207 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 do_div(delta, lru_pages + 1);
209 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800210 if (shrinker->nr < 0) {
211 printk(KERN_ERR "%s: nr=%ld\n",
212 __FUNCTION__, shrinker->nr);
213 shrinker->nr = max_pass;
214 }
215
216 /*
217 * Avoid risking looping forever due to too large nr value:
218 * never try to free more than twice the estimate number of
219 * freeable entries.
220 */
221 if (shrinker->nr > max_pass * 2)
222 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223
224 total_scan = shrinker->nr;
225 shrinker->nr = 0;
226
227 while (total_scan >= SHRINK_BATCH) {
228 long this_scan = SHRINK_BATCH;
229 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700230 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700232 nr_before = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
234 if (shrink_ret == -1)
235 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700236 if (shrink_ret < nr_before)
237 ret += nr_before - shrink_ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 mod_page_state(slabs_scanned, this_scan);
239 total_scan -= this_scan;
240
241 cond_resched();
242 }
243
244 shrinker->nr += total_scan;
245 }
246 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700247 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248}
249
250/* Called without lock on whether page is mapped, so answer is unstable */
251static inline int page_mapping_inuse(struct page *page)
252{
253 struct address_space *mapping;
254
255 /* Page is in somebody's page tables. */
256 if (page_mapped(page))
257 return 1;
258
259 /* Be more reluctant to reclaim swapcache than pagecache */
260 if (PageSwapCache(page))
261 return 1;
262
263 mapping = page_mapping(page);
264 if (!mapping)
265 return 0;
266
267 /* File is mmap'd by somebody? */
268 return mapping_mapped(mapping);
269}
270
271static inline int is_page_cache_freeable(struct page *page)
272{
273 return page_count(page) - !!PagePrivate(page) == 2;
274}
275
276static int may_write_to_queue(struct backing_dev_info *bdi)
277{
278 if (current_is_kswapd())
279 return 1;
280 if (current_is_pdflush()) /* This is unlikely, but why not... */
281 return 1;
282 if (!bdi_write_congested(bdi))
283 return 1;
284 if (bdi == current->backing_dev_info)
285 return 1;
286 return 0;
287}
288
289/*
290 * We detected a synchronous write error writing a page out. Probably
291 * -ENOSPC. We need to propagate that into the address_space for a subsequent
292 * fsync(), msync() or close().
293 *
294 * The tricky part is that after writepage we cannot touch the mapping: nothing
295 * prevents it from being freed up. But we have a ref on the page and once
296 * that page is locked, the mapping is pinned.
297 *
298 * We're allowed to run sleeping lock_page() here because we know the caller has
299 * __GFP_FS.
300 */
301static void handle_write_error(struct address_space *mapping,
302 struct page *page, int error)
303{
304 lock_page(page);
305 if (page_mapping(page) == mapping) {
306 if (error == -ENOSPC)
307 set_bit(AS_ENOSPC, &mapping->flags);
308 else
309 set_bit(AS_EIO, &mapping->flags);
310 }
311 unlock_page(page);
312}
313
314/*
315 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
316 */
317static pageout_t pageout(struct page *page, struct address_space *mapping)
318{
319 /*
320 * If the page is dirty, only perform writeback if that write
321 * will be non-blocking. To prevent this allocation from being
322 * stalled by pagecache activity. But note that there may be
323 * stalls if we need to run get_block(). We could test
324 * PagePrivate for that.
325 *
326 * If this process is currently in generic_file_write() against
327 * this page's queue, we can perform writeback even if that
328 * will block.
329 *
330 * If the page is swapcache, write it back even if that would
331 * block, for some throttling. This happens by accident, because
332 * swap_backing_dev_info is bust: it doesn't reflect the
333 * congestion state of the swapdevs. Easy to fix, if needed.
334 * See swapfile.c:page_queue_congested().
335 */
336 if (!is_page_cache_freeable(page))
337 return PAGE_KEEP;
338 if (!mapping) {
339 /*
340 * Some data journaling orphaned pages can have
341 * page->mapping == NULL while being dirty with clean buffers.
342 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700343 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 if (try_to_free_buffers(page)) {
345 ClearPageDirty(page);
346 printk("%s: orphaned page\n", __FUNCTION__);
347 return PAGE_CLEAN;
348 }
349 }
350 return PAGE_KEEP;
351 }
352 if (mapping->a_ops->writepage == NULL)
353 return PAGE_ACTIVATE;
354 if (!may_write_to_queue(mapping->backing_dev_info))
355 return PAGE_KEEP;
356
357 if (clear_page_dirty_for_io(page)) {
358 int res;
359 struct writeback_control wbc = {
360 .sync_mode = WB_SYNC_NONE,
361 .nr_to_write = SWAP_CLUSTER_MAX,
362 .nonblocking = 1,
363 .for_reclaim = 1,
364 };
365
366 SetPageReclaim(page);
367 res = mapping->a_ops->writepage(page, &wbc);
368 if (res < 0)
369 handle_write_error(mapping, page, res);
370 if (res == WRITEPAGE_ACTIVATE) {
371 ClearPageReclaim(page);
372 return PAGE_ACTIVATE;
373 }
374 if (!PageWriteback(page)) {
375 /* synchronous write or broken a_ops? */
376 ClearPageReclaim(page);
377 }
378
379 return PAGE_SUCCESS;
380 }
381
382 return PAGE_CLEAN;
383}
384
385/*
386 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
387 */
388static int shrink_list(struct list_head *page_list, struct scan_control *sc)
389{
390 LIST_HEAD(ret_pages);
391 struct pagevec freed_pvec;
392 int pgactivate = 0;
393 int reclaimed = 0;
394
395 cond_resched();
396
397 pagevec_init(&freed_pvec, 1);
398 while (!list_empty(page_list)) {
399 struct address_space *mapping;
400 struct page *page;
401 int may_enter_fs;
402 int referenced;
403
404 cond_resched();
405
406 page = lru_to_page(page_list);
407 list_del(&page->lru);
408
409 if (TestSetPageLocked(page))
410 goto keep;
411
412 BUG_ON(PageActive(page));
413
414 sc->nr_scanned++;
415 /* Double the slab pressure for mapped and swapcache pages */
416 if (page_mapped(page) || PageSwapCache(page))
417 sc->nr_scanned++;
418
419 if (PageWriteback(page))
420 goto keep_locked;
421
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800422 referenced = page_referenced(page, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 /* In active use or really unfreeable? Activate it. */
424 if (referenced && page_mapping_inuse(page))
425 goto activate_locked;
426
427#ifdef CONFIG_SWAP
428 /*
429 * Anonymous process memory has backing store?
430 * Try to allocate it some swap space here.
431 */
Lee Schermerhornc3400102005-10-29 18:15:51 -0700432 if (PageAnon(page) && !PageSwapCache(page)) {
433 if (!sc->may_swap)
434 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 if (!add_to_swap(page))
436 goto activate_locked;
437 }
438#endif /* CONFIG_SWAP */
439
440 mapping = page_mapping(page);
441 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
442 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
443
444 /*
445 * The page is mapped into the page tables of one or more
446 * processes. Try to unmap it here.
447 */
448 if (page_mapped(page) && mapping) {
449 switch (try_to_unmap(page)) {
450 case SWAP_FAIL:
451 goto activate_locked;
452 case SWAP_AGAIN:
453 goto keep_locked;
454 case SWAP_SUCCESS:
455 ; /* try to free the page below */
456 }
457 }
458
459 if (PageDirty(page)) {
460 if (referenced)
461 goto keep_locked;
462 if (!may_enter_fs)
463 goto keep_locked;
464 if (laptop_mode && !sc->may_writepage)
465 goto keep_locked;
466
467 /* Page is dirty, try to write it out here */
468 switch(pageout(page, mapping)) {
469 case PAGE_KEEP:
470 goto keep_locked;
471 case PAGE_ACTIVATE:
472 goto activate_locked;
473 case PAGE_SUCCESS:
474 if (PageWriteback(page) || PageDirty(page))
475 goto keep;
476 /*
477 * A synchronous write - probably a ramdisk. Go
478 * ahead and try to reclaim the page.
479 */
480 if (TestSetPageLocked(page))
481 goto keep;
482 if (PageDirty(page) || PageWriteback(page))
483 goto keep_locked;
484 mapping = page_mapping(page);
485 case PAGE_CLEAN:
486 ; /* try to free the page below */
487 }
488 }
489
490 /*
491 * If the page has buffers, try to free the buffer mappings
492 * associated with this page. If we succeed we try to free
493 * the page as well.
494 *
495 * We do this even if the page is PageDirty().
496 * try_to_release_page() does not perform I/O, but it is
497 * possible for a page to have PageDirty set, but it is actually
498 * clean (all its buffers are clean). This happens if the
499 * buffers were written out directly, with submit_bh(). ext3
500 * will do this, as well as the blockdev mapping.
501 * try_to_release_page() will discover that cleanness and will
502 * drop the buffers and mark the page clean - it can be freed.
503 *
504 * Rarely, pages can have buffers and no ->mapping. These are
505 * the pages which were not successfully invalidated in
506 * truncate_complete_page(). We try to drop those buffers here
507 * and if that worked, and the page is no longer mapped into
508 * process address space (page_count == 1) it can be freed.
509 * Otherwise, leave the page on the LRU so it is swappable.
510 */
511 if (PagePrivate(page)) {
512 if (!try_to_release_page(page, sc->gfp_mask))
513 goto activate_locked;
514 if (!mapping && page_count(page) == 1)
515 goto free_it;
516 }
517
518 if (!mapping)
519 goto keep_locked; /* truncate got there first */
520
521 write_lock_irq(&mapping->tree_lock);
522
523 /*
524 * The non-racy check for busy page. It is critical to check
525 * PageDirty _after_ making sure that the page is freeable and
526 * not in use by anybody. (pagecache + us == 2)
527 */
Linus Torvalds3d806362005-10-16 17:36:06 -0700528 if (unlikely(page_count(page) != 2))
529 goto cannot_free;
530 smp_rmb();
531 if (unlikely(PageDirty(page)))
532 goto cannot_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533
534#ifdef CONFIG_SWAP
535 if (PageSwapCache(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700536 swp_entry_t swap = { .val = page_private(page) };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 __delete_from_swap_cache(page);
538 write_unlock_irq(&mapping->tree_lock);
539 swap_free(swap);
540 __put_page(page); /* The pagecache ref */
541 goto free_it;
542 }
543#endif /* CONFIG_SWAP */
544
545 __remove_from_page_cache(page);
546 write_unlock_irq(&mapping->tree_lock);
547 __put_page(page);
548
549free_it:
550 unlock_page(page);
551 reclaimed++;
552 if (!pagevec_add(&freed_pvec, page))
553 __pagevec_release_nonlru(&freed_pvec);
554 continue;
555
Linus Torvalds3d806362005-10-16 17:36:06 -0700556cannot_free:
557 write_unlock_irq(&mapping->tree_lock);
558 goto keep_locked;
559
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560activate_locked:
561 SetPageActive(page);
562 pgactivate++;
563keep_locked:
564 unlock_page(page);
565keep:
566 list_add(&page->lru, &ret_pages);
567 BUG_ON(PageLRU(page));
568 }
569 list_splice(&ret_pages, page_list);
570 if (pagevec_count(&freed_pvec))
571 __pagevec_release_nonlru(&freed_pvec);
572 mod_page_state(pgactivate, pgactivate);
573 sc->nr_reclaimed += reclaimed;
574 return reclaimed;
575}
576
577/*
578 * zone->lru_lock is heavily contended. Some of the functions that
579 * shrink the lists perform better by taking out a batch of pages
580 * and working on them outside the LRU lock.
581 *
582 * For pagecache intensive workloads, this function is the hottest
583 * spot in the kernel (apart from copy_*_user functions).
584 *
585 * Appropriate locks must be held before calling this function.
586 *
587 * @nr_to_scan: The number of pages to look through on the list.
588 * @src: The LRU list to pull pages off.
589 * @dst: The temp list to put pages on to.
590 * @scanned: The number of pages that were scanned.
591 *
592 * returns how many pages were moved onto *@dst.
593 */
594static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
595 struct list_head *dst, int *scanned)
596{
597 int nr_taken = 0;
598 struct page *page;
599 int scan = 0;
600
601 while (scan++ < nr_to_scan && !list_empty(src)) {
602 page = lru_to_page(src);
603 prefetchw_prev_lru_page(page, src, flags);
604
605 if (!TestClearPageLRU(page))
606 BUG();
607 list_del(&page->lru);
608 if (get_page_testone(page)) {
609 /*
610 * It is being freed elsewhere
611 */
612 __put_page(page);
613 SetPageLRU(page);
614 list_add(&page->lru, src);
615 continue;
616 } else {
617 list_add(&page->lru, dst);
618 nr_taken++;
619 }
620 }
621
622 *scanned = scan;
623 return nr_taken;
624}
625
626/*
627 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
628 */
629static void shrink_cache(struct zone *zone, struct scan_control *sc)
630{
631 LIST_HEAD(page_list);
632 struct pagevec pvec;
633 int max_scan = sc->nr_to_scan;
634
635 pagevec_init(&pvec, 1);
636
637 lru_add_drain();
638 spin_lock_irq(&zone->lru_lock);
639 while (max_scan > 0) {
640 struct page *page;
641 int nr_taken;
642 int nr_scan;
643 int nr_freed;
644
645 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
646 &zone->inactive_list,
647 &page_list, &nr_scan);
648 zone->nr_inactive -= nr_taken;
649 zone->pages_scanned += nr_scan;
650 spin_unlock_irq(&zone->lru_lock);
651
652 if (nr_taken == 0)
653 goto done;
654
655 max_scan -= nr_scan;
656 if (current_is_kswapd())
657 mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
658 else
659 mod_page_state_zone(zone, pgscan_direct, nr_scan);
660 nr_freed = shrink_list(&page_list, sc);
661 if (current_is_kswapd())
662 mod_page_state(kswapd_steal, nr_freed);
663 mod_page_state_zone(zone, pgsteal, nr_freed);
664 sc->nr_to_reclaim -= nr_freed;
665
666 spin_lock_irq(&zone->lru_lock);
667 /*
668 * Put back any unfreeable pages.
669 */
670 while (!list_empty(&page_list)) {
671 page = lru_to_page(&page_list);
672 if (TestSetPageLRU(page))
673 BUG();
674 list_del(&page->lru);
675 if (PageActive(page))
676 add_page_to_active_list(zone, page);
677 else
678 add_page_to_inactive_list(zone, page);
679 if (!pagevec_add(&pvec, page)) {
680 spin_unlock_irq(&zone->lru_lock);
681 __pagevec_release(&pvec);
682 spin_lock_irq(&zone->lru_lock);
683 }
684 }
685 }
686 spin_unlock_irq(&zone->lru_lock);
687done:
688 pagevec_release(&pvec);
689}
690
691/*
692 * This moves pages from the active list to the inactive list.
693 *
694 * We move them the other way if the page is referenced by one or more
695 * processes, from rmap.
696 *
697 * If the pages are mostly unmapped, the processing is fast and it is
698 * appropriate to hold zone->lru_lock across the whole operation. But if
699 * the pages are mapped, the processing is slow (page_referenced()) so we
700 * should drop zone->lru_lock around each page. It's impossible to balance
701 * this, so instead we remove the pages from the LRU while processing them.
702 * It is safe to rely on PG_active against the non-LRU pages in here because
703 * nobody will play with that bit on a non-LRU page.
704 *
705 * The downside is that we have to touch page->_count against each page.
706 * But we had to alter page->flags anyway.
707 */
708static void
709refill_inactive_zone(struct zone *zone, struct scan_control *sc)
710{
711 int pgmoved;
712 int pgdeactivate = 0;
713 int pgscanned;
714 int nr_pages = sc->nr_to_scan;
715 LIST_HEAD(l_hold); /* The pages which were snipped off */
716 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
717 LIST_HEAD(l_active); /* Pages to go onto the active_list */
718 struct page *page;
719 struct pagevec pvec;
720 int reclaim_mapped = 0;
721 long mapped_ratio;
722 long distress;
723 long swap_tendency;
724
725 lru_add_drain();
726 spin_lock_irq(&zone->lru_lock);
727 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
728 &l_hold, &pgscanned);
729 zone->pages_scanned += pgscanned;
730 zone->nr_active -= pgmoved;
731 spin_unlock_irq(&zone->lru_lock);
732
733 /*
734 * `distress' is a measure of how much trouble we're having reclaiming
735 * pages. 0 -> no problems. 100 -> great trouble.
736 */
737 distress = 100 >> zone->prev_priority;
738
739 /*
740 * The point of this algorithm is to decide when to start reclaiming
741 * mapped memory instead of just pagecache. Work out how much memory
742 * is mapped.
743 */
744 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
745
746 /*
747 * Now decide how much we really want to unmap some pages. The mapped
748 * ratio is downgraded - just because there's a lot of mapped memory
749 * doesn't necessarily mean that page reclaim isn't succeeding.
750 *
751 * The distress ratio is important - we don't want to start going oom.
752 *
753 * A 100% value of vm_swappiness overrides this algorithm altogether.
754 */
755 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
756
757 /*
758 * Now use this metric to decide whether to start moving mapped memory
759 * onto the inactive list.
760 */
761 if (swap_tendency >= 100)
762 reclaim_mapped = 1;
763
764 while (!list_empty(&l_hold)) {
765 cond_resched();
766 page = lru_to_page(&l_hold);
767 list_del(&page->lru);
768 if (page_mapped(page)) {
769 if (!reclaim_mapped ||
770 (total_swap_pages == 0 && PageAnon(page)) ||
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800771 page_referenced(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 list_add(&page->lru, &l_active);
773 continue;
774 }
775 }
776 list_add(&page->lru, &l_inactive);
777 }
778
779 pagevec_init(&pvec, 1);
780 pgmoved = 0;
781 spin_lock_irq(&zone->lru_lock);
782 while (!list_empty(&l_inactive)) {
783 page = lru_to_page(&l_inactive);
784 prefetchw_prev_lru_page(page, &l_inactive, flags);
785 if (TestSetPageLRU(page))
786 BUG();
787 if (!TestClearPageActive(page))
788 BUG();
789 list_move(&page->lru, &zone->inactive_list);
790 pgmoved++;
791 if (!pagevec_add(&pvec, page)) {
792 zone->nr_inactive += pgmoved;
793 spin_unlock_irq(&zone->lru_lock);
794 pgdeactivate += pgmoved;
795 pgmoved = 0;
796 if (buffer_heads_over_limit)
797 pagevec_strip(&pvec);
798 __pagevec_release(&pvec);
799 spin_lock_irq(&zone->lru_lock);
800 }
801 }
802 zone->nr_inactive += pgmoved;
803 pgdeactivate += pgmoved;
804 if (buffer_heads_over_limit) {
805 spin_unlock_irq(&zone->lru_lock);
806 pagevec_strip(&pvec);
807 spin_lock_irq(&zone->lru_lock);
808 }
809
810 pgmoved = 0;
811 while (!list_empty(&l_active)) {
812 page = lru_to_page(&l_active);
813 prefetchw_prev_lru_page(page, &l_active, flags);
814 if (TestSetPageLRU(page))
815 BUG();
816 BUG_ON(!PageActive(page));
817 list_move(&page->lru, &zone->active_list);
818 pgmoved++;
819 if (!pagevec_add(&pvec, page)) {
820 zone->nr_active += pgmoved;
821 pgmoved = 0;
822 spin_unlock_irq(&zone->lru_lock);
823 __pagevec_release(&pvec);
824 spin_lock_irq(&zone->lru_lock);
825 }
826 }
827 zone->nr_active += pgmoved;
828 spin_unlock_irq(&zone->lru_lock);
829 pagevec_release(&pvec);
830
831 mod_page_state_zone(zone, pgrefill, pgscanned);
832 mod_page_state(pgdeactivate, pgdeactivate);
833}
834
835/*
836 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
837 */
838static void
839shrink_zone(struct zone *zone, struct scan_control *sc)
840{
841 unsigned long nr_active;
842 unsigned long nr_inactive;
843
Martin Hicks53e9a612005-09-03 15:54:51 -0700844 atomic_inc(&zone->reclaim_in_progress);
845
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 /*
847 * Add one to `nr_to_scan' just to make sure that the kernel will
848 * slowly sift through the active list.
849 */
850 zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
851 nr_active = zone->nr_scan_active;
852 if (nr_active >= sc->swap_cluster_max)
853 zone->nr_scan_active = 0;
854 else
855 nr_active = 0;
856
857 zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
858 nr_inactive = zone->nr_scan_inactive;
859 if (nr_inactive >= sc->swap_cluster_max)
860 zone->nr_scan_inactive = 0;
861 else
862 nr_inactive = 0;
863
864 sc->nr_to_reclaim = sc->swap_cluster_max;
865
866 while (nr_active || nr_inactive) {
867 if (nr_active) {
868 sc->nr_to_scan = min(nr_active,
869 (unsigned long)sc->swap_cluster_max);
870 nr_active -= sc->nr_to_scan;
871 refill_inactive_zone(zone, sc);
872 }
873
874 if (nr_inactive) {
875 sc->nr_to_scan = min(nr_inactive,
876 (unsigned long)sc->swap_cluster_max);
877 nr_inactive -= sc->nr_to_scan;
878 shrink_cache(zone, sc);
879 if (sc->nr_to_reclaim <= 0)
880 break;
881 }
882 }
883
884 throttle_vm_writeout();
Martin Hicks53e9a612005-09-03 15:54:51 -0700885
886 atomic_dec(&zone->reclaim_in_progress);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887}
888
889/*
890 * This is the direct reclaim path, for page-allocating processes. We only
891 * try to reclaim pages from zones which will satisfy the caller's allocation
892 * request.
893 *
894 * We reclaim from a zone even if that zone is over pages_high. Because:
895 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
896 * allocation or
897 * b) The zones may be over pages_high but they must go *over* pages_high to
898 * satisfy the `incremental min' zone defense algorithm.
899 *
900 * Returns the number of reclaimed pages.
901 *
902 * If a zone is deemed to be full of pinned pages then just give it a light
903 * scan then give up on it.
904 */
905static void
906shrink_caches(struct zone **zones, struct scan_control *sc)
907{
908 int i;
909
910 for (i = 0; zones[i] != NULL; i++) {
911 struct zone *zone = zones[i];
912
913 if (zone->present_pages == 0)
914 continue;
915
Paul Jackson9bf22292005-09-06 15:18:12 -0700916 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 continue;
918
919 zone->temp_priority = sc->priority;
920 if (zone->prev_priority > sc->priority)
921 zone->prev_priority = sc->priority;
922
923 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
924 continue; /* Let kswapd poll it */
925
926 shrink_zone(zone, sc);
927 }
928}
929
930/*
931 * This is the main entry point to direct page reclaim.
932 *
933 * If a full scan of the inactive list fails to free enough memory then we
934 * are "out of memory" and something needs to be killed.
935 *
936 * If the caller is !__GFP_FS then the probability of a failure is reasonably
937 * high - the zone may be full of dirty or under-writeback pages, which this
938 * caller can't do much about. We kick pdflush and take explicit naps in the
939 * hope that some of these pages can be written. But if the allocating task
940 * holds filesystem locks which prevent writeout this might not work, and the
941 * allocation attempt will fail.
942 */
Al Viro6daa0e22005-10-21 03:18:50 -0400943int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944{
945 int priority;
946 int ret = 0;
947 int total_scanned = 0, total_reclaimed = 0;
948 struct reclaim_state *reclaim_state = current->reclaim_state;
949 struct scan_control sc;
950 unsigned long lru_pages = 0;
951 int i;
952
953 sc.gfp_mask = gfp_mask;
954 sc.may_writepage = 0;
Martin Hicksbfbb38f2005-06-21 17:14:40 -0700955 sc.may_swap = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956
957 inc_page_state(allocstall);
958
959 for (i = 0; zones[i] != NULL; i++) {
960 struct zone *zone = zones[i];
961
Paul Jackson9bf22292005-09-06 15:18:12 -0700962 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 continue;
964
965 zone->temp_priority = DEF_PRIORITY;
966 lru_pages += zone->nr_active + zone->nr_inactive;
967 }
968
969 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
970 sc.nr_mapped = read_page_state(nr_mapped);
971 sc.nr_scanned = 0;
972 sc.nr_reclaimed = 0;
973 sc.priority = priority;
974 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800975 if (!priority)
976 disable_swap_token();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 shrink_caches(zones, &sc);
978 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
979 if (reclaim_state) {
980 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
981 reclaim_state->reclaimed_slab = 0;
982 }
983 total_scanned += sc.nr_scanned;
984 total_reclaimed += sc.nr_reclaimed;
985 if (total_reclaimed >= sc.swap_cluster_max) {
986 ret = 1;
987 goto out;
988 }
989
990 /*
991 * Try to write back as many pages as we just scanned. This
992 * tends to cause slow streaming writers to write data to the
993 * disk smoothly, at the dirtying rate, which is nice. But
994 * that's undesirable in laptop mode, where we *want* lumpy
995 * writeout. So in laptop mode, write out the whole world.
996 */
997 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -0700998 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 sc.may_writepage = 1;
1000 }
1001
1002 /* Take a nap, wait for some writeback to complete */
1003 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1004 blk_congestion_wait(WRITE, HZ/10);
1005 }
1006out:
1007 for (i = 0; zones[i] != 0; i++) {
1008 struct zone *zone = zones[i];
1009
Paul Jackson9bf22292005-09-06 15:18:12 -07001010 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 continue;
1012
1013 zone->prev_priority = zone->temp_priority;
1014 }
1015 return ret;
1016}
1017
1018/*
1019 * For kswapd, balance_pgdat() will work across all this node's zones until
1020 * they are all at pages_high.
1021 *
1022 * If `nr_pages' is non-zero then it is the number of pages which are to be
1023 * reclaimed, regardless of the zone occupancies. This is a software suspend
1024 * special.
1025 *
1026 * Returns the number of pages which were actually freed.
1027 *
1028 * There is special handling here for zones which are full of pinned pages.
1029 * This can happen if the pages are all mlocked, or if they are all used by
1030 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1031 * What we do is to detect the case where all pages in the zone have been
1032 * scanned twice and there has been zero successful reclaim. Mark the zone as
1033 * dead and from now on, only perform a short scan. Basically we're polling
1034 * the zone for when the problem goes away.
1035 *
1036 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1037 * zones which have free_pages > pages_high, but once a zone is found to have
1038 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1039 * of the number of free pages in the lower zones. This interoperates with
1040 * the page allocator fallback scheme to ensure that aging of pages is balanced
1041 * across the zones.
1042 */
1043static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1044{
1045 int to_free = nr_pages;
1046 int all_zones_ok;
1047 int priority;
1048 int i;
1049 int total_scanned, total_reclaimed;
1050 struct reclaim_state *reclaim_state = current->reclaim_state;
1051 struct scan_control sc;
1052
1053loop_again:
1054 total_scanned = 0;
1055 total_reclaimed = 0;
1056 sc.gfp_mask = GFP_KERNEL;
1057 sc.may_writepage = 0;
Martin Hicksbfbb38f2005-06-21 17:14:40 -07001058 sc.may_swap = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 sc.nr_mapped = read_page_state(nr_mapped);
1060
1061 inc_page_state(pageoutrun);
1062
1063 for (i = 0; i < pgdat->nr_zones; i++) {
1064 struct zone *zone = pgdat->node_zones + i;
1065
1066 zone->temp_priority = DEF_PRIORITY;
1067 }
1068
1069 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1070 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1071 unsigned long lru_pages = 0;
1072
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001073 /* The swap token gets in the way of swapout... */
1074 if (!priority)
1075 disable_swap_token();
1076
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 all_zones_ok = 1;
1078
1079 if (nr_pages == 0) {
1080 /*
1081 * Scan in the highmem->dma direction for the highest
1082 * zone which needs scanning
1083 */
1084 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1085 struct zone *zone = pgdat->node_zones + i;
1086
1087 if (zone->present_pages == 0)
1088 continue;
1089
1090 if (zone->all_unreclaimable &&
1091 priority != DEF_PRIORITY)
1092 continue;
1093
1094 if (!zone_watermark_ok(zone, order,
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001095 zone->pages_high, 0, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 end_zone = i;
1097 goto scan;
1098 }
1099 }
1100 goto out;
1101 } else {
1102 end_zone = pgdat->nr_zones - 1;
1103 }
1104scan:
1105 for (i = 0; i <= end_zone; i++) {
1106 struct zone *zone = pgdat->node_zones + i;
1107
1108 lru_pages += zone->nr_active + zone->nr_inactive;
1109 }
1110
1111 /*
1112 * Now scan the zone in the dma->highmem direction, stopping
1113 * at the last zone which needs scanning.
1114 *
1115 * We do this because the page allocator works in the opposite
1116 * direction. This prevents the page allocator from allocating
1117 * pages behind kswapd's direction of progress, which would
1118 * cause too much scanning of the lower zones.
1119 */
1120 for (i = 0; i <= end_zone; i++) {
1121 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001122 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123
1124 if (zone->present_pages == 0)
1125 continue;
1126
1127 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1128 continue;
1129
1130 if (nr_pages == 0) { /* Not software suspend */
1131 if (!zone_watermark_ok(zone, order,
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001132 zone->pages_high, end_zone, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 all_zones_ok = 0;
1134 }
1135 zone->temp_priority = priority;
1136 if (zone->prev_priority > priority)
1137 zone->prev_priority = priority;
1138 sc.nr_scanned = 0;
1139 sc.nr_reclaimed = 0;
1140 sc.priority = priority;
1141 sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001142 atomic_inc(&zone->reclaim_in_progress);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 shrink_zone(zone, &sc);
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001144 atomic_dec(&zone->reclaim_in_progress);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001146 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1147 lru_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1149 total_reclaimed += sc.nr_reclaimed;
1150 total_scanned += sc.nr_scanned;
1151 if (zone->all_unreclaimable)
1152 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001153 if (nr_slab == 0 && zone->pages_scanned >=
1154 (zone->nr_active + zone->nr_inactive) * 4)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 zone->all_unreclaimable = 1;
1156 /*
1157 * If we've done a decent amount of scanning and
1158 * the reclaim ratio is low, start doing writepage
1159 * even in laptop mode
1160 */
1161 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1162 total_scanned > total_reclaimed+total_reclaimed/2)
1163 sc.may_writepage = 1;
1164 }
1165 if (nr_pages && to_free > total_reclaimed)
1166 continue; /* swsusp: need to do more work */
1167 if (all_zones_ok)
1168 break; /* kswapd: all done */
1169 /*
1170 * OK, kswapd is getting into trouble. Take a nap, then take
1171 * another pass across the zones.
1172 */
1173 if (total_scanned && priority < DEF_PRIORITY - 2)
1174 blk_congestion_wait(WRITE, HZ/10);
1175
1176 /*
1177 * We do this so kswapd doesn't build up large priorities for
1178 * example when it is freeing in parallel with allocators. It
1179 * matches the direct reclaim path behaviour in terms of impact
1180 * on zone->*_priority.
1181 */
1182 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1183 break;
1184 }
1185out:
1186 for (i = 0; i < pgdat->nr_zones; i++) {
1187 struct zone *zone = pgdat->node_zones + i;
1188
1189 zone->prev_priority = zone->temp_priority;
1190 }
1191 if (!all_zones_ok) {
1192 cond_resched();
1193 goto loop_again;
1194 }
1195
1196 return total_reclaimed;
1197}
1198
1199/*
1200 * The background pageout daemon, started as a kernel thread
1201 * from the init process.
1202 *
1203 * This basically trickles out pages so that we have _some_
1204 * free memory available even if there is no other activity
1205 * that frees anything up. This is needed for things like routing
1206 * etc, where we otherwise might have all activity going on in
1207 * asynchronous contexts that cannot page things out.
1208 *
1209 * If there are applications that are active memory-allocators
1210 * (most normal use), this basically shouldn't matter.
1211 */
1212static int kswapd(void *p)
1213{
1214 unsigned long order;
1215 pg_data_t *pgdat = (pg_data_t*)p;
1216 struct task_struct *tsk = current;
1217 DEFINE_WAIT(wait);
1218 struct reclaim_state reclaim_state = {
1219 .reclaimed_slab = 0,
1220 };
1221 cpumask_t cpumask;
1222
1223 daemonize("kswapd%d", pgdat->node_id);
1224 cpumask = node_to_cpumask(pgdat->node_id);
1225 if (!cpus_empty(cpumask))
1226 set_cpus_allowed(tsk, cpumask);
1227 current->reclaim_state = &reclaim_state;
1228
1229 /*
1230 * Tell the memory management that we're a "memory allocator",
1231 * and that if we need more memory we should get access to it
1232 * regardless (see "__alloc_pages()"). "kswapd" should
1233 * never get caught in the normal page freeing logic.
1234 *
1235 * (Kswapd normally doesn't need memory anyway, but sometimes
1236 * you need a small amount of memory in order to be able to
1237 * page out something else, and this flag essentially protects
1238 * us from recursively trying to free more memory as we're
1239 * trying to free the first piece of memory in the first place).
1240 */
1241 tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1242
1243 order = 0;
1244 for ( ; ; ) {
1245 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001246
1247 try_to_freeze();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
1249 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1250 new_order = pgdat->kswapd_max_order;
1251 pgdat->kswapd_max_order = 0;
1252 if (order < new_order) {
1253 /*
1254 * Don't sleep if someone wants a larger 'order'
1255 * allocation
1256 */
1257 order = new_order;
1258 } else {
1259 schedule();
1260 order = pgdat->kswapd_max_order;
1261 }
1262 finish_wait(&pgdat->kswapd_wait, &wait);
1263
1264 balance_pgdat(pgdat, 0, order);
1265 }
1266 return 0;
1267}
1268
1269/*
1270 * A zone is low on free memory, so wake its kswapd task to service it.
1271 */
1272void wakeup_kswapd(struct zone *zone, int order)
1273{
1274 pg_data_t *pgdat;
1275
1276 if (zone->present_pages == 0)
1277 return;
1278
1279 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001280 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 return;
1282 if (pgdat->kswapd_max_order < order)
1283 pgdat->kswapd_max_order = order;
Paul Jackson9bf22292005-09-06 15:18:12 -07001284 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001286 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001288 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289}
1290
1291#ifdef CONFIG_PM
1292/*
1293 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1294 * pages.
1295 */
1296int shrink_all_memory(int nr_pages)
1297{
1298 pg_data_t *pgdat;
1299 int nr_to_free = nr_pages;
1300 int ret = 0;
1301 struct reclaim_state reclaim_state = {
1302 .reclaimed_slab = 0,
1303 };
1304
1305 current->reclaim_state = &reclaim_state;
1306 for_each_pgdat(pgdat) {
1307 int freed;
1308 freed = balance_pgdat(pgdat, nr_to_free, 0);
1309 ret += freed;
1310 nr_to_free -= freed;
1311 if (nr_to_free <= 0)
1312 break;
1313 }
1314 current->reclaim_state = NULL;
1315 return ret;
1316}
1317#endif
1318
1319#ifdef CONFIG_HOTPLUG_CPU
1320/* It's optimal to keep kswapds on the same CPUs as their memory, but
1321 not required for correctness. So if the last cpu in a node goes
1322 away, we get changed to run anywhere: as the first one comes back,
1323 restore their cpu bindings. */
1324static int __devinit cpu_callback(struct notifier_block *nfb,
1325 unsigned long action,
1326 void *hcpu)
1327{
1328 pg_data_t *pgdat;
1329 cpumask_t mask;
1330
1331 if (action == CPU_ONLINE) {
1332 for_each_pgdat(pgdat) {
1333 mask = node_to_cpumask(pgdat->node_id);
1334 if (any_online_cpu(mask) != NR_CPUS)
1335 /* One of our CPUs online: restore mask */
1336 set_cpus_allowed(pgdat->kswapd, mask);
1337 }
1338 }
1339 return NOTIFY_OK;
1340}
1341#endif /* CONFIG_HOTPLUG_CPU */
1342
1343static int __init kswapd_init(void)
1344{
1345 pg_data_t *pgdat;
1346 swap_setup();
1347 for_each_pgdat(pgdat)
1348 pgdat->kswapd
1349 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1350 total_memory = nr_free_pagecache_pages();
1351 hotcpu_notifier(cpu_callback, 0);
1352 return 0;
1353}
1354
1355module_init(kswapd_init)
Martin Hicks753ee722005-06-21 17:14:41 -07001356
1357
1358/*
1359 * Try to free up some pages from this zone through reclaim.
1360 */
Al Viro6daa0e22005-10-21 03:18:50 -04001361int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Martin Hicks753ee722005-06-21 17:14:41 -07001362{
1363 struct scan_control sc;
1364 int nr_pages = 1 << order;
1365 int total_reclaimed = 0;
1366
1367 /* The reclaim may sleep, so don't do it if sleep isn't allowed */
1368 if (!(gfp_mask & __GFP_WAIT))
1369 return 0;
1370 if (zone->all_unreclaimable)
1371 return 0;
1372
1373 sc.gfp_mask = gfp_mask;
1374 sc.may_writepage = 0;
1375 sc.may_swap = 0;
1376 sc.nr_mapped = read_page_state(nr_mapped);
1377 sc.nr_scanned = 0;
1378 sc.nr_reclaimed = 0;
1379 /* scan at the highest priority */
1380 sc.priority = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001381 disable_swap_token();
Martin Hicks753ee722005-06-21 17:14:41 -07001382
1383 if (nr_pages > SWAP_CLUSTER_MAX)
1384 sc.swap_cluster_max = nr_pages;
1385 else
1386 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1387
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001388 /* Don't reclaim the zone if there are other reclaimers active */
Martin Hicks53e9a612005-09-03 15:54:51 -07001389 if (atomic_read(&zone->reclaim_in_progress) > 0)
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001390 goto out;
1391
Martin Hicks753ee722005-06-21 17:14:41 -07001392 shrink_zone(zone, &sc);
1393 total_reclaimed = sc.nr_reclaimed;
1394
Martin Hicks1e7e5a92005-06-21 17:14:43 -07001395 out:
Martin Hicks753ee722005-06-21 17:14:41 -07001396 return total_reclaimed;
1397}
1398
1399asmlinkage long sys_set_zone_reclaim(unsigned int node, unsigned int zone,
1400 unsigned int state)
1401{
1402 struct zone *z;
1403 int i;
1404
Martin Hicksbce5f6ba2005-09-03 15:54:50 -07001405 if (!capable(CAP_SYS_ADMIN))
1406 return -EACCES;
1407
Martin Hicks753ee722005-06-21 17:14:41 -07001408 if (node >= MAX_NUMNODES || !node_online(node))
1409 return -EINVAL;
1410
1411 /* This will break if we ever add more zones */
1412 if (!(zone & (1<<ZONE_DMA|1<<ZONE_NORMAL|1<<ZONE_HIGHMEM)))
1413 return -EINVAL;
1414
1415 for (i = 0; i < MAX_NR_ZONES; i++) {
1416 if (!(zone & 1<<i))
1417 continue;
1418
1419 z = &NODE_DATA(node)->node_zones[i];
1420
1421 if (state)
1422 z->reclaim_pages = 1;
1423 else
1424 z->reclaim_pages = 0;
1425 }
1426
1427 return 0;
1428}