blob: d419e10e3daa2dea26c6da35fb8e418b27629c6f [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;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070069
70 int order;
Linus Torvalds1da177e2005-04-16 15:20:36 -070071};
72
Linus Torvalds1da177e2005-04-16 15:20:36 -070073#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
74
75#ifdef ARCH_HAS_PREFETCH
76#define prefetch_prev_lru_page(_page, _base, _field) \
77 do { \
78 if ((_page)->lru.prev != _base) { \
79 struct page *prev; \
80 \
81 prev = lru_to_page(&(_page->lru)); \
82 prefetch(&prev->_field); \
83 } \
84 } while (0)
85#else
86#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
87#endif
88
89#ifdef ARCH_HAS_PREFETCHW
90#define prefetchw_prev_lru_page(_page, _base, _field) \
91 do { \
92 if ((_page)->lru.prev != _base) { \
93 struct page *prev; \
94 \
95 prev = lru_to_page(&(_page->lru)); \
96 prefetchw(&prev->_field); \
97 } \
98 } while (0)
99#else
100#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
101#endif
102
103/*
104 * From 0 .. 100. Higher means more swappy.
105 */
106int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700107long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108
109static LIST_HEAD(shrinker_list);
110static DECLARE_RWSEM(shrinker_rwsem);
111
112/*
113 * Add a shrinker callback to be called from the vm
114 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700115void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700117 shrinker->nr = 0;
118 down_write(&shrinker_rwsem);
119 list_add_tail(&shrinker->list, &shrinker_list);
120 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700122EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
124/*
125 * Remove one
126 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700127void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128{
129 down_write(&shrinker_rwsem);
130 list_del(&shrinker->list);
131 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700133EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
135#define SHRINK_BATCH 128
136/*
137 * Call the shrink functions to age shrinkable caches
138 *
139 * Here we assume it costs one seek to replace a lru page and that it also
140 * takes a seek to recreate a cache object. With this in mind we age equal
141 * percentages of the lru and ageable caches. This should balance the seeks
142 * generated by these structures.
143 *
144 * If the vm encounted mapped pages on the LRU it increase the pressure on
145 * slab to avoid swapping.
146 *
147 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
148 *
149 * `lru_pages' represents the number of on-LRU pages in all the zones which
150 * are eligible for the caller's allocation attempt. It is used for balancing
151 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700152 *
153 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 */
Andrew Morton69e05942006-03-22 00:08:19 -0800155unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
156 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157{
158 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800159 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
161 if (scanned == 0)
162 scanned = SWAP_CLUSTER_MAX;
163
164 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700165 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
167 list_for_each_entry(shrinker, &shrinker_list, list) {
168 unsigned long long delta;
169 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700170 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
172 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800173 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 do_div(delta, lru_pages + 1);
175 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800176 if (shrinker->nr < 0) {
177 printk(KERN_ERR "%s: nr=%ld\n",
178 __FUNCTION__, shrinker->nr);
179 shrinker->nr = max_pass;
180 }
181
182 /*
183 * Avoid risking looping forever due to too large nr value:
184 * never try to free more than twice the estimate number of
185 * freeable entries.
186 */
187 if (shrinker->nr > max_pass * 2)
188 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189
190 total_scan = shrinker->nr;
191 shrinker->nr = 0;
192
193 while (total_scan >= SHRINK_BATCH) {
194 long this_scan = SHRINK_BATCH;
195 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700196 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197
Rusty Russell8e1f9362007-07-17 04:03:17 -0700198 nr_before = (*shrinker->shrink)(0, gfp_mask);
199 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 if (shrink_ret == -1)
201 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700202 if (shrink_ret < nr_before)
203 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700204 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 total_scan -= this_scan;
206
207 cond_resched();
208 }
209
210 shrinker->nr += total_scan;
211 }
212 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700213 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
215
216/* Called without lock on whether page is mapped, so answer is unstable */
217static inline int page_mapping_inuse(struct page *page)
218{
219 struct address_space *mapping;
220
221 /* Page is in somebody's page tables. */
222 if (page_mapped(page))
223 return 1;
224
225 /* Be more reluctant to reclaim swapcache than pagecache */
226 if (PageSwapCache(page))
227 return 1;
228
229 mapping = page_mapping(page);
230 if (!mapping)
231 return 0;
232
233 /* File is mmap'd by somebody? */
234 return mapping_mapped(mapping);
235}
236
237static inline int is_page_cache_freeable(struct page *page)
238{
239 return page_count(page) - !!PagePrivate(page) == 2;
240}
241
242static int may_write_to_queue(struct backing_dev_info *bdi)
243{
Christoph Lameter930d9152006-01-08 01:00:47 -0800244 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 return 1;
246 if (!bdi_write_congested(bdi))
247 return 1;
248 if (bdi == current->backing_dev_info)
249 return 1;
250 return 0;
251}
252
253/*
254 * We detected a synchronous write error writing a page out. Probably
255 * -ENOSPC. We need to propagate that into the address_space for a subsequent
256 * fsync(), msync() or close().
257 *
258 * The tricky part is that after writepage we cannot touch the mapping: nothing
259 * prevents it from being freed up. But we have a ref on the page and once
260 * that page is locked, the mapping is pinned.
261 *
262 * We're allowed to run sleeping lock_page() here because we know the caller has
263 * __GFP_FS.
264 */
265static void handle_write_error(struct address_space *mapping,
266 struct page *page, int error)
267{
268 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700269 if (page_mapping(page) == mapping)
270 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 unlock_page(page);
272}
273
Christoph Lameter04e62a22006-06-23 02:03:38 -0700274/* possible outcome of pageout() */
275typedef enum {
276 /* failed to write page out, page is locked */
277 PAGE_KEEP,
278 /* move page to the active list, page is locked */
279 PAGE_ACTIVATE,
280 /* page has been sent to the disk successfully, page is unlocked */
281 PAGE_SUCCESS,
282 /* page is clean and locked */
283 PAGE_CLEAN,
284} pageout_t;
285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286/*
Andrew Morton1742f192006-03-22 00:08:21 -0800287 * pageout is called by shrink_page_list() for each dirty page.
288 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700290static pageout_t pageout(struct page *page, struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291{
292 /*
293 * If the page is dirty, only perform writeback if that write
294 * will be non-blocking. To prevent this allocation from being
295 * stalled by pagecache activity. But note that there may be
296 * stalls if we need to run get_block(). We could test
297 * PagePrivate for that.
298 *
299 * If this process is currently in generic_file_write() against
300 * this page's queue, we can perform writeback even if that
301 * will block.
302 *
303 * If the page is swapcache, write it back even if that would
304 * block, for some throttling. This happens by accident, because
305 * swap_backing_dev_info is bust: it doesn't reflect the
306 * congestion state of the swapdevs. Easy to fix, if needed.
307 * See swapfile.c:page_queue_congested().
308 */
309 if (!is_page_cache_freeable(page))
310 return PAGE_KEEP;
311 if (!mapping) {
312 /*
313 * Some data journaling orphaned pages can have
314 * page->mapping == NULL while being dirty with clean buffers.
315 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700316 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317 if (try_to_free_buffers(page)) {
318 ClearPageDirty(page);
319 printk("%s: orphaned page\n", __FUNCTION__);
320 return PAGE_CLEAN;
321 }
322 }
323 return PAGE_KEEP;
324 }
325 if (mapping->a_ops->writepage == NULL)
326 return PAGE_ACTIVATE;
327 if (!may_write_to_queue(mapping->backing_dev_info))
328 return PAGE_KEEP;
329
330 if (clear_page_dirty_for_io(page)) {
331 int res;
332 struct writeback_control wbc = {
333 .sync_mode = WB_SYNC_NONE,
334 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700335 .range_start = 0,
336 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 .nonblocking = 1,
338 .for_reclaim = 1,
339 };
340
341 SetPageReclaim(page);
342 res = mapping->a_ops->writepage(page, &wbc);
343 if (res < 0)
344 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800345 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 ClearPageReclaim(page);
347 return PAGE_ACTIVATE;
348 }
349 if (!PageWriteback(page)) {
350 /* synchronous write or broken a_ops? */
351 ClearPageReclaim(page);
352 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700353 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 return PAGE_SUCCESS;
355 }
356
357 return PAGE_CLEAN;
358}
359
Andrew Mortona649fd92006-10-17 00:09:36 -0700360/*
361 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
362 * someone else has a ref on the page, abort and return 0. If it was
363 * successfully detached, return 1. Assumes the caller has a single ref on
364 * this page.
365 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800366int remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800367{
Nick Piggin28e4d962006-09-25 23:31:23 -0700368 BUG_ON(!PageLocked(page));
369 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800370
371 write_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800372 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700373 * The non racy check for a busy page.
374 *
375 * Must be careful with the order of the tests. When someone has
376 * a ref to the page, it may be possible that they dirty it then
377 * drop the reference. So if PageDirty is tested before page_count
378 * here, then the following race may occur:
379 *
380 * get_user_pages(&page);
381 * [user mapping goes away]
382 * write_to(page);
383 * !PageDirty(page) [good]
384 * SetPageDirty(page);
385 * put_page(page);
386 * !page_count(page) [good, discard it]
387 *
388 * [oops, our write_to data is lost]
389 *
390 * Reversing the order of the tests ensures such a situation cannot
391 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
392 * load is not satisfied before that of page->_count.
393 *
394 * Note that if SetPageDirty is always performed via set_page_dirty,
395 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800396 */
397 if (unlikely(page_count(page) != 2))
398 goto cannot_free;
399 smp_rmb();
400 if (unlikely(PageDirty(page)))
401 goto cannot_free;
402
403 if (PageSwapCache(page)) {
404 swp_entry_t swap = { .val = page_private(page) };
405 __delete_from_swap_cache(page);
406 write_unlock_irq(&mapping->tree_lock);
407 swap_free(swap);
408 __put_page(page); /* The pagecache ref */
409 return 1;
410 }
411
412 __remove_from_page_cache(page);
413 write_unlock_irq(&mapping->tree_lock);
414 __put_page(page);
415 return 1;
416
417cannot_free:
418 write_unlock_irq(&mapping->tree_lock);
419 return 0;
420}
421
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422/*
Andrew Morton1742f192006-03-22 00:08:21 -0800423 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 */
Andrew Morton1742f192006-03-22 00:08:21 -0800425static unsigned long shrink_page_list(struct list_head *page_list,
426 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427{
428 LIST_HEAD(ret_pages);
429 struct pagevec freed_pvec;
430 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800431 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
433 cond_resched();
434
435 pagevec_init(&freed_pvec, 1);
436 while (!list_empty(page_list)) {
437 struct address_space *mapping;
438 struct page *page;
439 int may_enter_fs;
440 int referenced;
441
442 cond_resched();
443
444 page = lru_to_page(page_list);
445 list_del(&page->lru);
446
447 if (TestSetPageLocked(page))
448 goto keep;
449
Nick Piggin725d7042006-09-25 23:30:55 -0700450 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451
452 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800453
454 if (!sc->may_swap && page_mapped(page))
455 goto keep_locked;
456
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 /* Double the slab pressure for mapped and swapcache pages */
458 if (page_mapped(page) || PageSwapCache(page))
459 sc->nr_scanned++;
460
461 if (PageWriteback(page))
462 goto keep_locked;
463
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800464 referenced = page_referenced(page, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 /* In active use or really unfreeable? Activate it. */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700466 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
467 referenced && page_mapping_inuse(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 goto activate_locked;
469
470#ifdef CONFIG_SWAP
471 /*
472 * Anonymous process memory has backing store?
473 * Try to allocate it some swap space here.
474 */
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800475 if (PageAnon(page) && !PageSwapCache(page))
Christoph Lameter1480a542006-01-08 01:00:53 -0800476 if (!add_to_swap(page, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478#endif /* CONFIG_SWAP */
479
480 mapping = page_mapping(page);
481 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
482 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
483
484 /*
485 * The page is mapped into the page tables of one or more
486 * processes. Try to unmap it here.
487 */
488 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800489 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 case SWAP_FAIL:
491 goto activate_locked;
492 case SWAP_AGAIN:
493 goto keep_locked;
494 case SWAP_SUCCESS:
495 ; /* try to free the page below */
496 }
497 }
498
499 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700500 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 goto keep_locked;
502 if (!may_enter_fs)
503 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800504 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 goto keep_locked;
506
507 /* Page is dirty, try to write it out here */
508 switch(pageout(page, mapping)) {
509 case PAGE_KEEP:
510 goto keep_locked;
511 case PAGE_ACTIVATE:
512 goto activate_locked;
513 case PAGE_SUCCESS:
514 if (PageWriteback(page) || PageDirty(page))
515 goto keep;
516 /*
517 * A synchronous write - probably a ramdisk. Go
518 * ahead and try to reclaim the page.
519 */
520 if (TestSetPageLocked(page))
521 goto keep;
522 if (PageDirty(page) || PageWriteback(page))
523 goto keep_locked;
524 mapping = page_mapping(page);
525 case PAGE_CLEAN:
526 ; /* try to free the page below */
527 }
528 }
529
530 /*
531 * If the page has buffers, try to free the buffer mappings
532 * associated with this page. If we succeed we try to free
533 * the page as well.
534 *
535 * We do this even if the page is PageDirty().
536 * try_to_release_page() does not perform I/O, but it is
537 * possible for a page to have PageDirty set, but it is actually
538 * clean (all its buffers are clean). This happens if the
539 * buffers were written out directly, with submit_bh(). ext3
540 * will do this, as well as the blockdev mapping.
541 * try_to_release_page() will discover that cleanness and will
542 * drop the buffers and mark the page clean - it can be freed.
543 *
544 * Rarely, pages can have buffers and no ->mapping. These are
545 * the pages which were not successfully invalidated in
546 * truncate_complete_page(). We try to drop those buffers here
547 * and if that worked, and the page is no longer mapped into
548 * process address space (page_count == 1) it can be freed.
549 * Otherwise, leave the page on the LRU so it is swappable.
550 */
551 if (PagePrivate(page)) {
552 if (!try_to_release_page(page, sc->gfp_mask))
553 goto activate_locked;
554 if (!mapping && page_count(page) == 1)
555 goto free_it;
556 }
557
Nick Piggin28e4d962006-09-25 23:31:23 -0700558 if (!mapping || !remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800559 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560
561free_it:
562 unlock_page(page);
Andrew Morton05ff5132006-03-22 00:08:20 -0800563 nr_reclaimed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 if (!pagevec_add(&freed_pvec, page))
565 __pagevec_release_nonlru(&freed_pvec);
566 continue;
567
568activate_locked:
569 SetPageActive(page);
570 pgactivate++;
571keep_locked:
572 unlock_page(page);
573keep:
574 list_add(&page->lru, &ret_pages);
Nick Piggin725d7042006-09-25 23:30:55 -0700575 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 }
577 list_splice(&ret_pages, page_list);
578 if (pagevec_count(&freed_pvec))
579 __pagevec_release_nonlru(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700580 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800581 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582}
583
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700584/* LRU Isolation modes. */
585#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
586#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
587#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
588
589/*
590 * Attempt to remove the specified page from its LRU. Only take this page
591 * if it is of the appropriate PageActive status. Pages which are being
592 * freed elsewhere are also ignored.
593 *
594 * page: page to consider
595 * mode: one of the LRU isolation modes defined above
596 *
597 * returns 0 on success, -ve errno on failure.
598 */
599static int __isolate_lru_page(struct page *page, int mode)
600{
601 int ret = -EINVAL;
602
603 /* Only take pages on the LRU. */
604 if (!PageLRU(page))
605 return ret;
606
607 /*
608 * When checking the active state, we need to be sure we are
609 * dealing with comparible boolean values. Take the logical not
610 * of each.
611 */
612 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
613 return ret;
614
615 ret = -EBUSY;
616 if (likely(get_page_unless_zero(page))) {
617 /*
618 * Be careful not to clear PageLRU until after we're
619 * sure the page is not being freed elsewhere -- the
620 * page release code relies on it.
621 */
622 ClearPageLRU(page);
623 ret = 0;
624 }
625
626 return ret;
627}
628
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800629/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 * zone->lru_lock is heavily contended. Some of the functions that
631 * shrink the lists perform better by taking out a batch of pages
632 * and working on them outside the LRU lock.
633 *
634 * For pagecache intensive workloads, this function is the hottest
635 * spot in the kernel (apart from copy_*_user functions).
636 *
637 * Appropriate locks must be held before calling this function.
638 *
639 * @nr_to_scan: The number of pages to look through on the list.
640 * @src: The LRU list to pull pages off.
641 * @dst: The temp list to put pages on to.
642 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700643 * @order: The caller's attempted allocation order
644 * @mode: One of the LRU isolation modes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 *
646 * returns how many pages were moved onto *@dst.
647 */
Andrew Morton69e05942006-03-22 00:08:19 -0800648static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
649 struct list_head *src, struct list_head *dst,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700650 unsigned long *scanned, int order, int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651{
Andrew Morton69e05942006-03-22 00:08:19 -0800652 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800653 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800655 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700656 struct page *page;
657 unsigned long pfn;
658 unsigned long end_pfn;
659 unsigned long page_pfn;
660 int zone_id;
661
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 page = lru_to_page(src);
663 prefetchw_prev_lru_page(page, src, flags);
664
Nick Piggin725d7042006-09-25 23:30:55 -0700665 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800666
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700667 switch (__isolate_lru_page(page, mode)) {
668 case 0:
669 list_move(&page->lru, dst);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800670 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700671 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800672
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700673 case -EBUSY:
674 /* else it is being freed elsewhere */
675 list_move(&page->lru, src);
676 continue;
677
678 default:
679 BUG();
680 }
681
682 if (!order)
683 continue;
684
685 /*
686 * Attempt to take all pages in the order aligned region
687 * surrounding the tag page. Only take those pages of
688 * the same active state as that tag page. We may safely
689 * round the target page pfn down to the requested order
690 * as the mem_map is guarenteed valid out to MAX_ORDER,
691 * where that page is in a different zone we will detect
692 * it from its zone id and abort this block scan.
693 */
694 zone_id = page_zone_id(page);
695 page_pfn = page_to_pfn(page);
696 pfn = page_pfn & ~((1 << order) - 1);
697 end_pfn = pfn + (1 << order);
698 for (; pfn < end_pfn; pfn++) {
699 struct page *cursor_page;
700
701 /* The target page is in the block, ignore it. */
702 if (unlikely(pfn == page_pfn))
703 continue;
704
705 /* Avoid holes within the zone. */
706 if (unlikely(!pfn_valid_within(pfn)))
707 break;
708
709 cursor_page = pfn_to_page(pfn);
710 /* Check that we have not crossed a zone boundary. */
711 if (unlikely(page_zone_id(cursor_page) != zone_id))
712 continue;
713 switch (__isolate_lru_page(cursor_page, mode)) {
714 case 0:
715 list_move(&cursor_page->lru, dst);
716 nr_taken++;
717 scan++;
718 break;
719
720 case -EBUSY:
721 /* else it is being freed elsewhere */
722 list_move(&cursor_page->lru, src);
723 default:
724 break;
725 }
726 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 }
728
729 *scanned = scan;
730 return nr_taken;
731}
732
733/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700734 * clear_active_flags() is a helper for shrink_active_list(), clearing
735 * any active bits from the pages in the list.
736 */
737static unsigned long clear_active_flags(struct list_head *page_list)
738{
739 int nr_active = 0;
740 struct page *page;
741
742 list_for_each_entry(page, page_list, lru)
743 if (PageActive(page)) {
744 ClearPageActive(page);
745 nr_active++;
746 }
747
748 return nr_active;
749}
750
751/*
Andrew Morton1742f192006-03-22 00:08:21 -0800752 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
753 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 */
Andrew Morton1742f192006-03-22 00:08:21 -0800755static unsigned long shrink_inactive_list(unsigned long max_scan,
756 struct zone *zone, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
758 LIST_HEAD(page_list);
759 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -0800760 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800761 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
763 pagevec_init(&pvec, 1);
764
765 lru_add_drain();
766 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -0800767 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -0800769 unsigned long nr_taken;
770 unsigned long nr_scan;
771 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700772 unsigned long nr_active;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
774 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700775 &zone->inactive_list,
776 &page_list, &nr_scan, sc->order,
777 (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
778 ISOLATE_BOTH : ISOLATE_INACTIVE);
779 nr_active = clear_active_flags(&page_list);
780
781 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
782 __mod_zone_page_state(zone, NR_INACTIVE,
783 -(nr_taken - nr_active));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 zone->pages_scanned += nr_scan;
785 spin_unlock_irq(&zone->lru_lock);
786
Andrew Morton69e05942006-03-22 00:08:19 -0800787 nr_scanned += nr_scan;
Andrew Morton1742f192006-03-22 00:08:21 -0800788 nr_freed = shrink_page_list(&page_list, sc);
Andrew Morton05ff5132006-03-22 00:08:20 -0800789 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -0800790 local_irq_disable();
791 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -0700792 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
793 __count_vm_events(KSWAPD_STEAL, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800794 } else
Christoph Lameterf8891e52006-06-30 01:55:45 -0700795 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
Shantanu Goel918d3f92006-12-29 16:48:59 -0800796 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800797
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800798 if (nr_taken == 0)
799 goto done;
800
Nick Piggina74609f2006-01-06 00:11:20 -0800801 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 /*
803 * Put back any unfreeable pages.
804 */
805 while (!list_empty(&page_list)) {
806 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -0700807 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800808 SetPageLRU(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 list_del(&page->lru);
810 if (PageActive(page))
811 add_page_to_active_list(zone, page);
812 else
813 add_page_to_inactive_list(zone, page);
814 if (!pagevec_add(&pvec, page)) {
815 spin_unlock_irq(&zone->lru_lock);
816 __pagevec_release(&pvec);
817 spin_lock_irq(&zone->lru_lock);
818 }
819 }
Andrew Morton69e05942006-03-22 00:08:19 -0800820 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800821 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800823 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -0800825 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826}
827
Martin Bligh3bb1a8522006-10-28 10:38:24 -0700828/*
829 * We are about to scan this zone at a certain priority level. If that priority
830 * level is smaller (ie: more urgent) than the previous priority, then note
831 * that priority level within the zone. This is done so that when the next
832 * process comes in to scan this zone, it will immediately start out at this
833 * priority level rather than having to build up its own scanning priority.
834 * Here, this priority affects only the reclaim-mapped threshold.
835 */
836static inline void note_zone_scanning_priority(struct zone *zone, int priority)
837{
838 if (priority < zone->prev_priority)
839 zone->prev_priority = priority;
840}
841
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700842static inline int zone_is_near_oom(struct zone *zone)
843{
Christoph Lameterc8785382007-02-10 01:43:01 -0800844 return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
845 + zone_page_state(zone, NR_INACTIVE))*3;
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700846}
847
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848/*
849 * This moves pages from the active list to the inactive list.
850 *
851 * We move them the other way if the page is referenced by one or more
852 * processes, from rmap.
853 *
854 * If the pages are mostly unmapped, the processing is fast and it is
855 * appropriate to hold zone->lru_lock across the whole operation. But if
856 * the pages are mapped, the processing is slow (page_referenced()) so we
857 * should drop zone->lru_lock around each page. It's impossible to balance
858 * this, so instead we remove the pages from the LRU while processing them.
859 * It is safe to rely on PG_active against the non-LRU pages in here because
860 * nobody will play with that bit on a non-LRU page.
861 *
862 * The downside is that we have to touch page->_count against each page.
863 * But we had to alter page->flags anyway.
864 */
Andrew Morton1742f192006-03-22 00:08:21 -0800865static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Martin Blighbbdb3962006-10-28 10:38:25 -0700866 struct scan_control *sc, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867{
Andrew Morton69e05942006-03-22 00:08:19 -0800868 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -0800870 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 LIST_HEAD(l_hold); /* The pages which were snipped off */
872 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
873 LIST_HEAD(l_active); /* Pages to go onto the active_list */
874 struct page *page;
875 struct pagevec pvec;
876 int reclaim_mapped = 0;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800877
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800878 if (sc->may_swap) {
Christoph Lameter2903fb12006-02-11 17:55:55 -0800879 long mapped_ratio;
880 long distress;
881 long swap_tendency;
882
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700883 if (zone_is_near_oom(zone))
884 goto force_reclaim_mapped;
885
Christoph Lameter2903fb12006-02-11 17:55:55 -0800886 /*
887 * `distress' is a measure of how much trouble we're having
888 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
889 */
Martin Blighbbdb3962006-10-28 10:38:25 -0700890 distress = 100 >> min(zone->prev_priority, priority);
Christoph Lameter2903fb12006-02-11 17:55:55 -0800891
892 /*
893 * The point of this algorithm is to decide when to start
894 * reclaiming mapped memory instead of just pagecache. Work out
895 * how much memory
896 * is mapped.
897 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700898 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
899 global_page_state(NR_ANON_PAGES)) * 100) /
Christoph Lameterbf02cf42006-06-30 01:55:36 -0700900 vm_total_pages;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800901
902 /*
903 * Now decide how much we really want to unmap some pages. The
904 * mapped ratio is downgraded - just because there's a lot of
905 * mapped memory doesn't necessarily mean that page reclaim
906 * isn't succeeding.
907 *
908 * The distress ratio is important - we don't want to start
909 * going oom.
910 *
911 * A 100% value of vm_swappiness overrides this algorithm
912 * altogether.
913 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -0700914 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800915
916 /*
917 * Now use this metric to decide whether to start moving mapped
918 * memory onto the inactive list.
919 */
920 if (swap_tendency >= 100)
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700921force_reclaim_mapped:
Christoph Lameter2903fb12006-02-11 17:55:55 -0800922 reclaim_mapped = 1;
923 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924
925 lru_add_drain();
926 spin_lock_irq(&zone->lru_lock);
927 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700928 &l_hold, &pgscanned, sc->order, ISOLATE_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 zone->pages_scanned += pgscanned;
Christoph Lameterc8785382007-02-10 01:43:01 -0800930 __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 spin_unlock_irq(&zone->lru_lock);
932
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 while (!list_empty(&l_hold)) {
934 cond_resched();
935 page = lru_to_page(&l_hold);
936 list_del(&page->lru);
937 if (page_mapped(page)) {
938 if (!reclaim_mapped ||
939 (total_swap_pages == 0 && PageAnon(page)) ||
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800940 page_referenced(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 list_add(&page->lru, &l_active);
942 continue;
943 }
944 }
945 list_add(&page->lru, &l_inactive);
946 }
947
948 pagevec_init(&pvec, 1);
949 pgmoved = 0;
950 spin_lock_irq(&zone->lru_lock);
951 while (!list_empty(&l_inactive)) {
952 page = lru_to_page(&l_inactive);
953 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin725d7042006-09-25 23:30:55 -0700954 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800955 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -0700956 VM_BUG_ON(!PageActive(page));
Nick Piggin4c84cac2006-03-22 00:08:00 -0800957 ClearPageActive(page);
958
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 list_move(&page->lru, &zone->inactive_list);
960 pgmoved++;
961 if (!pagevec_add(&pvec, page)) {
Christoph Lameterc8785382007-02-10 01:43:01 -0800962 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 spin_unlock_irq(&zone->lru_lock);
964 pgdeactivate += pgmoved;
965 pgmoved = 0;
966 if (buffer_heads_over_limit)
967 pagevec_strip(&pvec);
968 __pagevec_release(&pvec);
969 spin_lock_irq(&zone->lru_lock);
970 }
971 }
Christoph Lameterc8785382007-02-10 01:43:01 -0800972 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 pgdeactivate += pgmoved;
974 if (buffer_heads_over_limit) {
975 spin_unlock_irq(&zone->lru_lock);
976 pagevec_strip(&pvec);
977 spin_lock_irq(&zone->lru_lock);
978 }
979
980 pgmoved = 0;
981 while (!list_empty(&l_active)) {
982 page = lru_to_page(&l_active);
983 prefetchw_prev_lru_page(page, &l_active, flags);
Nick Piggin725d7042006-09-25 23:30:55 -0700984 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800985 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -0700986 VM_BUG_ON(!PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 list_move(&page->lru, &zone->active_list);
988 pgmoved++;
989 if (!pagevec_add(&pvec, page)) {
Christoph Lameterc8785382007-02-10 01:43:01 -0800990 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 pgmoved = 0;
992 spin_unlock_irq(&zone->lru_lock);
993 __pagevec_release(&pvec);
994 spin_lock_irq(&zone->lru_lock);
995 }
996 }
Christoph Lameterc8785382007-02-10 01:43:01 -0800997 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998
Christoph Lameterf8891e52006-06-30 01:55:45 -0700999 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1000 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1001 spin_unlock_irq(&zone->lru_lock);
Nick Piggina74609f2006-01-06 00:11:20 -08001002
1003 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004}
1005
1006/*
1007 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1008 */
Andrew Morton05ff5132006-03-22 00:08:20 -08001009static unsigned long shrink_zone(int priority, struct zone *zone,
1010 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011{
1012 unsigned long nr_active;
1013 unsigned long nr_inactive;
Christoph Lameter86959492006-03-22 00:08:18 -08001014 unsigned long nr_to_scan;
Andrew Morton05ff5132006-03-22 00:08:20 -08001015 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
Martin Hicks53e9a612005-09-03 15:54:51 -07001017 atomic_inc(&zone->reclaim_in_progress);
1018
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 /*
1020 * Add one to `nr_to_scan' just to make sure that the kernel will
1021 * slowly sift through the active list.
1022 */
Christoph Lameterc8785382007-02-10 01:43:01 -08001023 zone->nr_scan_active +=
1024 (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 nr_active = zone->nr_scan_active;
1026 if (nr_active >= sc->swap_cluster_max)
1027 zone->nr_scan_active = 0;
1028 else
1029 nr_active = 0;
1030
Christoph Lameterc8785382007-02-10 01:43:01 -08001031 zone->nr_scan_inactive +=
1032 (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 nr_inactive = zone->nr_scan_inactive;
1034 if (nr_inactive >= sc->swap_cluster_max)
1035 zone->nr_scan_inactive = 0;
1036 else
1037 nr_inactive = 0;
1038
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 while (nr_active || nr_inactive) {
1040 if (nr_active) {
Christoph Lameter86959492006-03-22 00:08:18 -08001041 nr_to_scan = min(nr_active,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -08001043 nr_active -= nr_to_scan;
Martin Blighbbdb3962006-10-28 10:38:25 -07001044 shrink_active_list(nr_to_scan, zone, sc, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 }
1046
1047 if (nr_inactive) {
Christoph Lameter86959492006-03-22 00:08:18 -08001048 nr_to_scan = min(nr_inactive,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -08001050 nr_inactive -= nr_to_scan;
Andrew Morton1742f192006-03-22 00:08:21 -08001051 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1052 sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053 }
1054 }
1055
Andrew Morton232ea4d2007-02-28 20:13:21 -08001056 throttle_vm_writeout(sc->gfp_mask);
Martin Hicks53e9a612005-09-03 15:54:51 -07001057
1058 atomic_dec(&zone->reclaim_in_progress);
Andrew Morton05ff5132006-03-22 00:08:20 -08001059 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060}
1061
1062/*
1063 * This is the direct reclaim path, for page-allocating processes. We only
1064 * try to reclaim pages from zones which will satisfy the caller's allocation
1065 * request.
1066 *
1067 * We reclaim from a zone even if that zone is over pages_high. Because:
1068 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1069 * allocation or
1070 * b) The zones may be over pages_high but they must go *over* pages_high to
1071 * satisfy the `incremental min' zone defense algorithm.
1072 *
1073 * Returns the number of reclaimed pages.
1074 *
1075 * If a zone is deemed to be full of pinned pages then just give it a light
1076 * scan then give up on it.
1077 */
Andrew Morton1742f192006-03-22 00:08:21 -08001078static unsigned long shrink_zones(int priority, struct zone **zones,
Andrew Morton05ff5132006-03-22 00:08:20 -08001079 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080{
Andrew Morton05ff5132006-03-22 00:08:20 -08001081 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082 int i;
1083
Nick Piggin408d8542006-09-25 23:31:27 -07001084 sc->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 for (i = 0; zones[i] != NULL; i++) {
1086 struct zone *zone = zones[i];
1087
Con Kolivasf3fe6512006-01-06 00:11:15 -08001088 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 continue;
1090
Paul Jackson02a0e532006-12-13 00:34:25 -08001091 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 continue;
1093
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001094 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095
Christoph Lameter86959492006-03-22 00:08:18 -08001096 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 continue; /* Let kswapd poll it */
1098
Nick Piggin408d8542006-09-25 23:31:27 -07001099 sc->all_unreclaimable = 0;
1100
Andrew Morton05ff5132006-03-22 00:08:20 -08001101 nr_reclaimed += shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 }
Andrew Morton05ff5132006-03-22 00:08:20 -08001103 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104}
1105
1106/*
1107 * This is the main entry point to direct page reclaim.
1108 *
1109 * If a full scan of the inactive list fails to free enough memory then we
1110 * are "out of memory" and something needs to be killed.
1111 *
1112 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1113 * high - the zone may be full of dirty or under-writeback pages, which this
1114 * caller can't do much about. We kick pdflush and take explicit naps in the
1115 * hope that some of these pages can be written. But if the allocating task
1116 * holds filesystem locks which prevent writeout this might not work, and the
1117 * allocation attempt will fail.
1118 */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001119unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120{
1121 int priority;
1122 int ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001123 unsigned long total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001124 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 unsigned long lru_pages = 0;
1127 int i;
Andrew Morton179e9632006-03-22 00:08:18 -08001128 struct scan_control sc = {
1129 .gfp_mask = gfp_mask,
1130 .may_writepage = !laptop_mode,
1131 .swap_cluster_max = SWAP_CLUSTER_MAX,
1132 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001133 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001134 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08001135 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
Christoph Lameterf8891e52006-06-30 01:55:45 -07001137 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138
1139 for (i = 0; zones[i] != NULL; i++) {
1140 struct zone *zone = zones[i];
1141
Paul Jackson02a0e532006-12-13 00:34:25 -08001142 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 continue;
1144
Christoph Lameterc8785382007-02-10 01:43:01 -08001145 lru_pages += zone_page_state(zone, NR_ACTIVE)
1146 + zone_page_state(zone, NR_INACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 }
1148
1149 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 sc.nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001151 if (!priority)
1152 disable_swap_token();
Andrew Morton1742f192006-03-22 00:08:21 -08001153 nr_reclaimed += shrink_zones(priority, zones, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1155 if (reclaim_state) {
Andrew Morton05ff5132006-03-22 00:08:20 -08001156 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 reclaim_state->reclaimed_slab = 0;
1158 }
1159 total_scanned += sc.nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001160 if (nr_reclaimed >= sc.swap_cluster_max) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 ret = 1;
1162 goto out;
1163 }
1164
1165 /*
1166 * Try to write back as many pages as we just scanned. This
1167 * tends to cause slow streaming writers to write data to the
1168 * disk smoothly, at the dirtying rate, which is nice. But
1169 * that's undesirable in laptop mode, where we *want* lumpy
1170 * writeout. So in laptop mode, write out the whole world.
1171 */
Andrew Morton179e9632006-03-22 00:08:18 -08001172 if (total_scanned > sc.swap_cluster_max +
1173 sc.swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001174 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 sc.may_writepage = 1;
1176 }
1177
1178 /* Take a nap, wait for some writeback to complete */
1179 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001180 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 }
Nick Piggin408d8542006-09-25 23:31:27 -07001182 /* top priority shrink_caches still had more to do? don't OOM, then */
1183 if (!sc.all_unreclaimable)
1184 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001186 /*
1187 * Now that we've scanned all the zones at this priority level, note
1188 * that level within the zone so that the next thread which performs
1189 * scanning of this zone will immediately start out at this priority
1190 * level. This affects only the decision whether or not to bring
1191 * mapped pages onto the inactive list.
1192 */
1193 if (priority < 0)
1194 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 for (i = 0; zones[i] != 0; i++) {
1196 struct zone *zone = zones[i];
1197
Paul Jackson02a0e532006-12-13 00:34:25 -08001198 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 continue;
1200
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001201 zone->prev_priority = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 }
1203 return ret;
1204}
1205
1206/*
1207 * For kswapd, balance_pgdat() will work across all this node's zones until
1208 * they are all at pages_high.
1209 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 * Returns the number of pages which were actually freed.
1211 *
1212 * There is special handling here for zones which are full of pinned pages.
1213 * This can happen if the pages are all mlocked, or if they are all used by
1214 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1215 * What we do is to detect the case where all pages in the zone have been
1216 * scanned twice and there has been zero successful reclaim. Mark the zone as
1217 * dead and from now on, only perform a short scan. Basically we're polling
1218 * the zone for when the problem goes away.
1219 *
1220 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1221 * zones which have free_pages > pages_high, but once a zone is found to have
1222 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1223 * of the number of free pages in the lower zones. This interoperates with
1224 * the page allocator fallback scheme to ensure that aging of pages is balanced
1225 * across the zones.
1226 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001227static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 int all_zones_ok;
1230 int priority;
1231 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001232 unsigned long total_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001233 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001235 struct scan_control sc = {
1236 .gfp_mask = GFP_KERNEL,
1237 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001238 .swap_cluster_max = SWAP_CLUSTER_MAX,
1239 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001240 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08001241 };
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001242 /*
1243 * temp_priority is used to remember the scanning priority at which
1244 * this zone was successfully refilled to free_pages == pages_high.
1245 */
1246 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
1248loop_again:
1249 total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001250 nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001251 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001252 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001254 for (i = 0; i < pgdat->nr_zones; i++)
1255 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
1257 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1258 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1259 unsigned long lru_pages = 0;
1260
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001261 /* The swap token gets in the way of swapout... */
1262 if (!priority)
1263 disable_swap_token();
1264
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 all_zones_ok = 1;
1266
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001267 /*
1268 * Scan in the highmem->dma direction for the highest
1269 * zone which needs scanning
1270 */
1271 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1272 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001274 if (!populated_zone(zone))
1275 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001277 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1278 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001280 if (!zone_watermark_ok(zone, order, zone->pages_high,
1281 0, 0)) {
1282 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001283 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001286 if (i < 0)
1287 goto out;
1288
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 for (i = 0; i <= end_zone; i++) {
1290 struct zone *zone = pgdat->node_zones + i;
1291
Christoph Lameterc8785382007-02-10 01:43:01 -08001292 lru_pages += zone_page_state(zone, NR_ACTIVE)
1293 + zone_page_state(zone, NR_INACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 }
1295
1296 /*
1297 * Now scan the zone in the dma->highmem direction, stopping
1298 * at the last zone which needs scanning.
1299 *
1300 * We do this because the page allocator works in the opposite
1301 * direction. This prevents the page allocator from allocating
1302 * pages behind kswapd's direction of progress, which would
1303 * cause too much scanning of the lower zones.
1304 */
1305 for (i = 0; i <= end_zone; i++) {
1306 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001307 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308
Con Kolivasf3fe6512006-01-06 00:11:15 -08001309 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 continue;
1311
1312 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1313 continue;
1314
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001315 if (!zone_watermark_ok(zone, order, zone->pages_high,
1316 end_zone, 0))
1317 all_zones_ok = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001318 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 sc.nr_scanned = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001320 note_zone_scanning_priority(zone, priority);
Andrew Morton05ff5132006-03-22 00:08:20 -08001321 nr_reclaimed += shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001323 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1324 lru_pages);
Andrew Morton05ff5132006-03-22 00:08:20 -08001325 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 total_scanned += sc.nr_scanned;
1327 if (zone->all_unreclaimable)
1328 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001329 if (nr_slab == 0 && zone->pages_scanned >=
Christoph Lameterc8785382007-02-10 01:43:01 -08001330 (zone_page_state(zone, NR_ACTIVE)
1331 + zone_page_state(zone, NR_INACTIVE)) * 6)
1332 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 /*
1334 * If we've done a decent amount of scanning and
1335 * the reclaim ratio is low, start doing writepage
1336 * even in laptop mode
1337 */
1338 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Andrew Morton05ff5132006-03-22 00:08:20 -08001339 total_scanned > nr_reclaimed + nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 sc.may_writepage = 1;
1341 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 if (all_zones_ok)
1343 break; /* kswapd: all done */
1344 /*
1345 * OK, kswapd is getting into trouble. Take a nap, then take
1346 * another pass across the zones.
1347 */
1348 if (total_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001349 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350
1351 /*
1352 * We do this so kswapd doesn't build up large priorities for
1353 * example when it is freeing in parallel with allocators. It
1354 * matches the direct reclaim path behaviour in terms of impact
1355 * on zone->*_priority.
1356 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001357 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 break;
1359 }
1360out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001361 /*
1362 * Note within each zone the priority level at which this zone was
1363 * brought into a happy state. So that the next thread which scans this
1364 * zone will start out at that priority level.
1365 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 for (i = 0; i < pgdat->nr_zones; i++) {
1367 struct zone *zone = pgdat->node_zones + i;
1368
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001369 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 }
1371 if (!all_zones_ok) {
1372 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08001373
1374 try_to_freeze();
1375
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 goto loop_again;
1377 }
1378
Andrew Morton05ff5132006-03-22 00:08:20 -08001379 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381
1382/*
1383 * The background pageout daemon, started as a kernel thread
1384 * from the init process.
1385 *
1386 * This basically trickles out pages so that we have _some_
1387 * free memory available even if there is no other activity
1388 * that frees anything up. This is needed for things like routing
1389 * etc, where we otherwise might have all activity going on in
1390 * asynchronous contexts that cannot page things out.
1391 *
1392 * If there are applications that are active memory-allocators
1393 * (most normal use), this basically shouldn't matter.
1394 */
1395static int kswapd(void *p)
1396{
1397 unsigned long order;
1398 pg_data_t *pgdat = (pg_data_t*)p;
1399 struct task_struct *tsk = current;
1400 DEFINE_WAIT(wait);
1401 struct reclaim_state reclaim_state = {
1402 .reclaimed_slab = 0,
1403 };
1404 cpumask_t cpumask;
1405
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 cpumask = node_to_cpumask(pgdat->node_id);
1407 if (!cpus_empty(cpumask))
1408 set_cpus_allowed(tsk, cpumask);
1409 current->reclaim_state = &reclaim_state;
1410
1411 /*
1412 * Tell the memory management that we're a "memory allocator",
1413 * and that if we need more memory we should get access to it
1414 * regardless (see "__alloc_pages()"). "kswapd" should
1415 * never get caught in the normal page freeing logic.
1416 *
1417 * (Kswapd normally doesn't need memory anyway, but sometimes
1418 * you need a small amount of memory in order to be able to
1419 * page out something else, and this flag essentially protects
1420 * us from recursively trying to free more memory as we're
1421 * trying to free the first piece of memory in the first place).
1422 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001423 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07001424 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
1426 order = 0;
1427 for ( ; ; ) {
1428 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001429
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1431 new_order = pgdat->kswapd_max_order;
1432 pgdat->kswapd_max_order = 0;
1433 if (order < new_order) {
1434 /*
1435 * Don't sleep if someone wants a larger 'order'
1436 * allocation
1437 */
1438 order = new_order;
1439 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001440 if (!freezing(current))
1441 schedule();
1442
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 order = pgdat->kswapd_max_order;
1444 }
1445 finish_wait(&pgdat->kswapd_wait, &wait);
1446
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001447 if (!try_to_freeze()) {
1448 /* We can speed up thawing tasks if we don't call
1449 * balance_pgdat after returning from the refrigerator
1450 */
1451 balance_pgdat(pgdat, order);
1452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 }
1454 return 0;
1455}
1456
1457/*
1458 * A zone is low on free memory, so wake its kswapd task to service it.
1459 */
1460void wakeup_kswapd(struct zone *zone, int order)
1461{
1462 pg_data_t *pgdat;
1463
Con Kolivasf3fe6512006-01-06 00:11:15 -08001464 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 return;
1466
1467 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001468 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 return;
1470 if (pgdat->kswapd_max_order < order)
1471 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08001472 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001474 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001476 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477}
1478
1479#ifdef CONFIG_PM
1480/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001481 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1482 * from LRU lists system-wide, for given pass and priority, and returns the
1483 * number of reclaimed pages
1484 *
1485 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1486 */
Nigel Cunninghame07aa052006-12-22 01:07:21 -08001487static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1488 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001489{
1490 struct zone *zone;
1491 unsigned long nr_to_scan, ret = 0;
1492
1493 for_each_zone(zone) {
1494
1495 if (!populated_zone(zone))
1496 continue;
1497
1498 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1499 continue;
1500
1501 /* For pass = 0 we don't shrink the active list */
1502 if (pass > 0) {
Christoph Lameterc8785382007-02-10 01:43:01 -08001503 zone->nr_scan_active +=
1504 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001505 if (zone->nr_scan_active >= nr_pages || pass > 3) {
1506 zone->nr_scan_active = 0;
Christoph Lameterc8785382007-02-10 01:43:01 -08001507 nr_to_scan = min(nr_pages,
1508 zone_page_state(zone, NR_ACTIVE));
Martin Blighbbdb3962006-10-28 10:38:25 -07001509 shrink_active_list(nr_to_scan, zone, sc, prio);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001510 }
1511 }
1512
Christoph Lameterc8785382007-02-10 01:43:01 -08001513 zone->nr_scan_inactive +=
1514 (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001515 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1516 zone->nr_scan_inactive = 0;
Christoph Lameterc8785382007-02-10 01:43:01 -08001517 nr_to_scan = min(nr_pages,
1518 zone_page_state(zone, NR_INACTIVE));
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001519 ret += shrink_inactive_list(nr_to_scan, zone, sc);
1520 if (ret >= nr_pages)
1521 return ret;
1522 }
1523 }
1524
1525 return ret;
1526}
1527
Andrew Morton76395d32007-01-05 16:37:05 -08001528static unsigned long count_lru_pages(void)
1529{
Christoph Lameterc8785382007-02-10 01:43:01 -08001530 return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
Andrew Morton76395d32007-01-05 16:37:05 -08001531}
1532
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001533/*
1534 * Try to free `nr_pages' of memory, system-wide, and return the number of
1535 * freed pages.
1536 *
1537 * Rather than trying to age LRUs the aim is to preserve the overall
1538 * LRU order by reclaiming preferentially
1539 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 */
Andrew Morton69e05942006-03-22 00:08:19 -08001541unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001543 unsigned long lru_pages, nr_slab;
Andrew Morton69e05942006-03-22 00:08:19 -08001544 unsigned long ret = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001545 int pass;
1546 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001547 struct scan_control sc = {
1548 .gfp_mask = GFP_KERNEL,
1549 .may_swap = 0,
1550 .swap_cluster_max = nr_pages,
1551 .may_writepage = 1,
1552 .swappiness = vm_swappiness,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 };
1554
1555 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08001556
Andrew Morton76395d32007-01-05 16:37:05 -08001557 lru_pages = count_lru_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07001558 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001559 /* If slab caches are huge, it's better to hit them first */
1560 while (nr_slab >= lru_pages) {
1561 reclaim_state.reclaimed_slab = 0;
1562 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1563 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001565
1566 ret += reclaim_state.reclaimed_slab;
1567 if (ret >= nr_pages)
1568 goto out;
1569
1570 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001572
1573 /*
1574 * We try to shrink LRUs in 5 passes:
1575 * 0 = Reclaim from inactive_list only
1576 * 1 = Reclaim from active list but don't reclaim mapped
1577 * 2 = 2nd pass of type 1
1578 * 3 = Reclaim mapped (normal reclaim)
1579 * 4 = 2nd pass of type 3
1580 */
1581 for (pass = 0; pass < 5; pass++) {
1582 int prio;
1583
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001584 /* Force reclaiming mapped pages in the passes #3 and #4 */
1585 if (pass > 2) {
1586 sc.may_swap = 1;
1587 sc.swappiness = 100;
1588 }
1589
1590 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1591 unsigned long nr_to_scan = nr_pages - ret;
1592
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001593 sc.nr_scanned = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001594 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1595 if (ret >= nr_pages)
1596 goto out;
1597
1598 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08001599 shrink_slab(sc.nr_scanned, sc.gfp_mask,
1600 count_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001601 ret += reclaim_state.reclaimed_slab;
1602 if (ret >= nr_pages)
1603 goto out;
1604
1605 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001606 congestion_wait(WRITE, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001607 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08001608 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001609
1610 /*
1611 * If ret = 0, we could not shrink LRUs, but there may be something
1612 * in slab caches
1613 */
Andrew Morton76395d32007-01-05 16:37:05 -08001614 if (!ret) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001615 do {
1616 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08001617 shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001618 ret += reclaim_state.reclaimed_slab;
1619 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08001620 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001621
1622out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001624
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 return ret;
1626}
1627#endif
1628
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629/* It's optimal to keep kswapds on the same CPUs as their memory, but
1630 not required for correctness. So if the last cpu in a node goes
1631 away, we get changed to run anywhere: as the first one comes back,
1632 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07001633static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08001634 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635{
1636 pg_data_t *pgdat;
1637 cpumask_t mask;
1638
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001639 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -08001640 for_each_online_pgdat(pgdat) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 mask = node_to_cpumask(pgdat->node_id);
1642 if (any_online_cpu(mask) != NR_CPUS)
1643 /* One of our CPUs online: restore mask */
1644 set_cpus_allowed(pgdat->kswapd, mask);
1645 }
1646 }
1647 return NOTIFY_OK;
1648}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Yasunori Goto3218ae12006-06-27 02:53:33 -07001650/*
1651 * This kswapd start function will be called by init and node-hot-add.
1652 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1653 */
1654int kswapd_run(int nid)
1655{
1656 pg_data_t *pgdat = NODE_DATA(nid);
1657 int ret = 0;
1658
1659 if (pgdat->kswapd)
1660 return 0;
1661
1662 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1663 if (IS_ERR(pgdat->kswapd)) {
1664 /* failure at boot is fatal */
1665 BUG_ON(system_state == SYSTEM_BOOTING);
1666 printk("Failed to start kswapd on node %d\n",nid);
1667 ret = -1;
1668 }
1669 return ret;
1670}
1671
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672static int __init kswapd_init(void)
1673{
Yasunori Goto3218ae12006-06-27 02:53:33 -07001674 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08001675
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 swap_setup();
Yasunori Goto3218ae12006-06-27 02:53:33 -07001677 for_each_online_node(nid)
1678 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 hotcpu_notifier(cpu_callback, 0);
1680 return 0;
1681}
1682
1683module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001684
1685#ifdef CONFIG_NUMA
1686/*
1687 * Zone reclaim mode
1688 *
1689 * If non-zero call zone_reclaim when the number of free pages falls below
1690 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08001691 */
1692int zone_reclaim_mode __read_mostly;
1693
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001694#define RECLAIM_OFF 0
1695#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1696#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1697#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1698
Christoph Lameter9eeff232006-01-18 17:42:31 -08001699/*
Christoph Lametera92f7122006-02-01 03:05:32 -08001700 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1701 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1702 * a zone.
1703 */
1704#define ZONE_RECLAIM_PRIORITY 4
1705
Christoph Lameter9eeff232006-01-18 17:42:31 -08001706/*
Christoph Lameter96146342006-07-03 00:24:13 -07001707 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1708 * occur.
1709 */
1710int sysctl_min_unmapped_ratio = 1;
1711
1712/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07001713 * If the number of slab pages in a zone grows beyond this percentage then
1714 * slab reclaim needs to occur.
1715 */
1716int sysctl_min_slab_ratio = 5;
1717
1718/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08001719 * Try to free up some pages from this zone through reclaim.
1720 */
Andrew Morton179e9632006-03-22 00:08:18 -08001721static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001722{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08001723 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08001724 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001725 struct task_struct *p = current;
1726 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08001727 int priority;
Andrew Morton05ff5132006-03-22 00:08:20 -08001728 unsigned long nr_reclaimed = 0;
Andrew Morton179e9632006-03-22 00:08:18 -08001729 struct scan_control sc = {
1730 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1731 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08001732 .swap_cluster_max = max_t(unsigned long, nr_pages,
1733 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08001734 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001735 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -08001736 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07001737 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001738
1739 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08001740 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08001741 /*
1742 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1743 * and we also need to be able to write out pages for RECLAIM_WRITE
1744 * and RECLAIM_SWAP.
1745 */
1746 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001747 reclaim_state.reclaimed_slab = 0;
1748 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08001749
Christoph Lameter0ff38492006-09-25 23:31:52 -07001750 if (zone_page_state(zone, NR_FILE_PAGES) -
1751 zone_page_state(zone, NR_FILE_MAPPED) >
1752 zone->min_unmapped_pages) {
1753 /*
1754 * Free memory by calling shrink zone with increasing
1755 * priorities until we have enough memory freed.
1756 */
1757 priority = ZONE_RECLAIM_PRIORITY;
1758 do {
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001759 note_zone_scanning_priority(zone, priority);
Christoph Lameter0ff38492006-09-25 23:31:52 -07001760 nr_reclaimed += shrink_zone(priority, zone, &sc);
1761 priority--;
1762 } while (priority >= 0 && nr_reclaimed < nr_pages);
1763 }
Christoph Lameterc84db232006-02-01 03:05:29 -08001764
Christoph Lameter83e33a42006-09-25 23:31:53 -07001765 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1766 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001767 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08001768 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07001769 * many pages were freed in this zone. So we take the current
1770 * number of slab pages and shake the slab until it is reduced
1771 * by the same nr_pages that we used for reclaiming unmapped
1772 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001773 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07001774 * Note that shrink_slab will free memory on all zones and may
1775 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001776 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07001777 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07001778 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1779 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07001780 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07001781
1782 /*
1783 * Update nr_reclaimed by the number of slab pages we
1784 * reclaimed from this zone.
1785 */
1786 nr_reclaimed += slab_reclaimable -
1787 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001788 }
1789
Christoph Lameter9eeff232006-01-18 17:42:31 -08001790 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08001791 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Andrew Morton05ff5132006-03-22 00:08:20 -08001792 return nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001793}
Andrew Morton179e9632006-03-22 00:08:18 -08001794
1795int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1796{
1797 cpumask_t mask;
1798 int node_id;
1799
1800 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07001801 * Zone reclaim reclaims unmapped file backed pages and
1802 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07001803 *
Christoph Lameter96146342006-07-03 00:24:13 -07001804 * A small portion of unmapped file backed pages is needed for
1805 * file I/O otherwise pages read by file I/O will be immediately
1806 * thrown out if the zone is overallocated. So we do not reclaim
1807 * if less than a specified percentage of the zone is used by
1808 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08001809 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07001810 if (zone_page_state(zone, NR_FILE_PAGES) -
Christoph Lameter0ff38492006-09-25 23:31:52 -07001811 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1812 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1813 <= zone->min_slab_pages)
Christoph Lameter96146342006-07-03 00:24:13 -07001814 return 0;
Andrew Morton179e9632006-03-22 00:08:18 -08001815
1816 /*
1817 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1818 * not have reclaimable pages and if we should not delay the allocation
1819 * then do not scan.
1820 */
1821 if (!(gfp_mask & __GFP_WAIT) ||
1822 zone->all_unreclaimable ||
1823 atomic_read(&zone->reclaim_in_progress) > 0 ||
1824 (current->flags & PF_MEMALLOC))
1825 return 0;
1826
1827 /*
1828 * Only run zone reclaim on the local zone or on zones that do not
1829 * have associated processors. This will favor the local processor
1830 * over remote processors and spread off node memory allocations
1831 * as wide as possible.
1832 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07001833 node_id = zone_to_nid(zone);
Andrew Morton179e9632006-03-22 00:08:18 -08001834 mask = node_to_cpumask(node_id);
1835 if (!cpus_empty(mask) && node_id != numa_node_id())
1836 return 0;
1837 return __zone_reclaim(zone, gfp_mask, order);
1838}
Christoph Lameter9eeff232006-01-18 17:42:31 -08001839#endif