blob: 093f5fe6dd7795cc5d0ca15f5918eedc822b87f4 [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>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41#include <asm/tlbflush.h>
42#include <asm/div64.h>
43
44#include <linux/swapops.h>
45
Nick Piggin0f8053a2006-03-22 00:08:33 -080046#include "internal.h"
47
Linus Torvalds1da177e2005-04-16 15:20:36 -070048struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070049 /* Incremented by the number of inactive pages that were scanned */
50 unsigned long nr_scanned;
51
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040053 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
55 int may_writepage;
56
Christoph Lameterf1fd1062006-01-18 17:42:30 -080057 /* Can pages be swapped as part of reclaim? */
58 int may_swap;
59
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
61 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
62 * In this context, it doesn't matter that we scan the
63 * whole list at once. */
64 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070065
66 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070067
68 int all_unreclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -070069};
70
71/*
72 * The list of shrinker callbacks used by to apply pressure to
73 * ageable caches.
74 */
75struct shrinker {
76 shrinker_t shrinker;
77 struct list_head list;
78 int seeks; /* seeks to recreate an obj */
79 long nr; /* objs pending delete */
80};
81
82#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
83
84#ifdef ARCH_HAS_PREFETCH
85#define prefetch_prev_lru_page(_page, _base, _field) \
86 do { \
87 if ((_page)->lru.prev != _base) { \
88 struct page *prev; \
89 \
90 prev = lru_to_page(&(_page->lru)); \
91 prefetch(&prev->_field); \
92 } \
93 } while (0)
94#else
95#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
96#endif
97
98#ifdef ARCH_HAS_PREFETCHW
99#define prefetchw_prev_lru_page(_page, _base, _field) \
100 do { \
101 if ((_page)->lru.prev != _base) { \
102 struct page *prev; \
103 \
104 prev = lru_to_page(&(_page->lru)); \
105 prefetchw(&prev->_field); \
106 } \
107 } while (0)
108#else
109#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
110#endif
111
112/*
113 * From 0 .. 100. Higher means more swappy.
114 */
115int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700116long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117
118static LIST_HEAD(shrinker_list);
119static DECLARE_RWSEM(shrinker_rwsem);
120
121/*
122 * Add a shrinker callback to be called from the vm
123 */
124struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
125{
126 struct shrinker *shrinker;
127
128 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
129 if (shrinker) {
130 shrinker->shrinker = theshrinker;
131 shrinker->seeks = seeks;
132 shrinker->nr = 0;
133 down_write(&shrinker_rwsem);
134 list_add_tail(&shrinker->list, &shrinker_list);
135 up_write(&shrinker_rwsem);
136 }
137 return shrinker;
138}
139EXPORT_SYMBOL(set_shrinker);
140
141/*
142 * Remove one
143 */
144void remove_shrinker(struct shrinker *shrinker)
145{
146 down_write(&shrinker_rwsem);
147 list_del(&shrinker->list);
148 up_write(&shrinker_rwsem);
149 kfree(shrinker);
150}
151EXPORT_SYMBOL(remove_shrinker);
152
153#define SHRINK_BATCH 128
154/*
155 * Call the shrink functions to age shrinkable caches
156 *
157 * Here we assume it costs one seek to replace a lru page and that it also
158 * takes a seek to recreate a cache object. With this in mind we age equal
159 * percentages of the lru and ageable caches. This should balance the seeks
160 * generated by these structures.
161 *
162 * If the vm encounted mapped pages on the LRU it increase the pressure on
163 * slab to avoid swapping.
164 *
165 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
166 *
167 * `lru_pages' represents the number of on-LRU pages in all the zones which
168 * are eligible for the caller's allocation attempt. It is used for balancing
169 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700170 *
171 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 */
Andrew Morton69e05942006-03-22 00:08:19 -0800173unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
174 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175{
176 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800177 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
179 if (scanned == 0)
180 scanned = SWAP_CLUSTER_MAX;
181
182 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700183 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184
185 list_for_each_entry(shrinker, &shrinker_list, list) {
186 unsigned long long delta;
187 unsigned long total_scan;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800188 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189
190 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800191 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 do_div(delta, lru_pages + 1);
193 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800194 if (shrinker->nr < 0) {
195 printk(KERN_ERR "%s: nr=%ld\n",
196 __FUNCTION__, shrinker->nr);
197 shrinker->nr = max_pass;
198 }
199
200 /*
201 * Avoid risking looping forever due to too large nr value:
202 * never try to free more than twice the estimate number of
203 * freeable entries.
204 */
205 if (shrinker->nr > max_pass * 2)
206 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207
208 total_scan = shrinker->nr;
209 shrinker->nr = 0;
210
211 while (total_scan >= SHRINK_BATCH) {
212 long this_scan = SHRINK_BATCH;
213 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700214 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700216 nr_before = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
218 if (shrink_ret == -1)
219 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700220 if (shrink_ret < nr_before)
221 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700222 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 total_scan -= this_scan;
224
225 cond_resched();
226 }
227
228 shrinker->nr += total_scan;
229 }
230 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700231 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232}
233
234/* Called without lock on whether page is mapped, so answer is unstable */
235static inline int page_mapping_inuse(struct page *page)
236{
237 struct address_space *mapping;
238
239 /* Page is in somebody's page tables. */
240 if (page_mapped(page))
241 return 1;
242
243 /* Be more reluctant to reclaim swapcache than pagecache */
244 if (PageSwapCache(page))
245 return 1;
246
247 mapping = page_mapping(page);
248 if (!mapping)
249 return 0;
250
251 /* File is mmap'd by somebody? */
252 return mapping_mapped(mapping);
253}
254
255static inline int is_page_cache_freeable(struct page *page)
256{
257 return page_count(page) - !!PagePrivate(page) == 2;
258}
259
260static int may_write_to_queue(struct backing_dev_info *bdi)
261{
Christoph Lameter930d9152006-01-08 01:00:47 -0800262 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 return 1;
264 if (!bdi_write_congested(bdi))
265 return 1;
266 if (bdi == current->backing_dev_info)
267 return 1;
268 return 0;
269}
270
271/*
272 * We detected a synchronous write error writing a page out. Probably
273 * -ENOSPC. We need to propagate that into the address_space for a subsequent
274 * fsync(), msync() or close().
275 *
276 * The tricky part is that after writepage we cannot touch the mapping: nothing
277 * prevents it from being freed up. But we have a ref on the page and once
278 * that page is locked, the mapping is pinned.
279 *
280 * We're allowed to run sleeping lock_page() here because we know the caller has
281 * __GFP_FS.
282 */
283static void handle_write_error(struct address_space *mapping,
284 struct page *page, int error)
285{
286 lock_page(page);
287 if (page_mapping(page) == mapping) {
288 if (error == -ENOSPC)
289 set_bit(AS_ENOSPC, &mapping->flags);
290 else
291 set_bit(AS_EIO, &mapping->flags);
292 }
293 unlock_page(page);
294}
295
Christoph Lameter04e62a22006-06-23 02:03:38 -0700296/* possible outcome of pageout() */
297typedef enum {
298 /* failed to write page out, page is locked */
299 PAGE_KEEP,
300 /* move page to the active list, page is locked */
301 PAGE_ACTIVATE,
302 /* page has been sent to the disk successfully, page is unlocked */
303 PAGE_SUCCESS,
304 /* page is clean and locked */
305 PAGE_CLEAN,
306} pageout_t;
307
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308/*
Andrew Morton1742f192006-03-22 00:08:21 -0800309 * pageout is called by shrink_page_list() for each dirty page.
310 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700312static pageout_t pageout(struct page *page, struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313{
314 /*
315 * If the page is dirty, only perform writeback if that write
316 * will be non-blocking. To prevent this allocation from being
317 * stalled by pagecache activity. But note that there may be
318 * stalls if we need to run get_block(). We could test
319 * PagePrivate for that.
320 *
321 * If this process is currently in generic_file_write() against
322 * this page's queue, we can perform writeback even if that
323 * will block.
324 *
325 * If the page is swapcache, write it back even if that would
326 * block, for some throttling. This happens by accident, because
327 * swap_backing_dev_info is bust: it doesn't reflect the
328 * congestion state of the swapdevs. Easy to fix, if needed.
329 * See swapfile.c:page_queue_congested().
330 */
331 if (!is_page_cache_freeable(page))
332 return PAGE_KEEP;
333 if (!mapping) {
334 /*
335 * Some data journaling orphaned pages can have
336 * page->mapping == NULL while being dirty with clean buffers.
337 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700338 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 if (try_to_free_buffers(page)) {
340 ClearPageDirty(page);
341 printk("%s: orphaned page\n", __FUNCTION__);
342 return PAGE_CLEAN;
343 }
344 }
345 return PAGE_KEEP;
346 }
347 if (mapping->a_ops->writepage == NULL)
348 return PAGE_ACTIVATE;
349 if (!may_write_to_queue(mapping->backing_dev_info))
350 return PAGE_KEEP;
351
352 if (clear_page_dirty_for_io(page)) {
353 int res;
354 struct writeback_control wbc = {
355 .sync_mode = WB_SYNC_NONE,
356 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700357 .range_start = 0,
358 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 .nonblocking = 1,
360 .for_reclaim = 1,
361 };
362
363 SetPageReclaim(page);
364 res = mapping->a_ops->writepage(page, &wbc);
365 if (res < 0)
366 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800367 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 ClearPageReclaim(page);
369 return PAGE_ACTIVATE;
370 }
371 if (!PageWriteback(page)) {
372 /* synchronous write or broken a_ops? */
373 ClearPageReclaim(page);
374 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700375 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 return PAGE_SUCCESS;
377 }
378
379 return PAGE_CLEAN;
380}
381
Andrew Mortona649fd92006-10-17 00:09:36 -0700382/*
383 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
384 * someone else has a ref on the page, abort and return 0. If it was
385 * successfully detached, return 1. Assumes the caller has a single ref on
386 * this page.
387 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800388int remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800389{
Nick Piggin28e4d962006-09-25 23:31:23 -0700390 BUG_ON(!PageLocked(page));
391 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800392
393 write_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800394 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700395 * The non racy check for a busy page.
396 *
397 * Must be careful with the order of the tests. When someone has
398 * a ref to the page, it may be possible that they dirty it then
399 * drop the reference. So if PageDirty is tested before page_count
400 * here, then the following race may occur:
401 *
402 * get_user_pages(&page);
403 * [user mapping goes away]
404 * write_to(page);
405 * !PageDirty(page) [good]
406 * SetPageDirty(page);
407 * put_page(page);
408 * !page_count(page) [good, discard it]
409 *
410 * [oops, our write_to data is lost]
411 *
412 * Reversing the order of the tests ensures such a situation cannot
413 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
414 * load is not satisfied before that of page->_count.
415 *
416 * Note that if SetPageDirty is always performed via set_page_dirty,
417 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800418 */
419 if (unlikely(page_count(page) != 2))
420 goto cannot_free;
421 smp_rmb();
422 if (unlikely(PageDirty(page)))
423 goto cannot_free;
424
425 if (PageSwapCache(page)) {
426 swp_entry_t swap = { .val = page_private(page) };
427 __delete_from_swap_cache(page);
428 write_unlock_irq(&mapping->tree_lock);
429 swap_free(swap);
430 __put_page(page); /* The pagecache ref */
431 return 1;
432 }
433
434 __remove_from_page_cache(page);
435 write_unlock_irq(&mapping->tree_lock);
436 __put_page(page);
437 return 1;
438
439cannot_free:
440 write_unlock_irq(&mapping->tree_lock);
441 return 0;
442}
443
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444/*
Andrew Morton1742f192006-03-22 00:08:21 -0800445 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 */
Andrew Morton1742f192006-03-22 00:08:21 -0800447static unsigned long shrink_page_list(struct list_head *page_list,
448 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449{
450 LIST_HEAD(ret_pages);
451 struct pagevec freed_pvec;
452 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800453 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
455 cond_resched();
456
457 pagevec_init(&freed_pvec, 1);
458 while (!list_empty(page_list)) {
459 struct address_space *mapping;
460 struct page *page;
461 int may_enter_fs;
462 int referenced;
463
464 cond_resched();
465
466 page = lru_to_page(page_list);
467 list_del(&page->lru);
468
469 if (TestSetPageLocked(page))
470 goto keep;
471
Nick Piggin725d7042006-09-25 23:30:55 -0700472 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
474 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800475
476 if (!sc->may_swap && page_mapped(page))
477 goto keep_locked;
478
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 /* Double the slab pressure for mapped and swapcache pages */
480 if (page_mapped(page) || PageSwapCache(page))
481 sc->nr_scanned++;
482
483 if (PageWriteback(page))
484 goto keep_locked;
485
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800486 referenced = page_referenced(page, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 /* In active use or really unfreeable? Activate it. */
488 if (referenced && page_mapping_inuse(page))
489 goto activate_locked;
490
491#ifdef CONFIG_SWAP
492 /*
493 * Anonymous process memory has backing store?
494 * Try to allocate it some swap space here.
495 */
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800496 if (PageAnon(page) && !PageSwapCache(page))
Christoph Lameter1480a542006-01-08 01:00:53 -0800497 if (!add_to_swap(page, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499#endif /* CONFIG_SWAP */
500
501 mapping = page_mapping(page);
502 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
503 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
504
505 /*
506 * The page is mapped into the page tables of one or more
507 * processes. Try to unmap it here.
508 */
509 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800510 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 case SWAP_FAIL:
512 goto activate_locked;
513 case SWAP_AGAIN:
514 goto keep_locked;
515 case SWAP_SUCCESS:
516 ; /* try to free the page below */
517 }
518 }
519
520 if (PageDirty(page)) {
521 if (referenced)
522 goto keep_locked;
523 if (!may_enter_fs)
524 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800525 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 goto keep_locked;
527
528 /* Page is dirty, try to write it out here */
529 switch(pageout(page, mapping)) {
530 case PAGE_KEEP:
531 goto keep_locked;
532 case PAGE_ACTIVATE:
533 goto activate_locked;
534 case PAGE_SUCCESS:
535 if (PageWriteback(page) || PageDirty(page))
536 goto keep;
537 /*
538 * A synchronous write - probably a ramdisk. Go
539 * ahead and try to reclaim the page.
540 */
541 if (TestSetPageLocked(page))
542 goto keep;
543 if (PageDirty(page) || PageWriteback(page))
544 goto keep_locked;
545 mapping = page_mapping(page);
546 case PAGE_CLEAN:
547 ; /* try to free the page below */
548 }
549 }
550
551 /*
552 * If the page has buffers, try to free the buffer mappings
553 * associated with this page. If we succeed we try to free
554 * the page as well.
555 *
556 * We do this even if the page is PageDirty().
557 * try_to_release_page() does not perform I/O, but it is
558 * possible for a page to have PageDirty set, but it is actually
559 * clean (all its buffers are clean). This happens if the
560 * buffers were written out directly, with submit_bh(). ext3
561 * will do this, as well as the blockdev mapping.
562 * try_to_release_page() will discover that cleanness and will
563 * drop the buffers and mark the page clean - it can be freed.
564 *
565 * Rarely, pages can have buffers and no ->mapping. These are
566 * the pages which were not successfully invalidated in
567 * truncate_complete_page(). We try to drop those buffers here
568 * and if that worked, and the page is no longer mapped into
569 * process address space (page_count == 1) it can be freed.
570 * Otherwise, leave the page on the LRU so it is swappable.
571 */
572 if (PagePrivate(page)) {
573 if (!try_to_release_page(page, sc->gfp_mask))
574 goto activate_locked;
575 if (!mapping && page_count(page) == 1)
576 goto free_it;
577 }
578
Nick Piggin28e4d962006-09-25 23:31:23 -0700579 if (!mapping || !remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800580 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581
582free_it:
583 unlock_page(page);
Andrew Morton05ff5132006-03-22 00:08:20 -0800584 nr_reclaimed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 if (!pagevec_add(&freed_pvec, page))
586 __pagevec_release_nonlru(&freed_pvec);
587 continue;
588
589activate_locked:
590 SetPageActive(page);
591 pgactivate++;
592keep_locked:
593 unlock_page(page);
594keep:
595 list_add(&page->lru, &ret_pages);
Nick Piggin725d7042006-09-25 23:30:55 -0700596 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 }
598 list_splice(&ret_pages, page_list);
599 if (pagevec_count(&freed_pvec))
600 __pagevec_release_nonlru(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700601 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800602 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603}
604
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800605/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 * zone->lru_lock is heavily contended. Some of the functions that
607 * shrink the lists perform better by taking out a batch of pages
608 * and working on them outside the LRU lock.
609 *
610 * For pagecache intensive workloads, this function is the hottest
611 * spot in the kernel (apart from copy_*_user functions).
612 *
613 * Appropriate locks must be held before calling this function.
614 *
615 * @nr_to_scan: The number of pages to look through on the list.
616 * @src: The LRU list to pull pages off.
617 * @dst: The temp list to put pages on to.
618 * @scanned: The number of pages that were scanned.
619 *
620 * returns how many pages were moved onto *@dst.
621 */
Andrew Morton69e05942006-03-22 00:08:19 -0800622static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
623 struct list_head *src, struct list_head *dst,
624 unsigned long *scanned)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625{
Andrew Morton69e05942006-03-22 00:08:19 -0800626 unsigned long nr_taken = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 struct page *page;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800628 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800630 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800631 struct list_head *target;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 page = lru_to_page(src);
633 prefetchw_prev_lru_page(page, src, flags);
634
Nick Piggin725d7042006-09-25 23:30:55 -0700635 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800636
Nick Piggin053837f2006-01-18 17:42:27 -0800637 list_del(&page->lru);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800638 target = src;
639 if (likely(get_page_unless_zero(page))) {
Nick Piggin053837f2006-01-18 17:42:27 -0800640 /*
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800641 * Be careful not to clear PageLRU until after we're
642 * sure the page is not being freed elsewhere -- the
643 * page release code relies on it.
Nick Piggin053837f2006-01-18 17:42:27 -0800644 */
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800645 ClearPageLRU(page);
646 target = dst;
647 nr_taken++;
648 } /* else it is being freed elsewhere */
Nick Piggin46453a62006-03-22 00:07:58 -0800649
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800650 list_add(&page->lru, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 }
652
653 *scanned = scan;
654 return nr_taken;
655}
656
657/*
Andrew Morton1742f192006-03-22 00:08:21 -0800658 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
659 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 */
Andrew Morton1742f192006-03-22 00:08:21 -0800661static unsigned long shrink_inactive_list(unsigned long max_scan,
662 struct zone *zone, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663{
664 LIST_HEAD(page_list);
665 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -0800666 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800667 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
669 pagevec_init(&pvec, 1);
670
671 lru_add_drain();
672 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -0800673 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -0800675 unsigned long nr_taken;
676 unsigned long nr_scan;
677 unsigned long nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
679 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
680 &zone->inactive_list,
681 &page_list, &nr_scan);
682 zone->nr_inactive -= nr_taken;
683 zone->pages_scanned += nr_scan;
684 spin_unlock_irq(&zone->lru_lock);
685
Andrew Morton69e05942006-03-22 00:08:19 -0800686 nr_scanned += nr_scan;
Andrew Morton1742f192006-03-22 00:08:21 -0800687 nr_freed = shrink_page_list(&page_list, sc);
Andrew Morton05ff5132006-03-22 00:08:20 -0800688 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -0800689 local_irq_disable();
690 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -0700691 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
692 __count_vm_events(KSWAPD_STEAL, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800693 } else
Christoph Lameterf8891e52006-06-30 01:55:45 -0700694 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
695 __count_vm_events(PGACTIVATE, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800696
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800697 if (nr_taken == 0)
698 goto done;
699
Nick Piggina74609f2006-01-06 00:11:20 -0800700 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 /*
702 * Put back any unfreeable pages.
703 */
704 while (!list_empty(&page_list)) {
705 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -0700706 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800707 SetPageLRU(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 list_del(&page->lru);
709 if (PageActive(page))
710 add_page_to_active_list(zone, page);
711 else
712 add_page_to_inactive_list(zone, page);
713 if (!pagevec_add(&pvec, page)) {
714 spin_unlock_irq(&zone->lru_lock);
715 __pagevec_release(&pvec);
716 spin_lock_irq(&zone->lru_lock);
717 }
718 }
Andrew Morton69e05942006-03-22 00:08:19 -0800719 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800720 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800722 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -0800724 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725}
726
Martin Bligh3bb1a8522006-10-28 10:38:24 -0700727/*
728 * We are about to scan this zone at a certain priority level. If that priority
729 * level is smaller (ie: more urgent) than the previous priority, then note
730 * that priority level within the zone. This is done so that when the next
731 * process comes in to scan this zone, it will immediately start out at this
732 * priority level rather than having to build up its own scanning priority.
733 * Here, this priority affects only the reclaim-mapped threshold.
734 */
735static inline void note_zone_scanning_priority(struct zone *zone, int priority)
736{
737 if (priority < zone->prev_priority)
738 zone->prev_priority = priority;
739}
740
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700741static inline int zone_is_near_oom(struct zone *zone)
742{
743 return zone->pages_scanned >= (zone->nr_active + zone->nr_inactive)*3;
744}
745
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746/*
747 * This moves pages from the active list to the inactive list.
748 *
749 * We move them the other way if the page is referenced by one or more
750 * processes, from rmap.
751 *
752 * If the pages are mostly unmapped, the processing is fast and it is
753 * appropriate to hold zone->lru_lock across the whole operation. But if
754 * the pages are mapped, the processing is slow (page_referenced()) so we
755 * should drop zone->lru_lock around each page. It's impossible to balance
756 * this, so instead we remove the pages from the LRU while processing them.
757 * It is safe to rely on PG_active against the non-LRU pages in here because
758 * nobody will play with that bit on a non-LRU page.
759 *
760 * The downside is that we have to touch page->_count against each page.
761 * But we had to alter page->flags anyway.
762 */
Andrew Morton1742f192006-03-22 00:08:21 -0800763static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Martin Blighbbdb3962006-10-28 10:38:25 -0700764 struct scan_control *sc, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765{
Andrew Morton69e05942006-03-22 00:08:19 -0800766 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -0800768 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 LIST_HEAD(l_hold); /* The pages which were snipped off */
770 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
771 LIST_HEAD(l_active); /* Pages to go onto the active_list */
772 struct page *page;
773 struct pagevec pvec;
774 int reclaim_mapped = 0;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800775
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800776 if (sc->may_swap) {
Christoph Lameter2903fb12006-02-11 17:55:55 -0800777 long mapped_ratio;
778 long distress;
779 long swap_tendency;
780
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700781 if (zone_is_near_oom(zone))
782 goto force_reclaim_mapped;
783
Christoph Lameter2903fb12006-02-11 17:55:55 -0800784 /*
785 * `distress' is a measure of how much trouble we're having
786 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
787 */
Martin Blighbbdb3962006-10-28 10:38:25 -0700788 distress = 100 >> min(zone->prev_priority, priority);
Christoph Lameter2903fb12006-02-11 17:55:55 -0800789
790 /*
791 * The point of this algorithm is to decide when to start
792 * reclaiming mapped memory instead of just pagecache. Work out
793 * how much memory
794 * is mapped.
795 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700796 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
797 global_page_state(NR_ANON_PAGES)) * 100) /
Christoph Lameterbf02cf42006-06-30 01:55:36 -0700798 vm_total_pages;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800799
800 /*
801 * Now decide how much we really want to unmap some pages. The
802 * mapped ratio is downgraded - just because there's a lot of
803 * mapped memory doesn't necessarily mean that page reclaim
804 * isn't succeeding.
805 *
806 * The distress ratio is important - we don't want to start
807 * going oom.
808 *
809 * A 100% value of vm_swappiness overrides this algorithm
810 * altogether.
811 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -0700812 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800813
814 /*
815 * Now use this metric to decide whether to start moving mapped
816 * memory onto the inactive list.
817 */
818 if (swap_tendency >= 100)
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700819force_reclaim_mapped:
Christoph Lameter2903fb12006-02-11 17:55:55 -0800820 reclaim_mapped = 1;
821 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
823 lru_add_drain();
824 spin_lock_irq(&zone->lru_lock);
825 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
826 &l_hold, &pgscanned);
827 zone->pages_scanned += pgscanned;
828 zone->nr_active -= pgmoved;
829 spin_unlock_irq(&zone->lru_lock);
830
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 while (!list_empty(&l_hold)) {
832 cond_resched();
833 page = lru_to_page(&l_hold);
834 list_del(&page->lru);
835 if (page_mapped(page)) {
836 if (!reclaim_mapped ||
837 (total_swap_pages == 0 && PageAnon(page)) ||
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800838 page_referenced(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 list_add(&page->lru, &l_active);
840 continue;
841 }
842 }
843 list_add(&page->lru, &l_inactive);
844 }
845
846 pagevec_init(&pvec, 1);
847 pgmoved = 0;
848 spin_lock_irq(&zone->lru_lock);
849 while (!list_empty(&l_inactive)) {
850 page = lru_to_page(&l_inactive);
851 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin725d7042006-09-25 23:30:55 -0700852 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800853 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -0700854 VM_BUG_ON(!PageActive(page));
Nick Piggin4c84cac2006-03-22 00:08:00 -0800855 ClearPageActive(page);
856
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 list_move(&page->lru, &zone->inactive_list);
858 pgmoved++;
859 if (!pagevec_add(&pvec, page)) {
860 zone->nr_inactive += pgmoved;
861 spin_unlock_irq(&zone->lru_lock);
862 pgdeactivate += pgmoved;
863 pgmoved = 0;
864 if (buffer_heads_over_limit)
865 pagevec_strip(&pvec);
866 __pagevec_release(&pvec);
867 spin_lock_irq(&zone->lru_lock);
868 }
869 }
870 zone->nr_inactive += pgmoved;
871 pgdeactivate += pgmoved;
872 if (buffer_heads_over_limit) {
873 spin_unlock_irq(&zone->lru_lock);
874 pagevec_strip(&pvec);
875 spin_lock_irq(&zone->lru_lock);
876 }
877
878 pgmoved = 0;
879 while (!list_empty(&l_active)) {
880 page = lru_to_page(&l_active);
881 prefetchw_prev_lru_page(page, &l_active, flags);
Nick Piggin725d7042006-09-25 23:30:55 -0700882 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800883 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -0700884 VM_BUG_ON(!PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 list_move(&page->lru, &zone->active_list);
886 pgmoved++;
887 if (!pagevec_add(&pvec, page)) {
888 zone->nr_active += pgmoved;
889 pgmoved = 0;
890 spin_unlock_irq(&zone->lru_lock);
891 __pagevec_release(&pvec);
892 spin_lock_irq(&zone->lru_lock);
893 }
894 }
895 zone->nr_active += pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
Christoph Lameterf8891e52006-06-30 01:55:45 -0700897 __count_zone_vm_events(PGREFILL, zone, pgscanned);
898 __count_vm_events(PGDEACTIVATE, pgdeactivate);
899 spin_unlock_irq(&zone->lru_lock);
Nick Piggina74609f2006-01-06 00:11:20 -0800900
901 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902}
903
904/*
905 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
906 */
Andrew Morton05ff5132006-03-22 00:08:20 -0800907static unsigned long shrink_zone(int priority, struct zone *zone,
908 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909{
910 unsigned long nr_active;
911 unsigned long nr_inactive;
Christoph Lameter86959492006-03-22 00:08:18 -0800912 unsigned long nr_to_scan;
Andrew Morton05ff5132006-03-22 00:08:20 -0800913 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
Martin Hicks53e9a612005-09-03 15:54:51 -0700915 atomic_inc(&zone->reclaim_in_progress);
916
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 /*
918 * Add one to `nr_to_scan' just to make sure that the kernel will
919 * slowly sift through the active list.
920 */
Christoph Lameter86959492006-03-22 00:08:18 -0800921 zone->nr_scan_active += (zone->nr_active >> priority) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 nr_active = zone->nr_scan_active;
923 if (nr_active >= sc->swap_cluster_max)
924 zone->nr_scan_active = 0;
925 else
926 nr_active = 0;
927
Christoph Lameter86959492006-03-22 00:08:18 -0800928 zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 nr_inactive = zone->nr_scan_inactive;
930 if (nr_inactive >= sc->swap_cluster_max)
931 zone->nr_scan_inactive = 0;
932 else
933 nr_inactive = 0;
934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 while (nr_active || nr_inactive) {
936 if (nr_active) {
Christoph Lameter86959492006-03-22 00:08:18 -0800937 nr_to_scan = min(nr_active,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -0800939 nr_active -= nr_to_scan;
Martin Blighbbdb3962006-10-28 10:38:25 -0700940 shrink_active_list(nr_to_scan, zone, sc, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 }
942
943 if (nr_inactive) {
Christoph Lameter86959492006-03-22 00:08:18 -0800944 nr_to_scan = min(nr_inactive,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -0800946 nr_inactive -= nr_to_scan;
Andrew Morton1742f192006-03-22 00:08:21 -0800947 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
948 sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 }
950 }
951
952 throttle_vm_writeout();
Martin Hicks53e9a612005-09-03 15:54:51 -0700953
954 atomic_dec(&zone->reclaim_in_progress);
Andrew Morton05ff5132006-03-22 00:08:20 -0800955 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956}
957
958/*
959 * This is the direct reclaim path, for page-allocating processes. We only
960 * try to reclaim pages from zones which will satisfy the caller's allocation
961 * request.
962 *
963 * We reclaim from a zone even if that zone is over pages_high. Because:
964 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
965 * allocation or
966 * b) The zones may be over pages_high but they must go *over* pages_high to
967 * satisfy the `incremental min' zone defense algorithm.
968 *
969 * Returns the number of reclaimed pages.
970 *
971 * If a zone is deemed to be full of pinned pages then just give it a light
972 * scan then give up on it.
973 */
Andrew Morton1742f192006-03-22 00:08:21 -0800974static unsigned long shrink_zones(int priority, struct zone **zones,
Andrew Morton05ff5132006-03-22 00:08:20 -0800975 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976{
Andrew Morton05ff5132006-03-22 00:08:20 -0800977 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 int i;
979
Nick Piggin408d8542006-09-25 23:31:27 -0700980 sc->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 for (i = 0; zones[i] != NULL; i++) {
982 struct zone *zone = zones[i];
983
Con Kolivasf3fe6512006-01-06 00:11:15 -0800984 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 continue;
986
Paul Jackson9bf22292005-09-06 15:18:12 -0700987 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 continue;
989
Martin Bligh3bb1a8522006-10-28 10:38:24 -0700990 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991
Christoph Lameter86959492006-03-22 00:08:18 -0800992 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 continue; /* Let kswapd poll it */
994
Nick Piggin408d8542006-09-25 23:31:27 -0700995 sc->all_unreclaimable = 0;
996
Andrew Morton05ff5132006-03-22 00:08:20 -0800997 nr_reclaimed += shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 }
Andrew Morton05ff5132006-03-22 00:08:20 -0800999 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000}
1001
1002/*
1003 * This is the main entry point to direct page reclaim.
1004 *
1005 * If a full scan of the inactive list fails to free enough memory then we
1006 * are "out of memory" and something needs to be killed.
1007 *
1008 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1009 * high - the zone may be full of dirty or under-writeback pages, which this
1010 * caller can't do much about. We kick pdflush and take explicit naps in the
1011 * hope that some of these pages can be written. But if the allocating task
1012 * holds filesystem locks which prevent writeout this might not work, and the
1013 * allocation attempt will fail.
1014 */
Andrew Morton69e05942006-03-22 00:08:19 -08001015unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016{
1017 int priority;
1018 int ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001019 unsigned long total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001020 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 unsigned long lru_pages = 0;
1023 int i;
Andrew Morton179e9632006-03-22 00:08:18 -08001024 struct scan_control sc = {
1025 .gfp_mask = gfp_mask,
1026 .may_writepage = !laptop_mode,
1027 .swap_cluster_max = SWAP_CLUSTER_MAX,
1028 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001029 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -08001030 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031
Christoph Lameterf8891e52006-06-30 01:55:45 -07001032 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033
1034 for (i = 0; zones[i] != NULL; i++) {
1035 struct zone *zone = zones[i];
1036
Paul Jackson9bf22292005-09-06 15:18:12 -07001037 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 continue;
1039
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 lru_pages += zone->nr_active + zone->nr_inactive;
1041 }
1042
1043 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 sc.nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001045 if (!priority)
1046 disable_swap_token();
Andrew Morton1742f192006-03-22 00:08:21 -08001047 nr_reclaimed += shrink_zones(priority, zones, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1049 if (reclaim_state) {
Andrew Morton05ff5132006-03-22 00:08:20 -08001050 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 reclaim_state->reclaimed_slab = 0;
1052 }
1053 total_scanned += sc.nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001054 if (nr_reclaimed >= sc.swap_cluster_max) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 ret = 1;
1056 goto out;
1057 }
1058
1059 /*
1060 * Try to write back as many pages as we just scanned. This
1061 * tends to cause slow streaming writers to write data to the
1062 * disk smoothly, at the dirtying rate, which is nice. But
1063 * that's undesirable in laptop mode, where we *want* lumpy
1064 * writeout. So in laptop mode, write out the whole world.
1065 */
Andrew Morton179e9632006-03-22 00:08:18 -08001066 if (total_scanned > sc.swap_cluster_max +
1067 sc.swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001068 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 sc.may_writepage = 1;
1070 }
1071
1072 /* Take a nap, wait for some writeback to complete */
1073 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001074 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 }
Nick Piggin408d8542006-09-25 23:31:27 -07001076 /* top priority shrink_caches still had more to do? don't OOM, then */
1077 if (!sc.all_unreclaimable)
1078 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001080 /*
1081 * Now that we've scanned all the zones at this priority level, note
1082 * that level within the zone so that the next thread which performs
1083 * scanning of this zone will immediately start out at this priority
1084 * level. This affects only the decision whether or not to bring
1085 * mapped pages onto the inactive list.
1086 */
1087 if (priority < 0)
1088 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 for (i = 0; zones[i] != 0; i++) {
1090 struct zone *zone = zones[i];
1091
Paul Jackson9bf22292005-09-06 15:18:12 -07001092 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 continue;
1094
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001095 zone->prev_priority = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 }
1097 return ret;
1098}
1099
1100/*
1101 * For kswapd, balance_pgdat() will work across all this node's zones until
1102 * they are all at pages_high.
1103 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 * Returns the number of pages which were actually freed.
1105 *
1106 * There is special handling here for zones which are full of pinned pages.
1107 * This can happen if the pages are all mlocked, or if they are all used by
1108 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1109 * What we do is to detect the case where all pages in the zone have been
1110 * scanned twice and there has been zero successful reclaim. Mark the zone as
1111 * dead and from now on, only perform a short scan. Basically we're polling
1112 * the zone for when the problem goes away.
1113 *
1114 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1115 * zones which have free_pages > pages_high, but once a zone is found to have
1116 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1117 * of the number of free pages in the lower zones. This interoperates with
1118 * the page allocator fallback scheme to ensure that aging of pages is balanced
1119 * across the zones.
1120 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001121static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 int all_zones_ok;
1124 int priority;
1125 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001126 unsigned long total_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001127 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001129 struct scan_control sc = {
1130 .gfp_mask = GFP_KERNEL,
1131 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001132 .swap_cluster_max = SWAP_CLUSTER_MAX,
1133 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -08001134 };
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001135 /*
1136 * temp_priority is used to remember the scanning priority at which
1137 * this zone was successfully refilled to free_pages == pages_high.
1138 */
1139 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140
1141loop_again:
1142 total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001143 nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001144 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001145 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001147 for (i = 0; i < pgdat->nr_zones; i++)
1148 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
1150 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1151 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1152 unsigned long lru_pages = 0;
1153
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001154 /* The swap token gets in the way of swapout... */
1155 if (!priority)
1156 disable_swap_token();
1157
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 all_zones_ok = 1;
1159
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001160 /*
1161 * Scan in the highmem->dma direction for the highest
1162 * zone which needs scanning
1163 */
1164 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1165 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001167 if (!populated_zone(zone))
1168 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001170 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1171 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001173 if (!zone_watermark_ok(zone, order, zone->pages_high,
1174 0, 0)) {
1175 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001176 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001179 if (i < 0)
1180 goto out;
1181
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 for (i = 0; i <= end_zone; i++) {
1183 struct zone *zone = pgdat->node_zones + i;
1184
1185 lru_pages += zone->nr_active + zone->nr_inactive;
1186 }
1187
1188 /*
1189 * Now scan the zone in the dma->highmem direction, stopping
1190 * at the last zone which needs scanning.
1191 *
1192 * We do this because the page allocator works in the opposite
1193 * direction. This prevents the page allocator from allocating
1194 * pages behind kswapd's direction of progress, which would
1195 * cause too much scanning of the lower zones.
1196 */
1197 for (i = 0; i <= end_zone; i++) {
1198 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001199 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
Con Kolivasf3fe6512006-01-06 00:11:15 -08001201 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 continue;
1203
1204 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1205 continue;
1206
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001207 if (!zone_watermark_ok(zone, order, zone->pages_high,
1208 end_zone, 0))
1209 all_zones_ok = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001210 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 sc.nr_scanned = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001212 note_zone_scanning_priority(zone, priority);
Andrew Morton05ff5132006-03-22 00:08:20 -08001213 nr_reclaimed += shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001215 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1216 lru_pages);
Andrew Morton05ff5132006-03-22 00:08:20 -08001217 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 total_scanned += sc.nr_scanned;
1219 if (zone->all_unreclaimable)
1220 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001221 if (nr_slab == 0 && zone->pages_scanned >=
Nick Piggin4ff1ffb2006-09-25 23:31:28 -07001222 (zone->nr_active + zone->nr_inactive) * 6)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 zone->all_unreclaimable = 1;
1224 /*
1225 * If we've done a decent amount of scanning and
1226 * the reclaim ratio is low, start doing writepage
1227 * even in laptop mode
1228 */
1229 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Andrew Morton05ff5132006-03-22 00:08:20 -08001230 total_scanned > nr_reclaimed + nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 sc.may_writepage = 1;
1232 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 if (all_zones_ok)
1234 break; /* kswapd: all done */
1235 /*
1236 * OK, kswapd is getting into trouble. Take a nap, then take
1237 * another pass across the zones.
1238 */
1239 if (total_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001240 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241
1242 /*
1243 * We do this so kswapd doesn't build up large priorities for
1244 * example when it is freeing in parallel with allocators. It
1245 * matches the direct reclaim path behaviour in terms of impact
1246 * on zone->*_priority.
1247 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001248 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 break;
1250 }
1251out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001252 /*
1253 * Note within each zone the priority level at which this zone was
1254 * brought into a happy state. So that the next thread which scans this
1255 * zone will start out at that priority level.
1256 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 for (i = 0; i < pgdat->nr_zones; i++) {
1258 struct zone *zone = pgdat->node_zones + i;
1259
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001260 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 }
1262 if (!all_zones_ok) {
1263 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08001264
1265 try_to_freeze();
1266
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 goto loop_again;
1268 }
1269
Andrew Morton05ff5132006-03-22 00:08:20 -08001270 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271}
1272
1273/*
1274 * The background pageout daemon, started as a kernel thread
1275 * from the init process.
1276 *
1277 * This basically trickles out pages so that we have _some_
1278 * free memory available even if there is no other activity
1279 * that frees anything up. This is needed for things like routing
1280 * etc, where we otherwise might have all activity going on in
1281 * asynchronous contexts that cannot page things out.
1282 *
1283 * If there are applications that are active memory-allocators
1284 * (most normal use), this basically shouldn't matter.
1285 */
1286static int kswapd(void *p)
1287{
1288 unsigned long order;
1289 pg_data_t *pgdat = (pg_data_t*)p;
1290 struct task_struct *tsk = current;
1291 DEFINE_WAIT(wait);
1292 struct reclaim_state reclaim_state = {
1293 .reclaimed_slab = 0,
1294 };
1295 cpumask_t cpumask;
1296
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 cpumask = node_to_cpumask(pgdat->node_id);
1298 if (!cpus_empty(cpumask))
1299 set_cpus_allowed(tsk, cpumask);
1300 current->reclaim_state = &reclaim_state;
1301
1302 /*
1303 * Tell the memory management that we're a "memory allocator",
1304 * and that if we need more memory we should get access to it
1305 * regardless (see "__alloc_pages()"). "kswapd" should
1306 * never get caught in the normal page freeing logic.
1307 *
1308 * (Kswapd normally doesn't need memory anyway, but sometimes
1309 * you need a small amount of memory in order to be able to
1310 * page out something else, and this flag essentially protects
1311 * us from recursively trying to free more memory as we're
1312 * trying to free the first piece of memory in the first place).
1313 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001314 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315
1316 order = 0;
1317 for ( ; ; ) {
1318 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001319
1320 try_to_freeze();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321
1322 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1323 new_order = pgdat->kswapd_max_order;
1324 pgdat->kswapd_max_order = 0;
1325 if (order < new_order) {
1326 /*
1327 * Don't sleep if someone wants a larger 'order'
1328 * allocation
1329 */
1330 order = new_order;
1331 } else {
1332 schedule();
1333 order = pgdat->kswapd_max_order;
1334 }
1335 finish_wait(&pgdat->kswapd_wait, &wait);
1336
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001337 balance_pgdat(pgdat, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 }
1339 return 0;
1340}
1341
1342/*
1343 * A zone is low on free memory, so wake its kswapd task to service it.
1344 */
1345void wakeup_kswapd(struct zone *zone, int order)
1346{
1347 pg_data_t *pgdat;
1348
Con Kolivasf3fe6512006-01-06 00:11:15 -08001349 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 return;
1351
1352 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001353 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 return;
1355 if (pgdat->kswapd_max_order < order)
1356 pgdat->kswapd_max_order = order;
Paul Jackson9bf22292005-09-06 15:18:12 -07001357 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001359 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001361 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362}
1363
1364#ifdef CONFIG_PM
1365/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001366 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1367 * from LRU lists system-wide, for given pass and priority, and returns the
1368 * number of reclaimed pages
1369 *
1370 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1371 */
1372static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1373 int prio, struct scan_control *sc)
1374{
1375 struct zone *zone;
1376 unsigned long nr_to_scan, ret = 0;
1377
1378 for_each_zone(zone) {
1379
1380 if (!populated_zone(zone))
1381 continue;
1382
1383 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1384 continue;
1385
1386 /* For pass = 0 we don't shrink the active list */
1387 if (pass > 0) {
1388 zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1389 if (zone->nr_scan_active >= nr_pages || pass > 3) {
1390 zone->nr_scan_active = 0;
1391 nr_to_scan = min(nr_pages, zone->nr_active);
Martin Blighbbdb3962006-10-28 10:38:25 -07001392 shrink_active_list(nr_to_scan, zone, sc, prio);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001393 }
1394 }
1395
1396 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1397 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1398 zone->nr_scan_inactive = 0;
1399 nr_to_scan = min(nr_pages, zone->nr_inactive);
1400 ret += shrink_inactive_list(nr_to_scan, zone, sc);
1401 if (ret >= nr_pages)
1402 return ret;
1403 }
1404 }
1405
1406 return ret;
1407}
1408
1409/*
1410 * Try to free `nr_pages' of memory, system-wide, and return the number of
1411 * freed pages.
1412 *
1413 * Rather than trying to age LRUs the aim is to preserve the overall
1414 * LRU order by reclaiming preferentially
1415 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 */
Andrew Morton69e05942006-03-22 00:08:19 -08001417unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001419 unsigned long lru_pages, nr_slab;
Andrew Morton69e05942006-03-22 00:08:19 -08001420 unsigned long ret = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001421 int pass;
1422 struct reclaim_state reclaim_state;
1423 struct zone *zone;
1424 struct scan_control sc = {
1425 .gfp_mask = GFP_KERNEL,
1426 .may_swap = 0,
1427 .swap_cluster_max = nr_pages,
1428 .may_writepage = 1,
1429 .swappiness = vm_swappiness,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 };
1431
1432 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08001433
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001434 lru_pages = 0;
1435 for_each_zone(zone)
1436 lru_pages += zone->nr_active + zone->nr_inactive;
1437
Christoph Lameter972d1a72006-09-25 23:31:51 -07001438 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001439 /* If slab caches are huge, it's better to hit them first */
1440 while (nr_slab >= lru_pages) {
1441 reclaim_state.reclaimed_slab = 0;
1442 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1443 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001445
1446 ret += reclaim_state.reclaimed_slab;
1447 if (ret >= nr_pages)
1448 goto out;
1449
1450 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001452
1453 /*
1454 * We try to shrink LRUs in 5 passes:
1455 * 0 = Reclaim from inactive_list only
1456 * 1 = Reclaim from active list but don't reclaim mapped
1457 * 2 = 2nd pass of type 1
1458 * 3 = Reclaim mapped (normal reclaim)
1459 * 4 = 2nd pass of type 3
1460 */
1461 for (pass = 0; pass < 5; pass++) {
1462 int prio;
1463
1464 /* Needed for shrinking slab caches later on */
1465 if (!lru_pages)
1466 for_each_zone(zone) {
1467 lru_pages += zone->nr_active;
1468 lru_pages += zone->nr_inactive;
1469 }
1470
1471 /* Force reclaiming mapped pages in the passes #3 and #4 */
1472 if (pass > 2) {
1473 sc.may_swap = 1;
1474 sc.swappiness = 100;
1475 }
1476
1477 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1478 unsigned long nr_to_scan = nr_pages - ret;
1479
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001480 sc.nr_scanned = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001481 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1482 if (ret >= nr_pages)
1483 goto out;
1484
1485 reclaim_state.reclaimed_slab = 0;
1486 shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1487 ret += reclaim_state.reclaimed_slab;
1488 if (ret >= nr_pages)
1489 goto out;
1490
1491 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001492 congestion_wait(WRITE, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001493 }
1494
1495 lru_pages = 0;
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08001496 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001497
1498 /*
1499 * If ret = 0, we could not shrink LRUs, but there may be something
1500 * in slab caches
1501 */
1502 if (!ret)
1503 do {
1504 reclaim_state.reclaimed_slab = 0;
1505 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1506 ret += reclaim_state.reclaimed_slab;
1507 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1508
1509out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001511
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 return ret;
1513}
1514#endif
1515
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516/* It's optimal to keep kswapds on the same CPUs as their memory, but
1517 not required for correctness. So if the last cpu in a node goes
1518 away, we get changed to run anywhere: as the first one comes back,
1519 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07001520static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08001521 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522{
1523 pg_data_t *pgdat;
1524 cpumask_t mask;
1525
1526 if (action == CPU_ONLINE) {
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -08001527 for_each_online_pgdat(pgdat) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 mask = node_to_cpumask(pgdat->node_id);
1529 if (any_online_cpu(mask) != NR_CPUS)
1530 /* One of our CPUs online: restore mask */
1531 set_cpus_allowed(pgdat->kswapd, mask);
1532 }
1533 }
1534 return NOTIFY_OK;
1535}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Yasunori Goto3218ae12006-06-27 02:53:33 -07001537/*
1538 * This kswapd start function will be called by init and node-hot-add.
1539 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1540 */
1541int kswapd_run(int nid)
1542{
1543 pg_data_t *pgdat = NODE_DATA(nid);
1544 int ret = 0;
1545
1546 if (pgdat->kswapd)
1547 return 0;
1548
1549 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1550 if (IS_ERR(pgdat->kswapd)) {
1551 /* failure at boot is fatal */
1552 BUG_ON(system_state == SYSTEM_BOOTING);
1553 printk("Failed to start kswapd on node %d\n",nid);
1554 ret = -1;
1555 }
1556 return ret;
1557}
1558
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559static int __init kswapd_init(void)
1560{
Yasunori Goto3218ae12006-06-27 02:53:33 -07001561 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08001562
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 swap_setup();
Yasunori Goto3218ae12006-06-27 02:53:33 -07001564 for_each_online_node(nid)
1565 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 hotcpu_notifier(cpu_callback, 0);
1567 return 0;
1568}
1569
1570module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001571
1572#ifdef CONFIG_NUMA
1573/*
1574 * Zone reclaim mode
1575 *
1576 * If non-zero call zone_reclaim when the number of free pages falls below
1577 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08001578 */
1579int zone_reclaim_mode __read_mostly;
1580
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001581#define RECLAIM_OFF 0
1582#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1583#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1584#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1585
Christoph Lameter9eeff232006-01-18 17:42:31 -08001586/*
Christoph Lametera92f7122006-02-01 03:05:32 -08001587 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1588 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1589 * a zone.
1590 */
1591#define ZONE_RECLAIM_PRIORITY 4
1592
Christoph Lameter9eeff232006-01-18 17:42:31 -08001593/*
Christoph Lameter96146342006-07-03 00:24:13 -07001594 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1595 * occur.
1596 */
1597int sysctl_min_unmapped_ratio = 1;
1598
1599/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07001600 * If the number of slab pages in a zone grows beyond this percentage then
1601 * slab reclaim needs to occur.
1602 */
1603int sysctl_min_slab_ratio = 5;
1604
1605/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08001606 * Try to free up some pages from this zone through reclaim.
1607 */
Andrew Morton179e9632006-03-22 00:08:18 -08001608static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001609{
Christoph Lameter7fb2d46d2006-03-22 00:08:22 -08001610 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08001611 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001612 struct task_struct *p = current;
1613 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08001614 int priority;
Andrew Morton05ff5132006-03-22 00:08:20 -08001615 unsigned long nr_reclaimed = 0;
Andrew Morton179e9632006-03-22 00:08:18 -08001616 struct scan_control sc = {
1617 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1618 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08001619 .swap_cluster_max = max_t(unsigned long, nr_pages,
1620 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08001621 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001622 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -08001623 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07001624 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001625
1626 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08001627 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08001628 /*
1629 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1630 * and we also need to be able to write out pages for RECLAIM_WRITE
1631 * and RECLAIM_SWAP.
1632 */
1633 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001634 reclaim_state.reclaimed_slab = 0;
1635 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08001636
Christoph Lameter0ff38492006-09-25 23:31:52 -07001637 if (zone_page_state(zone, NR_FILE_PAGES) -
1638 zone_page_state(zone, NR_FILE_MAPPED) >
1639 zone->min_unmapped_pages) {
1640 /*
1641 * Free memory by calling shrink zone with increasing
1642 * priorities until we have enough memory freed.
1643 */
1644 priority = ZONE_RECLAIM_PRIORITY;
1645 do {
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001646 note_zone_scanning_priority(zone, priority);
Christoph Lameter0ff38492006-09-25 23:31:52 -07001647 nr_reclaimed += shrink_zone(priority, zone, &sc);
1648 priority--;
1649 } while (priority >= 0 && nr_reclaimed < nr_pages);
1650 }
Christoph Lameterc84db232006-02-01 03:05:29 -08001651
Christoph Lameter83e33a42006-09-25 23:31:53 -07001652 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1653 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001654 /*
Christoph Lameter7fb2d46d2006-03-22 00:08:22 -08001655 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07001656 * many pages were freed in this zone. So we take the current
1657 * number of slab pages and shake the slab until it is reduced
1658 * by the same nr_pages that we used for reclaiming unmapped
1659 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001660 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07001661 * Note that shrink_slab will free memory on all zones and may
1662 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001663 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07001664 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07001665 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1666 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07001667 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07001668
1669 /*
1670 * Update nr_reclaimed by the number of slab pages we
1671 * reclaimed from this zone.
1672 */
1673 nr_reclaimed += slab_reclaimable -
1674 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001675 }
1676
Christoph Lameter9eeff232006-01-18 17:42:31 -08001677 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08001678 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Andrew Morton05ff5132006-03-22 00:08:20 -08001679 return nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001680}
Andrew Morton179e9632006-03-22 00:08:18 -08001681
1682int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1683{
1684 cpumask_t mask;
1685 int node_id;
1686
1687 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07001688 * Zone reclaim reclaims unmapped file backed pages and
1689 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07001690 *
Christoph Lameter96146342006-07-03 00:24:13 -07001691 * A small portion of unmapped file backed pages is needed for
1692 * file I/O otherwise pages read by file I/O will be immediately
1693 * thrown out if the zone is overallocated. So we do not reclaim
1694 * if less than a specified percentage of the zone is used by
1695 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08001696 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07001697 if (zone_page_state(zone, NR_FILE_PAGES) -
Christoph Lameter0ff38492006-09-25 23:31:52 -07001698 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1699 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1700 <= zone->min_slab_pages)
Christoph Lameter96146342006-07-03 00:24:13 -07001701 return 0;
Andrew Morton179e9632006-03-22 00:08:18 -08001702
1703 /*
1704 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1705 * not have reclaimable pages and if we should not delay the allocation
1706 * then do not scan.
1707 */
1708 if (!(gfp_mask & __GFP_WAIT) ||
1709 zone->all_unreclaimable ||
1710 atomic_read(&zone->reclaim_in_progress) > 0 ||
1711 (current->flags & PF_MEMALLOC))
1712 return 0;
1713
1714 /*
1715 * Only run zone reclaim on the local zone or on zones that do not
1716 * have associated processors. This will favor the local processor
1717 * over remote processors and spread off node memory allocations
1718 * as wide as possible.
1719 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07001720 node_id = zone_to_nid(zone);
Andrew Morton179e9632006-03-22 00:08:18 -08001721 mask = node_to_cpumask(node_id);
1722 if (!cpus_empty(mask) && node_id != numa_node_id())
1723 return 0;
1724 return __zone_reclaim(zone, gfp_mask, order);
1725}
Christoph Lameter9eeff232006-01-18 17:42:31 -08001726#endif